56:I[191445,["/_next/static/immutable/chunks/42yzn81wlo94m.js","/_next/static/immutable/chunks/2bm6_f2j__8vm.js","/_next/static/immutable/chunks/3i5397s96cu0a.js","/_next/static/immutable/chunks/2x03jtn4ys_wx.js","/_next/static/immutable/chunks/2f4lnnvm42_vp.js","/_next/static/immutable/chunks/16-tr_iilxrhb.js","/_next/static/immutable/chunks/2888ntn3nn8dg.js","/_next/static/immutable/chunks/2xqpa5c7ogclt.js","/_next/static/immutable/chunks/41cng7f4va9y-.js","/_next/static/immutable/chunks/2r_ddychp2ui5.js","/_next/static/immutable/chunks/1g9smhz0fks9x.js","/_next/static/immutable/chunks/2o71xrq9b0yu6.js","/_next/static/immutable/chunks/3qs9vw68v38q9.js","/_next/static/immutable/chunks/1rm5ppn9o7oew.js","/_next/static/immutable/chunks/2hbgrpnh1fmap.js","/_next/static/immutable/chunks/0tfe231gghziz.js","/_next/static/immutable/chunks/3ol7h3r_z2pit.js","/_next/static/immutable/chunks/3skty2nb9a-2-.js","/_next/static/immutable/chunks/3oekgz5eyywh-.js","/_next/static/immutable/chunks/1e0jyw4aho8bn.js","/_next/static/immutable/chunks/2ri0u_hhdt48v.js","/_next/static/immutable/chunks/25buewscho2w5.js","/_next/static/immutable/chunks/3u4carzbf5fig.js","/_next/static/immutable/chunks/0ewa-4giwctvp.js","/_next/static/immutable/chunks/0fopioox00_wn.js","/_next/static/immutable/chunks/17sbo0v8uw3y7.js","/_next/static/immutable/chunks/05727lidoh_64.js","/_next/static/immutable/chunks/0b0jg9r2u87k8.js","/_next/static/immutable/chunks/459bt34i9w_4u.js","/_next/static/immutable/chunks/2xi3oa1rgz4zi.js","/_next/static/immutable/chunks/0nuoj_60cpek_.js","/_next/static/immutable/chunks/25x_rwsdb74m3.js","/_next/static/immutable/chunks/0p2ndf8a03w3d.js","/_next/static/immutable/chunks/3rpw8gofk-lvz.js","/_next/static/immutable/chunks/0yqqzq4qo78e-.js","/_next/static/immutable/chunks/1lvgxqj71539d.js","/_next/static/immutable/chunks/412rzi0r7ty5b.js","/_next/static/immutable/chunks/2m_mu4_r75f07.js","/_next/static/immutable/chunks/084wa77nfb_vx.js","/_next/static/immutable/chunks/3mt6t4u0f9pd2.js","/_next/static/immutable/chunks/1lgj52kxs241t.js","/_next/static/immutable/chunks/2_9cwii8mx0pv.js","/_next/static/immutable/chunks/4579_u3pt_bl4.js","/_next/static/immutable/chunks/3-y1wg0q2kip1.js","/_next/static/immutable/chunks/3du8tbb5hc0uo.js"],"WorkTabProvider"] 57:I[649892,["/_next/static/immutable/chunks/42yzn81wlo94m.js","/_next/static/immutable/chunks/2bm6_f2j__8vm.js","/_next/static/immutable/chunks/3i5397s96cu0a.js","/_next/static/immutable/chunks/2x03jtn4ys_wx.js","/_next/static/immutable/chunks/2f4lnnvm42_vp.js","/_next/static/immutable/chunks/16-tr_iilxrhb.js","/_next/static/immutable/chunks/2888ntn3nn8dg.js","/_next/static/immutable/chunks/2xqpa5c7ogclt.js","/_next/static/immutable/chunks/41cng7f4va9y-.js","/_next/static/immutable/chunks/2r_ddychp2ui5.js","/_next/static/immutable/chunks/1g9smhz0fks9x.js","/_next/static/immutable/chunks/2o71xrq9b0yu6.js","/_next/static/immutable/chunks/3qs9vw68v38q9.js","/_next/static/immutable/chunks/1rm5ppn9o7oew.js","/_next/static/immutable/chunks/2hbgrpnh1fmap.js","/_next/static/immutable/chunks/0tfe231gghziz.js","/_next/static/immutable/chunks/3ol7h3r_z2pit.js","/_next/static/immutable/chunks/3skty2nb9a-2-.js","/_next/static/immutable/chunks/3oekgz5eyywh-.js","/_next/static/immutable/chunks/1e0jyw4aho8bn.js","/_next/static/immutable/chunks/2ri0u_hhdt48v.js","/_next/static/immutable/chunks/25buewscho2w5.js","/_next/static/immutable/chunks/3u4carzbf5fig.js","/_next/static/immutable/chunks/0ewa-4giwctvp.js","/_next/static/immutable/chunks/0fopioox00_wn.js","/_next/static/immutable/chunks/17sbo0v8uw3y7.js","/_next/static/immutable/chunks/05727lidoh_64.js","/_next/static/immutable/chunks/0b0jg9r2u87k8.js","/_next/static/immutable/chunks/459bt34i9w_4u.js","/_next/static/immutable/chunks/2xi3oa1rgz4zi.js","/_next/static/immutable/chunks/0nuoj_60cpek_.js","/_next/static/immutable/chunks/25x_rwsdb74m3.js","/_next/static/immutable/chunks/0p2ndf8a03w3d.js","/_next/static/immutable/chunks/3rpw8gofk-lvz.js","/_next/static/immutable/chunks/0yqqzq4qo78e-.js","/_next/static/immutable/chunks/1lvgxqj71539d.js","/_next/static/immutable/chunks/412rzi0r7ty5b.js","/_next/static/immutable/chunks/2m_mu4_r75f07.js","/_next/static/immutable/chunks/084wa77nfb_vx.js","/_next/static/immutable/chunks/3mt6t4u0f9pd2.js","/_next/static/immutable/chunks/1lgj52kxs241t.js","/_next/static/immutable/chunks/2_9cwii8mx0pv.js","/_next/static/immutable/chunks/4579_u3pt_bl4.js","/_next/static/immutable/chunks/3-y1wg0q2kip1.js","/_next/static/immutable/chunks/3du8tbb5hc0uo.js"],"WorkHeader"] 58:T8539,GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome 1. Overview Scientific Rationale: A search for aging-related terms ("aging", "senescence", "ageing", "lifespan", "longevity") on the NCBI Gene Expression Omnibus returns over 30,000 dataset entries, with approximately 10,000 involving human samples (NCBI E-utilities query, February 2026). Yet no study has systematically mined this resource at scale. Existing aging meta-analyses are limited to manually curated subsets of 25 to 50 datasets (1,2), leaving the vast majority of public data unexploited. Two compounding bottlenecks explain why. First, GEO metadata is notoriously inconsistent: sample ages are buried in free-text descriptions, tissue annotations use non-standard vocabularies, and experimental conditions are logged in idiosyncratic formats (3). Second, and critically, keyword-based GEO searches massively overcount relevant datasets: a study that merely reports upregulation of a senescence-associated pathway (e.g., p16/CDKN2A) in its results section will appear in a "senescence" search even if the study has nothing to do with aging or cellular senescence. This false-positive problem means the true number of directly aging-relevant transcriptomic datasets is likely a fraction of the raw search total, but no systematic triage has ever been performed. This project deploys an LLM-powered curation pipeline, built on Anthropic's Claude 4.5 and above model family, to systematically extract, standardize, and quality-control metadata across all candidate GEO entries, intelligently distinguishing genuinely aging-focused datasets from incidental keyword matches. This enables the first mega meta-analysis of human aging transcriptomes at unprecedented scale. The resulting resource, a curated, machine-readable atlas of aging gene expression, will be released as an open community tool. Timeline: 6 months Month 1-2: LLM pipeline development, GEO-wide metadata extraction, and validation. Month 3-4: Expression data download, harmonization, batch correction, and quality control. Month 5: Meta-analysis and statistical modeling (Random-effects meta-analysis, subgroup analyses, and meta-regression.) Month 6: Cross-validation, interactive visualization, manuscript preparation, and submission Data Sharing: All curated metadata, processed expression matrices, analysis code, and the LLM curation pipeline will be deposited in open repositories (GitHub under MIT license, Zenodo under CC-BY-4.0). The curated metadata catalogue itself will be released as a standalone resource for community reanalysis. An interactive Shiny web application will be deployed for the community to explore meta-analytic results, query gene-level aging signatures across tissues and conditions, and download subsets of interest. The study protocol will be preregistered on ResearchHub prior to analysis. The final manuscript will be published as a ResearchHub Registered Report with regular progress updates shared via the Updates tab throughout the project. 2. Introduction Aging is the dominant risk factor for cardiovascular disease, neurodegeneration, cancer, and metabolic dysfunction, yet the molecular programs that drive age-dependent decline remain incompletely mapped (4,5). Over the past two decades, transcriptomic profiling has emerged as the primary tool for characterizing aging at the molecular level. Individual studies have identified age-associated changes in inflammation, mitochondrial function, proteostasis, and DNA damage response pathways. However, these findings exhibit surprisingly limited overlap across studies, a problem that has persisted despite the growing volume of available data (1,6). Two landmark meta-analyses illustrate both the promise and the ceiling of current approaches. Shokhirev and Johnson reprocessed approximately 6,000 raw RNA-seq samples from the Sequence Read Archive, ultimately retaining 3,060 high-quality human samples after stringent filtering, and achieved remarkable predictive accuracy for transcriptomic age (R² = 0.96) (1). Voisin et al. assembled 3,176 human skeletal muscle samples across methylome and transcriptome datasets to demonstrate that exercise training shifts molecular profiles toward a younger state (7). Both studies relied on extensive manual curation and were restricted to specific tissues or contexts. A recent preprint used only 25 GEO datasets across multiple species to identify six causal aging driver genes, explicitly noting that their small dataset count was a limitation and calling for expanded meta-analyses as more data become available (19). Meanwhile, Palmer et al. demonstrated that transcriptomic aging signatures show minimal overlap across organs, raising the fundamental question of whether truly universal aging markers exist or whether the apparent inconsistency is an artifact of small, heterogeneous datasets (6). The core problem is scale, and specifically the curatorial bottleneck that prevents achieving it. GEO currently hosts over 200,000 series entries comprising more than 6 million individual samples (9). A broad keyword search for aging-related terms returns over 30,000 dataset entries, of which approximately 10,000 involve human samples. However, these raw counts are profoundly misleading. GEO's full-text search indexes not only study titles and designs but also result descriptions, pathway annotations, and associated publications. A cancer genomics study that incidentally reports upregulation of CDKN2A or other senescence-associated genes will appear in a "senescence" search despite having no relevance to aging biology whatsoever. Similarly, studies examining age as a covariate rather than a primary variable, or studies using "aging" only in the context of sample storage or reagent shelf life, inflate the counts. The true number of datasets whose primary focus is aging, cellular senescence, DNA damage in the context of aging, or lifespan modulation is likely an order of magnitude smaller than the raw search suggests, but no systematic triage has been performed, because manual screening of 10,000+ entries is prohibitively labor-intensive. This is precisely the problem that LLM-powered curation pipelines can solve. Large language models excel at reading unstructured text descriptions, understanding experimental intent from context, and making nuanced classification decisions that go beyond keyword matching (11,12). The Critical Path Institute has demonstrated that LLMs can perform arbitrary data cleaning and knowledge extraction tasks dynamically through prompt engineering, achieving results that would otherwise require weeks of manual curation (14). By deploying an LLM pipeline to read every GEO entry's title, summary, experimental design, and sample annotations, we can distinguish genuinely aging-focused datasets from incidental keyword matches while simultaneously extracting and standardizing the metadata needed for downstream integration. This project combines these two threads, the unmet need for large-scale human aging meta-analysis and the emerging capability of LLMs for intelligent data triage and curation, into a unified pipeline we call GEO-Age. The pipeline, built on Anthropic's Claude 4.5 and above model family, will systematically process all candidate GEO entries, classify each for aging relevance, extract and standardize metadata into a machine-readable schema, and feed the curated catalogue into a multi-level random-effects meta-analysis. The scope is restricted to human samples, including primary tissues, primary cell cultures, and cell lines (both normal and cancer-derived), to maximize translational relevance and minimize the confounds introduced by cross-species ortholog mapping. Cancer-derived cell lines are included specifically for the senescence and DNA damage analyses (RQ2, RQ5), where therapy-induced and oncogene-induced senescence are biologically relevant phenomena, but are excluded from the primary chronological aging meta-analysis (RQ1, RQ3) to avoid confounding by baseline transcriptomic aberrations (telomerase reactivation, p53 loss-of-function, aneuploidy). A particular focus will be the relationship between DNA damage responses and chronological aging. Genomic instability is a canonical hallmark of aging, and recent work has shown that genome-wide DNA damage profiles can predict mammalian age with accuracy exceeding transcriptome-based models (15). Hundreds of GEO datasets examine DNA damage induction (via irradiation, oxidative stress, genotoxic agents, or replicative senescence) in human cell lines, yet no systematic comparison has tested whether these acute damage signatures recapitulate the molecular program of chronological aging or represent a distinct biological response. The scale of our meta-analysis is uniquely positioned to resolve this question. Research Questions. RQ1. What molecular signatures of aging are robust across tissues, cell types, and experimental contexts? RQ2. Which aging-associated changes are specific to particular biological systems or perturbations? RQ3. Can meta-analysis identify high-confidence aging biomarkers that replicate across independent studies? RQ4. What factors (technical vs. biological) explain heterogeneity in reported aging signatures? RQ5. To what extent do DNA damage and genotoxic stress signatures overlap with, or diverge from, chronological aging programs? This is a discovery-driven study. Rather than testing a fixed set of hypotheses, the project is designed to generate a comprehensive descriptive atlas of human aging transcriptomics at unprecedented scale, from which specific mechanistic hypotheses can be derived and subsequently tested by the community. The analysis plan (Section 3.3) specifies pre-registered analytical frameworks and decision rules, but the primary output is a curated resource and empirical characterization rather than hypothesis confirmation. Impact: This project delivers three high-value outputs to the aging research community regardless of any specific finding. First, the curated metadata catalogue, a structured, machine-readable annotation of every aging-relevant human GEO dataset, constitutes a standalone resource that will accelerate future meta-analyses across any aging subfield. Second, the GEO-Age LLM pipeline establishes a generalizable methodology for LLM-powered data curation that can be applied to any domain with heterogeneous public data. Third, the mega meta-analysis itself will either identify genuinely universal aging markers (advancing biomarker development) or definitively demonstrate that human aging transcriptomes are context-dependent (redirecting the field toward tissue-specific intervention strategies). The interactive Shiny server will ensure these outputs remain accessible and explorable by the broader community well beyond the funding period. 3. Methods Participants/Sample Target population. All publicly available high-throughput transcriptomic datasets in GEO involving human samples that measure age-dependent molecular changes, cellular senescence, DNA damage responses, or lifespan-modulating interventions in primary tissues, primary cell cultures, or cell lines (normal and cancer-derived). Inclusion criteria. (i) Human samples (primary tissues, primary cells, or established cell lines). (ii) Study's primary focus involves aging, cellular senescence, DNA damage response, or age-modulating intervention, as determined by the GEO-Age pipeline's relevance classification, not merely keyword co-occurrence. (iii) At least 3 biological replicates per experimental condition. (iv) Raw or processed expression data available for download. (v) Sufficient metadata to assign age/condition groups (after AI-assisted extraction). Exclusion criteria. (i) Non-human samples. (ii) Datasets where aging/senescence terms appear only incidentally (e.g., pathway annotations in results, age as a minor covariate) rather than as the study's primary focus. (iii) Fewer than 3 biological replicates per group. (iv) Metadata insufficient for age/condition assignment even after AI extraction and manual review of the original publication. (v) Studies of age-associated diseases (e.g., Alzheimer's, cancer prognosis) without age-matched healthy controls or explicit aging comparisons. (vi) Datasets flagged for known technical failures in GEO. (vii) Single-cell RNA-seq datasets, which require distinct analytical frameworks (pseudobulk aggregation, cell-type deconvolution) that are outside the scope of this bulk transcriptomics meta-analysis; extension to scRNA-seq is planned as future work pending proof of concept. Expected sample size. Approximately 10,000 human GEO entries match aging-related keyword searches. Based on the high rate of false-positive keyword matches inherent to GEO's full-text indexing, we estimate that 1,000 to 2,500 of these represent genuinely aging-focused studies after AI-assisted relevance filtering. After applying the remaining inclusion/exclusion criteria (replication, data availability), we conservatively estimate 500 to 1,500 high-quality transcriptomic datasets representing 20,000 to 60,000 individual samples. For meta-analysis, a minimum of 20 independent studies per subgroup analysis is required to achieve stable random-effects estimates; we expect to exceed this threshold for all major tissue categories. 3.2 Materials and Procedures Phase 1: LLM Pipeline for Metadata Extraction and Relevance Classification (Months 1 to 2) The GEO-Age pipeline is a multi-step LLM classification and extraction system built on Anthropic's Claude 4.5 and above model family that performs two critical functions: (a) relevance classification, meaning determining whether a GEO entry is genuinely focused on aging/senescence versus incidentally matching keywords, and (b) structured metadata extraction for entries classified as relevant. Step 1: Retrieval. Use the NCBI E-utilities API (esearch + efetch) and the GEOquery R package to programmatically download all human GEO Series (GSE) records matching aging-related search terms ("aging", "ageing", "senescence", "cellular senescence", "lifespan", "longevity", "gerontology", "DNA damage aging", "replicative senescence", "oxidative stress aging"). For each entry, retrieve: title, summary, overall design, sample characteristics (GSM-level), platform annotation, publication DOI, and supplementary file links. Store raw metadata in a structured JSON schema. Step 2: AI-Powered Relevance Classification. For each GEO entry, submit the concatenated metadata fields to the Claude API with a classification prompt that instructs the model to assign one of four relevance categories: Primary aging study: aging, senescence, or DNA damage in the context of aging is the central research question. Secondary aging study: age is a major covariate or secondary variable, with sufficient aging data for extraction. Incidental mention: aging/senescence terms appear only in pathway annotations, background literature, or results interpretation, but the study does not investigate aging. Ambiguous: insufficient information to classify; flagged for manual review. Only entries classified as "primary" or "secondary" proceed to metadata extraction. "Incidental mention" entries are excluded. "Ambiguous" entries are flagged for manual review by the research assistant, with the original publication abstract retrieved via DOI to aid classification. Step 3: Structured Metadata Extraction. For entries passing relevance classification, submit to the Claude API with a structured extraction prompt specifying the following target fields: Tissue/cell type (mapped to Cell Ontology for cell types, Uberon for tissues, or CLO for cell lines; with explicit annotation of whether the sample is primary tissue, primary cells, or a cell line, and for cell lines whether normal or cancer-derived) Age or time point (normalized to years for humans) Sex (male, female, mixed, not reported) Experimental condition (natural aging, accelerated aging/progeroid model, senescence induction, DNA damage treatment, lifespan intervention, control) Intervention details (agent, dose, duration where applicable) Sample size (number of biological replicates per condition) Sequencing platform (microarray model or RNA-seq chemistry) Batch structure (single batch, multi-batch, unknown) The prompt enforces JSON output conforming to a predefined schema with constrained value sets for categorical fields. When the model cannot confidently extract a field, it returns "uncertain" with a free-text explanation, flagging the entry for manual review. Step 4: Validation. Three validation layers ensure accuracy, targeting ≥95% field-level agreement with human annotation and aiming for ≥99% on critical fields (relevance classification, tissue type, age): (a) Automated consistency checks: Cross-validate extracted organism against the GEO platform annotation. Verify that reported ages fall within plausible human ranges. Flag entries where the model's output contradicts structured GEO fields. (b) Independent LLM re-classification: For entries flagged as uncertain or where automated checks detect inconsistencies, re-submit to a separate API call with the original publication abstract (retrieved via DOI) appended to the prompt, enabling the model to resolve ambiguities from richer context. Because each API call operates without memory of prior classifications, this constitutes a functionally independent re-classification rather than self-correction. Concordance rates between first-pass and re-classification will be reported as an empirical measure of extraction reliability. (c) Human spot-check: A random 10% sample of extracted entries will be manually validated against the original publications by a trained research assistant with graduate-level experience in aging biology and GEO data structures. If field-level accuracy falls below 95% on any target field, the extraction prompt will be iteratively refined and the affected batch re-processed. All ambiguous entries flagged by the pipeline will be manually reviewed regardless of the spot-check results. Cohen's kappa will be computed between LLM and human annotations for the four-category relevance classification to quantify inter-rater reliability. Step 5: Ontology Mapping and Standardization. All extracted terms are mapped to controlled vocabularies using the EMBL-EBI Ontology Lookup Service API: tissue terms to Uberon, cell types to Cell Ontology, diseases to Disease Ontology, experimental interventions to the Experimental Factor Ontology (EFO). Unmapped terms are flagged for manual resolution. Phase 2: Expression Data Harmonization (Months 3 to 4) Step 1: Data Download. Download raw or normalized expression matrices for all included datasets. For microarray data, download CEL files where available; otherwise, use pre-processed series matrices. For RNA-seq, download count matrices from GEO Supplementary Files or re-quantify from FASTQ via the Recount3 resource where available. Step 2: Preprocessing and Normalization. Apply platform-appropriate normalization. For Affymetrix microarrays, apply RMA normalization via the oligo R package. For two-color arrays, apply loess within-array normalization followed by quantile between-array normalization via limma. For RNA-seq count data, apply the limma-voom transformation pipeline (20), which converts raw counts to log-counts-per-million with observation-level precision weights. We use limma-voom rather than DESeq2 as the primary RNA-seq framework because DESeq2's negative binomial GLM becomes computationally prohibitive beyond approximately 10,000 samples (21), whereas limma-voom scales efficiently to mega-analysis (20,22). Probe-to-gene mapping uses manufacturer annotation files with ambiguous probes removed. Gene identifiers are harmonized to HGNC symbols with Ensembl gene IDs as a secondary namespace. Step 3: Quality Control. For each dataset: (a) Remove outlier samples with inter-sample correlation < 0.8 (Pearson r on global expression profile). (b) Assess batch effects via principal component analysis; flag datasets where the first principal component separates by technical batch rather than biological condition. (c) Apply ComBat-seq for within-study batch correction where batch information is available. Step 4: Per-Dataset Differential Expression (Stage 1 of Two-Stage Meta-Analysis). Each dataset is analyzed independently to extract gene-level effect sizes (23). For microarray datasets, fit gene-wise linear models using limma's empirical Bayes moderated t-statistics (24). For RNA-seq datasets, apply the limma-voom pipeline: filter lowly expressed genes (requiring at least 10 counts in a minimum number of samples), apply the voom transformation, and fit gene-wise weighted linear models with empirical Bayes moderation (20). For studies with discrete age groups (young vs. old), the model is a simple contrast between groups. For studies with continuous age, fit a linear regression of expression on age. Where metadata permits, include sex, batch, and tissue subregion as covariates in the per-dataset model to reduce confounding. For each gene, extract the log2 fold-change (logFC) and its standard error. To combine continuous and categorical contrasts in a single meta-analysis, all effect sizes are converted to a common scale: the standardized mean difference (Hedges' g), computed from the logFC divided by the pooled within-group standard deviation, with a small-sample correction factor (25). For discrete age-group comparisons, this conversion is direct. For continuous age designs (regression of expression on age), the regression slope is converted to an equivalent standardized mean difference using the age range of each study to define the contrast (youngest versus oldest quartile), with the residual standard deviation of the regression as the denominator. Where the age range or residual SD cannot be recovered from GEO metadata, the dataset is flagged for manual extraction from the original publication. A sensitivity analysis will compare meta-analytic results when restricted to discrete-group studies only versus including converted continuous designs, to assess robustness to conversion assumptions. Phase 3: Random-Effects Meta-Analysis (Month 5) Stage 2: Combining Effect Sizes Across Datasets. For each gene, combine the per-dataset effect sizes using a random-effects meta-analysis with REML estimation of between-study variance via the metafor R package (26). REML is preferred over the DerSimonian-Laird moment estimator, which underestimates between-study variance under substantial heterogeneity (23,26). For each gene, compute: (a) the summary effect size (weighted mean Hedges' g) and its 95% confidence interval, (b) Cochran's Q statistic and I² for heterogeneity, and (c) the prediction interval. Apply Benjamini-Hochberg FDR correction across all tested genes (q < 0.05). Computational Strategy. Both stages are embarrassingly parallel. The per-dataset pipeline is orchestrated via Snakemake on the IISc HPC cluster (or equivalent GCP instances). Stage 2 parallelizes trivially across genes via BiocParallel. The computational bottleneck is I/O (downloading and parsing thousands of GEO files), not model fitting. Analysis for RQ1 and RQ3: conserved aging signatures and biomarker identification. Using the genome-wide REML random-effects results from Stage 2, restricted to primary tissues, primary cell cultures, and normal cell lines (excluding cancer-derived cell lines to avoid confounding by oncogenic transcriptomic aberrations), identify genes with consistent direction of effect and q < 0.05 across a high proportion of datasets as candidate conserved aging markers. Compare against the Human Ageing Genomic Resources (GenAge, CellAge) and the Aging Atlas (16,17,18). Assess replication by leave-one-study-out cross-validation and by testing whether top markers replicate in held-out datasets not included in the meta-analysis. Analysis for RQ2: Tissue-specific and perturbation-specific signatures. Stratify datasets by sample type: (a) primary tissues by category (immune/hematopoietic, epithelial, neural, muscle, adipose, other), (b) primary cell cultures, (c) normal cell lines, (d) cancer cell lines. Perform separate meta-analyses within each stratum. Compute pairwise Spearman correlations of gene-level effect sizes between strata. Apply hierarchical clustering on the correlation matrix to identify natural groupings of aging signatures. Perform pathway enrichment (KEGG, Reactome, GO) within each cluster to characterize biological themes. Analysis for RQ4: Sources of heterogeneity. Perform subgroup meta-analyses stratified by: sequencing platform (microarray vs. RNA-seq), tissue category, sample size per group (above/below median), and publication year (pre-2015 vs. post-2015). Report I² within each stratum to assess whether heterogeneity is driven by technical or biological factors. Use leave-one-study-out sensitivity analysis to identify influential datasets. Analysis for RQ5: DNA damage versus chronological aging. Separate datasets into two groups: (a) chronological aging studies (natural aging across timepoints) and (b) DNA damage/genotoxic stress studies (irradiation, oxidative stress, replicative senescence), including cancer-derived cell lines undergoing therapy-induced or DNA damage-induced senescence. Perform meta-analysis within each group. To compare the two signatures, apply three complementary approaches: (i) rank-rank hypergeometric overlap (RRHO) analysis, which provides a threshold-free visualization of concordance and discordance across the full ranked gene lists without arbitrary significance cutoffs; (ii) Spearman ρ restricted to genes reaching nominal significance (p < 0.05) in at least one of the two categories, to avoid inflating the correlation through shared null effects across thousands of uninformative genes; and (iii) pathway-level comparison using gene set enrichment analysis (GSEA) against KEGG and Reactome, testing whether shared and divergent pathways align with predictions from the hallmarks of aging framework (4). Sensitivity analyses. (i) Restrict to studies with n ≥ 5 per group. (ii) Restrict to primary tissues only (exclude cell lines). (iii) Stratify by publication year (pre-2015 microarray era vs. post-2015 RNA-seq era). (iv) Funnel plots and Egger's test for publication bias. (v) Restrict to discrete age-group designs only (excluding continuous-age regression designs) to assess robustness to the Hedges' g conversion procedure. Phase 4: Validation and Deliverables (Month 6) Cross-reference validation. Compare meta-analytic aging gene sets against: (a) the GenAge database of aging-related genes (16), (b) the CellAge database of cellular senescence genes (18), (c) the Aging Atlas multi-omics resource (17), and (d) recently published independent aging cohorts not included in the meta-analysis. Quantify overlap using hypergeometric enrichment tests. Interactive Shiny server. Develop and deploy an interactive Shiny web application enabling users to: query the curated metadata catalogue; visualize gene-level meta-analytic results across tissues, cell types, and conditions; explore tissue-specific and perturbation-specific aging signatures; and download subsets of interest for their own analyses. The server will be deployed via shinyapps.io or a containerized instance and maintained as a community resource. Open resource release. Deposit the following on Zenodo with DOI: (a) complete curated metadata catalogue (JSON + CSV), (b) processed expression matrices with harmonized annotations, (c) full meta-analytic results for all genes across all subgroup analyses, (d) the GEO-Age LLM pipeline code and prompt library (GitHub, MIT license), (e) Docker container for full pipeline reproducibility. 3.3 Ethics and Data Management Ethics. This study exclusively reanalyzes publicly available, de-identified datasets from GEO. No new human subjects research, animal experimentation, or data collection involving identifiable individuals is conducted. All source datasets comply with the original studies' ethics approvals and GEO's data sharing policies. Data Management. All analysis code will be version-controlled on GitHub with MIT license from project initiation. The curated metadata catalogue and processed datasets will be deposited on Zenodo with a persistent DOI upon completion. Analysis pipelines will be containerized using Docker for full reproducibility. The final manuscript will be published as a ResearchHub Registered Report. The study protocol will be preregistered on ResearchHub prior to initiating the meta-analysis phase, per the RFP requirements. 4. Budget Request: $10,000 USD (in ResearchCoin) Item Cost Justification Cloud computing (GCP) $3,000 Data storage, bulk expression data processing, meta-analysis computation across thousands of datasets LLM API costs (Claude 4.5 Sonnet + API) $2,000 Relevance classification and metadata extraction for 10,000 GEO entries with multi-step prompting ($0.20/dataset including validation passes) Research assistant (100 hrs × $25/hr) $2,500 Graduate-level research assistant with aging biology expertise for manual validation of 10% metadata sample, review of ambiguous entries flagged by pipeline, literature cross-referencing Interactive Shiny server (development + deployment) $1,000 Web application development and hosting for community access to curated atlas and meta-analytic results ResearchHub Registered Report publication $1,000 Open-access publication of final manuscript as ResearchHub Registered Report Contingency $500 Additional compute for expanded analyses, unforeseen API cost overruns Budget is flexible and will be adjusted based on Phase 1 results. If metadata extraction proves more efficient than estimated, surplus funds will be allocated to expanding the analysis to additional sample types, scRNA-Seq or deeper pathway-level analyses. 6. References Shokhirev MN, Johnson AA. Modeling the human aging transcriptome across tissues, health status, and sex. Aging Cell. 2021;20(1):e13280. doi:10.1111/acel.13280 de Magalhães JP, Curado J, Church GM. Meta-analysis of age-related gene expression profiles identifies common signatures of aging. Bioinformatics. 2009;25(7):875-881. doi:10.1093/bioinformatics/btp073 Sansone, S.A. et al. Toward interoperable bioscience data. Nat. Genet. 44, 121-126 (2012). López-Otín, C., Blasco, M.A., Partridge, L., Serrano, M. & Kroemer, G. Hallmarks of aging: an expanding universe. Cell 186, 243-278 (2023). Aramillo Irizar, P. et al. Transcriptomic alterations during ageing reflect the shift from cancer to degenerative diseases in the elderly. Nat. Commun. 9, 327 (2018). Palmer, D. et al. Ageing transcriptome meta-analysis reveals similarities and differences between key mammalian tissues. Aging 13, 3313-3341 (2021). Voisin, S. et al. Exercise is associated with younger methylome and transcriptome profiles in human skeletal muscle. Aging Cell 23, e13859 (2024). Winterhalter, P.R. et al. The circadian rhythm: a key variable in aging? Aging Cell 23, e14268 (2024). Barrett, T. et al. NCBI GEO: archive for functional genomics data sets, update. Nucleic Acids Res. 41, D991-D995 (2013). Brazma, A. et al. Minimum information about a microarray experiment (MIAME): toward standards for microarray data. Nat. Genet. 29, 365-371 (2001). Lee, P. et al. Benefits, limits, and risks of GPT-4 as an AI chatbot for medicine. N. Engl. J. Med. 388, 1233-1239 (2023). Jiang, L.Y. et al. Health system-scale language models are all-purpose prediction engines. Nature 619, 357-362 (2023). Churchill, A. et al. GPT models for text annotation: an empirical exploration in public policy research. Policy Stud. J. (2025). Anderson, W. et al. Unlocking the capabilities of large language models for accelerating drug development. Clin. Pharmacol. Ther. 116, 38-41 (2024). Cao, H. et al. Genome-wide profiles of DNA damage represent highly accurate predictors of mammalian age. Aging Cell 23, e14122 (2024). Tacutu, R. et al. Human Ageing Genomic Resources: new and updated databases. Nucleic Acids Res. 46, D1083-D1090 (2018). Aging Atlas Consortium. Aging Atlas: a multi-omics database for aging biology. Nucleic Acids Res. 49, D825-D830 (2021). Avelar, R.A. et al. A multidimensional systems biology analysis of cellular senescence in aging and disease. Genome Biol. 21, 91 (2020). Lopes, I. et al. Six drivers of aging identified among genes differentially expressed with age. bioRxiv 2024.08.02.606402 (2024). Law, C.W., Chen, Y., Shi, W. & Smyth, G.K. voom: precision weights unlock linear model analysis tools for RNA-seq read counts. Genome Biol. 15, R29 (2014). Love, M.I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550 (2014). Ritchie, M.E. et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 43, e47 (2015). Tseng, G.C., Ghosh, D. & Feingold, E. Comprehensive literature review and statistical considerations for microarray meta-analysis. Nucleic Acids Res. 40, 3785-3799 (2012). Smyth, G.K. Linear models and empirical Bayes methods for assessing differential expression in microarray experiments. Stat. Appl. Genet. Mol. Biol. 3, Article 3 (2004). Hedges, L.V. & Olkin, I. Statistical Methods for Meta-Analysis. Academic Press (1985). Viechtbauer, W. Conducting meta-analyses in R with the metafor package. J. Stat. Softw. 36, 1-48 (2010). AI Disclosure Gemini 3 Pro Nano Banana tool was used to generate the graphical abstract based on the proposal given as input. Claude Opus 4.6 was used to draft the initial proposal. The idea, methodology and workplan are completely original ideas of the author. The llm was used for initial draft after which the author then thoroughly vetted and heavily edited the proposal for accuracy and flow of thought and added crucial details. All references have also been cross-validated.59:T8650,GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome 1. Overview Scientific Rationale: A search for aging-related terms ("aging", "senescence", "ageing", "lifespan", "longevity") on the NCBI Gene Expression Omnibus returns over 30,000 dataset entries, with approximately 10,000 involving human samples (NCBI E-utilities query, February 2026). Yet no study has systematically mined this resource at scale. Existing aging meta-analyses are limited to manually curated subsets of 25 to 50 datasets (1,2), leaving the vast majority of public data unexploited. Two compounding bottlenecks explain why. First, GEO metadata is notoriously inconsistent: sample ages are buried in free-text descriptions, tissue annotations use non-standard vocabularies, and experimental conditions are logged in idiosyncratic formats (3). Second, and critically, keyword-based GEO searches massively overcount relevant datasets: a study that merely reports upregulation of a senescence-associated pathway (e.g., p16/CDKN2A) in its results section will appear in a "senescence" search even if the study has nothing to do with aging or cellular senescence. This false-positive problem means the true number of directly aging-relevant transcriptomic datasets is likely a fraction of the raw search total, but no systematic triage has ever been performed. This project deploys an LLM-powered curation pipeline, built on Anthropic's Claude 4.5 and above model family, to systematically extract, standardize, and quality-control metadata across all candidate GEO entries, intelligently distinguishing genuinely aging-focused datasets from incidental keyword matches. This enables the first mega meta-analysis of human aging transcriptomes at unprecedented scale. The resulting resource, a curated, machine-readable atlas of aging gene expression, will be released as an open community tool. Timeline: 6 months Month 1-2: LLM pipeline development, GEO-wide metadata extraction, and validation. Month 3-4: Expression data download, harmonization, batch correction, and quality control. Month 5: Meta-analysis and statistical modeling (Random-effects meta-analysis, subgroup analyses, and meta-regression.) Month 6: Cross-validation, interactive visualization, manuscript preparation, and submission Data Sharing: All curated metadata, processed expression matrices, analysis code, and the LLM curation pipeline will be deposited in open repositories (GitHub under MIT license, Zenodo under CC-BY-4.0). The curated metadata catalogue itself will be released as a standalone resource for community reanalysis. An interactive Shiny web application will be deployed for the community to explore meta-analytic results, query gene-level aging signatures across tissues and conditions, and download subsets of interest. The study protocol will be preregistered on ResearchHub prior to analysis. The final manuscript will be published as a ResearchHub Registered Report with regular progress updates shared via the Updates tab throughout the project. 2. Introduction Aging is the dominant risk factor for cardiovascular disease, neurodegeneration, cancer, and metabolic dysfunction, yet the molecular programs that drive age-dependent decline remain incompletely mapped (4,5). Over the past two decades, transcriptomic profiling has emerged as the primary tool for characterizing aging at the molecular level. Individual studies have identified age-associated changes in inflammation, mitochondrial function, proteostasis, and DNA damage response pathways. However, these findings exhibit surprisingly limited overlap across studies, a problem that has persisted despite the growing volume of available data (1,6). Two landmark meta-analyses illustrate both the promise and the ceiling of current approaches. Shokhirev and Johnson reprocessed approximately 6,000 raw RNA-seq samples from the Sequence Read Archive, ultimately retaining 3,060 high-quality human samples after stringent filtering, and achieved remarkable predictive accuracy for transcriptomic age (R² = 0.96) (1). Voisin et al. assembled 3,176 human skeletal muscle samples across methylome and transcriptome datasets to demonstrate that exercise training shifts molecular profiles toward a younger state (7). Both studies relied on extensive manual curation and were restricted to specific tissues or contexts. A recent preprint used only 25 GEO datasets across multiple species to identify six causal aging driver genes, explicitly noting that their small dataset count was a limitation and calling for expanded meta-analyses as more data become available (19). Meanwhile, Palmer et al. demonstrated that transcriptomic aging signatures show minimal overlap across organs, raising the fundamental question of whether truly universal aging markers exist or whether the apparent inconsistency is an artifact of small, heterogeneous datasets (6). The core problem is scale, and specifically the curatorial bottleneck that prevents achieving it. GEO currently hosts over 200,000 series entries comprising more than 6 million individual samples (9). A broad keyword search for aging-related terms returns over 30,000 dataset entries, of which approximately 10,000 involve human samples. However, these raw counts are profoundly misleading. GEO's full-text search indexes not only study titles and designs but also result descriptions, pathway annotations, and associated publications. A cancer genomics study that incidentally reports upregulation of CDKN2A or other senescence-associated genes will appear in a "senescence" search despite having no relevance to aging biology whatsoever. Similarly, studies examining age as a covariate rather than a primary variable, or studies using "aging" only in the context of sample storage or reagent shelf life, inflate the counts. The true number of datasets whose primary focus is aging, cellular senescence, DNA damage in the context of aging, or lifespan modulation is likely an order of magnitude smaller than the raw search suggests, but no systematic triage has been performed, because manual screening of 10,000+ entries is prohibitively labor-intensive. This is precisely the problem that LLM-powered curation pipelines can solve. Large language models excel at reading unstructured text descriptions, understanding experimental intent from context, and making nuanced classification decisions that go beyond keyword matching (11,12). The Critical Path Institute has demonstrated that LLMs can perform arbitrary data cleaning and knowledge extraction tasks dynamically through prompt engineering, achieving results that would otherwise require weeks of manual curation (14). By deploying an LLM pipeline to read every GEO entry's title, summary, experimental design, and sample annotations, we can distinguish genuinely aging-focused datasets from incidental keyword matches while simultaneously extracting and standardizing the metadata needed for downstream integration. This project combines these two threads, the unmet need for large-scale human aging meta-analysis and the emerging capability of LLMs for intelligent data triage and curation, into a unified pipeline we call GEO-Age. The pipeline, built on Anthropic's Claude 4.5 and above model family, will systematically process all candidate GEO entries, classify each for aging relevance, extract and standardize metadata into a machine-readable schema, and feed the curated catalogue into a multi-level random-effects meta-analysis. The scope is restricted to human samples, including primary tissues, primary cell cultures, and cell lines (both normal and cancer-derived), to maximize translational relevance and minimize the confounds introduced by cross-species ortholog mapping. Cancer-derived cell lines are included specifically for the senescence and DNA damage analyses (RQ2, RQ5), where therapy-induced and oncogene-induced senescence are biologically relevant phenomena, but are excluded from the primary chronological aging meta-analysis (RQ1, RQ3) to avoid confounding by baseline transcriptomic aberrations (telomerase reactivation, p53 loss-of-function, aneuploidy). A particular focus will be the relationship between DNA damage responses and chronological aging. Genomic instability is a canonical hallmark of aging, and recent work has shown that genome-wide DNA damage profiles can predict mammalian age with accuracy exceeding transcriptome-based models (15). Hundreds of GEO datasets examine DNA damage induction (via irradiation, oxidative stress, genotoxic agents, or replicative senescence) in human cell lines, yet no systematic comparison has tested whether these acute damage signatures recapitulate the molecular program of chronological aging or represent a distinct biological response. The scale of our meta-analysis is uniquely positioned to resolve this question. Research Questions. RQ1. What molecular signatures of aging are robust across tissues, cell types, and experimental contexts? RQ2. Which aging-associated changes are specific to particular biological systems or perturbations? RQ3. Can meta-analysis identify high-confidence aging biomarkers that replicate across independent studies? RQ4. What factors (technical vs. biological) explain heterogeneity in reported aging signatures? RQ5. To what extent do DNA damage and genotoxic stress signatures overlap with, or diverge from, chronological aging programs? This is a discovery-driven study. Rather than testing a fixed set of hypotheses, the project is designed to generate a comprehensive descriptive atlas of human aging transcriptomics at unprecedented scale, from which specific mechanistic hypotheses can be derived and subsequently tested by the community. The analysis plan (Section 3.3) specifies pre-registered analytical frameworks and decision rules, but the primary output is a curated resource and empirical characterization rather than hypothesis confirmation. Impact: This project delivers three high-value outputs to the aging research community regardless of any specific finding. First, the curated metadata catalogue, a structured, machine-readable annotation of every aging-relevant human GEO dataset, constitutes a standalone resource that will accelerate future meta-analyses across any aging subfield. Second, the GEO-Age LLM pipeline establishes a generalizable methodology for LLM-powered data curation that can be applied to any domain with heterogeneous public data. Third, the mega meta-analysis itself will either identify genuinely universal aging markers (advancing biomarker development) or definitively demonstrate that human aging transcriptomes are context-dependent (redirecting the field toward tissue-specific intervention strategies). The interactive Shiny server will ensure these outputs remain accessible and explorable by the broader community well beyond the funding period. 3. Methods Participants/Sample Target population. All publicly available high-throughput transcriptomic datasets in GEO involving human samples that measure age-dependent molecular changes, cellular senescence, DNA damage responses, or lifespan-modulating interventions in primary tissues, primary cell cultures, or cell lines (normal and cancer-derived). Inclusion criteria. (i) Human samples (primary tissues, primary cells, or established cell lines). (ii) Study's primary focus involves aging, cellular senescence, DNA damage response, or age-modulating intervention, as determined by the GEO-Age pipeline's relevance classification, not merely keyword co-occurrence. (iii) At least 3 biological replicates per experimental condition. (iv) Raw or processed expression data available for download. (v) Sufficient metadata to assign age/condition groups (after AI-assisted extraction). Exclusion criteria. (i) Non-human samples. (ii) Datasets where aging/senescence terms appear only incidentally (e.g., pathway annotations in results, age as a minor covariate) rather than as the study's primary focus. (iii) Fewer than 3 biological replicates per group. (iv) Metadata insufficient for age/condition assignment even after AI extraction and manual review of the original publication. (v) Studies of age-associated diseases (e.g., Alzheimer's, cancer prognosis) without age-matched healthy controls or explicit aging comparisons. (vi) Datasets flagged for known technical failures in GEO. (vii) Single-cell RNA-seq datasets, which require distinct analytical frameworks (pseudobulk aggregation, cell-type deconvolution) that are outside the scope of this bulk transcriptomics meta-analysis; extension to scRNA-seq is planned as future work pending proof of concept. Expected sample size. Approximately 10,000 human GEO entries match aging-related keyword searches. Based on the high rate of false-positive keyword matches inherent to GEO's full-text indexing, we estimate that 1,000 to 2,500 of these represent genuinely aging-focused studies after AI-assisted relevance filtering. After applying the remaining inclusion/exclusion criteria (replication, data availability), we conservatively estimate 500 to 1,500 high-quality transcriptomic datasets representing 20,000 to 60,000 individual samples. For meta-analysis, a minimum of 20 independent studies per subgroup analysis is required to achieve stable random-effects estimates; we expect to exceed this threshold for all major tissue categories. 3.2 Materials and Procedures Phase 1: LLM Pipeline for Metadata Extraction and Relevance Classification (Months 1 to 2) The GEO-Age pipeline is a multi-step LLM classification and extraction system built on Anthropic's Claude 4.5 and above model family that performs two critical functions: (a) relevance classification, meaning determining whether a GEO entry is genuinely focused on aging/senescence versus incidentally matching keywords, and (b) structured metadata extraction for entries classified as relevant. Step 1: Retrieval. Use the NCBI E-utilities API (esearch + efetch) and the GEOquery R package to programmatically download all human GEO Series (GSE) records matching aging-related search terms ("aging", "ageing", "senescence", "cellular senescence", "lifespan", "longevity", "gerontology", "DNA damage aging", "replicative senescence", "oxidative stress aging"). For each entry, retrieve: title, summary, overall design, sample characteristics (GSM-level), platform annotation, publication DOI, and supplementary file links. Store raw metadata in a structured JSON schema. Step 2: AI-Powered Relevance Classification. For each GEO entry, submit the concatenated metadata fields to the Claude API with a classification prompt that instructs the model to assign one of four relevance categories: Primary aging study: aging, senescence, or DNA damage in the context of aging is the central research question. Secondary aging study: age is a major covariate or secondary variable, with sufficient aging data for extraction. Incidental mention: aging/senescence terms appear only in pathway annotations, background literature, or results interpretation, but the study does not investigate aging. Ambiguous: insufficient information to classify; flagged for manual review. Only entries classified as "primary" or "secondary" proceed to metadata extraction. "Incidental mention" entries are excluded. "Ambiguous" entries are flagged for manual review by the research assistant, with the original publication abstract retrieved via DOI to aid classification. Step 3: Structured Metadata Extraction. For entries passing relevance classification, submit to the Claude API with a structured extraction prompt specifying the following target fields: Tissue/cell type (mapped to Cell Ontology for cell types, Uberon for tissues, or CLO for cell lines; with explicit annotation of whether the sample is primary tissue, primary cells, or a cell line, and for cell lines whether normal or cancer-derived) Age or time point (normalized to years for humans) Sex (male, female, mixed, not reported) Experimental condition (natural aging, accelerated aging/progeroid model, senescence induction, DNA damage treatment, lifespan intervention, control) Intervention details (agent, dose, duration where applicable) Sample size (number of biological replicates per condition) Sequencing platform (microarray model or RNA-seq chemistry) Batch structure (single batch, multi-batch, unknown) The prompt enforces JSON output conforming to a predefined schema with constrained value sets for categorical fields. When the model cannot confidently extract a field, it returns "uncertain" with a free-text explanation, flagging the entry for manual review. Step 4: Validation. Three validation layers ensure accuracy, targeting ≥95% field-level agreement with human annotation and aiming for ≥99% on critical fields (relevance classification, tissue type, age): (a) Automated consistency checks: Cross-validate extracted organism against the GEO platform annotation. Verify that reported ages fall within plausible human ranges. Flag entries where the model's output contradicts structured GEO fields. (b) Independent LLM re-classification: For entries flagged as uncertain or where automated checks detect inconsistencies, re-submit to a separate API call with the original publication abstract (retrieved via DOI) appended to the prompt, enabling the model to resolve ambiguities from richer context. Because each API call operates without memory of prior classifications, this constitutes a functionally independent re-classification rather than self-correction. Concordance rates between first-pass and re-classification will be reported as an empirical measure of extraction reliability. (c) Human spot-check: A random 10% sample of extracted entries will be manually validated against the original publications by a trained research assistant with graduate-level experience in aging biology and GEO data structures. If field-level accuracy falls below 95% on any target field, the extraction prompt will be iteratively refined and the affected batch re-processed. All ambiguous entries flagged by the pipeline will be manually reviewed regardless of the spot-check results. Cohen's kappa will be computed between LLM and human annotations for the four-category relevance classification to quantify inter-rater reliability. Step 5: Ontology Mapping and Standardization. All extracted terms are mapped to controlled vocabularies using the EMBL-EBI Ontology Lookup Service API: tissue terms to Uberon, cell types to Cell Ontology, diseases to Disease Ontology, experimental interventions to the Experimental Factor Ontology (EFO). Unmapped terms are flagged for manual resolution. Phase 2: Expression Data Harmonization (Months 3 to 4) Step 1: Data Download. Download raw or normalized expression matrices for all included datasets. For microarray data, download CEL files where available; otherwise, use pre-processed series matrices. For RNA-seq, download count matrices from GEO Supplementary Files or re-quantify from FASTQ via the Recount3 resource where available. Step 2: Preprocessing and Normalization. Apply platform-appropriate normalization. For Affymetrix microarrays, apply RMA normalization via the oligo R package. For two-color arrays, apply loess within-array normalization followed by quantile between-array normalization via limma. For RNA-seq count data, apply the limma-voom transformation pipeline (20), which converts raw counts to log-counts-per-million with observation-level precision weights. We use limma-voom rather than DESeq2 as the primary RNA-seq framework because DESeq2's negative binomial GLM becomes computationally prohibitive beyond approximately 10,000 samples (21), whereas limma-voom scales efficiently to mega-analysis (20,22). Probe-to-gene mapping uses manufacturer annotation files with ambiguous probes removed. Gene identifiers are harmonized to HGNC symbols with Ensembl gene IDs as a secondary namespace. Step 3: Quality Control. For each dataset: (a) Remove outlier samples with inter-sample correlation < 0.8 (Pearson r on global expression profile). (b) Assess batch effects via principal component analysis; flag datasets where the first principal component separates by technical batch rather than biological condition. (c) Apply ComBat-seq for within-study batch correction where batch information is available. Step 4: Per-Dataset Differential Expression (Stage 1 of Two-Stage Meta-Analysis). Each dataset is analyzed independently to extract gene-level effect sizes (23). For microarray datasets, fit gene-wise linear models using limma's empirical Bayes moderated t-statistics (24). For RNA-seq datasets, apply the limma-voom pipeline: filter lowly expressed genes (requiring at least 10 counts in a minimum number of samples), apply the voom transformation, and fit gene-wise weighted linear models with empirical Bayes moderation (20). For studies with discrete age groups (young vs. old), the model is a simple contrast between groups. For studies with continuous age, fit a linear regression of expression on age. Where metadata permits, include sex, batch, and tissue subregion as covariates in the per-dataset model to reduce confounding. For each gene, extract the log2 fold-change (logFC) and its standard error. To combine continuous and categorical contrasts in a single meta-analysis, all effect sizes are converted to a common scale: the standardized mean difference (Hedges' g), computed from the logFC divided by the pooled within-group standard deviation, with a small-sample correction factor (25). For discrete age-group comparisons, this conversion is direct. For continuous age designs (regression of expression on age), the regression slope is converted to an equivalent standardized mean difference using the age range of each study to define the contrast (youngest versus oldest quartile), with the residual standard deviation of the regression as the denominator. Where the age range or residual SD cannot be recovered from GEO metadata, the dataset is flagged for manual extraction from the original publication. A sensitivity analysis will compare meta-analytic results when restricted to discrete-group studies only versus including converted continuous designs, to assess robustness to conversion assumptions. Phase 3: Random-Effects Meta-Analysis (Month 5) Stage 2: Combining Effect Sizes Across Datasets. For each gene, combine the per-dataset effect sizes using a random-effects meta-analysis with REML estimation of between-study variance via the metafor R package (26). REML is preferred over the DerSimonian-Laird moment estimator, which underestimates between-study variance under substantial heterogeneity (23,26). For each gene, compute: (a) the summary effect size (weighted mean Hedges' g) and its 95% confidence interval, (b) Cochran's Q statistic and I² for heterogeneity, and (c) the prediction interval. Apply Benjamini-Hochberg FDR correction across all tested genes (q < 0.05). Computational Strategy. Both stages are embarrassingly parallel. The per-dataset pipeline is orchestrated via Snakemake on the IISc HPC cluster (or equivalent GCP instances). Stage 2 parallelizes trivially across genes via BiocParallel. The computational bottleneck is I/O (downloading and parsing thousands of GEO files), not model fitting. Analysis for RQ1 and RQ3: conserved aging signatures and biomarker identification. Using the genome-wide REML random-effects results from Stage 2, restricted to primary tissues, primary cell cultures, and normal cell lines (excluding cancer-derived cell lines to avoid confounding by oncogenic transcriptomic aberrations), identify genes with consistent direction of effect and q < 0.05 across a high proportion of datasets as candidate conserved aging markers. Compare against the Human Ageing Genomic Resources (GenAge, CellAge) and the Aging Atlas (16,17,18). Assess replication by leave-one-study-out cross-validation and by testing whether top markers replicate in held-out datasets not included in the meta-analysis. Analysis for RQ2: Tissue-specific and perturbation-specific signatures. Stratify datasets by sample type: (a) primary tissues by category (immune/hematopoietic, epithelial, neural, muscle, adipose, other), (b) primary cell cultures, (c) normal cell lines, (d) cancer cell lines. Perform separate meta-analyses within each stratum. Compute pairwise Spearman correlations of gene-level effect sizes between strata. Apply hierarchical clustering on the correlation matrix to identify natural groupings of aging signatures. Perform pathway enrichment (KEGG, Reactome, GO) within each cluster to characterize biological themes. Analysis for RQ4: Sources of heterogeneity. Perform subgroup meta-analyses stratified by: sequencing platform (microarray vs. RNA-seq), tissue category, sample size per group (above/below median), and publication year (pre-2015 vs. post-2015). Report I² within each stratum to assess whether heterogeneity is driven by technical or biological factors. Use leave-one-study-out sensitivity analysis to identify influential datasets. Analysis for RQ5: DNA damage versus chronological aging. Separate datasets into two groups: (a) chronological aging studies (natural aging across timepoints) and (b) DNA damage/genotoxic stress studies (irradiation, oxidative stress, replicative senescence), including cancer-derived cell lines undergoing therapy-induced or DNA damage-induced senescence. Perform meta-analysis within each group. To compare the two signatures, apply three complementary approaches: (i) rank-rank hypergeometric overlap (RRHO) analysis, which provides a threshold-free visualization of concordance and discordance across the full ranked gene lists without arbitrary significance cutoffs; (ii) Spearman ρ restricted to genes reaching nominal significance (p < 0.05) in at least one of the two categories, to avoid inflating the correlation through shared null effects across thousands of uninformative genes; and (iii) pathway-level comparison using gene set enrichment analysis (GSEA) against KEGG and Reactome, testing whether shared and divergent pathways align with predictions from the hallmarks of aging framework (4). Sensitivity analyses. (i) Restrict to studies with n ≥ 5 per group. (ii) Restrict to primary tissues only (exclude cell lines). (iii) Stratify by publication year (pre-2015 microarray era vs. post-2015 RNA-seq era). (iv) Funnel plots and Egger's test for publication bias. (v) Restrict to discrete age-group designs only (excluding continuous-age regression designs) to assess robustness to the Hedges' g conversion procedure. Phase 4: Validation and Deliverables (Month 6) Cross-reference validation. Compare meta-analytic aging gene sets against: (a) the GenAge database of aging-related genes (16), (b) the CellAge database of cellular senescence genes (18), (c) the Aging Atlas multi-omics resource (17), and (d) recently published independent aging cohorts not included in the meta-analysis. Quantify overlap using hypergeometric enrichment tests. Interactive Shiny server. Develop and deploy an interactive Shiny web application enabling users to: query the curated metadata catalogue; visualize gene-level meta-analytic results across tissues, cell types, and conditions; explore tissue-specific and perturbation-specific aging signatures; and download subsets of interest for their own analyses. The server will be deployed via shinyapps.io or a containerized instance and maintained as a community resource. Open resource release. Deposit the following on Zenodo with DOI: (a) complete curated metadata catalogue (JSON + CSV), (b) processed expression matrices with harmonized annotations, (c) full meta-analytic results for all genes across all subgroup analyses, (d) the GEO-Age LLM pipeline code and prompt library (GitHub, MIT license), (e) Docker container for full pipeline reproducibility. 3.3 Ethics and Data Management Ethics. This study exclusively reanalyzes publicly available, de-identified datasets from GEO. No new human subjects research, animal experimentation, or data collection involving identifiable individuals is conducted. All source datasets comply with the original studies' ethics approvals and GEO's data sharing policies. Data Management. All analysis code will be version-controlled on GitHub with MIT license from project initiation. The curated metadata catalogue and processed datasets will be deposited on Zenodo with a persistent DOI upon completion. Analysis pipelines will be containerized using Docker for full reproducibility. The final manuscript will be published as a ResearchHub Registered Report. The study protocol will be preregistered on ResearchHub prior to initiating the meta-analysis phase, per the RFP requirements. 4. Budget Request: $10,000 USD (in ResearchCoin) Item Cost Justification Cloud computing (GCP) $3,000 Data storage, bulk expression data processing, meta-analysis computation across thousands of datasets LLM API costs (Claude 4.5 Sonnet + API) $2,000 Relevance classification and metadata extraction for 10,000 GEO entries with multi-step prompting ($0.20/dataset including validation passes) Research assistant (100 hrs × $25/hr) $2,500 Graduate-level research assistant with aging biology expertise for manual validation of 10% metadata sample, review of ambiguous entries flagged by pipeline, literature cross-referencing Interactive Shiny server (development + deployment) $1,000 Web application development and hosting for community access to curated atlas and meta-analytic results ResearchHub Registered Report publication $1,000 Open-access publication of final manuscript as ResearchHub Registered Report Contingency $500 Additional compute for expanded analyses, unforeseen API cost overruns Budget is flexible and will be adjusted based on Phase 1 results. If metadata extraction proves more efficient than estimated, surplus funds will be allocated to expanding the analysis to additional sample types, scRNA-Seq or deeper pathway-level analyses. 6. References Shokhirev MN, Johnson AA. Modeling the human aging transcriptome across tissues, health status, and sex. Aging Cell. 2021;20(1):e13280. doi:10.1111/acel.13280 de Magalhães JP, Curado J, Church GM. Meta-analysis of age-related gene expression profiles identifies common signatures of aging. Bioinformatics. 2009;25(7):875-881. doi:10.1093/bioinformatics/btp073 Sansone, S.A. et al. Toward interoperable bioscience data. Nat. Genet. 44, 121-126 (2012). López-Otín, C., Blasco, M.A., Partridge, L., Serrano, M. & Kroemer, G. Hallmarks of aging: an expanding universe. Cell 186, 243-278 (2023). Aramillo Irizar, P. et al. Transcriptomic alterations during ageing reflect the shift from cancer to degenerative diseases in the elderly. Nat. Commun. 9, 327 (2018). Palmer, D. et al. Ageing transcriptome meta-analysis reveals similarities and differences between key mammalian tissues. Aging 13, 3313-3341 (2021). Voisin, S. et al. Exercise is associated with younger methylome and transcriptome profiles in human skeletal muscle. Aging Cell 23, e13859 (2024). Winterhalter, P.R. et al. The circadian rhythm: a key variable in aging? Aging Cell 23, e14268 (2024). Barrett, T. et al. NCBI GEO: archive for functional genomics data sets, update. Nucleic Acids Res. 41, D991-D995 (2013). Brazma, A. et al. Minimum information about a microarray experiment (MIAME): toward standards for microarray data. Nat. Genet. 29, 365-371 (2001). Lee, P. et al. Benefits, limits, and risks of GPT-4 as an AI chatbot for medicine. N. Engl. J. Med. 388, 1233-1239 (2023). Jiang, L.Y. et al. Health system-scale language models are all-purpose prediction engines. Nature 619, 357-362 (2023). Churchill, A. et al. GPT models for text annotation: an empirical exploration in public policy research. Policy Stud. J. (2025). Anderson, W. et al. Unlocking the capabilities of large language models for accelerating drug development. Clin. Pharmacol. Ther. 116, 38-41 (2024). Cao, H. et al. Genome-wide profiles of DNA damage represent highly accurate predictors of mammalian age. Aging Cell 23, e14122 (2024). Tacutu, R. et al. Human Ageing Genomic Resources: new and updated databases. Nucleic Acids Res. 46, D1083-D1090 (2018). Aging Atlas Consortium. Aging Atlas: a multi-omics database for aging biology. Nucleic Acids Res. 49, D825-D830 (2021). Avelar, R.A. et al. A multidimensional systems biology analysis of cellular senescence in aging and disease. Genome Biol. 21, 91 (2020). Lopes, I. et al. Six drivers of aging identified among genes differentially expressed with age. bioRxiv 2024.08.02.606402 (2024). Law, C.W., Chen, Y., Shi, W. & Smyth, G.K. voom: precision weights unlock linear model analysis tools for RNA-seq read counts. Genome Biol. 15, R29 (2014). Love, M.I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550 (2014). Ritchie, M.E. et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 43, e47 (2015). Tseng, G.C., Ghosh, D. & Feingold, E. Comprehensive literature review and statistical considerations for microarray meta-analysis. Nucleic Acids Res. 40, 3785-3799 (2012). Smyth, G.K. Linear models and empirical Bayes methods for assessing differential expression in microarray experiments. Stat. Appl. Genet. Mol. Biol. 3, Article 3 (2004). Hedges, L.V. & Olkin, I. Statistical Methods for Meta-Analysis. Academic Press (1985). Viechtbauer, W. Conducting meta-analyses in R with the metafor package. J. Stat. Softw. 36, 1-48 (2010). AI Disclosure Gemini 3 Pro Nano Banana tool was used to generate the graphical abstract based on the proposal given as input. Claude Opus 4.6 was used to draft the initial proposal. The idea, methodology and workplan are completely original ideas of the author. The llm was used for initial draft after which the author then thoroughly vetted and heavily edited the proposal for accuracy and flow of thought and added crucial details. All references have also been cross-validated. 5a:T107eb,{"type":"doc","content":[{"type":"heading","attrs":{"id":"b59c6651-cf57-49db-bbc7-e22266af2b29","data-toc-id":"b59c6651-cf57-49db-bbc7-e22266af2b29","textAlign":"left","level":1},"content":[{"type":"text","text":"GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome"}]},{"type":"heading","attrs":{"id":"55c7df49-e63f-48de-8009-a916d718ef08","data-toc-id":"55c7df49-e63f-48de-8009-a916d718ef08","textAlign":"left","level":2},"content":[{"type":"text","text":"1. Overview"}]},{"type":"paragraph","attrs":{"id":"77650712-e750-4513-a159-6f543c0997de","class":null,"textAlign":"left"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Scientific Rationale:"},{"type":"text","text":" A search for aging-related terms (\"aging\", \"senescence\", \"ageing\", \"lifespan\", \"longevity\") on the NCBI Gene Expression Omnibus returns over 30,000 dataset entries, with approximately 10,000 involving human samples (NCBI E-utilities query, February 2026). Yet no study has systematically mined this resource at scale. Existing aging meta-analyses are limited to manually curated subsets of 25 to 50 datasets (1,2), leaving the vast majority of public data unexploited. Two compounding bottlenecks explain why. First, GEO metadata is notoriously inconsistent: sample ages are buried in free-text descriptions, tissue annotations use non-standard vocabularies, and experimental conditions are logged in idiosyncratic formats (3). Second, and critically, keyword-based GEO searches massively overcount relevant datasets: a study that merely reports upregulation of a senescence-associated pathway (e.g., p16/CDKN2A) in its results section will appear in a \"senescence\" search even if the study has nothing to do with aging or cellular senescence. This false-positive problem means the true number of directly aging-relevant transcriptomic datasets is likely a fraction of the raw search total, but no systematic triage has ever been performed. This project deploys an LLM-powered curation pipeline, built on Anthropic's Claude 4.5 and above model family, to systematically extract, standardize, and quality-control metadata across all candidate GEO entries, intelligently distinguishing genuinely aging-focused datasets from incidental keyword matches. This enables the first mega meta-analysis of human aging transcriptomes at unprecedented scale. The resulting resource, a curated, machine-readable atlas of aging gene expression, will be released as an open community tool."}]},{"type":"paragraph","attrs":{"id":"46dd1b14-1162-499d-be14-0e0027917948","class":null,"textAlign":"left"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Timeline:"},{"type":"text","text":" 6 months"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"638e6adf-aa73-44ca-9096-9a90f90c3e6f","class":null,"textAlign":null},"content":[{"type":"text","text":"Month 1-2: LLM pipeline development, GEO-wide metadata extraction, and validation."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"3855c507-f68d-4782-9d2a-981d5581de79","class":null,"textAlign":null},"content":[{"type":"text","text":"Month 3-4: Expression data download, harmonization, batch correction, and quality control."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"4dc20135-c59b-4525-8918-1bafeffa588a","class":null,"textAlign":null},"content":[{"type":"text","text":"Month 5: Meta-analysis and statistical modeling (Random-effects meta-analysis, subgroup analyses, and meta-regression.)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"eec655e6-aaa1-4cbc-8dc7-e4e0e57ce4cb","class":null,"textAlign":null},"content":[{"type":"text","text":"Month 6: Cross-validation, interactive visualization, manuscript preparation, and submission"}]}]}]},{"type":"paragraph","attrs":{"id":"8a57baff-8eef-48ba-a171-7d41284ddad7","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Data Sharing:"},{"type":"text","text":" All curated metadata, processed expression matrices, analysis code, and the LLM curation pipeline will be deposited in open repositories (GitHub under MIT license, Zenodo under CC-BY-4.0). The curated metadata catalogue itself will be released as a standalone resource for community reanalysis. An interactive Shiny web application will be deployed for the community to explore meta-analytic results, query gene-level aging signatures across tissues and conditions, and download subsets of interest. The study protocol will be preregistered on ResearchHub prior to analysis. The final manuscript will be published as a ResearchHub Registered Report with regular progress updates shared via the Updates tab throughout the project."}]},{"type":"heading","attrs":{"id":"983b6250-2e54-44bc-b9a4-037cc0fcc3bc","data-toc-id":"983b6250-2e54-44bc-b9a4-037cc0fcc3bc","textAlign":"left","level":2},"content":[{"type":"text","text":"2. Introduction"}]},{"type":"paragraph","attrs":{"id":"33cac8a3-842b-4799-8106-c72987020ab1","class":null,"textAlign":"left"},"content":[{"type":"text","text":"Aging is the dominant risk factor for cardiovascular disease, neurodegeneration, cancer, and metabolic dysfunction, yet the molecular programs that drive age-dependent decline remain incompletely mapped (4,5). Over the past two decades, transcriptomic profiling has emerged as the primary tool for characterizing aging at the molecular level. Individual studies have identified age-associated changes in inflammation, mitochondrial function, proteostasis, and DNA damage response pathways. However, these findings exhibit surprisingly limited overlap across studies, a problem that has persisted despite the growing volume of available data (1,6)."}]},{"type":"paragraph","attrs":{"id":"1bdcf51e-e987-4706-b60e-091698880d55","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"Two landmark meta-analyses illustrate both the promise and the ceiling of current approaches. Shokhirev and Johnson reprocessed approximately 6,000 raw RNA-seq samples from the Sequence Read Archive, ultimately retaining 3,060 high-quality human samples after stringent filtering, and achieved remarkable predictive accuracy for transcriptomic age (R² = 0.96) (1). Voisin et al. assembled 3,176 human skeletal muscle samples across methylome and transcriptome datasets to demonstrate that exercise training shifts molecular profiles toward a younger state (7). Both studies relied on extensive manual curation and were restricted to specific tissues or contexts. A recent preprint used only 25 GEO datasets across multiple species to identify six causal aging driver genes, explicitly noting that their small dataset count was a limitation and calling for expanded meta-analyses as more data become available (19). Meanwhile, Palmer et al. demonstrated that transcriptomic aging signatures show minimal overlap across organs, raising the fundamental question of whether truly universal aging markers exist or whether the apparent inconsistency is an artifact of small, heterogeneous datasets (6)."}]},{"type":"paragraph","attrs":{"id":"336da42a-3218-44f5-ba47-96057a4da700","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"The core problem is scale, and specifically the curatorial bottleneck that prevents achieving it. GEO currently hosts over 200,000 series entries comprising more than 6 million individual samples (9). A broad keyword search for aging-related terms returns over 30,000 dataset entries, of which approximately 10,000 involve human samples. However, these raw counts are profoundly misleading. GEO's full-text search indexes not only study titles and designs but also result descriptions, pathway annotations, and associated publications. A cancer genomics study that incidentally reports upregulation of CDKN2A or other senescence-associated genes will appear in a \"senescence\" search despite having no relevance to aging biology whatsoever. Similarly, studies examining age as a covariate rather than a primary variable, or studies using \"aging\" only in the context of sample storage or reagent shelf life, inflate the counts. The true number of datasets whose primary focus is aging, cellular senescence, DNA damage in the context of aging, or lifespan modulation is likely an order of magnitude smaller than the raw search suggests, but no systematic triage has been performed, because manual screening of 10,000+ entries is prohibitively labor-intensive."}]},{"type":"paragraph","attrs":{"id":"8de9212e-7917-4965-87c7-1442e3faebd8","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"This is precisely the problem that LLM-powered curation pipelines can solve. Large language models excel at reading unstructured text descriptions, understanding experimental intent from context, and making nuanced classification decisions that go beyond keyword matching (11,12). The Critical Path Institute has demonstrated that LLMs can perform arbitrary data cleaning and knowledge extraction tasks dynamically through prompt engineering, achieving results that would otherwise require weeks of manual curation (14). By deploying an LLM pipeline to read every GEO entry's title, summary, experimental design, and sample annotations, we can distinguish genuinely aging-focused datasets from incidental keyword matches while simultaneously extracting and standardizing the metadata needed for downstream integration."}]},{"type":"paragraph","attrs":{"id":"4b12838e-7e03-47ae-a960-0dd57bca9581","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"This project combines these two threads, the unmet need for large-scale human aging meta-analysis and the emerging capability of LLMs for intelligent data triage and curation, into a unified pipeline we call "},{"type":"text","marks":[{"type":"bold"}],"text":"GEO-Age"},{"type":"text","text":". The pipeline, built on Anthropic's Claude 4.5 and above model family, will systematically process all candidate GEO entries, classify each for aging relevance, extract and standardize metadata into a machine-readable schema, and feed the curated catalogue into a multi-level random-effects meta-analysis. The scope is restricted to human samples, including primary tissues, primary cell cultures, and cell lines (both normal and cancer-derived), to maximize translational relevance and minimize the confounds introduced by cross-species ortholog mapping. Cancer-derived cell lines are included specifically for the senescence and DNA damage analyses (RQ2, RQ5), where therapy-induced and oncogene-induced senescence are biologically relevant phenomena, but are excluded from the primary chronological aging meta-analysis (RQ1, RQ3) to avoid confounding by baseline transcriptomic aberrations (telomerase reactivation, p53 loss-of-function, aneuploidy)."}]},{"type":"paragraph","attrs":{"id":"2edb4406-f619-4f6e-b05f-64ea4c988ee0","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"A particular focus will be the relationship between DNA damage responses and chronological aging. Genomic instability is a canonical hallmark of aging, and recent work has shown that genome-wide DNA damage profiles can predict mammalian age with accuracy exceeding transcriptome-based models (15). Hundreds of GEO datasets examine DNA damage induction (via irradiation, oxidative stress, genotoxic agents, or replicative senescence) in human cell lines, yet no systematic comparison has tested whether these acute damage signatures recapitulate the molecular program of chronological aging or represent a distinct biological response. The scale of our meta-analysis is uniquely positioned to resolve this question."}]},{"type":"paragraph","attrs":{"id":"560c1f14-fa8b-4014-90c5-ff5301939233","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Research Questions."}]},{"type":"paragraph","attrs":{"id":"749c02e7-c7dd-4c28-9c37-a1e1c9e8c258","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"RQ1. What molecular signatures of aging are robust across tissues, cell types, and experimental contexts?"}]},{"type":"paragraph","attrs":{"id":"d6a5caaa-6ce2-4657-850c-02daa4c64930","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"RQ2. Which aging-associated changes are specific to particular biological systems or perturbations?"}]},{"type":"paragraph","attrs":{"id":"6a41e894-ac2f-4793-9365-019dd1e0469e","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"RQ3. Can meta-analysis identify high-confidence aging biomarkers that replicate across independent studies?"}]},{"type":"paragraph","attrs":{"id":"a8566f0a-45ba-472e-9887-709ded377e36","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"RQ4. What factors (technical vs. biological) explain heterogeneity in reported aging signatures?"}]},{"type":"paragraph","attrs":{"id":"6bc32221-c1e5-4753-99aa-93199578ca2c","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"RQ5. To what extent do DNA damage and genotoxic stress signatures overlap with, or diverge from, chronological aging programs?"}]},{"type":"paragraph","attrs":{"id":"7cd5568e-ff0d-4a1f-889e-1b54775442f6","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"This is a discovery-driven study. Rather than testing a fixed set of hypotheses, the project is designed to generate a comprehensive descriptive atlas of human aging transcriptomics at unprecedented scale, from which specific mechanistic hypotheses can be derived and subsequently tested by the community. The analysis plan (Section 3.3) specifies pre-registered analytical frameworks and decision rules, but the primary output is a curated resource and empirical characterization rather than hypothesis confirmation."}]},{"type":"paragraph","attrs":{"id":"64298fff-d0a8-4cd8-922f-a9a00820f243","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Impact:"},{"type":"text","text":" This project delivers three high-value outputs to the aging research community regardless of any specific finding. First, the curated metadata catalogue, a structured, machine-readable annotation of every aging-relevant human GEO dataset, constitutes a standalone resource that will accelerate future meta-analyses across any aging subfield. Second, the GEO-Age LLM pipeline establishes a generalizable methodology for LLM-powered data curation that can be applied to any domain with heterogeneous public data. Third, the mega meta-analysis itself will either identify genuinely universal aging markers (advancing biomarker development) or definitively demonstrate that human aging transcriptomes are context-dependent (redirecting the field toward tissue-specific intervention strategies). The interactive Shiny server will ensure these outputs remain accessible and explorable by the broader community well beyond the funding period."}]},{"type":"heading","attrs":{"id":"b7d5e9cd-19b2-4ab0-b478-4d80cf9ac123","data-toc-id":"b7d5e9cd-19b2-4ab0-b478-4d80cf9ac123","textAlign":"left","level":2},"content":[{"type":"text","text":"3. Methods"}]},{"type":"heading","attrs":{"id":"4e97bc7b-b93b-4613-8a26-fbb19b78c721","data-toc-id":"4e97bc7b-b93b-4613-8a26-fbb19b78c721","textAlign":"left","level":3},"content":[{"type":"text","text":"Participants/Sample"}]},{"type":"paragraph","attrs":{"id":"77be43ec-69c4-4b1c-b3ae-d4490721342c","class":null,"textAlign":"left"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Target population. "},{"type":"text","text":"All publicly available high-throughput transcriptomic datasets in GEO involving human samples that measure age-dependent molecular changes, cellular senescence, DNA damage responses, or lifespan-modulating interventions in primary tissues, primary cell cultures, or cell lines (normal and cancer-derived)."}]},{"type":"paragraph","attrs":{"id":"c2eb7768-590f-4061-9164-22716a891954","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Inclusion criteria."},{"type":"text","text":" (i) Human samples (primary tissues, primary cells, or established cell lines). (ii) Study's primary focus involves aging, cellular senescence, DNA damage response, or age-modulating intervention, as determined by the GEO-Age pipeline's relevance classification, not merely keyword co-occurrence. (iii) At least 3 biological replicates per experimental condition. (iv) Raw or processed expression data available for download. (v) Sufficient metadata to assign age/condition groups (after AI-assisted extraction)."}]},{"type":"paragraph","attrs":{"id":"07deff5a-b6d6-4fb9-8e37-fc087e5f4dcf","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Exclusion criteria."},{"type":"text","text":" (i) Non-human samples. (ii) Datasets where aging/senescence terms appear only incidentally (e.g., pathway annotations in results, age as a minor covariate) rather than as the study's primary focus. (iii) Fewer than 3 biological replicates per group. (iv) Metadata insufficient for age/condition assignment even after AI extraction and manual review of the original publication. (v) Studies of age-associated diseases (e.g., Alzheimer's, cancer prognosis) without age-matched healthy controls or explicit aging comparisons. (vi) Datasets flagged for known technical failures in GEO. (vii) Single-cell RNA-seq datasets, which require distinct analytical frameworks (pseudobulk aggregation, cell-type deconvolution) that are outside the scope of this bulk transcriptomics meta-analysis; extension to scRNA-seq is planned as future work pending proof of concept."}]},{"type":"paragraph","attrs":{"id":"e937efc5-cfc7-46a0-9c8a-e5ecf4f6c3da","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Expected sample size."},{"type":"text","text":" Approximately 10,000 human GEO entries match aging-related keyword searches. Based on the high rate of false-positive keyword matches inherent to GEO's full-text indexing, we estimate that 1,000 to 2,500 of these represent genuinely aging-focused studies after AI-assisted relevance filtering. After applying the remaining inclusion/exclusion criteria (replication, data availability), we conservatively estimate 500 to 1,500 high-quality transcriptomic datasets representing 20,000 to 60,000 individual samples. For meta-analysis, a minimum of 20 independent studies per subgroup analysis is required to achieve stable random-effects estimates; we expect to exceed this threshold for all major tissue categories."}]},{"type":"heading","attrs":{"id":"cc2fa152-2040-4448-abda-52c549c4fd96","data-toc-id":"cc2fa152-2040-4448-abda-52c549c4fd96","textAlign":null,"level":3},"content":[{"type":"text","text":"3.2 Materials and Procedures"}]},{"type":"paragraph","attrs":{"id":"92b86816-f5b8-430e-892b-4eb090825dc3","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Phase 1: LLM Pipeline for Metadata Extraction and Relevance Classification (Months 1 to 2)"}]},{"type":"paragraph","attrs":{"id":"3ce468a0-8c7e-4ca9-9a24-99a17f1509a2","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"The GEO-Age pipeline is a multi-step LLM classification and extraction system built on Anthropic's Claude 4.5 and above model family that performs two critical functions: (a) relevance classification, meaning determining whether a GEO entry is genuinely focused on aging/senescence versus incidentally matching keywords, and (b) structured metadata extraction for entries classified as relevant."}]},{"type":"paragraph","attrs":{"id":"daf7de2a-ef12-44a3-ae6a-6a0ecaf6c5de","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 1: Retrieval."},{"type":"text","text":" Use the NCBI E-utilities API (esearch + efetch) and the GEOquery R package to programmatically download all human GEO Series (GSE) records matching aging-related search terms (\"aging\", \"ageing\", \"senescence\", \"cellular senescence\", \"lifespan\", \"longevity\", \"gerontology\", \"DNA damage aging\", \"replicative senescence\", \"oxidative stress aging\"). For each entry, retrieve: title, summary, overall design, sample characteristics (GSM-level), platform annotation, publication DOI, and supplementary file links. Store raw metadata in a structured JSON schema."}]},{"type":"paragraph","attrs":{"id":"95a4f3ac-81fe-41d8-badc-628d5facae6e","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 2: AI-Powered Relevance Classification."},{"type":"text","text":" For each GEO entry, submit the concatenated metadata fields to the Claude API with a classification prompt that instructs the model to assign one of four relevance categories:"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"ece197b3-349f-4058-a3c6-2f3ecba6552b","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Primary aging study:"},{"type":"text","text":" aging, senescence, or DNA damage in the context of aging is the central research question."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"897a5934-88bd-40ce-a712-e265a6745da0","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Secondary aging study:"},{"type":"text","text":" age is a major covariate or secondary variable, with sufficient aging data for extraction."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"936de766-d6d6-4628-8e4c-56808d0a4739","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Incidental mention:"},{"type":"text","text":" aging/senescence terms appear only in pathway annotations, background literature, or results interpretation, but the study does not investigate aging."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"1af66e1c-a5aa-4b49-91a2-a027862533bf","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Ambiguous:"},{"type":"text","text":" insufficient information to classify; flagged for manual review."}]}]}]},{"type":"paragraph","attrs":{"id":"868bb795-afbe-42f4-b11d-69de254326f8","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"Only entries classified as \"primary\" or \"secondary\" proceed to metadata extraction. \"Incidental mention\" entries are excluded. \"Ambiguous\" entries are flagged for manual review by the research assistant, with the original publication abstract retrieved via DOI to aid classification."}]},{"type":"paragraph","attrs":{"id":"89a094ad-5487-4bce-a278-e362d2542bb7","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 3: Structured Metadata Extraction."},{"type":"text","text":" For entries passing relevance classification, submit to the Claude API with a structured extraction prompt specifying the following target fields:"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"a713428d-9d00-4cd5-842b-414249d9bccf","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Tissue/cell type"},{"type":"text","text":" (mapped to Cell Ontology for cell types, Uberon for tissues, or CLO for cell lines; with explicit annotation of whether the sample is primary tissue, primary cells, or a cell line, and for cell lines whether normal or cancer-derived)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"5811c32c-eb77-48a0-9c3a-15ae8932b38a","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Age or time point"},{"type":"text","text":" (normalized to years for humans)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"dbfac0b1-3127-449c-b739-5f3a05004a38","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Sex"},{"type":"text","text":" (male, female, mixed, not reported)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"b658da4a-2b19-4b5e-a0d6-f90ee276b5ba","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Experimental condition"},{"type":"text","text":" (natural aging, accelerated aging/progeroid model, senescence induction, DNA damage treatment, lifespan intervention, control)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"377187a3-5414-4f89-98d0-968a2393fb84","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Intervention details"},{"type":"text","text":" (agent, dose, duration where applicable)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"2978fe9e-915f-4b0f-b059-26ef6bf3afae","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Sample size"},{"type":"text","text":" (number of biological replicates per condition)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"48635e63-7563-4928-b821-567d26893db1","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Sequencing platform"},{"type":"text","text":" (microarray model or RNA-seq chemistry)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"860f1762-432f-4c0b-9a9e-f722913d3e08","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Batch structure"},{"type":"text","text":" (single batch, multi-batch, unknown)"}]}]}]},{"type":"paragraph","attrs":{"id":"8f5cfced-5e24-47e6-acde-bad84af9ff79","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"The prompt enforces JSON output conforming to a predefined schema with constrained value sets for categorical fields. When the model cannot confidently extract a field, it returns \"uncertain\" with a free-text explanation, flagging the entry for manual review."}]},{"type":"paragraph","attrs":{"id":"e56bd011-23bc-4e2e-b7ad-93f3071db3a3","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 4: Validation."},{"type":"text","text":" Three validation layers ensure accuracy, targeting ≥95% field-level agreement with human annotation and aiming for ≥99% on critical fields (relevance classification, tissue type, age):"}]},{"type":"paragraph","attrs":{"id":"884ddf31-0252-40eb-b487-5e69f8b7e7bd","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"(a) "},{"type":"text","marks":[{"type":"italic"}],"text":"Automated consistency checks:"},{"type":"text","text":" Cross-validate extracted organism against the GEO platform annotation. Verify that reported ages fall within plausible human ranges. Flag entries where the model's output contradicts structured GEO fields."}]},{"type":"paragraph","attrs":{"id":"d087dae3-a85f-4d7a-bade-9fbf36f29162","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"(b) "},{"type":"text","marks":[{"type":"italic"}],"text":"Independent LLM re-classification:"},{"type":"text","text":" For entries flagged as uncertain or where automated checks detect inconsistencies, re-submit to a separate API call with the original publication abstract (retrieved via DOI) appended to the prompt, enabling the model to resolve ambiguities from richer context. Because each API call operates without memory of prior classifications, this constitutes a functionally independent re-classification rather than self-correction. Concordance rates between first-pass and re-classification will be reported as an empirical measure of extraction reliability."}]},{"type":"paragraph","attrs":{"id":"9376e777-b9dd-4d47-ab47-c5b5772aab23","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"(c) "},{"type":"text","marks":[{"type":"italic"}],"text":"Human spot-check:"},{"type":"text","text":" A random 10% sample of extracted entries will be manually validated against the original publications by a trained research assistant with graduate-level experience in aging biology and GEO data structures. If field-level accuracy falls below 95% on any target field, the extraction prompt will be iteratively refined and the affected batch re-processed. All ambiguous entries flagged by the pipeline will be manually reviewed regardless of the spot-check results. Cohen's kappa will be computed between LLM and human annotations for the four-category relevance classification to quantify inter-rater reliability."}]},{"type":"paragraph","attrs":{"id":"793be019-1746-4bbb-96f3-b923568a16cd","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 5: Ontology Mapping and Standardization."},{"type":"text","text":" All extracted terms are mapped to controlled vocabularies using the EMBL-EBI Ontology Lookup Service API: tissue terms to Uberon, cell types to Cell Ontology, diseases to Disease Ontology, experimental interventions to the Experimental Factor Ontology (EFO). Unmapped terms are flagged for manual resolution."}]},{"type":"paragraph","attrs":{"id":"d2bac98f-2a90-4df1-9368-7ee07d94e653","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Phase 2: Expression Data Harmonization (Months 3 to 4)"}]},{"type":"paragraph","attrs":{"id":"b8629ad0-c959-475d-8aad-063f6295d13a","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 1: Data Download."},{"type":"text","text":" Download raw or normalized expression matrices for all included datasets. For microarray data, download CEL files where available; otherwise, use pre-processed series matrices. For RNA-seq, download count matrices from GEO Supplementary Files or re-quantify from FASTQ via the Recount3 resource where available."}]},{"type":"paragraph","attrs":{"id":"abb4776d-c955-4462-9f79-813dbe0d400d","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 2: Preprocessing and Normalization."},{"type":"text","text":" Apply platform-appropriate normalization. For Affymetrix microarrays, apply RMA normalization via the oligo R package. For two-color arrays, apply loess within-array normalization followed by quantile between-array normalization via limma. For RNA-seq count data, apply the limma-voom transformation pipeline (20), which converts raw counts to log-counts-per-million with observation-level precision weights. We use limma-voom rather than DESeq2 as the primary RNA-seq framework because DESeq2's negative binomial GLM becomes computationally prohibitive beyond approximately 10,000 samples (21), whereas limma-voom scales efficiently to mega-analysis (20,22). Probe-to-gene mapping uses manufacturer annotation files with ambiguous probes removed. Gene identifiers are harmonized to HGNC symbols with Ensembl gene IDs as a secondary namespace."}]},{"type":"paragraph","attrs":{"id":"1b9e54d5-4d6c-4c01-88af-df47ad4b0ff9","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 3: Quality Control."},{"type":"text","text":" For each dataset: (a) Remove outlier samples with inter-sample correlation < 0.8 (Pearson r on global expression profile). (b) Assess batch effects via principal component analysis; flag datasets where the first principal component separates by technical batch rather than biological condition. (c) Apply ComBat-seq for within-study batch correction where batch information is available."}]},{"type":"paragraph","attrs":{"id":"1e685a25-0d51-4865-91ec-8ab5f4a5b16a","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 4: Per-Dataset Differential Expression (Stage 1 of Two-Stage Meta-Analysis)."},{"type":"text","text":" Each dataset is analyzed independently to extract gene-level effect sizes (23). For microarray datasets, fit gene-wise linear models using limma's empirical Bayes moderated t-statistics (24). For RNA-seq datasets, apply the limma-voom pipeline: filter lowly expressed genes (requiring at least 10 counts in a minimum number of samples), apply the voom transformation, and fit gene-wise weighted linear models with empirical Bayes moderation (20). For studies with discrete age groups (young vs. old), the model is a simple contrast between groups. For studies with continuous age, fit a linear regression of expression on age. Where metadata permits, include sex, batch, and tissue subregion as covariates in the per-dataset model to reduce confounding. For each gene, extract the log2 fold-change (logFC) and its standard error. To combine continuous and categorical contrasts in a single meta-analysis, all effect sizes are converted to a common scale: the standardized mean difference (Hedges' g), computed from the logFC divided by the pooled within-group standard deviation, with a small-sample correction factor (25). For discrete age-group comparisons, this conversion is direct. For continuous age designs (regression of expression on age), the regression slope is converted to an equivalent standardized mean difference using the age range of each study to define the contrast (youngest versus oldest quartile), with the residual standard deviation of the regression as the denominator. Where the age range or residual SD cannot be recovered from GEO metadata, the dataset is flagged for manual extraction from the original publication. A sensitivity analysis will compare meta-analytic results when restricted to discrete-group studies only versus including converted continuous designs, to assess robustness to conversion assumptions."}]},{"type":"paragraph","attrs":{"id":"b99f571f-f612-43e0-be95-f337f7c570c3","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Phase 3: Random-Effects Meta-Analysis (Month 5)"}]},{"type":"paragraph","attrs":{"id":"0ebfe726-ec4e-432e-bfbc-a522c3d803a8","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Stage 2: Combining Effect Sizes Across Datasets."},{"type":"text","text":" For each gene, combine the per-dataset effect sizes using a random-effects meta-analysis with REML estimation of between-study variance via the metafor R package (26). REML is preferred over the DerSimonian-Laird moment estimator, which underestimates between-study variance under substantial heterogeneity (23,26). For each gene, compute: (a) the summary effect size (weighted mean Hedges' g) and its 95% confidence interval, (b) Cochran's Q statistic and I² for heterogeneity, and (c) the prediction interval. Apply Benjamini-Hochberg FDR correction across all tested genes (q < 0.05)."}]},{"type":"paragraph","attrs":{"id":"89335312-623c-488e-aff2-037141fb58d6","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Computational Strategy."},{"type":"text","text":" Both stages are embarrassingly parallel. The per-dataset pipeline is orchestrated via Snakemake on the IISc HPC cluster (or equivalent GCP instances). Stage 2 parallelizes trivially across genes via BiocParallel. The computational bottleneck is I/O (downloading and parsing thousands of GEO files), not model fitting."}]},{"type":"paragraph","attrs":{"id":"639aa201-f345-4f95-abfc-01874a7f3ae2","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Analysis for RQ1 and RQ3: conserved aging signatures and biomarker identification."},{"type":"text","text":" Using the genome-wide REML random-effects results from Stage 2, restricted to primary tissues, primary cell cultures, and normal cell lines (excluding cancer-derived cell lines to avoid confounding by oncogenic transcriptomic aberrations), identify genes with consistent direction of effect and q < 0.05 across a high proportion of datasets as candidate conserved aging markers. Compare against the Human Ageing Genomic Resources (GenAge, CellAge) and the Aging Atlas (16,17,18). Assess replication by leave-one-study-out cross-validation and by testing whether top markers replicate in held-out datasets not included in the meta-analysis."}]},{"type":"paragraph","attrs":{"id":"c5c70110-1537-4e1d-bac5-62dc8accb7eb","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Analysis for RQ2: Tissue-specific and perturbation-specific signatures."},{"type":"text","text":" Stratify datasets by sample type: (a) primary tissues by category (immune/hematopoietic, epithelial, neural, muscle, adipose, other), (b) primary cell cultures, (c) normal cell lines, (d) cancer cell lines. Perform separate meta-analyses within each stratum. Compute pairwise Spearman correlations of gene-level effect sizes between strata. Apply hierarchical clustering on the correlation matrix to identify natural groupings of aging signatures. Perform pathway enrichment (KEGG, Reactome, GO) within each cluster to characterize biological themes."}]},{"type":"paragraph","attrs":{"id":"3a920d57-6b54-4ee6-98e2-3cb06d51da37","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Analysis for RQ4: Sources of heterogeneity."},{"type":"text","text":" Perform subgroup meta-analyses stratified by: sequencing platform (microarray vs. RNA-seq), tissue category, sample size per group (above/below median), and publication year (pre-2015 vs. post-2015). Report I² within each stratum to assess whether heterogeneity is driven by technical or biological factors. Use leave-one-study-out sensitivity analysis to identify influential datasets."}]},{"type":"paragraph","attrs":{"id":"8860731c-4ff6-4fa9-aa81-b744acf6740b","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Analysis for RQ5: DNA damage versus chronological aging."},{"type":"text","text":" Separate datasets into two groups: (a) chronological aging studies (natural aging across timepoints) and (b) DNA damage/genotoxic stress studies (irradiation, oxidative stress, replicative senescence), including cancer-derived cell lines undergoing therapy-induced or DNA damage-induced senescence. Perform meta-analysis within each group. To compare the two signatures, apply three complementary approaches: (i) rank-rank hypergeometric overlap (RRHO) analysis, which provides a threshold-free visualization of concordance and discordance across the full ranked gene lists without arbitrary significance cutoffs; (ii) Spearman ρ restricted to genes reaching nominal significance (p < 0.05) in at least one of the two categories, to avoid inflating the correlation through shared null effects across thousands of uninformative genes; and (iii) pathway-level comparison using gene set enrichment analysis (GSEA) against KEGG and Reactome, testing whether shared and divergent pathways align with predictions from the hallmarks of aging framework (4)."}]},{"type":"paragraph","attrs":{"id":"6ff06830-a963-41d7-a7b1-45a8701c6574","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Sensitivity analyses."},{"type":"text","text":" (i) Restrict to studies with n ≥ 5 per group. (ii) Restrict to primary tissues only (exclude cell lines). (iii) Stratify by publication year (pre-2015 microarray era vs. post-2015 RNA-seq era). (iv) Funnel plots and Egger's test for publication bias. (v) Restrict to discrete age-group designs only (excluding continuous-age regression designs) to assess robustness to the Hedges' g conversion procedure."}]},{"type":"paragraph","attrs":{"id":"84193dd6-d73a-4668-92b2-cfb23811b231","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Phase 4: Validation and Deliverables (Month 6)"}]},{"type":"paragraph","attrs":{"id":"145350f7-7e15-41d5-a58d-717705bd8164","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Cross-reference validation."},{"type":"text","text":" Compare meta-analytic aging gene sets against: (a) the GenAge database of aging-related genes (16), (b) the CellAge database of cellular senescence genes (18), (c) the Aging Atlas multi-omics resource (17), and (d) recently published independent aging cohorts not included in the meta-analysis. Quantify overlap using hypergeometric enrichment tests."}]},{"type":"paragraph","attrs":{"id":"0fab8a05-4e43-4863-83aa-5ec507678d6e","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Interactive Shiny server."},{"type":"text","text":" Develop and deploy an interactive Shiny web application enabling users to: query the curated metadata catalogue; visualize gene-level meta-analytic results across tissues, cell types, and conditions; explore tissue-specific and perturbation-specific aging signatures; and download subsets of interest for their own analyses. The server will be deployed via "},{"type":"text","marks":[{"type":"link","attrs":{"href":"http://shinyapps.io","target":"_blank","rel":"noopener noreferrer nofollow","class":"link"}}],"text":"shinyapps.io"},{"type":"text","text":" or a containerized instance and maintained as a community resource."}]},{"type":"paragraph","attrs":{"id":"11c57257-15f7-4009-8ea5-fc213ae5da5e","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Open resource release."},{"type":"text","text":" Deposit the following on Zenodo with DOI: (a) complete curated metadata catalogue (JSON + CSV), (b) processed expression matrices with harmonized annotations, (c) full meta-analytic results for all genes across all subgroup analyses, (d) the GEO-Age LLM pipeline code and prompt library (GitHub, MIT license), (e) Docker container for full pipeline reproducibility."}]},{"type":"heading","attrs":{"id":"90864883-1598-421a-8c92-66cdf635ab0c","data-toc-id":"90864883-1598-421a-8c92-66cdf635ab0c","textAlign":null,"level":3},"content":[{"type":"text","text":"3.3 Ethics and Data Management"}]},{"type":"paragraph","attrs":{"id":"3c0f8687-4bf7-4672-9e7c-8678146c4e80","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Ethics."},{"type":"text","text":" This study exclusively reanalyzes publicly available, de-identified datasets from GEO. No new human subjects research, animal experimentation, or data collection involving identifiable individuals is conducted. All source datasets comply with the original studies' ethics approvals and GEO's data sharing policies."}]},{"type":"paragraph","attrs":{"id":"7556735e-8ebf-4f0b-a7b9-f26c0ae6cf8e","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Data Management."},{"type":"text","text":" All analysis code will be version-controlled on GitHub with MIT license from project initiation. The curated metadata catalogue and processed datasets will be deposited on Zenodo with a persistent DOI upon completion. Analysis pipelines will be containerized using Docker for full reproducibility. The final manuscript will be published as a ResearchHub Registered Report. The study protocol will be preregistered on ResearchHub prior to initiating the meta-analysis phase, per the RFP requirements."}]},{"type":"heading","attrs":{"id":"09ced967-c854-423d-b321-4da26cd0c3e3","data-toc-id":"09ced967-c854-423d-b321-4da26cd0c3e3","textAlign":"left","level":2},"content":[{"type":"text","text":"4. Budget"}]},{"type":"paragraph","attrs":{"id":"3f0dc63e-8a35-4259-aa1b-9f98077d3597","class":null,"textAlign":"left"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Request: $10,000 USD (in ResearchCoin)"}]},{"type":"table","attrs":{"id":"99900f72-8bbd-4db1-a65f-2b8b67060079"},"content":[{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"6570149c-0907-4c85-bd2d-cf0d53d41da5","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Item"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"0e50d782-814a-4460-8c5a-46c5dfa2bf2c","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Cost"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"6b3439bb-a4a2-4890-8c2b-d4bbee5f2701","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Justification"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"829e4229-2f96-4f75-a102-3f537b87bc33","class":null,"textAlign":null},"content":[{"type":"text","text":"Cloud computing (GCP)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"f866c2de-dbde-411e-a653-7d83b3ea8114","class":null,"textAlign":null},"content":[{"type":"text","text":"$3,000"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"b4099417-504b-42e8-bd02-e63d167f15ef","class":null,"textAlign":null},"content":[{"type":"text","text":"Data storage, bulk expression data processing, meta-analysis computation across thousands of datasets"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"16827215-4849-47b0-95e1-52026ba9ef9a","class":null,"textAlign":null},"content":[{"type":"text","text":"LLM API costs (Claude 4.5 Sonnet + API)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"95aac98e-c054-49bc-a0ea-d7c46adb3a6a","class":null,"textAlign":null},"content":[{"type":"text","text":"$2,000"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"3950a2d5-5b95-47e5-ab4a-c0280c872a9d","class":null,"textAlign":null},"content":[{"type":"text","text":"Relevance classification and metadata extraction for 10,000 GEO entries with multi-step prompting ($0.20/dataset including validation passes)"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"5491b557-54df-4dfb-9daf-0a289481c7df","class":null,"textAlign":null},"content":[{"type":"text","text":"Research assistant (100 hrs × $25/hr)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"57cf83ca-7545-42b4-a23d-aaca6f3ec24b","class":null,"textAlign":null},"content":[{"type":"text","text":"$2,500"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"23d36ec9-8bba-4ce6-9c98-2158d7850c32","class":null,"textAlign":null},"content":[{"type":"text","text":"Graduate-level research assistant with aging biology expertise for manual validation of 10% metadata sample, review of ambiguous entries flagged by pipeline, literature cross-referencing"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"577eabd5-fabf-4d59-a913-16a11e7bb6e2","class":null,"textAlign":null},"content":[{"type":"text","text":"Interactive Shiny server (development + deployment)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"98fb1e02-fd23-4c6d-ad7e-bb761cc7021a","class":null,"textAlign":null},"content":[{"type":"text","text":"$1,000"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"abbec98d-c24e-4d9d-951d-c8bdca81762e","class":null,"textAlign":null},"content":[{"type":"text","text":"Web application development and hosting for community access to curated atlas and meta-analytic results"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"646ca841-6dda-4367-8500-eaeb8e47bf77","class":null,"textAlign":null},"content":[{"type":"text","text":"ResearchHub Registered Report publication"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"a7a9ba14-a644-49e9-bd9b-1a4279b5c023","class":null,"textAlign":null},"content":[{"type":"text","text":"$1,000"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"97cdcc80-ac60-4f61-927d-339841c66b18","class":null,"textAlign":null},"content":[{"type":"text","text":"Open-access publication of final manuscript as ResearchHub Registered Report"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"a9b6a0e5-43f6-42ac-a2e4-3764ab7d863f","class":null,"textAlign":null},"content":[{"type":"text","text":"Contingency"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"e6e4d3df-47c4-47bc-9344-62778ccb286c","class":null,"textAlign":null},"content":[{"type":"text","text":"$500"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"6f0c9075-e633-4963-bd0f-03c3a8e4b6d2","class":null,"textAlign":null},"content":[{"type":"text","text":"Additional compute for expanded analyses, unforeseen API cost overruns"}]}]}]}]},{"type":"paragraph","attrs":{"id":"c9d67022-283a-4aa6-90db-36af5c147dd1","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"Budget is flexible and will be adjusted based on Phase 1 results. If metadata extraction proves more efficient than estimated, surplus funds will be allocated to expanding the analysis to additional sample types, scRNA-Seq or deeper pathway-level analyses."}]},{"type":"heading","attrs":{"id":"08664ef8-a028-404e-806a-dbdccddb8a27","data-toc-id":"08664ef8-a028-404e-806a-dbdccddb8a27","textAlign":"left","level":2},"content":[{"type":"text","text":"6. References"}]},{"type":"orderedList","attrs":{"start":1,"type":null},"content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"ceadae81-4bfb-4b24-a4ab-3299cba8e403","class":null,"textAlign":"left"},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"fontSize":"16px","fontFamily":"BlinkMacSystemFont, -apple-system, \"Segoe UI\", Roboto, Oxygen, Ubuntu, Cantarell, \"Fira Sans\", \"Droid Sans\", \"Helvetica Neue\", sans-serif","color":"rgb(33, 33, 33)"}}],"text":"Shokhirev MN, Johnson AA. Modeling the human aging transcriptome across tissues, health status, and sex. "},{"type":"text","marks":[{"type":"italic"}],"text":"Aging Cell"},{"type":"text","marks":[{"type":"textStyle","attrs":{"fontSize":"16px","fontFamily":"BlinkMacSystemFont, -apple-system, \"Segoe UI\", Roboto, Oxygen, Ubuntu, Cantarell, \"Fira Sans\", \"Droid Sans\", \"Helvetica Neue\", sans-serif","color":"rgb(33, 33, 33)"}}],"text":". 2021;20(1):e13280. doi:10.1111/acel.13280"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"31e54f2d-1137-470f-af51-88ad3d994af8","class":null,"textAlign":"left"},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"fontSize":"16px","fontFamily":"BlinkMacSystemFont, -apple-system, \"Segoe UI\", Roboto, Oxygen, Ubuntu, Cantarell, \"Fira Sans\", \"Droid Sans\", \"Helvetica Neue\", sans-serif","color":"rgb(33, 33, 33)"}}],"text":"de Magalhães JP, Curado J, Church GM. Meta-analysis of age-related gene expression profiles identifies common signatures of aging. "},{"type":"text","marks":[{"type":"italic"}],"text":"Bioinformatics"},{"type":"text","marks":[{"type":"textStyle","attrs":{"fontSize":"16px","fontFamily":"BlinkMacSystemFont, -apple-system, \"Segoe UI\", Roboto, Oxygen, Ubuntu, Cantarell, \"Fira Sans\", \"Droid Sans\", \"Helvetica Neue\", sans-serif","color":"rgb(33, 33, 33)"}}],"text":". 2009;25(7):875-881. doi:10.1093/bioinformatics/btp073"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"0f41f2fb-63be-4dc5-9b24-7fbcbda77123","class":null,"textAlign":"left"},"content":[{"type":"text","text":"Sansone, S.A. et al. Toward interoperable bioscience data. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nat. Genet."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"44"},{"type":"text","text":", 121-126 (2012)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"fc34f1cb-36f1-436a-a2c6-b3b68ee9cf68","class":null,"textAlign":null},"content":[{"type":"text","text":"López-Otín, C., Blasco, M.A., Partridge, L., Serrano, M. & Kroemer, G. Hallmarks of aging: an expanding universe. "},{"type":"text","marks":[{"type":"italic"}],"text":"Cell"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"186"},{"type":"text","text":", 243-278 (2023)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"33ddd45c-6d98-46ab-8071-281876209994","class":null,"textAlign":null},"content":[{"type":"text","text":"Aramillo Irizar, P. et al. Transcriptomic alterations during ageing reflect the shift from cancer to degenerative diseases in the elderly. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nat. Commun."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"9"},{"type":"text","text":", 327 (2018)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"cb79f42e-6c21-4768-b5e2-47b744a4a7a2","class":null,"textAlign":null},"content":[{"type":"text","text":"Palmer, D. et al. Ageing transcriptome meta-analysis reveals similarities and differences between key mammalian tissues. "},{"type":"text","marks":[{"type":"italic"}],"text":"Aging"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"13"},{"type":"text","text":", 3313-3341 (2021)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"2622922a-c892-4129-aaf0-f8dd9f6e8e0f","class":null,"textAlign":null},"content":[{"type":"text","text":"Voisin, S. et al. Exercise is associated with younger methylome and transcriptome profiles in human skeletal muscle. "},{"type":"text","marks":[{"type":"italic"}],"text":"Aging Cell"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"23"},{"type":"text","text":", e13859 (2024)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"19cf1a4b-6b7d-44bf-a97f-fc02c2e572de","class":null,"textAlign":null},"content":[{"type":"text","text":"Winterhalter, P.R. et al. The circadian rhythm: a key variable in aging? "},{"type":"text","marks":[{"type":"italic"}],"text":"Aging Cell"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"23"},{"type":"text","text":", e14268 (2024)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"558a6bb5-b34c-4f23-b202-e8e0cd405b59","class":null,"textAlign":null},"content":[{"type":"text","text":"Barrett, T. et al. NCBI GEO: archive for functional genomics data sets, update. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nucleic Acids Res."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"41"},{"type":"text","text":", D991-D995 (2013)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"b9ce284f-e543-41f8-9a36-9b966a5ac61d","class":null,"textAlign":null},"content":[{"type":"text","text":"Brazma, A. et al. Minimum information about a microarray experiment (MIAME): toward standards for microarray data. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nat. Genet."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"29"},{"type":"text","text":", 365-371 (2001)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"9af06d9d-a1f3-4585-98d4-6d88a3867d78","class":null,"textAlign":null},"content":[{"type":"text","text":"Lee, P. et al. Benefits, limits, and risks of GPT-4 as an AI chatbot for medicine. "},{"type":"text","marks":[{"type":"italic"}],"text":"N. Engl. J. Med."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"388"},{"type":"text","text":", 1233-1239 (2023)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"335ea508-27b5-40ee-adcf-b8622f81a416","class":null,"textAlign":null},"content":[{"type":"text","text":"Jiang, L.Y. et al. Health system-scale language models are all-purpose prediction engines. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nature"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"619"},{"type":"text","text":", 357-362 (2023)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"1d05a461-e86d-4420-92c4-41c2254ac55e","class":null,"textAlign":null},"content":[{"type":"text","text":"Churchill, A. et al. GPT models for text annotation: an empirical exploration in public policy research. "},{"type":"text","marks":[{"type":"italic"}],"text":"Policy Stud. J."},{"type":"text","text":" (2025)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"3a1f9ed0-adde-4fd1-8f35-a774c2ccfb13","class":null,"textAlign":null},"content":[{"type":"text","text":"Anderson, W. et al. Unlocking the capabilities of large language models for accelerating drug development. "},{"type":"text","marks":[{"type":"italic"}],"text":"Clin. Pharmacol. Ther."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"116"},{"type":"text","text":", 38-41 (2024)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"a2a3814c-f113-4a5b-9420-5aa69341548d","class":null,"textAlign":null},"content":[{"type":"text","text":"Cao, H. et al. Genome-wide profiles of DNA damage represent highly accurate predictors of mammalian age. "},{"type":"text","marks":[{"type":"italic"}],"text":"Aging Cell"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"23"},{"type":"text","text":", e14122 (2024)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"247c1774-2112-4b20-914a-f021010eefe4","class":null,"textAlign":null},"content":[{"type":"text","text":"Tacutu, R. et al. Human Ageing Genomic Resources: new and updated databases. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nucleic Acids Res."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"46"},{"type":"text","text":", D1083-D1090 (2018)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"1b605685-8ee7-4b8c-ab7f-681c4e875442","class":null,"textAlign":null},"content":[{"type":"text","text":"Aging Atlas Consortium. Aging Atlas: a multi-omics database for aging biology. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nucleic Acids Res."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"49"},{"type":"text","text":", D825-D830 (2021)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"e0385b0c-2780-4acd-a6a2-b9529d02e579","class":null,"textAlign":null},"content":[{"type":"text","text":"Avelar, R.A. et al. A multidimensional systems biology analysis of cellular senescence in aging and disease. "},{"type":"text","marks":[{"type":"italic"}],"text":"Genome Biol."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"21"},{"type":"text","text":", 91 (2020)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"ed709362-8256-4a1e-be19-a43dffe4f62c","class":null,"textAlign":null},"content":[{"type":"text","text":"Lopes, I. et al. Six drivers of aging identified among genes differentially expressed with age. "},{"type":"text","marks":[{"type":"italic"}],"text":"bioRxiv"},{"type":"text","text":" 2024.08.02.606402 (2024)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"05a0a4cf-f9ef-4703-bf6d-a20b3841d961","class":null,"textAlign":null},"content":[{"type":"text","text":"Law, C.W., Chen, Y., Shi, W. & Smyth, G.K. voom: precision weights unlock linear model analysis tools for RNA-seq read counts. "},{"type":"text","marks":[{"type":"italic"}],"text":"Genome Biol."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"15"},{"type":"text","text":", R29 (2014)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"7442bc8e-12d2-4def-8eb2-f2b9b894e9f9","class":null,"textAlign":null},"content":[{"type":"text","text":"Love, M.I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. "},{"type":"text","marks":[{"type":"italic"}],"text":"Genome Biol."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"15"},{"type":"text","text":", 550 (2014)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"7f090c71-c4d1-4622-814f-d15d472738d2","class":null,"textAlign":null},"content":[{"type":"text","text":"Ritchie, M.E. et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nucleic Acids Res."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"43"},{"type":"text","text":", e47 (2015)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"82ae0bde-b5c2-4fcf-add5-6804264fefc4","class":null,"textAlign":null},"content":[{"type":"text","text":"Tseng, G.C., Ghosh, D. & Feingold, E. Comprehensive literature review and statistical considerations for microarray meta-analysis. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nucleic Acids Res."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"40"},{"type":"text","text":", 3785-3799 (2012)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"265167e8-3d40-40bd-8e88-c7f1d001c7d5","class":null,"textAlign":null},"content":[{"type":"text","text":"Smyth, G.K. Linear models and empirical Bayes methods for assessing differential expression in microarray experiments. "},{"type":"text","marks":[{"type":"italic"}],"text":"Stat. Appl. Genet. Mol. Biol."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"3"},{"type":"text","text":", Article 3 (2004)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"f1585334-a12c-411d-a806-c1d3ef26b905","class":null,"textAlign":null},"content":[{"type":"text","text":"Hedges, L.V. & Olkin, I. "},{"type":"text","marks":[{"type":"italic"}],"text":"Statistical Methods for Meta-Analysis."},{"type":"text","text":" Academic Press (1985)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"7ede985c-c0c7-41cf-8785-380226524bcb","class":null,"textAlign":null},"content":[{"type":"text","text":"Viechtbauer, W. Conducting meta-analyses in R with the metafor package. "},{"type":"text","marks":[{"type":"italic"}],"text":"J. Stat. Softw."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"36"},{"type":"text","text":", 1-48 (2010)."}]}]}]},{"type":"heading","attrs":{"id":"26bf4bf5-4330-492a-8c6a-de997c1f2e08","data-toc-id":"26bf4bf5-4330-492a-8c6a-de997c1f2e08","textAlign":null,"level":2},"content":[{"type":"text","text":"AI Disclosure"}]},{"type":"paragraph","attrs":{"id":"49c8dd7d-923f-423c-9927-491a67650b12","class":null,"textAlign":null},"content":[{"type":"text","text":"Gemini 3 Pro Nano Banana tool was used to generate the graphical abstract based on the proposal given as input. Claude Opus 4.6 was used to draft the initial proposal. The idea, methodology and workplan are completely original ideas of the author. The llm was used for initial draft after which the author then thoroughly vetted and heavily edited the proposal for accuracy and flow of thought and added crucial details. All references have also been cross-validated."}]},{"type":"paragraph","attrs":{"id":"4d625cc6-01b3-4ce1-a1ab-cd1a64a3ed4c","class":null,"textAlign":null}}]}5b:T8650,GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome 1. Overview Scientific Rationale: A search for aging-related terms ("aging", "senescence", "ageing", "lifespan", "longevity") on the NCBI Gene Expression Omnibus returns over 30,000 dataset entries, with approximately 10,000 involving human samples (NCBI E-utilities query, February 2026). Yet no study has systematically mined this resource at scale. Existing aging meta-analyses are limited to manually curated subsets of 25 to 50 datasets (1,2), leaving the vast majority of public data unexploited. Two compounding bottlenecks explain why. First, GEO metadata is notoriously inconsistent: sample ages are buried in free-text descriptions, tissue annotations use non-standard vocabularies, and experimental conditions are logged in idiosyncratic formats (3). Second, and critically, keyword-based GEO searches massively overcount relevant datasets: a study that merely reports upregulation of a senescence-associated pathway (e.g., p16/CDKN2A) in its results section will appear in a "senescence" search even if the study has nothing to do with aging or cellular senescence. This false-positive problem means the true number of directly aging-relevant transcriptomic datasets is likely a fraction of the raw search total, but no systematic triage has ever been performed. This project deploys an LLM-powered curation pipeline, built on Anthropic's Claude 4.5 and above model family, to systematically extract, standardize, and quality-control metadata across all candidate GEO entries, intelligently distinguishing genuinely aging-focused datasets from incidental keyword matches. This enables the first mega meta-analysis of human aging transcriptomes at unprecedented scale. The resulting resource, a curated, machine-readable atlas of aging gene expression, will be released as an open community tool. Timeline: 6 months Month 1-2: LLM pipeline development, GEO-wide metadata extraction, and validation. Month 3-4: Expression data download, harmonization, batch correction, and quality control. Month 5: Meta-analysis and statistical modeling (Random-effects meta-analysis, subgroup analyses, and meta-regression.) Month 6: Cross-validation, interactive visualization, manuscript preparation, and submission Data Sharing: All curated metadata, processed expression matrices, analysis code, and the LLM curation pipeline will be deposited in open repositories (GitHub under MIT license, Zenodo under CC-BY-4.0). The curated metadata catalogue itself will be released as a standalone resource for community reanalysis. An interactive Shiny web application will be deployed for the community to explore meta-analytic results, query gene-level aging signatures across tissues and conditions, and download subsets of interest. The study protocol will be preregistered on ResearchHub prior to analysis. The final manuscript will be published as a ResearchHub Registered Report with regular progress updates shared via the Updates tab throughout the project. 2. Introduction Aging is the dominant risk factor for cardiovascular disease, neurodegeneration, cancer, and metabolic dysfunction, yet the molecular programs that drive age-dependent decline remain incompletely mapped (4,5). Over the past two decades, transcriptomic profiling has emerged as the primary tool for characterizing aging at the molecular level. Individual studies have identified age-associated changes in inflammation, mitochondrial function, proteostasis, and DNA damage response pathways. However, these findings exhibit surprisingly limited overlap across studies, a problem that has persisted despite the growing volume of available data (1,6). Two landmark meta-analyses illustrate both the promise and the ceiling of current approaches. Shokhirev and Johnson reprocessed approximately 6,000 raw RNA-seq samples from the Sequence Read Archive, ultimately retaining 3,060 high-quality human samples after stringent filtering, and achieved remarkable predictive accuracy for transcriptomic age (R² = 0.96) (1). Voisin et al. assembled 3,176 human skeletal muscle samples across methylome and transcriptome datasets to demonstrate that exercise training shifts molecular profiles toward a younger state (7). Both studies relied on extensive manual curation and were restricted to specific tissues or contexts. A recent preprint used only 25 GEO datasets across multiple species to identify six causal aging driver genes, explicitly noting that their small dataset count was a limitation and calling for expanded meta-analyses as more data become available (19). Meanwhile, Palmer et al. demonstrated that transcriptomic aging signatures show minimal overlap across organs, raising the fundamental question of whether truly universal aging markers exist or whether the apparent inconsistency is an artifact of small, heterogeneous datasets (6). The core problem is scale, and specifically the curatorial bottleneck that prevents achieving it. GEO currently hosts over 200,000 series entries comprising more than 6 million individual samples (9). A broad keyword search for aging-related terms returns over 30,000 dataset entries, of which approximately 10,000 involve human samples. However, these raw counts are profoundly misleading. GEO's full-text search indexes not only study titles and designs but also result descriptions, pathway annotations, and associated publications. A cancer genomics study that incidentally reports upregulation of CDKN2A or other senescence-associated genes will appear in a "senescence" search despite having no relevance to aging biology whatsoever. Similarly, studies examining age as a covariate rather than a primary variable, or studies using "aging" only in the context of sample storage or reagent shelf life, inflate the counts. The true number of datasets whose primary focus is aging, cellular senescence, DNA damage in the context of aging, or lifespan modulation is likely an order of magnitude smaller than the raw search suggests, but no systematic triage has been performed, because manual screening of 10,000+ entries is prohibitively labor-intensive. This is precisely the problem that LLM-powered curation pipelines can solve. Large language models excel at reading unstructured text descriptions, understanding experimental intent from context, and making nuanced classification decisions that go beyond keyword matching (11,12). The Critical Path Institute has demonstrated that LLMs can perform arbitrary data cleaning and knowledge extraction tasks dynamically through prompt engineering, achieving results that would otherwise require weeks of manual curation (14). By deploying an LLM pipeline to read every GEO entry's title, summary, experimental design, and sample annotations, we can distinguish genuinely aging-focused datasets from incidental keyword matches while simultaneously extracting and standardizing the metadata needed for downstream integration. This project combines these two threads, the unmet need for large-scale human aging meta-analysis and the emerging capability of LLMs for intelligent data triage and curation, into a unified pipeline we call GEO-Age. The pipeline, built on Anthropic's Claude 4.5 and above model family, will systematically process all candidate GEO entries, classify each for aging relevance, extract and standardize metadata into a machine-readable schema, and feed the curated catalogue into a multi-level random-effects meta-analysis. The scope is restricted to human samples, including primary tissues, primary cell cultures, and cell lines (both normal and cancer-derived), to maximize translational relevance and minimize the confounds introduced by cross-species ortholog mapping. Cancer-derived cell lines are included specifically for the senescence and DNA damage analyses (RQ2, RQ5), where therapy-induced and oncogene-induced senescence are biologically relevant phenomena, but are excluded from the primary chronological aging meta-analysis (RQ1, RQ3) to avoid confounding by baseline transcriptomic aberrations (telomerase reactivation, p53 loss-of-function, aneuploidy). A particular focus will be the relationship between DNA damage responses and chronological aging. Genomic instability is a canonical hallmark of aging, and recent work has shown that genome-wide DNA damage profiles can predict mammalian age with accuracy exceeding transcriptome-based models (15). Hundreds of GEO datasets examine DNA damage induction (via irradiation, oxidative stress, genotoxic agents, or replicative senescence) in human cell lines, yet no systematic comparison has tested whether these acute damage signatures recapitulate the molecular program of chronological aging or represent a distinct biological response. The scale of our meta-analysis is uniquely positioned to resolve this question. Research Questions. RQ1. What molecular signatures of aging are robust across tissues, cell types, and experimental contexts? RQ2. Which aging-associated changes are specific to particular biological systems or perturbations? RQ3. Can meta-analysis identify high-confidence aging biomarkers that replicate across independent studies? RQ4. What factors (technical vs. biological) explain heterogeneity in reported aging signatures? RQ5. To what extent do DNA damage and genotoxic stress signatures overlap with, or diverge from, chronological aging programs? This is a discovery-driven study. Rather than testing a fixed set of hypotheses, the project is designed to generate a comprehensive descriptive atlas of human aging transcriptomics at unprecedented scale, from which specific mechanistic hypotheses can be derived and subsequently tested by the community. The analysis plan (Section 3.3) specifies pre-registered analytical frameworks and decision rules, but the primary output is a curated resource and empirical characterization rather than hypothesis confirmation. Impact: This project delivers three high-value outputs to the aging research community regardless of any specific finding. First, the curated metadata catalogue, a structured, machine-readable annotation of every aging-relevant human GEO dataset, constitutes a standalone resource that will accelerate future meta-analyses across any aging subfield. Second, the GEO-Age LLM pipeline establishes a generalizable methodology for LLM-powered data curation that can be applied to any domain with heterogeneous public data. Third, the mega meta-analysis itself will either identify genuinely universal aging markers (advancing biomarker development) or definitively demonstrate that human aging transcriptomes are context-dependent (redirecting the field toward tissue-specific intervention strategies). The interactive Shiny server will ensure these outputs remain accessible and explorable by the broader community well beyond the funding period. 3. Methods Participants/Sample Target population. All publicly available high-throughput transcriptomic datasets in GEO involving human samples that measure age-dependent molecular changes, cellular senescence, DNA damage responses, or lifespan-modulating interventions in primary tissues, primary cell cultures, or cell lines (normal and cancer-derived). Inclusion criteria. (i) Human samples (primary tissues, primary cells, or established cell lines). (ii) Study's primary focus involves aging, cellular senescence, DNA damage response, or age-modulating intervention, as determined by the GEO-Age pipeline's relevance classification, not merely keyword co-occurrence. (iii) At least 3 biological replicates per experimental condition. (iv) Raw or processed expression data available for download. (v) Sufficient metadata to assign age/condition groups (after AI-assisted extraction). Exclusion criteria. (i) Non-human samples. (ii) Datasets where aging/senescence terms appear only incidentally (e.g., pathway annotations in results, age as a minor covariate) rather than as the study's primary focus. (iii) Fewer than 3 biological replicates per group. (iv) Metadata insufficient for age/condition assignment even after AI extraction and manual review of the original publication. (v) Studies of age-associated diseases (e.g., Alzheimer's, cancer prognosis) without age-matched healthy controls or explicit aging comparisons. (vi) Datasets flagged for known technical failures in GEO. (vii) Single-cell RNA-seq datasets, which require distinct analytical frameworks (pseudobulk aggregation, cell-type deconvolution) that are outside the scope of this bulk transcriptomics meta-analysis; extension to scRNA-seq is planned as future work pending proof of concept. Expected sample size. Approximately 10,000 human GEO entries match aging-related keyword searches. Based on the high rate of false-positive keyword matches inherent to GEO's full-text indexing, we estimate that 1,000 to 2,500 of these represent genuinely aging-focused studies after AI-assisted relevance filtering. After applying the remaining inclusion/exclusion criteria (replication, data availability), we conservatively estimate 500 to 1,500 high-quality transcriptomic datasets representing 20,000 to 60,000 individual samples. For meta-analysis, a minimum of 20 independent studies per subgroup analysis is required to achieve stable random-effects estimates; we expect to exceed this threshold for all major tissue categories. 3.2 Materials and Procedures Phase 1: LLM Pipeline for Metadata Extraction and Relevance Classification (Months 1 to 2) The GEO-Age pipeline is a multi-step LLM classification and extraction system built on Anthropic's Claude 4.5 and above model family that performs two critical functions: (a) relevance classification, meaning determining whether a GEO entry is genuinely focused on aging/senescence versus incidentally matching keywords, and (b) structured metadata extraction for entries classified as relevant. Step 1: Retrieval. Use the NCBI E-utilities API (esearch + efetch) and the GEOquery R package to programmatically download all human GEO Series (GSE) records matching aging-related search terms ("aging", "ageing", "senescence", "cellular senescence", "lifespan", "longevity", "gerontology", "DNA damage aging", "replicative senescence", "oxidative stress aging"). For each entry, retrieve: title, summary, overall design, sample characteristics (GSM-level), platform annotation, publication DOI, and supplementary file links. Store raw metadata in a structured JSON schema. Step 2: AI-Powered Relevance Classification. For each GEO entry, submit the concatenated metadata fields to the Claude API with a classification prompt that instructs the model to assign one of four relevance categories: Primary aging study: aging, senescence, or DNA damage in the context of aging is the central research question. Secondary aging study: age is a major covariate or secondary variable, with sufficient aging data for extraction. Incidental mention: aging/senescence terms appear only in pathway annotations, background literature, or results interpretation, but the study does not investigate aging. Ambiguous: insufficient information to classify; flagged for manual review. Only entries classified as "primary" or "secondary" proceed to metadata extraction. "Incidental mention" entries are excluded. "Ambiguous" entries are flagged for manual review by the research assistant, with the original publication abstract retrieved via DOI to aid classification. Step 3: Structured Metadata Extraction. For entries passing relevance classification, submit to the Claude API with a structured extraction prompt specifying the following target fields: Tissue/cell type (mapped to Cell Ontology for cell types, Uberon for tissues, or CLO for cell lines; with explicit annotation of whether the sample is primary tissue, primary cells, or a cell line, and for cell lines whether normal or cancer-derived) Age or time point (normalized to years for humans) Sex (male, female, mixed, not reported) Experimental condition (natural aging, accelerated aging/progeroid model, senescence induction, DNA damage treatment, lifespan intervention, control) Intervention details (agent, dose, duration where applicable) Sample size (number of biological replicates per condition) Sequencing platform (microarray model or RNA-seq chemistry) Batch structure (single batch, multi-batch, unknown) The prompt enforces JSON output conforming to a predefined schema with constrained value sets for categorical fields. When the model cannot confidently extract a field, it returns "uncertain" with a free-text explanation, flagging the entry for manual review. Step 4: Validation. Three validation layers ensure accuracy, targeting ≥95% field-level agreement with human annotation and aiming for ≥99% on critical fields (relevance classification, tissue type, age): (a) Automated consistency checks: Cross-validate extracted organism against the GEO platform annotation. Verify that reported ages fall within plausible human ranges. Flag entries where the model's output contradicts structured GEO fields. (b) Independent LLM re-classification: For entries flagged as uncertain or where automated checks detect inconsistencies, re-submit to a separate API call with the original publication abstract (retrieved via DOI) appended to the prompt, enabling the model to resolve ambiguities from richer context. Because each API call operates without memory of prior classifications, this constitutes a functionally independent re-classification rather than self-correction. Concordance rates between first-pass and re-classification will be reported as an empirical measure of extraction reliability. (c) Human spot-check: A random 10% sample of extracted entries will be manually validated against the original publications by a trained research assistant with graduate-level experience in aging biology and GEO data structures. If field-level accuracy falls below 95% on any target field, the extraction prompt will be iteratively refined and the affected batch re-processed. All ambiguous entries flagged by the pipeline will be manually reviewed regardless of the spot-check results. Cohen's kappa will be computed between LLM and human annotations for the four-category relevance classification to quantify inter-rater reliability. Step 5: Ontology Mapping and Standardization. All extracted terms are mapped to controlled vocabularies using the EMBL-EBI Ontology Lookup Service API: tissue terms to Uberon, cell types to Cell Ontology, diseases to Disease Ontology, experimental interventions to the Experimental Factor Ontology (EFO). Unmapped terms are flagged for manual resolution. Phase 2: Expression Data Harmonization (Months 3 to 4) Step 1: Data Download. Download raw or normalized expression matrices for all included datasets. For microarray data, download CEL files where available; otherwise, use pre-processed series matrices. For RNA-seq, download count matrices from GEO Supplementary Files or re-quantify from FASTQ via the Recount3 resource where available. Step 2: Preprocessing and Normalization. Apply platform-appropriate normalization. For Affymetrix microarrays, apply RMA normalization via the oligo R package. For two-color arrays, apply loess within-array normalization followed by quantile between-array normalization via limma. For RNA-seq count data, apply the limma-voom transformation pipeline (20), which converts raw counts to log-counts-per-million with observation-level precision weights. We use limma-voom rather than DESeq2 as the primary RNA-seq framework because DESeq2's negative binomial GLM becomes computationally prohibitive beyond approximately 10,000 samples (21), whereas limma-voom scales efficiently to mega-analysis (20,22). Probe-to-gene mapping uses manufacturer annotation files with ambiguous probes removed. Gene identifiers are harmonized to HGNC symbols with Ensembl gene IDs as a secondary namespace. Step 3: Quality Control. For each dataset: (a) Remove outlier samples with inter-sample correlation < 0.8 (Pearson r on global expression profile). (b) Assess batch effects via principal component analysis; flag datasets where the first principal component separates by technical batch rather than biological condition. (c) Apply ComBat-seq for within-study batch correction where batch information is available. Step 4: Per-Dataset Differential Expression (Stage 1 of Two-Stage Meta-Analysis). Each dataset is analyzed independently to extract gene-level effect sizes (23). For microarray datasets, fit gene-wise linear models using limma's empirical Bayes moderated t-statistics (24). For RNA-seq datasets, apply the limma-voom pipeline: filter lowly expressed genes (requiring at least 10 counts in a minimum number of samples), apply the voom transformation, and fit gene-wise weighted linear models with empirical Bayes moderation (20). For studies with discrete age groups (young vs. old), the model is a simple contrast between groups. For studies with continuous age, fit a linear regression of expression on age. Where metadata permits, include sex, batch, and tissue subregion as covariates in the per-dataset model to reduce confounding. For each gene, extract the log2 fold-change (logFC) and its standard error. To combine continuous and categorical contrasts in a single meta-analysis, all effect sizes are converted to a common scale: the standardized mean difference (Hedges' g), computed from the logFC divided by the pooled within-group standard deviation, with a small-sample correction factor (25). For discrete age-group comparisons, this conversion is direct. For continuous age designs (regression of expression on age), the regression slope is converted to an equivalent standardized mean difference using the age range of each study to define the contrast (youngest versus oldest quartile), with the residual standard deviation of the regression as the denominator. Where the age range or residual SD cannot be recovered from GEO metadata, the dataset is flagged for manual extraction from the original publication. A sensitivity analysis will compare meta-analytic results when restricted to discrete-group studies only versus including converted continuous designs, to assess robustness to conversion assumptions. Phase 3: Random-Effects Meta-Analysis (Month 5) Stage 2: Combining Effect Sizes Across Datasets. For each gene, combine the per-dataset effect sizes using a random-effects meta-analysis with REML estimation of between-study variance via the metafor R package (26). REML is preferred over the DerSimonian-Laird moment estimator, which underestimates between-study variance under substantial heterogeneity (23,26). For each gene, compute: (a) the summary effect size (weighted mean Hedges' g) and its 95% confidence interval, (b) Cochran's Q statistic and I² for heterogeneity, and (c) the prediction interval. Apply Benjamini-Hochberg FDR correction across all tested genes (q < 0.05). Computational Strategy. Both stages are embarrassingly parallel. The per-dataset pipeline is orchestrated via Snakemake on the IISc HPC cluster (or equivalent GCP instances). Stage 2 parallelizes trivially across genes via BiocParallel. The computational bottleneck is I/O (downloading and parsing thousands of GEO files), not model fitting. Analysis for RQ1 and RQ3: conserved aging signatures and biomarker identification. Using the genome-wide REML random-effects results from Stage 2, restricted to primary tissues, primary cell cultures, and normal cell lines (excluding cancer-derived cell lines to avoid confounding by oncogenic transcriptomic aberrations), identify genes with consistent direction of effect and q < 0.05 across a high proportion of datasets as candidate conserved aging markers. Compare against the Human Ageing Genomic Resources (GenAge, CellAge) and the Aging Atlas (16,17,18). Assess replication by leave-one-study-out cross-validation and by testing whether top markers replicate in held-out datasets not included in the meta-analysis. Analysis for RQ2: Tissue-specific and perturbation-specific signatures. Stratify datasets by sample type: (a) primary tissues by category (immune/hematopoietic, epithelial, neural, muscle, adipose, other), (b) primary cell cultures, (c) normal cell lines, (d) cancer cell lines. Perform separate meta-analyses within each stratum. Compute pairwise Spearman correlations of gene-level effect sizes between strata. Apply hierarchical clustering on the correlation matrix to identify natural groupings of aging signatures. Perform pathway enrichment (KEGG, Reactome, GO) within each cluster to characterize biological themes. Analysis for RQ4: Sources of heterogeneity. Perform subgroup meta-analyses stratified by: sequencing platform (microarray vs. RNA-seq), tissue category, sample size per group (above/below median), and publication year (pre-2015 vs. post-2015). Report I² within each stratum to assess whether heterogeneity is driven by technical or biological factors. Use leave-one-study-out sensitivity analysis to identify influential datasets. Analysis for RQ5: DNA damage versus chronological aging. Separate datasets into two groups: (a) chronological aging studies (natural aging across timepoints) and (b) DNA damage/genotoxic stress studies (irradiation, oxidative stress, replicative senescence), including cancer-derived cell lines undergoing therapy-induced or DNA damage-induced senescence. Perform meta-analysis within each group. To compare the two signatures, apply three complementary approaches: (i) rank-rank hypergeometric overlap (RRHO) analysis, which provides a threshold-free visualization of concordance and discordance across the full ranked gene lists without arbitrary significance cutoffs; (ii) Spearman ρ restricted to genes reaching nominal significance (p < 0.05) in at least one of the two categories, to avoid inflating the correlation through shared null effects across thousands of uninformative genes; and (iii) pathway-level comparison using gene set enrichment analysis (GSEA) against KEGG and Reactome, testing whether shared and divergent pathways align with predictions from the hallmarks of aging framework (4). Sensitivity analyses. (i) Restrict to studies with n ≥ 5 per group. (ii) Restrict to primary tissues only (exclude cell lines). (iii) Stratify by publication year (pre-2015 microarray era vs. post-2015 RNA-seq era). (iv) Funnel plots and Egger's test for publication bias. (v) Restrict to discrete age-group designs only (excluding continuous-age regression designs) to assess robustness to the Hedges' g conversion procedure. Phase 4: Validation and Deliverables (Month 6) Cross-reference validation. Compare meta-analytic aging gene sets against: (a) the GenAge database of aging-related genes (16), (b) the CellAge database of cellular senescence genes (18), (c) the Aging Atlas multi-omics resource (17), and (d) recently published independent aging cohorts not included in the meta-analysis. Quantify overlap using hypergeometric enrichment tests. Interactive Shiny server. Develop and deploy an interactive Shiny web application enabling users to: query the curated metadata catalogue; visualize gene-level meta-analytic results across tissues, cell types, and conditions; explore tissue-specific and perturbation-specific aging signatures; and download subsets of interest for their own analyses. The server will be deployed via shinyapps.io or a containerized instance and maintained as a community resource. Open resource release. Deposit the following on Zenodo with DOI: (a) complete curated metadata catalogue (JSON + CSV), (b) processed expression matrices with harmonized annotations, (c) full meta-analytic results for all genes across all subgroup analyses, (d) the GEO-Age LLM pipeline code and prompt library (GitHub, MIT license), (e) Docker container for full pipeline reproducibility. 3.3 Ethics and Data Management Ethics. This study exclusively reanalyzes publicly available, de-identified datasets from GEO. No new human subjects research, animal experimentation, or data collection involving identifiable individuals is conducted. All source datasets comply with the original studies' ethics approvals and GEO's data sharing policies. Data Management. All analysis code will be version-controlled on GitHub with MIT license from project initiation. The curated metadata catalogue and processed datasets will be deposited on Zenodo with a persistent DOI upon completion. Analysis pipelines will be containerized using Docker for full reproducibility. The final manuscript will be published as a ResearchHub Registered Report. The study protocol will be preregistered on ResearchHub prior to initiating the meta-analysis phase, per the RFP requirements. 4. Budget Request: $10,000 USD (in ResearchCoin) Item Cost Justification Cloud computing (GCP) $3,000 Data storage, bulk expression data processing, meta-analysis computation across thousands of datasets LLM API costs (Claude 4.5 Sonnet + API) $2,000 Relevance classification and metadata extraction for 10,000 GEO entries with multi-step prompting ($0.20/dataset including validation passes) Research assistant (100 hrs × $25/hr) $2,500 Graduate-level research assistant with aging biology expertise for manual validation of 10% metadata sample, review of ambiguous entries flagged by pipeline, literature cross-referencing Interactive Shiny server (development + deployment) $1,000 Web application development and hosting for community access to curated atlas and meta-analytic results ResearchHub Registered Report publication $1,000 Open-access publication of final manuscript as ResearchHub Registered Report Contingency $500 Additional compute for expanded analyses, unforeseen API cost overruns Budget is flexible and will be adjusted based on Phase 1 results. If metadata extraction proves more efficient than estimated, surplus funds will be allocated to expanding the analysis to additional sample types, scRNA-Seq or deeper pathway-level analyses. 6. References Shokhirev MN, Johnson AA. Modeling the human aging transcriptome across tissues, health status, and sex. Aging Cell. 2021;20(1):e13280. doi:10.1111/acel.13280 de Magalhães JP, Curado J, Church GM. Meta-analysis of age-related gene expression profiles identifies common signatures of aging. Bioinformatics. 2009;25(7):875-881. doi:10.1093/bioinformatics/btp073 Sansone, S.A. et al. Toward interoperable bioscience data. Nat. Genet. 44, 121-126 (2012). López-Otín, C., Blasco, M.A., Partridge, L., Serrano, M. & Kroemer, G. Hallmarks of aging: an expanding universe. Cell 186, 243-278 (2023). Aramillo Irizar, P. et al. Transcriptomic alterations during ageing reflect the shift from cancer to degenerative diseases in the elderly. Nat. Commun. 9, 327 (2018). Palmer, D. et al. Ageing transcriptome meta-analysis reveals similarities and differences between key mammalian tissues. Aging 13, 3313-3341 (2021). Voisin, S. et al. Exercise is associated with younger methylome and transcriptome profiles in human skeletal muscle. Aging Cell 23, e13859 (2024). Winterhalter, P.R. et al. The circadian rhythm: a key variable in aging? Aging Cell 23, e14268 (2024). Barrett, T. et al. NCBI GEO: archive for functional genomics data sets, update. Nucleic Acids Res. 41, D991-D995 (2013). Brazma, A. et al. Minimum information about a microarray experiment (MIAME): toward standards for microarray data. Nat. Genet. 29, 365-371 (2001). Lee, P. et al. Benefits, limits, and risks of GPT-4 as an AI chatbot for medicine. N. Engl. J. Med. 388, 1233-1239 (2023). Jiang, L.Y. et al. Health system-scale language models are all-purpose prediction engines. Nature 619, 357-362 (2023). Churchill, A. et al. GPT models for text annotation: an empirical exploration in public policy research. Policy Stud. J. (2025). Anderson, W. et al. Unlocking the capabilities of large language models for accelerating drug development. Clin. Pharmacol. Ther. 116, 38-41 (2024). Cao, H. et al. Genome-wide profiles of DNA damage represent highly accurate predictors of mammalian age. Aging Cell 23, e14122 (2024). Tacutu, R. et al. Human Ageing Genomic Resources: new and updated databases. Nucleic Acids Res. 46, D1083-D1090 (2018). Aging Atlas Consortium. Aging Atlas: a multi-omics database for aging biology. Nucleic Acids Res. 49, D825-D830 (2021). Avelar, R.A. et al. A multidimensional systems biology analysis of cellular senescence in aging and disease. Genome Biol. 21, 91 (2020). Lopes, I. et al. Six drivers of aging identified among genes differentially expressed with age. bioRxiv 2024.08.02.606402 (2024). Law, C.W., Chen, Y., Shi, W. & Smyth, G.K. voom: precision weights unlock linear model analysis tools for RNA-seq read counts. Genome Biol. 15, R29 (2014). Love, M.I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550 (2014). Ritchie, M.E. et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 43, e47 (2015). Tseng, G.C., Ghosh, D. & Feingold, E. Comprehensive literature review and statistical considerations for microarray meta-analysis. Nucleic Acids Res. 40, 3785-3799 (2012). Smyth, G.K. Linear models and empirical Bayes methods for assessing differential expression in microarray experiments. Stat. Appl. Genet. Mol. Biol. 3, Article 3 (2004). Hedges, L.V. & Olkin, I. Statistical Methods for Meta-Analysis. Academic Press (1985). Viechtbauer, W. Conducting meta-analyses in R with the metafor package. J. Stat. Softw. 36, 1-48 (2010). AI Disclosure Gemini 3 Pro Nano Banana tool was used to generate the graphical abstract based on the proposal given as input. Claude Opus 4.6 was used to draft the initial proposal. The idea, methodology and workplan are completely original ideas of the author. The llm was used for initial draft after which the author then thoroughly vetted and heavily edited the proposal for accuracy and flow of thought and added crucial details. All references have also been cross-validated. 5c:T107eb,{"type":"doc","content":[{"type":"heading","attrs":{"id":"b59c6651-cf57-49db-bbc7-e22266af2b29","data-toc-id":"b59c6651-cf57-49db-bbc7-e22266af2b29","textAlign":"left","level":1},"content":[{"type":"text","text":"GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome"}]},{"type":"heading","attrs":{"id":"55c7df49-e63f-48de-8009-a916d718ef08","data-toc-id":"55c7df49-e63f-48de-8009-a916d718ef08","textAlign":"left","level":2},"content":[{"type":"text","text":"1. Overview"}]},{"type":"paragraph","attrs":{"id":"77650712-e750-4513-a159-6f543c0997de","class":null,"textAlign":"left"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Scientific Rationale:"},{"type":"text","text":" A search for aging-related terms (\"aging\", \"senescence\", \"ageing\", \"lifespan\", \"longevity\") on the NCBI Gene Expression Omnibus returns over 30,000 dataset entries, with approximately 10,000 involving human samples (NCBI E-utilities query, February 2026). Yet no study has systematically mined this resource at scale. Existing aging meta-analyses are limited to manually curated subsets of 25 to 50 datasets (1,2), leaving the vast majority of public data unexploited. Two compounding bottlenecks explain why. First, GEO metadata is notoriously inconsistent: sample ages are buried in free-text descriptions, tissue annotations use non-standard vocabularies, and experimental conditions are logged in idiosyncratic formats (3). Second, and critically, keyword-based GEO searches massively overcount relevant datasets: a study that merely reports upregulation of a senescence-associated pathway (e.g., p16/CDKN2A) in its results section will appear in a \"senescence\" search even if the study has nothing to do with aging or cellular senescence. This false-positive problem means the true number of directly aging-relevant transcriptomic datasets is likely a fraction of the raw search total, but no systematic triage has ever been performed. This project deploys an LLM-powered curation pipeline, built on Anthropic's Claude 4.5 and above model family, to systematically extract, standardize, and quality-control metadata across all candidate GEO entries, intelligently distinguishing genuinely aging-focused datasets from incidental keyword matches. This enables the first mega meta-analysis of human aging transcriptomes at unprecedented scale. The resulting resource, a curated, machine-readable atlas of aging gene expression, will be released as an open community tool."}]},{"type":"paragraph","attrs":{"id":"46dd1b14-1162-499d-be14-0e0027917948","class":null,"textAlign":"left"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Timeline:"},{"type":"text","text":" 6 months"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"638e6adf-aa73-44ca-9096-9a90f90c3e6f","class":null,"textAlign":null},"content":[{"type":"text","text":"Month 1-2: LLM pipeline development, GEO-wide metadata extraction, and validation."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"3855c507-f68d-4782-9d2a-981d5581de79","class":null,"textAlign":null},"content":[{"type":"text","text":"Month 3-4: Expression data download, harmonization, batch correction, and quality control."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"4dc20135-c59b-4525-8918-1bafeffa588a","class":null,"textAlign":null},"content":[{"type":"text","text":"Month 5: Meta-analysis and statistical modeling (Random-effects meta-analysis, subgroup analyses, and meta-regression.)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"eec655e6-aaa1-4cbc-8dc7-e4e0e57ce4cb","class":null,"textAlign":null},"content":[{"type":"text","text":"Month 6: Cross-validation, interactive visualization, manuscript preparation, and submission"}]}]}]},{"type":"paragraph","attrs":{"id":"8a57baff-8eef-48ba-a171-7d41284ddad7","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Data Sharing:"},{"type":"text","text":" All curated metadata, processed expression matrices, analysis code, and the LLM curation pipeline will be deposited in open repositories (GitHub under MIT license, Zenodo under CC-BY-4.0). The curated metadata catalogue itself will be released as a standalone resource for community reanalysis. An interactive Shiny web application will be deployed for the community to explore meta-analytic results, query gene-level aging signatures across tissues and conditions, and download subsets of interest. The study protocol will be preregistered on ResearchHub prior to analysis. The final manuscript will be published as a ResearchHub Registered Report with regular progress updates shared via the Updates tab throughout the project."}]},{"type":"heading","attrs":{"id":"983b6250-2e54-44bc-b9a4-037cc0fcc3bc","data-toc-id":"983b6250-2e54-44bc-b9a4-037cc0fcc3bc","textAlign":"left","level":2},"content":[{"type":"text","text":"2. Introduction"}]},{"type":"paragraph","attrs":{"id":"33cac8a3-842b-4799-8106-c72987020ab1","class":null,"textAlign":"left"},"content":[{"type":"text","text":"Aging is the dominant risk factor for cardiovascular disease, neurodegeneration, cancer, and metabolic dysfunction, yet the molecular programs that drive age-dependent decline remain incompletely mapped (4,5). Over the past two decades, transcriptomic profiling has emerged as the primary tool for characterizing aging at the molecular level. Individual studies have identified age-associated changes in inflammation, mitochondrial function, proteostasis, and DNA damage response pathways. However, these findings exhibit surprisingly limited overlap across studies, a problem that has persisted despite the growing volume of available data (1,6)."}]},{"type":"paragraph","attrs":{"id":"1bdcf51e-e987-4706-b60e-091698880d55","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"Two landmark meta-analyses illustrate both the promise and the ceiling of current approaches. Shokhirev and Johnson reprocessed approximately 6,000 raw RNA-seq samples from the Sequence Read Archive, ultimately retaining 3,060 high-quality human samples after stringent filtering, and achieved remarkable predictive accuracy for transcriptomic age (R² = 0.96) (1). Voisin et al. assembled 3,176 human skeletal muscle samples across methylome and transcriptome datasets to demonstrate that exercise training shifts molecular profiles toward a younger state (7). Both studies relied on extensive manual curation and were restricted to specific tissues or contexts. A recent preprint used only 25 GEO datasets across multiple species to identify six causal aging driver genes, explicitly noting that their small dataset count was a limitation and calling for expanded meta-analyses as more data become available (19). Meanwhile, Palmer et al. demonstrated that transcriptomic aging signatures show minimal overlap across organs, raising the fundamental question of whether truly universal aging markers exist or whether the apparent inconsistency is an artifact of small, heterogeneous datasets (6)."}]},{"type":"paragraph","attrs":{"id":"336da42a-3218-44f5-ba47-96057a4da700","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"The core problem is scale, and specifically the curatorial bottleneck that prevents achieving it. GEO currently hosts over 200,000 series entries comprising more than 6 million individual samples (9). A broad keyword search for aging-related terms returns over 30,000 dataset entries, of which approximately 10,000 involve human samples. However, these raw counts are profoundly misleading. GEO's full-text search indexes not only study titles and designs but also result descriptions, pathway annotations, and associated publications. A cancer genomics study that incidentally reports upregulation of CDKN2A or other senescence-associated genes will appear in a \"senescence\" search despite having no relevance to aging biology whatsoever. Similarly, studies examining age as a covariate rather than a primary variable, or studies using \"aging\" only in the context of sample storage or reagent shelf life, inflate the counts. The true number of datasets whose primary focus is aging, cellular senescence, DNA damage in the context of aging, or lifespan modulation is likely an order of magnitude smaller than the raw search suggests, but no systematic triage has been performed, because manual screening of 10,000+ entries is prohibitively labor-intensive."}]},{"type":"paragraph","attrs":{"id":"8de9212e-7917-4965-87c7-1442e3faebd8","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"This is precisely the problem that LLM-powered curation pipelines can solve. Large language models excel at reading unstructured text descriptions, understanding experimental intent from context, and making nuanced classification decisions that go beyond keyword matching (11,12). The Critical Path Institute has demonstrated that LLMs can perform arbitrary data cleaning and knowledge extraction tasks dynamically through prompt engineering, achieving results that would otherwise require weeks of manual curation (14). By deploying an LLM pipeline to read every GEO entry's title, summary, experimental design, and sample annotations, we can distinguish genuinely aging-focused datasets from incidental keyword matches while simultaneously extracting and standardizing the metadata needed for downstream integration."}]},{"type":"paragraph","attrs":{"id":"4b12838e-7e03-47ae-a960-0dd57bca9581","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"This project combines these two threads, the unmet need for large-scale human aging meta-analysis and the emerging capability of LLMs for intelligent data triage and curation, into a unified pipeline we call "},{"type":"text","marks":[{"type":"bold"}],"text":"GEO-Age"},{"type":"text","text":". The pipeline, built on Anthropic's Claude 4.5 and above model family, will systematically process all candidate GEO entries, classify each for aging relevance, extract and standardize metadata into a machine-readable schema, and feed the curated catalogue into a multi-level random-effects meta-analysis. The scope is restricted to human samples, including primary tissues, primary cell cultures, and cell lines (both normal and cancer-derived), to maximize translational relevance and minimize the confounds introduced by cross-species ortholog mapping. Cancer-derived cell lines are included specifically for the senescence and DNA damage analyses (RQ2, RQ5), where therapy-induced and oncogene-induced senescence are biologically relevant phenomena, but are excluded from the primary chronological aging meta-analysis (RQ1, RQ3) to avoid confounding by baseline transcriptomic aberrations (telomerase reactivation, p53 loss-of-function, aneuploidy)."}]},{"type":"paragraph","attrs":{"id":"2edb4406-f619-4f6e-b05f-64ea4c988ee0","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"A particular focus will be the relationship between DNA damage responses and chronological aging. Genomic instability is a canonical hallmark of aging, and recent work has shown that genome-wide DNA damage profiles can predict mammalian age with accuracy exceeding transcriptome-based models (15). Hundreds of GEO datasets examine DNA damage induction (via irradiation, oxidative stress, genotoxic agents, or replicative senescence) in human cell lines, yet no systematic comparison has tested whether these acute damage signatures recapitulate the molecular program of chronological aging or represent a distinct biological response. The scale of our meta-analysis is uniquely positioned to resolve this question."}]},{"type":"paragraph","attrs":{"id":"560c1f14-fa8b-4014-90c5-ff5301939233","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Research Questions."}]},{"type":"paragraph","attrs":{"id":"749c02e7-c7dd-4c28-9c37-a1e1c9e8c258","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"RQ1. What molecular signatures of aging are robust across tissues, cell types, and experimental contexts?"}]},{"type":"paragraph","attrs":{"id":"d6a5caaa-6ce2-4657-850c-02daa4c64930","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"RQ2. Which aging-associated changes are specific to particular biological systems or perturbations?"}]},{"type":"paragraph","attrs":{"id":"6a41e894-ac2f-4793-9365-019dd1e0469e","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"RQ3. Can meta-analysis identify high-confidence aging biomarkers that replicate across independent studies?"}]},{"type":"paragraph","attrs":{"id":"a8566f0a-45ba-472e-9887-709ded377e36","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"RQ4. What factors (technical vs. biological) explain heterogeneity in reported aging signatures?"}]},{"type":"paragraph","attrs":{"id":"6bc32221-c1e5-4753-99aa-93199578ca2c","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"RQ5. To what extent do DNA damage and genotoxic stress signatures overlap with, or diverge from, chronological aging programs?"}]},{"type":"paragraph","attrs":{"id":"7cd5568e-ff0d-4a1f-889e-1b54775442f6","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"This is a discovery-driven study. Rather than testing a fixed set of hypotheses, the project is designed to generate a comprehensive descriptive atlas of human aging transcriptomics at unprecedented scale, from which specific mechanistic hypotheses can be derived and subsequently tested by the community. The analysis plan (Section 3.3) specifies pre-registered analytical frameworks and decision rules, but the primary output is a curated resource and empirical characterization rather than hypothesis confirmation."}]},{"type":"paragraph","attrs":{"id":"64298fff-d0a8-4cd8-922f-a9a00820f243","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Impact:"},{"type":"text","text":" This project delivers three high-value outputs to the aging research community regardless of any specific finding. First, the curated metadata catalogue, a structured, machine-readable annotation of every aging-relevant human GEO dataset, constitutes a standalone resource that will accelerate future meta-analyses across any aging subfield. Second, the GEO-Age LLM pipeline establishes a generalizable methodology for LLM-powered data curation that can be applied to any domain with heterogeneous public data. Third, the mega meta-analysis itself will either identify genuinely universal aging markers (advancing biomarker development) or definitively demonstrate that human aging transcriptomes are context-dependent (redirecting the field toward tissue-specific intervention strategies). The interactive Shiny server will ensure these outputs remain accessible and explorable by the broader community well beyond the funding period."}]},{"type":"heading","attrs":{"id":"b7d5e9cd-19b2-4ab0-b478-4d80cf9ac123","data-toc-id":"b7d5e9cd-19b2-4ab0-b478-4d80cf9ac123","textAlign":"left","level":2},"content":[{"type":"text","text":"3. Methods"}]},{"type":"heading","attrs":{"id":"4e97bc7b-b93b-4613-8a26-fbb19b78c721","data-toc-id":"4e97bc7b-b93b-4613-8a26-fbb19b78c721","textAlign":"left","level":3},"content":[{"type":"text","text":"Participants/Sample"}]},{"type":"paragraph","attrs":{"id":"77be43ec-69c4-4b1c-b3ae-d4490721342c","class":null,"textAlign":"left"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Target population. "},{"type":"text","text":"All publicly available high-throughput transcriptomic datasets in GEO involving human samples that measure age-dependent molecular changes, cellular senescence, DNA damage responses, or lifespan-modulating interventions in primary tissues, primary cell cultures, or cell lines (normal and cancer-derived)."}]},{"type":"paragraph","attrs":{"id":"c2eb7768-590f-4061-9164-22716a891954","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Inclusion criteria."},{"type":"text","text":" (i) Human samples (primary tissues, primary cells, or established cell lines). (ii) Study's primary focus involves aging, cellular senescence, DNA damage response, or age-modulating intervention, as determined by the GEO-Age pipeline's relevance classification, not merely keyword co-occurrence. (iii) At least 3 biological replicates per experimental condition. (iv) Raw or processed expression data available for download. (v) Sufficient metadata to assign age/condition groups (after AI-assisted extraction)."}]},{"type":"paragraph","attrs":{"id":"07deff5a-b6d6-4fb9-8e37-fc087e5f4dcf","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Exclusion criteria."},{"type":"text","text":" (i) Non-human samples. (ii) Datasets where aging/senescence terms appear only incidentally (e.g., pathway annotations in results, age as a minor covariate) rather than as the study's primary focus. (iii) Fewer than 3 biological replicates per group. (iv) Metadata insufficient for age/condition assignment even after AI extraction and manual review of the original publication. (v) Studies of age-associated diseases (e.g., Alzheimer's, cancer prognosis) without age-matched healthy controls or explicit aging comparisons. (vi) Datasets flagged for known technical failures in GEO. (vii) Single-cell RNA-seq datasets, which require distinct analytical frameworks (pseudobulk aggregation, cell-type deconvolution) that are outside the scope of this bulk transcriptomics meta-analysis; extension to scRNA-seq is planned as future work pending proof of concept."}]},{"type":"paragraph","attrs":{"id":"e937efc5-cfc7-46a0-9c8a-e5ecf4f6c3da","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Expected sample size."},{"type":"text","text":" Approximately 10,000 human GEO entries match aging-related keyword searches. Based on the high rate of false-positive keyword matches inherent to GEO's full-text indexing, we estimate that 1,000 to 2,500 of these represent genuinely aging-focused studies after AI-assisted relevance filtering. After applying the remaining inclusion/exclusion criteria (replication, data availability), we conservatively estimate 500 to 1,500 high-quality transcriptomic datasets representing 20,000 to 60,000 individual samples. For meta-analysis, a minimum of 20 independent studies per subgroup analysis is required to achieve stable random-effects estimates; we expect to exceed this threshold for all major tissue categories."}]},{"type":"heading","attrs":{"id":"cc2fa152-2040-4448-abda-52c549c4fd96","data-toc-id":"cc2fa152-2040-4448-abda-52c549c4fd96","textAlign":null,"level":3},"content":[{"type":"text","text":"3.2 Materials and Procedures"}]},{"type":"paragraph","attrs":{"id":"92b86816-f5b8-430e-892b-4eb090825dc3","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Phase 1: LLM Pipeline for Metadata Extraction and Relevance Classification (Months 1 to 2)"}]},{"type":"paragraph","attrs":{"id":"3ce468a0-8c7e-4ca9-9a24-99a17f1509a2","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"The GEO-Age pipeline is a multi-step LLM classification and extraction system built on Anthropic's Claude 4.5 and above model family that performs two critical functions: (a) relevance classification, meaning determining whether a GEO entry is genuinely focused on aging/senescence versus incidentally matching keywords, and (b) structured metadata extraction for entries classified as relevant."}]},{"type":"paragraph","attrs":{"id":"daf7de2a-ef12-44a3-ae6a-6a0ecaf6c5de","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 1: Retrieval."},{"type":"text","text":" Use the NCBI E-utilities API (esearch + efetch) and the GEOquery R package to programmatically download all human GEO Series (GSE) records matching aging-related search terms (\"aging\", \"ageing\", \"senescence\", \"cellular senescence\", \"lifespan\", \"longevity\", \"gerontology\", \"DNA damage aging\", \"replicative senescence\", \"oxidative stress aging\"). For each entry, retrieve: title, summary, overall design, sample characteristics (GSM-level), platform annotation, publication DOI, and supplementary file links. Store raw metadata in a structured JSON schema."}]},{"type":"paragraph","attrs":{"id":"95a4f3ac-81fe-41d8-badc-628d5facae6e","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 2: AI-Powered Relevance Classification."},{"type":"text","text":" For each GEO entry, submit the concatenated metadata fields to the Claude API with a classification prompt that instructs the model to assign one of four relevance categories:"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"ece197b3-349f-4058-a3c6-2f3ecba6552b","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Primary aging study:"},{"type":"text","text":" aging, senescence, or DNA damage in the context of aging is the central research question."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"897a5934-88bd-40ce-a712-e265a6745da0","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Secondary aging study:"},{"type":"text","text":" age is a major covariate or secondary variable, with sufficient aging data for extraction."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"936de766-d6d6-4628-8e4c-56808d0a4739","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Incidental mention:"},{"type":"text","text":" aging/senescence terms appear only in pathway annotations, background literature, or results interpretation, but the study does not investigate aging."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"1af66e1c-a5aa-4b49-91a2-a027862533bf","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Ambiguous:"},{"type":"text","text":" insufficient information to classify; flagged for manual review."}]}]}]},{"type":"paragraph","attrs":{"id":"868bb795-afbe-42f4-b11d-69de254326f8","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"Only entries classified as \"primary\" or \"secondary\" proceed to metadata extraction. \"Incidental mention\" entries are excluded. \"Ambiguous\" entries are flagged for manual review by the research assistant, with the original publication abstract retrieved via DOI to aid classification."}]},{"type":"paragraph","attrs":{"id":"89a094ad-5487-4bce-a278-e362d2542bb7","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 3: Structured Metadata Extraction."},{"type":"text","text":" For entries passing relevance classification, submit to the Claude API with a structured extraction prompt specifying the following target fields:"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"a713428d-9d00-4cd5-842b-414249d9bccf","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Tissue/cell type"},{"type":"text","text":" (mapped to Cell Ontology for cell types, Uberon for tissues, or CLO for cell lines; with explicit annotation of whether the sample is primary tissue, primary cells, or a cell line, and for cell lines whether normal or cancer-derived)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"5811c32c-eb77-48a0-9c3a-15ae8932b38a","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Age or time point"},{"type":"text","text":" (normalized to years for humans)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"dbfac0b1-3127-449c-b739-5f3a05004a38","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Sex"},{"type":"text","text":" (male, female, mixed, not reported)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"b658da4a-2b19-4b5e-a0d6-f90ee276b5ba","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Experimental condition"},{"type":"text","text":" (natural aging, accelerated aging/progeroid model, senescence induction, DNA damage treatment, lifespan intervention, control)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"377187a3-5414-4f89-98d0-968a2393fb84","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Intervention details"},{"type":"text","text":" (agent, dose, duration where applicable)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"2978fe9e-915f-4b0f-b059-26ef6bf3afae","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Sample size"},{"type":"text","text":" (number of biological replicates per condition)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"48635e63-7563-4928-b821-567d26893db1","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Sequencing platform"},{"type":"text","text":" (microarray model or RNA-seq chemistry)"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"860f1762-432f-4c0b-9a9e-f722913d3e08","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Batch structure"},{"type":"text","text":" (single batch, multi-batch, unknown)"}]}]}]},{"type":"paragraph","attrs":{"id":"8f5cfced-5e24-47e6-acde-bad84af9ff79","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"The prompt enforces JSON output conforming to a predefined schema with constrained value sets for categorical fields. When the model cannot confidently extract a field, it returns \"uncertain\" with a free-text explanation, flagging the entry for manual review."}]},{"type":"paragraph","attrs":{"id":"e56bd011-23bc-4e2e-b7ad-93f3071db3a3","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 4: Validation."},{"type":"text","text":" Three validation layers ensure accuracy, targeting ≥95% field-level agreement with human annotation and aiming for ≥99% on critical fields (relevance classification, tissue type, age):"}]},{"type":"paragraph","attrs":{"id":"884ddf31-0252-40eb-b487-5e69f8b7e7bd","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"(a) "},{"type":"text","marks":[{"type":"italic"}],"text":"Automated consistency checks:"},{"type":"text","text":" Cross-validate extracted organism against the GEO platform annotation. Verify that reported ages fall within plausible human ranges. Flag entries where the model's output contradicts structured GEO fields."}]},{"type":"paragraph","attrs":{"id":"d087dae3-a85f-4d7a-bade-9fbf36f29162","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"(b) "},{"type":"text","marks":[{"type":"italic"}],"text":"Independent LLM re-classification:"},{"type":"text","text":" For entries flagged as uncertain or where automated checks detect inconsistencies, re-submit to a separate API call with the original publication abstract (retrieved via DOI) appended to the prompt, enabling the model to resolve ambiguities from richer context. Because each API call operates without memory of prior classifications, this constitutes a functionally independent re-classification rather than self-correction. Concordance rates between first-pass and re-classification will be reported as an empirical measure of extraction reliability."}]},{"type":"paragraph","attrs":{"id":"9376e777-b9dd-4d47-ab47-c5b5772aab23","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"(c) "},{"type":"text","marks":[{"type":"italic"}],"text":"Human spot-check:"},{"type":"text","text":" A random 10% sample of extracted entries will be manually validated against the original publications by a trained research assistant with graduate-level experience in aging biology and GEO data structures. If field-level accuracy falls below 95% on any target field, the extraction prompt will be iteratively refined and the affected batch re-processed. All ambiguous entries flagged by the pipeline will be manually reviewed regardless of the spot-check results. Cohen's kappa will be computed between LLM and human annotations for the four-category relevance classification to quantify inter-rater reliability."}]},{"type":"paragraph","attrs":{"id":"793be019-1746-4bbb-96f3-b923568a16cd","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 5: Ontology Mapping and Standardization."},{"type":"text","text":" All extracted terms are mapped to controlled vocabularies using the EMBL-EBI Ontology Lookup Service API: tissue terms to Uberon, cell types to Cell Ontology, diseases to Disease Ontology, experimental interventions to the Experimental Factor Ontology (EFO). Unmapped terms are flagged for manual resolution."}]},{"type":"paragraph","attrs":{"id":"d2bac98f-2a90-4df1-9368-7ee07d94e653","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Phase 2: Expression Data Harmonization (Months 3 to 4)"}]},{"type":"paragraph","attrs":{"id":"b8629ad0-c959-475d-8aad-063f6295d13a","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 1: Data Download."},{"type":"text","text":" Download raw or normalized expression matrices for all included datasets. For microarray data, download CEL files where available; otherwise, use pre-processed series matrices. For RNA-seq, download count matrices from GEO Supplementary Files or re-quantify from FASTQ via the Recount3 resource where available."}]},{"type":"paragraph","attrs":{"id":"abb4776d-c955-4462-9f79-813dbe0d400d","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 2: Preprocessing and Normalization."},{"type":"text","text":" Apply platform-appropriate normalization. For Affymetrix microarrays, apply RMA normalization via the oligo R package. For two-color arrays, apply loess within-array normalization followed by quantile between-array normalization via limma. For RNA-seq count data, apply the limma-voom transformation pipeline (20), which converts raw counts to log-counts-per-million with observation-level precision weights. We use limma-voom rather than DESeq2 as the primary RNA-seq framework because DESeq2's negative binomial GLM becomes computationally prohibitive beyond approximately 10,000 samples (21), whereas limma-voom scales efficiently to mega-analysis (20,22). Probe-to-gene mapping uses manufacturer annotation files with ambiguous probes removed. Gene identifiers are harmonized to HGNC symbols with Ensembl gene IDs as a secondary namespace."}]},{"type":"paragraph","attrs":{"id":"1b9e54d5-4d6c-4c01-88af-df47ad4b0ff9","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 3: Quality Control."},{"type":"text","text":" For each dataset: (a) Remove outlier samples with inter-sample correlation < 0.8 (Pearson r on global expression profile). (b) Assess batch effects via principal component analysis; flag datasets where the first principal component separates by technical batch rather than biological condition. (c) Apply ComBat-seq for within-study batch correction where batch information is available."}]},{"type":"paragraph","attrs":{"id":"1e685a25-0d51-4865-91ec-8ab5f4a5b16a","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Step 4: Per-Dataset Differential Expression (Stage 1 of Two-Stage Meta-Analysis)."},{"type":"text","text":" Each dataset is analyzed independently to extract gene-level effect sizes (23). For microarray datasets, fit gene-wise linear models using limma's empirical Bayes moderated t-statistics (24). For RNA-seq datasets, apply the limma-voom pipeline: filter lowly expressed genes (requiring at least 10 counts in a minimum number of samples), apply the voom transformation, and fit gene-wise weighted linear models with empirical Bayes moderation (20). For studies with discrete age groups (young vs. old), the model is a simple contrast between groups. For studies with continuous age, fit a linear regression of expression on age. Where metadata permits, include sex, batch, and tissue subregion as covariates in the per-dataset model to reduce confounding. For each gene, extract the log2 fold-change (logFC) and its standard error. To combine continuous and categorical contrasts in a single meta-analysis, all effect sizes are converted to a common scale: the standardized mean difference (Hedges' g), computed from the logFC divided by the pooled within-group standard deviation, with a small-sample correction factor (25). For discrete age-group comparisons, this conversion is direct. For continuous age designs (regression of expression on age), the regression slope is converted to an equivalent standardized mean difference using the age range of each study to define the contrast (youngest versus oldest quartile), with the residual standard deviation of the regression as the denominator. Where the age range or residual SD cannot be recovered from GEO metadata, the dataset is flagged for manual extraction from the original publication. A sensitivity analysis will compare meta-analytic results when restricted to discrete-group studies only versus including converted continuous designs, to assess robustness to conversion assumptions."}]},{"type":"paragraph","attrs":{"id":"b99f571f-f612-43e0-be95-f337f7c570c3","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Phase 3: Random-Effects Meta-Analysis (Month 5)"}]},{"type":"paragraph","attrs":{"id":"0ebfe726-ec4e-432e-bfbc-a522c3d803a8","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Stage 2: Combining Effect Sizes Across Datasets."},{"type":"text","text":" For each gene, combine the per-dataset effect sizes using a random-effects meta-analysis with REML estimation of between-study variance via the metafor R package (26). REML is preferred over the DerSimonian-Laird moment estimator, which underestimates between-study variance under substantial heterogeneity (23,26). For each gene, compute: (a) the summary effect size (weighted mean Hedges' g) and its 95% confidence interval, (b) Cochran's Q statistic and I² for heterogeneity, and (c) the prediction interval. Apply Benjamini-Hochberg FDR correction across all tested genes (q < 0.05)."}]},{"type":"paragraph","attrs":{"id":"89335312-623c-488e-aff2-037141fb58d6","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Computational Strategy."},{"type":"text","text":" Both stages are embarrassingly parallel. The per-dataset pipeline is orchestrated via Snakemake on the IISc HPC cluster (or equivalent GCP instances). Stage 2 parallelizes trivially across genes via BiocParallel. The computational bottleneck is I/O (downloading and parsing thousands of GEO files), not model fitting."}]},{"type":"paragraph","attrs":{"id":"639aa201-f345-4f95-abfc-01874a7f3ae2","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Analysis for RQ1 and RQ3: conserved aging signatures and biomarker identification."},{"type":"text","text":" Using the genome-wide REML random-effects results from Stage 2, restricted to primary tissues, primary cell cultures, and normal cell lines (excluding cancer-derived cell lines to avoid confounding by oncogenic transcriptomic aberrations), identify genes with consistent direction of effect and q < 0.05 across a high proportion of datasets as candidate conserved aging markers. Compare against the Human Ageing Genomic Resources (GenAge, CellAge) and the Aging Atlas (16,17,18). Assess replication by leave-one-study-out cross-validation and by testing whether top markers replicate in held-out datasets not included in the meta-analysis."}]},{"type":"paragraph","attrs":{"id":"c5c70110-1537-4e1d-bac5-62dc8accb7eb","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Analysis for RQ2: Tissue-specific and perturbation-specific signatures."},{"type":"text","text":" Stratify datasets by sample type: (a) primary tissues by category (immune/hematopoietic, epithelial, neural, muscle, adipose, other), (b) primary cell cultures, (c) normal cell lines, (d) cancer cell lines. Perform separate meta-analyses within each stratum. Compute pairwise Spearman correlations of gene-level effect sizes between strata. Apply hierarchical clustering on the correlation matrix to identify natural groupings of aging signatures. Perform pathway enrichment (KEGG, Reactome, GO) within each cluster to characterize biological themes."}]},{"type":"paragraph","attrs":{"id":"3a920d57-6b54-4ee6-98e2-3cb06d51da37","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Analysis for RQ4: Sources of heterogeneity."},{"type":"text","text":" Perform subgroup meta-analyses stratified by: sequencing platform (microarray vs. RNA-seq), tissue category, sample size per group (above/below median), and publication year (pre-2015 vs. post-2015). Report I² within each stratum to assess whether heterogeneity is driven by technical or biological factors. Use leave-one-study-out sensitivity analysis to identify influential datasets."}]},{"type":"paragraph","attrs":{"id":"8860731c-4ff6-4fa9-aa81-b744acf6740b","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Analysis for RQ5: DNA damage versus chronological aging."},{"type":"text","text":" Separate datasets into two groups: (a) chronological aging studies (natural aging across timepoints) and (b) DNA damage/genotoxic stress studies (irradiation, oxidative stress, replicative senescence), including cancer-derived cell lines undergoing therapy-induced or DNA damage-induced senescence. Perform meta-analysis within each group. To compare the two signatures, apply three complementary approaches: (i) rank-rank hypergeometric overlap (RRHO) analysis, which provides a threshold-free visualization of concordance and discordance across the full ranked gene lists without arbitrary significance cutoffs; (ii) Spearman ρ restricted to genes reaching nominal significance (p < 0.05) in at least one of the two categories, to avoid inflating the correlation through shared null effects across thousands of uninformative genes; and (iii) pathway-level comparison using gene set enrichment analysis (GSEA) against KEGG and Reactome, testing whether shared and divergent pathways align with predictions from the hallmarks of aging framework (4)."}]},{"type":"paragraph","attrs":{"id":"6ff06830-a963-41d7-a7b1-45a8701c6574","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Sensitivity analyses."},{"type":"text","text":" (i) Restrict to studies with n ≥ 5 per group. (ii) Restrict to primary tissues only (exclude cell lines). (iii) Stratify by publication year (pre-2015 microarray era vs. post-2015 RNA-seq era). (iv) Funnel plots and Egger's test for publication bias. (v) Restrict to discrete age-group designs only (excluding continuous-age regression designs) to assess robustness to the Hedges' g conversion procedure."}]},{"type":"paragraph","attrs":{"id":"84193dd6-d73a-4668-92b2-cfb23811b231","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Phase 4: Validation and Deliverables (Month 6)"}]},{"type":"paragraph","attrs":{"id":"145350f7-7e15-41d5-a58d-717705bd8164","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Cross-reference validation."},{"type":"text","text":" Compare meta-analytic aging gene sets against: (a) the GenAge database of aging-related genes (16), (b) the CellAge database of cellular senescence genes (18), (c) the Aging Atlas multi-omics resource (17), and (d) recently published independent aging cohorts not included in the meta-analysis. Quantify overlap using hypergeometric enrichment tests."}]},{"type":"paragraph","attrs":{"id":"0fab8a05-4e43-4863-83aa-5ec507678d6e","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Interactive Shiny server."},{"type":"text","text":" Develop and deploy an interactive Shiny web application enabling users to: query the curated metadata catalogue; visualize gene-level meta-analytic results across tissues, cell types, and conditions; explore tissue-specific and perturbation-specific aging signatures; and download subsets of interest for their own analyses. The server will be deployed via "},{"type":"text","marks":[{"type":"link","attrs":{"href":"http://shinyapps.io","target":"_blank","rel":"noopener noreferrer nofollow","class":"link"}}],"text":"shinyapps.io"},{"type":"text","text":" or a containerized instance and maintained as a community resource."}]},{"type":"paragraph","attrs":{"id":"11c57257-15f7-4009-8ea5-fc213ae5da5e","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"Open resource release."},{"type":"text","text":" Deposit the following on Zenodo with DOI: (a) complete curated metadata catalogue (JSON + CSV), (b) processed expression matrices with harmonized annotations, (c) full meta-analytic results for all genes across all subgroup analyses, (d) the GEO-Age LLM pipeline code and prompt library (GitHub, MIT license), (e) Docker container for full pipeline reproducibility."}]},{"type":"heading","attrs":{"id":"90864883-1598-421a-8c92-66cdf635ab0c","data-toc-id":"90864883-1598-421a-8c92-66cdf635ab0c","textAlign":null,"level":3},"content":[{"type":"text","text":"3.3 Ethics and Data Management"}]},{"type":"paragraph","attrs":{"id":"3c0f8687-4bf7-4672-9e7c-8678146c4e80","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Ethics."},{"type":"text","text":" This study exclusively reanalyzes publicly available, de-identified datasets from GEO. No new human subjects research, animal experimentation, or data collection involving identifiable individuals is conducted. All source datasets comply with the original studies' ethics approvals and GEO's data sharing policies."}]},{"type":"paragraph","attrs":{"id":"7556735e-8ebf-4f0b-a7b9-f26c0ae6cf8e","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Data Management."},{"type":"text","text":" All analysis code will be version-controlled on GitHub with MIT license from project initiation. The curated metadata catalogue and processed datasets will be deposited on Zenodo with a persistent DOI upon completion. Analysis pipelines will be containerized using Docker for full reproducibility. The final manuscript will be published as a ResearchHub Registered Report. The study protocol will be preregistered on ResearchHub prior to initiating the meta-analysis phase, per the RFP requirements."}]},{"type":"heading","attrs":{"id":"09ced967-c854-423d-b321-4da26cd0c3e3","data-toc-id":"09ced967-c854-423d-b321-4da26cd0c3e3","textAlign":"left","level":2},"content":[{"type":"text","text":"4. Budget"}]},{"type":"paragraph","attrs":{"id":"3f0dc63e-8a35-4259-aa1b-9f98077d3597","class":null,"textAlign":"left"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Request: $10,000 USD (in ResearchCoin)"}]},{"type":"table","attrs":{"id":"99900f72-8bbd-4db1-a65f-2b8b67060079"},"content":[{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"6570149c-0907-4c85-bd2d-cf0d53d41da5","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Item"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"0e50d782-814a-4460-8c5a-46c5dfa2bf2c","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Cost"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"6b3439bb-a4a2-4890-8c2b-d4bbee5f2701","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Justification"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"829e4229-2f96-4f75-a102-3f537b87bc33","class":null,"textAlign":null},"content":[{"type":"text","text":"Cloud computing (GCP)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"f866c2de-dbde-411e-a653-7d83b3ea8114","class":null,"textAlign":null},"content":[{"type":"text","text":"$3,000"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"b4099417-504b-42e8-bd02-e63d167f15ef","class":null,"textAlign":null},"content":[{"type":"text","text":"Data storage, bulk expression data processing, meta-analysis computation across thousands of datasets"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"16827215-4849-47b0-95e1-52026ba9ef9a","class":null,"textAlign":null},"content":[{"type":"text","text":"LLM API costs (Claude 4.5 Sonnet + API)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"95aac98e-c054-49bc-a0ea-d7c46adb3a6a","class":null,"textAlign":null},"content":[{"type":"text","text":"$2,000"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"3950a2d5-5b95-47e5-ab4a-c0280c872a9d","class":null,"textAlign":null},"content":[{"type":"text","text":"Relevance classification and metadata extraction for 10,000 GEO entries with multi-step prompting ($0.20/dataset including validation passes)"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"5491b557-54df-4dfb-9daf-0a289481c7df","class":null,"textAlign":null},"content":[{"type":"text","text":"Research assistant (100 hrs × $25/hr)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"57cf83ca-7545-42b4-a23d-aaca6f3ec24b","class":null,"textAlign":null},"content":[{"type":"text","text":"$2,500"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"23d36ec9-8bba-4ce6-9c98-2158d7850c32","class":null,"textAlign":null},"content":[{"type":"text","text":"Graduate-level research assistant with aging biology expertise for manual validation of 10% metadata sample, review of ambiguous entries flagged by pipeline, literature cross-referencing"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"577eabd5-fabf-4d59-a913-16a11e7bb6e2","class":null,"textAlign":null},"content":[{"type":"text","text":"Interactive Shiny server (development + deployment)"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"98fb1e02-fd23-4c6d-ad7e-bb761cc7021a","class":null,"textAlign":null},"content":[{"type":"text","text":"$1,000"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"abbec98d-c24e-4d9d-951d-c8bdca81762e","class":null,"textAlign":null},"content":[{"type":"text","text":"Web application development and hosting for community access to curated atlas and meta-analytic results"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"646ca841-6dda-4367-8500-eaeb8e47bf77","class":null,"textAlign":null},"content":[{"type":"text","text":"ResearchHub Registered Report publication"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"a7a9ba14-a644-49e9-bd9b-1a4279b5c023","class":null,"textAlign":null},"content":[{"type":"text","text":"$1,000"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"97cdcc80-ac60-4f61-927d-339841c66b18","class":null,"textAlign":null},"content":[{"type":"text","text":"Open-access publication of final manuscript as ResearchHub Registered Report"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"a9b6a0e5-43f6-42ac-a2e4-3764ab7d863f","class":null,"textAlign":null},"content":[{"type":"text","text":"Contingency"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"e6e4d3df-47c4-47bc-9344-62778ccb286c","class":null,"textAlign":null},"content":[{"type":"text","text":"$500"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"6f0c9075-e633-4963-bd0f-03c3a8e4b6d2","class":null,"textAlign":null},"content":[{"type":"text","text":"Additional compute for expanded analyses, unforeseen API cost overruns"}]}]}]}]},{"type":"paragraph","attrs":{"id":"c9d67022-283a-4aa6-90db-36af5c147dd1","class":"font-claude-response-body break-words whitespace-normal leading-[1.7]","textAlign":null},"content":[{"type":"text","text":"Budget is flexible and will be adjusted based on Phase 1 results. If metadata extraction proves more efficient than estimated, surplus funds will be allocated to expanding the analysis to additional sample types, scRNA-Seq or deeper pathway-level analyses."}]},{"type":"heading","attrs":{"id":"08664ef8-a028-404e-806a-dbdccddb8a27","data-toc-id":"08664ef8-a028-404e-806a-dbdccddb8a27","textAlign":"left","level":2},"content":[{"type":"text","text":"6. References"}]},{"type":"orderedList","attrs":{"start":1,"type":null},"content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"ceadae81-4bfb-4b24-a4ab-3299cba8e403","class":null,"textAlign":"left"},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"fontSize":"16px","fontFamily":"BlinkMacSystemFont, -apple-system, \"Segoe UI\", Roboto, Oxygen, Ubuntu, Cantarell, \"Fira Sans\", \"Droid Sans\", \"Helvetica Neue\", sans-serif","color":"rgb(33, 33, 33)"}}],"text":"Shokhirev MN, Johnson AA. Modeling the human aging transcriptome across tissues, health status, and sex. "},{"type":"text","marks":[{"type":"italic"}],"text":"Aging Cell"},{"type":"text","marks":[{"type":"textStyle","attrs":{"fontSize":"16px","fontFamily":"BlinkMacSystemFont, -apple-system, \"Segoe UI\", Roboto, Oxygen, Ubuntu, Cantarell, \"Fira Sans\", \"Droid Sans\", \"Helvetica Neue\", sans-serif","color":"rgb(33, 33, 33)"}}],"text":". 2021;20(1):e13280. doi:10.1111/acel.13280"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"31e54f2d-1137-470f-af51-88ad3d994af8","class":null,"textAlign":"left"},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"fontSize":"16px","fontFamily":"BlinkMacSystemFont, -apple-system, \"Segoe UI\", Roboto, Oxygen, Ubuntu, Cantarell, \"Fira Sans\", \"Droid Sans\", \"Helvetica Neue\", sans-serif","color":"rgb(33, 33, 33)"}}],"text":"de Magalhães JP, Curado J, Church GM. Meta-analysis of age-related gene expression profiles identifies common signatures of aging. "},{"type":"text","marks":[{"type":"italic"}],"text":"Bioinformatics"},{"type":"text","marks":[{"type":"textStyle","attrs":{"fontSize":"16px","fontFamily":"BlinkMacSystemFont, -apple-system, \"Segoe UI\", Roboto, Oxygen, Ubuntu, Cantarell, \"Fira Sans\", \"Droid Sans\", \"Helvetica Neue\", sans-serif","color":"rgb(33, 33, 33)"}}],"text":". 2009;25(7):875-881. doi:10.1093/bioinformatics/btp073"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"0f41f2fb-63be-4dc5-9b24-7fbcbda77123","class":null,"textAlign":"left"},"content":[{"type":"text","text":"Sansone, S.A. et al. Toward interoperable bioscience data. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nat. Genet."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"44"},{"type":"text","text":", 121-126 (2012)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"fc34f1cb-36f1-436a-a2c6-b3b68ee9cf68","class":null,"textAlign":null},"content":[{"type":"text","text":"López-Otín, C., Blasco, M.A., Partridge, L., Serrano, M. & Kroemer, G. Hallmarks of aging: an expanding universe. "},{"type":"text","marks":[{"type":"italic"}],"text":"Cell"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"186"},{"type":"text","text":", 243-278 (2023)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"33ddd45c-6d98-46ab-8071-281876209994","class":null,"textAlign":null},"content":[{"type":"text","text":"Aramillo Irizar, P. et al. Transcriptomic alterations during ageing reflect the shift from cancer to degenerative diseases in the elderly. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nat. Commun."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"9"},{"type":"text","text":", 327 (2018)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"cb79f42e-6c21-4768-b5e2-47b744a4a7a2","class":null,"textAlign":null},"content":[{"type":"text","text":"Palmer, D. et al. Ageing transcriptome meta-analysis reveals similarities and differences between key mammalian tissues. "},{"type":"text","marks":[{"type":"italic"}],"text":"Aging"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"13"},{"type":"text","text":", 3313-3341 (2021)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"2622922a-c892-4129-aaf0-f8dd9f6e8e0f","class":null,"textAlign":null},"content":[{"type":"text","text":"Voisin, S. et al. Exercise is associated with younger methylome and transcriptome profiles in human skeletal muscle. "},{"type":"text","marks":[{"type":"italic"}],"text":"Aging Cell"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"23"},{"type":"text","text":", e13859 (2024)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"19cf1a4b-6b7d-44bf-a97f-fc02c2e572de","class":null,"textAlign":null},"content":[{"type":"text","text":"Winterhalter, P.R. et al. The circadian rhythm: a key variable in aging? "},{"type":"text","marks":[{"type":"italic"}],"text":"Aging Cell"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"23"},{"type":"text","text":", e14268 (2024)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"558a6bb5-b34c-4f23-b202-e8e0cd405b59","class":null,"textAlign":null},"content":[{"type":"text","text":"Barrett, T. et al. NCBI GEO: archive for functional genomics data sets, update. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nucleic Acids Res."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"41"},{"type":"text","text":", D991-D995 (2013)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"b9ce284f-e543-41f8-9a36-9b966a5ac61d","class":null,"textAlign":null},"content":[{"type":"text","text":"Brazma, A. et al. Minimum information about a microarray experiment (MIAME): toward standards for microarray data. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nat. Genet."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"29"},{"type":"text","text":", 365-371 (2001)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"9af06d9d-a1f3-4585-98d4-6d88a3867d78","class":null,"textAlign":null},"content":[{"type":"text","text":"Lee, P. et al. Benefits, limits, and risks of GPT-4 as an AI chatbot for medicine. "},{"type":"text","marks":[{"type":"italic"}],"text":"N. Engl. J. Med."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"388"},{"type":"text","text":", 1233-1239 (2023)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"335ea508-27b5-40ee-adcf-b8622f81a416","class":null,"textAlign":null},"content":[{"type":"text","text":"Jiang, L.Y. et al. Health system-scale language models are all-purpose prediction engines. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nature"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"619"},{"type":"text","text":", 357-362 (2023)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"1d05a461-e86d-4420-92c4-41c2254ac55e","class":null,"textAlign":null},"content":[{"type":"text","text":"Churchill, A. et al. GPT models for text annotation: an empirical exploration in public policy research. "},{"type":"text","marks":[{"type":"italic"}],"text":"Policy Stud. J."},{"type":"text","text":" (2025)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"3a1f9ed0-adde-4fd1-8f35-a774c2ccfb13","class":null,"textAlign":null},"content":[{"type":"text","text":"Anderson, W. et al. Unlocking the capabilities of large language models for accelerating drug development. "},{"type":"text","marks":[{"type":"italic"}],"text":"Clin. Pharmacol. Ther."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"116"},{"type":"text","text":", 38-41 (2024)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"a2a3814c-f113-4a5b-9420-5aa69341548d","class":null,"textAlign":null},"content":[{"type":"text","text":"Cao, H. et al. Genome-wide profiles of DNA damage represent highly accurate predictors of mammalian age. "},{"type":"text","marks":[{"type":"italic"}],"text":"Aging Cell"},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"23"},{"type":"text","text":", e14122 (2024)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"247c1774-2112-4b20-914a-f021010eefe4","class":null,"textAlign":null},"content":[{"type":"text","text":"Tacutu, R. et al. Human Ageing Genomic Resources: new and updated databases. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nucleic Acids Res."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"46"},{"type":"text","text":", D1083-D1090 (2018)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"1b605685-8ee7-4b8c-ab7f-681c4e875442","class":null,"textAlign":null},"content":[{"type":"text","text":"Aging Atlas Consortium. Aging Atlas: a multi-omics database for aging biology. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nucleic Acids Res."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"49"},{"type":"text","text":", D825-D830 (2021)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"e0385b0c-2780-4acd-a6a2-b9529d02e579","class":null,"textAlign":null},"content":[{"type":"text","text":"Avelar, R.A. et al. A multidimensional systems biology analysis of cellular senescence in aging and disease. "},{"type":"text","marks":[{"type":"italic"}],"text":"Genome Biol."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"21"},{"type":"text","text":", 91 (2020)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"ed709362-8256-4a1e-be19-a43dffe4f62c","class":null,"textAlign":null},"content":[{"type":"text","text":"Lopes, I. et al. Six drivers of aging identified among genes differentially expressed with age. "},{"type":"text","marks":[{"type":"italic"}],"text":"bioRxiv"},{"type":"text","text":" 2024.08.02.606402 (2024)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"05a0a4cf-f9ef-4703-bf6d-a20b3841d961","class":null,"textAlign":null},"content":[{"type":"text","text":"Law, C.W., Chen, Y., Shi, W. & Smyth, G.K. voom: precision weights unlock linear model analysis tools for RNA-seq read counts. "},{"type":"text","marks":[{"type":"italic"}],"text":"Genome Biol."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"15"},{"type":"text","text":", R29 (2014)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"7442bc8e-12d2-4def-8eb2-f2b9b894e9f9","class":null,"textAlign":null},"content":[{"type":"text","text":"Love, M.I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. "},{"type":"text","marks":[{"type":"italic"}],"text":"Genome Biol."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"15"},{"type":"text","text":", 550 (2014)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"7f090c71-c4d1-4622-814f-d15d472738d2","class":null,"textAlign":null},"content":[{"type":"text","text":"Ritchie, M.E. et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nucleic Acids Res."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"43"},{"type":"text","text":", e47 (2015)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"82ae0bde-b5c2-4fcf-add5-6804264fefc4","class":null,"textAlign":null},"content":[{"type":"text","text":"Tseng, G.C., Ghosh, D. & Feingold, E. Comprehensive literature review and statistical considerations for microarray meta-analysis. "},{"type":"text","marks":[{"type":"italic"}],"text":"Nucleic Acids Res."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"40"},{"type":"text","text":", 3785-3799 (2012)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"265167e8-3d40-40bd-8e88-c7f1d001c7d5","class":null,"textAlign":null},"content":[{"type":"text","text":"Smyth, G.K. Linear models and empirical Bayes methods for assessing differential expression in microarray experiments. "},{"type":"text","marks":[{"type":"italic"}],"text":"Stat. Appl. Genet. Mol. Biol."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"3"},{"type":"text","text":", Article 3 (2004)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"f1585334-a12c-411d-a806-c1d3ef26b905","class":null,"textAlign":null},"content":[{"type":"text","text":"Hedges, L.V. & Olkin, I. "},{"type":"text","marks":[{"type":"italic"}],"text":"Statistical Methods for Meta-Analysis."},{"type":"text","text":" Academic Press (1985)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"7ede985c-c0c7-41cf-8785-380226524bcb","class":null,"textAlign":null},"content":[{"type":"text","text":"Viechtbauer, W. Conducting meta-analyses in R with the metafor package. "},{"type":"text","marks":[{"type":"italic"}],"text":"J. Stat. Softw."},{"type":"text","text":" "},{"type":"text","marks":[{"type":"bold"}],"text":"36"},{"type":"text","text":", 1-48 (2010)."}]}]}]},{"type":"heading","attrs":{"id":"26bf4bf5-4330-492a-8c6a-de997c1f2e08","data-toc-id":"26bf4bf5-4330-492a-8c6a-de997c1f2e08","textAlign":null,"level":2},"content":[{"type":"text","text":"AI Disclosure"}]},{"type":"paragraph","attrs":{"id":"49c8dd7d-923f-423c-9927-491a67650b12","class":null,"textAlign":null},"content":[{"type":"text","text":"Gemini 3 Pro Nano Banana tool was used to generate the graphical abstract based on the proposal given as input. Claude Opus 4.6 was used to draft the initial proposal. The idea, methodology and workplan are completely original ideas of the author. The llm was used for initial draft after which the author then thoroughly vetted and heavily edited the proposal for accuracy and flow of thought and added crucial details. All references have also been cross-validated."}]},{"type":"paragraph","attrs":{"id":"4d625cc6-01b3-4ce1-a1ab-cd1a64a3ed4c","class":null,"textAlign":null}}]}5d:Tc2e4,

GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome

1. Overview

Scientific Rationale: A search for aging-related terms ("aging", "senescence", "ageing", "lifespan", "longevity") on the NCBI Gene Expression Omnibus returns over 30,000 dataset entries, with approximately 10,000 involving human samples (NCBI E-utilities query, February 2026). Yet no study has systematically mined this resource at scale. Existing aging meta-analyses are limited to manually curated subsets of 25 to 50 datasets (1,2), leaving the vast majority of public data unexploited. Two compounding bottlenecks explain why. First, GEO metadata is notoriously inconsistent: sample ages are buried in free-text descriptions, tissue annotations use non-standard vocabularies, and experimental conditions are logged in idiosyncratic formats (3). Second, and critically, keyword-based GEO searches massively overcount relevant datasets: a study that merely reports upregulation of a senescence-associated pathway (e.g., p16/CDKN2A) in its results section will appear in a "senescence" search even if the study has nothing to do with aging or cellular senescence. This false-positive problem means the true number of directly aging-relevant transcriptomic datasets is likely a fraction of the raw search total, but no systematic triage has ever been performed. This project deploys an LLM-powered curation pipeline, built on Anthropic's Claude 4.5 and above model family, to systematically extract, standardize, and quality-control metadata across all candidate GEO entries, intelligently distinguishing genuinely aging-focused datasets from incidental keyword matches. This enables the first mega meta-analysis of human aging transcriptomes at unprecedented scale. The resulting resource, a curated, machine-readable atlas of aging gene expression, will be released as an open community tool.

Timeline: 6 months

Month 1-2: LLM pipeline development, GEO-wide metadata extraction, and validation.
Month 3-4: Expression data download, harmonization, batch correction, and quality control.
Month 5: Meta-analysis and statistical modeling (Random-effects meta-analysis, subgroup analyses, and meta-regression.)
Month 6: Cross-validation, interactive visualization, manuscript preparation, and submission

Data Sharing: All curated metadata, processed expression matrices, analysis code, and the LLM curation pipeline will be deposited in open repositories (GitHub under MIT license, Zenodo under CC-BY-4.0). The curated metadata catalogue itself will be released as a standalone resource for community reanalysis. An interactive Shiny web application will be deployed for the community to explore meta-analytic results, query gene-level aging signatures across tissues and conditions, and download subsets of interest. The study protocol will be preregistered on ResearchHub prior to analysis. The final manuscript will be published as a ResearchHub Registered Report with regular progress updates shared via the Updates tab throughout the project.

2. Introduction

Aging is the dominant risk factor for cardiovascular disease, neurodegeneration, cancer, and metabolic dysfunction, yet the molecular programs that drive age-dependent decline remain incompletely mapped (4,5). Over the past two decades, transcriptomic profiling has emerged as the primary tool for characterizing aging at the molecular level. Individual studies have identified age-associated changes in inflammation, mitochondrial function, proteostasis, and DNA damage response pathways. However, these findings exhibit surprisingly limited overlap across studies, a problem that has persisted despite the growing volume of available data (1,6).

Two landmark meta-analyses illustrate both the promise and the ceiling of current approaches. Shokhirev and Johnson reprocessed approximately 6,000 raw RNA-seq samples from the Sequence Read Archive, ultimately retaining 3,060 high-quality human samples after stringent filtering, and achieved remarkable predictive accuracy for transcriptomic age (R² = 0.96) (1). Voisin et al. assembled 3,176 human skeletal muscle samples across methylome and transcriptome datasets to demonstrate that exercise training shifts molecular profiles toward a younger state (7). Both studies relied on extensive manual curation and were restricted to specific tissues or contexts. A recent preprint used only 25 GEO datasets across multiple species to identify six causal aging driver genes, explicitly noting that their small dataset count was a limitation and calling for expanded meta-analyses as more data become available (19). Meanwhile, Palmer et al. demonstrated that transcriptomic aging signatures show minimal overlap across organs, raising the fundamental question of whether truly universal aging markers exist or whether the apparent inconsistency is an artifact of small, heterogeneous datasets (6).

The core problem is scale, and specifically the curatorial bottleneck that prevents achieving it. GEO currently hosts over 200,000 series entries comprising more than 6 million individual samples (9). A broad keyword search for aging-related terms returns over 30,000 dataset entries, of which approximately 10,000 involve human samples. However, these raw counts are profoundly misleading. GEO's full-text search indexes not only study titles and designs but also result descriptions, pathway annotations, and associated publications. A cancer genomics study that incidentally reports upregulation of CDKN2A or other senescence-associated genes will appear in a "senescence" search despite having no relevance to aging biology whatsoever. Similarly, studies examining age as a covariate rather than a primary variable, or studies using "aging" only in the context of sample storage or reagent shelf life, inflate the counts. The true number of datasets whose primary focus is aging, cellular senescence, DNA damage in the context of aging, or lifespan modulation is likely an order of magnitude smaller than the raw search suggests, but no systematic triage has been performed, because manual screening of 10,000+ entries is prohibitively labor-intensive.

This is precisely the problem that LLM-powered curation pipelines can solve. Large language models excel at reading unstructured text descriptions, understanding experimental intent from context, and making nuanced classification decisions that go beyond keyword matching (11,12). The Critical Path Institute has demonstrated that LLMs can perform arbitrary data cleaning and knowledge extraction tasks dynamically through prompt engineering, achieving results that would otherwise require weeks of manual curation (14). By deploying an LLM pipeline to read every GEO entry's title, summary, experimental design, and sample annotations, we can distinguish genuinely aging-focused datasets from incidental keyword matches while simultaneously extracting and standardizing the metadata needed for downstream integration.

This project combines these two threads, the unmet need for large-scale human aging meta-analysis and the emerging capability of LLMs for intelligent data triage and curation, into a unified pipeline we call GEO-Age. The pipeline, built on Anthropic's Claude 4.5 and above model family, will systematically process all candidate GEO entries, classify each for aging relevance, extract and standardize metadata into a machine-readable schema, and feed the curated catalogue into a multi-level random-effects meta-analysis. The scope is restricted to human samples, including primary tissues, primary cell cultures, and cell lines (both normal and cancer-derived), to maximize translational relevance and minimize the confounds introduced by cross-species ortholog mapping. Cancer-derived cell lines are included specifically for the senescence and DNA damage analyses (RQ2, RQ5), where therapy-induced and oncogene-induced senescence are biologically relevant phenomena, but are excluded from the primary chronological aging meta-analysis (RQ1, RQ3) to avoid confounding by baseline transcriptomic aberrations (telomerase reactivation, p53 loss-of-function, aneuploidy).

A particular focus will be the relationship between DNA damage responses and chronological aging. Genomic instability is a canonical hallmark of aging, and recent work has shown that genome-wide DNA damage profiles can predict mammalian age with accuracy exceeding transcriptome-based models (15). Hundreds of GEO datasets examine DNA damage induction (via irradiation, oxidative stress, genotoxic agents, or replicative senescence) in human cell lines, yet no systematic comparison has tested whether these acute damage signatures recapitulate the molecular program of chronological aging or represent a distinct biological response. The scale of our meta-analysis is uniquely positioned to resolve this question.

Research Questions.

RQ1. What molecular signatures of aging are robust across tissues, cell types, and experimental contexts?

RQ2. Which aging-associated changes are specific to particular biological systems or perturbations?

RQ3. Can meta-analysis identify high-confidence aging biomarkers that replicate across independent studies?

RQ4. What factors (technical vs. biological) explain heterogeneity in reported aging signatures?

RQ5. To what extent do DNA damage and genotoxic stress signatures overlap with, or diverge from, chronological aging programs?

This is a discovery-driven study. Rather than testing a fixed set of hypotheses, the project is designed to generate a comprehensive descriptive atlas of human aging transcriptomics at unprecedented scale, from which specific mechanistic hypotheses can be derived and subsequently tested by the community. The analysis plan (Section 3.3) specifies pre-registered analytical frameworks and decision rules, but the primary output is a curated resource and empirical characterization rather than hypothesis confirmation.

Impact: This project delivers three high-value outputs to the aging research community regardless of any specific finding. First, the curated metadata catalogue, a structured, machine-readable annotation of every aging-relevant human GEO dataset, constitutes a standalone resource that will accelerate future meta-analyses across any aging subfield. Second, the GEO-Age LLM pipeline establishes a generalizable methodology for LLM-powered data curation that can be applied to any domain with heterogeneous public data. Third, the mega meta-analysis itself will either identify genuinely universal aging markers (advancing biomarker development) or definitively demonstrate that human aging transcriptomes are context-dependent (redirecting the field toward tissue-specific intervention strategies). The interactive Shiny server will ensure these outputs remain accessible and explorable by the broader community well beyond the funding period.

3. Methods

Participants/Sample

Target population. All publicly available high-throughput transcriptomic datasets in GEO involving human samples that measure age-dependent molecular changes, cellular senescence, DNA damage responses, or lifespan-modulating interventions in primary tissues, primary cell cultures, or cell lines (normal and cancer-derived).

Inclusion criteria. (i) Human samples (primary tissues, primary cells, or established cell lines). (ii) Study's primary focus involves aging, cellular senescence, DNA damage response, or age-modulating intervention, as determined by the GEO-Age pipeline's relevance classification, not merely keyword co-occurrence. (iii) At least 3 biological replicates per experimental condition. (iv) Raw or processed expression data available for download. (v) Sufficient metadata to assign age/condition groups (after AI-assisted extraction).

Exclusion criteria. (i) Non-human samples. (ii) Datasets where aging/senescence terms appear only incidentally (e.g., pathway annotations in results, age as a minor covariate) rather than as the study's primary focus. (iii) Fewer than 3 biological replicates per group. (iv) Metadata insufficient for age/condition assignment even after AI extraction and manual review of the original publication. (v) Studies of age-associated diseases (e.g., Alzheimer's, cancer prognosis) without age-matched healthy controls or explicit aging comparisons. (vi) Datasets flagged for known technical failures in GEO. (vii) Single-cell RNA-seq datasets, which require distinct analytical frameworks (pseudobulk aggregation, cell-type deconvolution) that are outside the scope of this bulk transcriptomics meta-analysis; extension to scRNA-seq is planned as future work pending proof of concept.

Expected sample size. Approximately 10,000 human GEO entries match aging-related keyword searches. Based on the high rate of false-positive keyword matches inherent to GEO's full-text indexing, we estimate that 1,000 to 2,500 of these represent genuinely aging-focused studies after AI-assisted relevance filtering. After applying the remaining inclusion/exclusion criteria (replication, data availability), we conservatively estimate 500 to 1,500 high-quality transcriptomic datasets representing 20,000 to 60,000 individual samples. For meta-analysis, a minimum of 20 independent studies per subgroup analysis is required to achieve stable random-effects estimates; we expect to exceed this threshold for all major tissue categories.

3.2 Materials and Procedures

Phase 1: LLM Pipeline for Metadata Extraction and Relevance Classification (Months 1 to 2)

The GEO-Age pipeline is a multi-step LLM classification and extraction system built on Anthropic's Claude 4.5 and above model family that performs two critical functions: (a) relevance classification, meaning determining whether a GEO entry is genuinely focused on aging/senescence versus incidentally matching keywords, and (b) structured metadata extraction for entries classified as relevant.

Step 1: Retrieval. Use the NCBI E-utilities API (esearch + efetch) and the GEOquery R package to programmatically download all human GEO Series (GSE) records matching aging-related search terms ("aging", "ageing", "senescence", "cellular senescence", "lifespan", "longevity", "gerontology", "DNA damage aging", "replicative senescence", "oxidative stress aging"). For each entry, retrieve: title, summary, overall design, sample characteristics (GSM-level), platform annotation, publication DOI, and supplementary file links. Store raw metadata in a structured JSON schema.

Step 2: AI-Powered Relevance Classification. For each GEO entry, submit the concatenated metadata fields to the Claude API with a classification prompt that instructs the model to assign one of four relevance categories:

Primary aging study: aging, senescence, or DNA damage in the context of aging is the central research question.
Secondary aging study: age is a major covariate or secondary variable, with sufficient aging data for extraction.
Incidental mention: aging/senescence terms appear only in pathway annotations, background literature, or results interpretation, but the study does not investigate aging.
Ambiguous: insufficient information to classify; flagged for manual review.

Only entries classified as "primary" or "secondary" proceed to metadata extraction. "Incidental mention" entries are excluded. "Ambiguous" entries are flagged for manual review by the research assistant, with the original publication abstract retrieved via DOI to aid classification.

Step 3: Structured Metadata Extraction. For entries passing relevance classification, submit to the Claude API with a structured extraction prompt specifying the following target fields:

Tissue/cell type (mapped to Cell Ontology for cell types, Uberon for tissues, or CLO for cell lines; with explicit annotation of whether the sample is primary tissue, primary cells, or a cell line, and for cell lines whether normal or cancer-derived)
Age or time point (normalized to years for humans)
Sex (male, female, mixed, not reported)
Experimental condition (natural aging, accelerated aging/progeroid model, senescence induction, DNA damage treatment, lifespan intervention, control)
Intervention details (agent, dose, duration where applicable)
Sample size (number of biological replicates per condition)
Sequencing platform (microarray model or RNA-seq chemistry)
Batch structure (single batch, multi-batch, unknown)

The prompt enforces JSON output conforming to a predefined schema with constrained value sets for categorical fields. When the model cannot confidently extract a field, it returns "uncertain" with a free-text explanation, flagging the entry for manual review.

Step 4: Validation. Three validation layers ensure accuracy, targeting ≥95% field-level agreement with human annotation and aiming for ≥99% on critical fields (relevance classification, tissue type, age):

(a) Automated consistency checks: Cross-validate extracted organism against the GEO platform annotation. Verify that reported ages fall within plausible human ranges. Flag entries where the model's output contradicts structured GEO fields.

(b) Independent LLM re-classification: For entries flagged as uncertain or where automated checks detect inconsistencies, re-submit to a separate API call with the original publication abstract (retrieved via DOI) appended to the prompt, enabling the model to resolve ambiguities from richer context. Because each API call operates without memory of prior classifications, this constitutes a functionally independent re-classification rather than self-correction. Concordance rates between first-pass and re-classification will be reported as an empirical measure of extraction reliability.

(c) Human spot-check: A random 10% sample of extracted entries will be manually validated against the original publications by a trained research assistant with graduate-level experience in aging biology and GEO data structures. If field-level accuracy falls below 95% on any target field, the extraction prompt will be iteratively refined and the affected batch re-processed. All ambiguous entries flagged by the pipeline will be manually reviewed regardless of the spot-check results. Cohen's kappa will be computed between LLM and human annotations for the four-category relevance classification to quantify inter-rater reliability.

Step 5: Ontology Mapping and Standardization. All extracted terms are mapped to controlled vocabularies using the EMBL-EBI Ontology Lookup Service API: tissue terms to Uberon, cell types to Cell Ontology, diseases to Disease Ontology, experimental interventions to the Experimental Factor Ontology (EFO). Unmapped terms are flagged for manual resolution.

Phase 2: Expression Data Harmonization (Months 3 to 4)

Step 1: Data Download. Download raw or normalized expression matrices for all included datasets. For microarray data, download CEL files where available; otherwise, use pre-processed series matrices. For RNA-seq, download count matrices from GEO Supplementary Files or re-quantify from FASTQ via the Recount3 resource where available.

Step 2: Preprocessing and Normalization. Apply platform-appropriate normalization. For Affymetrix microarrays, apply RMA normalization via the oligo R package. For two-color arrays, apply loess within-array normalization followed by quantile between-array normalization via limma. For RNA-seq count data, apply the limma-voom transformation pipeline (20), which converts raw counts to log-counts-per-million with observation-level precision weights. We use limma-voom rather than DESeq2 as the primary RNA-seq framework because DESeq2's negative binomial GLM becomes computationally prohibitive beyond approximately 10,000 samples (21), whereas limma-voom scales efficiently to mega-analysis (20,22). Probe-to-gene mapping uses manufacturer annotation files with ambiguous probes removed. Gene identifiers are harmonized to HGNC symbols with Ensembl gene IDs as a secondary namespace.

Step 3: Quality Control. For each dataset: (a) Remove outlier samples with inter-sample correlation < 0.8 (Pearson r on global expression profile). (b) Assess batch effects via principal component analysis; flag datasets where the first principal component separates by technical batch rather than biological condition. (c) Apply ComBat-seq for within-study batch correction where batch information is available.

Step 4: Per-Dataset Differential Expression (Stage 1 of Two-Stage Meta-Analysis). Each dataset is analyzed independently to extract gene-level effect sizes (23). For microarray datasets, fit gene-wise linear models using limma's empirical Bayes moderated t-statistics (24). For RNA-seq datasets, apply the limma-voom pipeline: filter lowly expressed genes (requiring at least 10 counts in a minimum number of samples), apply the voom transformation, and fit gene-wise weighted linear models with empirical Bayes moderation (20). For studies with discrete age groups (young vs. old), the model is a simple contrast between groups. For studies with continuous age, fit a linear regression of expression on age. Where metadata permits, include sex, batch, and tissue subregion as covariates in the per-dataset model to reduce confounding. For each gene, extract the log2 fold-change (logFC) and its standard error. To combine continuous and categorical contrasts in a single meta-analysis, all effect sizes are converted to a common scale: the standardized mean difference (Hedges' g), computed from the logFC divided by the pooled within-group standard deviation, with a small-sample correction factor (25). For discrete age-group comparisons, this conversion is direct. For continuous age designs (regression of expression on age), the regression slope is converted to an equivalent standardized mean difference using the age range of each study to define the contrast (youngest versus oldest quartile), with the residual standard deviation of the regression as the denominator. Where the age range or residual SD cannot be recovered from GEO metadata, the dataset is flagged for manual extraction from the original publication. A sensitivity analysis will compare meta-analytic results when restricted to discrete-group studies only versus including converted continuous designs, to assess robustness to conversion assumptions.

Phase 3: Random-Effects Meta-Analysis (Month 5)

Stage 2: Combining Effect Sizes Across Datasets. For each gene, combine the per-dataset effect sizes using a random-effects meta-analysis with REML estimation of between-study variance via the metafor R package (26). REML is preferred over the DerSimonian-Laird moment estimator, which underestimates between-study variance under substantial heterogeneity (23,26). For each gene, compute: (a) the summary effect size (weighted mean Hedges' g) and its 95% confidence interval, (b) Cochran's Q statistic and I² for heterogeneity, and (c) the prediction interval. Apply Benjamini-Hochberg FDR correction across all tested genes (q < 0.05).

Computational Strategy. Both stages are embarrassingly parallel. The per-dataset pipeline is orchestrated via Snakemake on the IISc HPC cluster (or equivalent GCP instances). Stage 2 parallelizes trivially across genes via BiocParallel. The computational bottleneck is I/O (downloading and parsing thousands of GEO files), not model fitting.

Analysis for RQ1 and RQ3: conserved aging signatures and biomarker identification. Using the genome-wide REML random-effects results from Stage 2, restricted to primary tissues, primary cell cultures, and normal cell lines (excluding cancer-derived cell lines to avoid confounding by oncogenic transcriptomic aberrations), identify genes with consistent direction of effect and q < 0.05 across a high proportion of datasets as candidate conserved aging markers. Compare against the Human Ageing Genomic Resources (GenAge, CellAge) and the Aging Atlas (16,17,18). Assess replication by leave-one-study-out cross-validation and by testing whether top markers replicate in held-out datasets not included in the meta-analysis.

Analysis for RQ2: Tissue-specific and perturbation-specific signatures. Stratify datasets by sample type: (a) primary tissues by category (immune/hematopoietic, epithelial, neural, muscle, adipose, other), (b) primary cell cultures, (c) normal cell lines, (d) cancer cell lines. Perform separate meta-analyses within each stratum. Compute pairwise Spearman correlations of gene-level effect sizes between strata. Apply hierarchical clustering on the correlation matrix to identify natural groupings of aging signatures. Perform pathway enrichment (KEGG, Reactome, GO) within each cluster to characterize biological themes.

Analysis for RQ4: Sources of heterogeneity. Perform subgroup meta-analyses stratified by: sequencing platform (microarray vs. RNA-seq), tissue category, sample size per group (above/below median), and publication year (pre-2015 vs. post-2015). Report I² within each stratum to assess whether heterogeneity is driven by technical or biological factors. Use leave-one-study-out sensitivity analysis to identify influential datasets.

Analysis for RQ5: DNA damage versus chronological aging. Separate datasets into two groups: (a) chronological aging studies (natural aging across timepoints) and (b) DNA damage/genotoxic stress studies (irradiation, oxidative stress, replicative senescence), including cancer-derived cell lines undergoing therapy-induced or DNA damage-induced senescence. Perform meta-analysis within each group. To compare the two signatures, apply three complementary approaches: (i) rank-rank hypergeometric overlap (RRHO) analysis, which provides a threshold-free visualization of concordance and discordance across the full ranked gene lists without arbitrary significance cutoffs; (ii) Spearman ρ restricted to genes reaching nominal significance (p < 0.05) in at least one of the two categories, to avoid inflating the correlation through shared null effects across thousands of uninformative genes; and (iii) pathway-level comparison using gene set enrichment analysis (GSEA) against KEGG and Reactome, testing whether shared and divergent pathways align with predictions from the hallmarks of aging framework (4).

Sensitivity analyses. (i) Restrict to studies with n ≥ 5 per group. (ii) Restrict to primary tissues only (exclude cell lines). (iii) Stratify by publication year (pre-2015 microarray era vs. post-2015 RNA-seq era). (iv) Funnel plots and Egger's test for publication bias. (v) Restrict to discrete age-group designs only (excluding continuous-age regression designs) to assess robustness to the Hedges' g conversion procedure.

Phase 4: Validation and Deliverables (Month 6)

Cross-reference validation. Compare meta-analytic aging gene sets against: (a) the GenAge database of aging-related genes (16), (b) the CellAge database of cellular senescence genes (18), (c) the Aging Atlas multi-omics resource (17), and (d) recently published independent aging cohorts not included in the meta-analysis. Quantify overlap using hypergeometric enrichment tests.

Interactive Shiny server. Develop and deploy an interactive Shiny web application enabling users to: query the curated metadata catalogue; visualize gene-level meta-analytic results across tissues, cell types, and conditions; explore tissue-specific and perturbation-specific aging signatures; and download subsets of interest for their own analyses. The server will be deployed via shinyapps.io or a containerized instance and maintained as a community resource.

Open resource release. Deposit the following on Zenodo with DOI: (a) complete curated metadata catalogue (JSON + CSV), (b) processed expression matrices with harmonized annotations, (c) full meta-analytic results for all genes across all subgroup analyses, (d) the GEO-Age LLM pipeline code and prompt library (GitHub, MIT license), (e) Docker container for full pipeline reproducibility.

3.3 Ethics and Data Management

Ethics. This study exclusively reanalyzes publicly available, de-identified datasets from GEO. No new human subjects research, animal experimentation, or data collection involving identifiable individuals is conducted. All source datasets comply with the original studies' ethics approvals and GEO's data sharing policies.

Data Management. All analysis code will be version-controlled on GitHub with MIT license from project initiation. The curated metadata catalogue and processed datasets will be deposited on Zenodo with a persistent DOI upon completion. Analysis pipelines will be containerized using Docker for full reproducibility. The final manuscript will be published as a ResearchHub Registered Report. The study protocol will be preregistered on ResearchHub prior to initiating the meta-analysis phase, per the RFP requirements.

4. Budget

Request: $10,000 USD (in ResearchCoin)

Item	Cost	Justification
Cloud computing (GCP)	$3,000	Data storage, bulk expression data processing, meta-analysis computation across thousands of datasets
LLM API costs (Claude 4.5 Sonnet + API)	$2,000	Relevance classification and metadata extraction for 10,000 GEO entries with multi-step prompting ($0.20/dataset including validation passes)
Research assistant (100 hrs × $25/hr)	$2,500	Graduate-level research assistant with aging biology expertise for manual validation of 10% metadata sample, review of ambiguous entries flagged by pipeline, literature cross-referencing
Interactive Shiny server (development + deployment)	$1,000	Web application development and hosting for community access to curated atlas and meta-analytic results
ResearchHub Registered Report publication	$1,000	Open-access publication of final manuscript as ResearchHub Registered Report
Contingency	$500	Additional compute for expanded analyses, unforeseen API cost overruns

Budget is flexible and will be adjusted based on Phase 1 results. If metadata extraction proves more efficient than estimated, surplus funds will be allocated to expanding the analysis to additional sample types, scRNA-Seq or deeper pathway-level analyses.

6. References

Shokhirev MN, Johnson AA. Modeling the human aging transcriptome across tissues, health status, and sex. Aging Cell. 2021;20(1):e13280. doi:10.1111/acel.13280
de Magalhães JP, Curado J, Church GM. Meta-analysis of age-related gene expression profiles identifies common signatures of aging. Bioinformatics. 2009;25(7):875-881. doi:10.1093/bioinformatics/btp073
Sansone, S.A. et al. Toward interoperable bioscience data. Nat. Genet. 44, 121-126 (2012).
López-Otín, C., Blasco, M.A., Partridge, L., Serrano, M. & Kroemer, G. Hallmarks of aging: an expanding universe. Cell 186, 243-278 (2023).
Aramillo Irizar, P. et al. Transcriptomic alterations during ageing reflect the shift from cancer to degenerative diseases in the elderly. Nat. Commun. 9, 327 (2018).
Palmer, D. et al. Ageing transcriptome meta-analysis reveals similarities and differences between key mammalian tissues. Aging 13, 3313-3341 (2021).
Voisin, S. et al. Exercise is associated with younger methylome and transcriptome profiles in human skeletal muscle. Aging Cell 23, e13859 (2024).
Winterhalter, P.R. et al. The circadian rhythm: a key variable in aging? Aging Cell 23, e14268 (2024).
Barrett, T. et al. NCBI GEO: archive for functional genomics data sets, update. Nucleic Acids Res. 41, D991-D995 (2013).
Brazma, A. et al. Minimum information about a microarray experiment (MIAME): toward standards for microarray data. Nat. Genet. 29, 365-371 (2001).
Lee, P. et al. Benefits, limits, and risks of GPT-4 as an AI chatbot for medicine. N. Engl. J. Med. 388, 1233-1239 (2023).
Jiang, L.Y. et al. Health system-scale language models are all-purpose prediction engines. Nature 619, 357-362 (2023).
Churchill, A. et al. GPT models for text annotation: an empirical exploration in public policy research. Policy Stud. J. (2025).
Anderson, W. et al. Unlocking the capabilities of large language models for accelerating drug development. Clin. Pharmacol. Ther. 116, 38-41 (2024).
Cao, H. et al. Genome-wide profiles of DNA damage represent highly accurate predictors of mammalian age. Aging Cell 23, e14122 (2024).
Tacutu, R. et al. Human Ageing Genomic Resources: new and updated databases. Nucleic Acids Res. 46, D1083-D1090 (2018).
Aging Atlas Consortium. Aging Atlas: a multi-omics database for aging biology. Nucleic Acids Res. 49, D825-D830 (2021).
Avelar, R.A. et al. A multidimensional systems biology analysis of cellular senescence in aging and disease. Genome Biol. 21, 91 (2020).
Lopes, I. et al. Six drivers of aging identified among genes differentially expressed with age. bioRxiv 2024.08.02.606402 (2024).
Law, C.W., Chen, Y., Shi, W. & Smyth, G.K. voom: precision weights unlock linear model analysis tools for RNA-seq read counts. Genome Biol. 15, R29 (2014).
Love, M.I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550 (2014).
Ritchie, M.E. et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 43, e47 (2015).
Tseng, G.C., Ghosh, D. & Feingold, E. Comprehensive literature review and statistical considerations for microarray meta-analysis. Nucleic Acids Res. 40, 3785-3799 (2012).
Smyth, G.K. Linear models and empirical Bayes methods for assessing differential expression in microarray experiments. Stat. Appl. Genet. Mol. Biol. 3, Article 3 (2004).
Hedges, L.V. & Olkin, I. Statistical Methods for Meta-Analysis. Academic Press (1985).
Viechtbauer, W. Conducting meta-analyses in R with the metafor package. J. Stat. Softw. 36, 1-48 (2010).

AI Disclosure

Gemini 3 Pro Nano Banana tool was used to generate the graphical abstract based on the proposal given as input. Claude Opus 4.6 was used to draft the initial proposal. The idea, methodology and workplan are completely original ideas of the author. The llm was used for initial draft after which the author then thoroughly vetted and heavily edited the proposal for accuracy and flow of thought and added crucial details. All references have also been cross-validated.

5e:Tc2e4,

GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome

1. Overview

Timeline: 6 months

Month 1-2: LLM pipeline development, GEO-wide metadata extraction, and validation.
Month 3-4: Expression data download, harmonization, batch correction, and quality control.
Month 5: Meta-analysis and statistical modeling (Random-effects meta-analysis, subgroup analyses, and meta-regression.)
Month 6: Cross-validation, interactive visualization, manuscript preparation, and submission

2. Introduction

Research Questions.

RQ1. What molecular signatures of aging are robust across tissues, cell types, and experimental contexts?

RQ2. Which aging-associated changes are specific to particular biological systems or perturbations?

RQ3. Can meta-analysis identify high-confidence aging biomarkers that replicate across independent studies?

RQ4. What factors (technical vs. biological) explain heterogeneity in reported aging signatures?

RQ5. To what extent do DNA damage and genotoxic stress signatures overlap with, or diverge from, chronological aging programs?

3. Methods

Participants/Sample

3.2 Materials and Procedures

Phase 1: LLM Pipeline for Metadata Extraction and Relevance Classification (Months 1 to 2)

Primary aging study: aging, senescence, or DNA damage in the context of aging is the central research question.
Secondary aging study: age is a major covariate or secondary variable, with sufficient aging data for extraction.
Incidental mention: aging/senescence terms appear only in pathway annotations, background literature, or results interpretation, but the study does not investigate aging.
Ambiguous: insufficient information to classify; flagged for manual review.

Step 3: Structured Metadata Extraction. For entries passing relevance classification, submit to the Claude API with a structured extraction prompt specifying the following target fields:

Tissue/cell type (mapped to Cell Ontology for cell types, Uberon for tissues, or CLO for cell lines; with explicit annotation of whether the sample is primary tissue, primary cells, or a cell line, and for cell lines whether normal or cancer-derived)
Age or time point (normalized to years for humans)
Sex (male, female, mixed, not reported)
Experimental condition (natural aging, accelerated aging/progeroid model, senescence induction, DNA damage treatment, lifespan intervention, control)
Intervention details (agent, dose, duration where applicable)
Sample size (number of biological replicates per condition)
Sequencing platform (microarray model or RNA-seq chemistry)
Batch structure (single batch, multi-batch, unknown)

Phase 2: Expression Data Harmonization (Months 3 to 4)

Phase 3: Random-Effects Meta-Analysis (Month 5)

Phase 4: Validation and Deliverables (Month 6)

3.3 Ethics and Data Management

4. Budget

Request: $10,000 USD (in ResearchCoin)

Item	Cost	Justification
Cloud computing (GCP)	$3,000	Data storage, bulk expression data processing, meta-analysis computation across thousands of datasets
LLM API costs (Claude 4.5 Sonnet + API)	$2,000	Relevance classification and metadata extraction for 10,000 GEO entries with multi-step prompting ($0.20/dataset including validation passes)
Research assistant (100 hrs × $25/hr)	$2,500	Graduate-level research assistant with aging biology expertise for manual validation of 10% metadata sample, review of ambiguous entries flagged by pipeline, literature cross-referencing
Interactive Shiny server (development + deployment)	$1,000	Web application development and hosting for community access to curated atlas and meta-analytic results
ResearchHub Registered Report publication	$1,000	Open-access publication of final manuscript as ResearchHub Registered Report
Contingency	$500	Additional compute for expanded analyses, unforeseen API cost overruns

6. References

Shokhirev MN, Johnson AA. Modeling the human aging transcriptome across tissues, health status, and sex. Aging Cell. 2021;20(1):e13280. doi:10.1111/acel.13280
de Magalhães JP, Curado J, Church GM. Meta-analysis of age-related gene expression profiles identifies common signatures of aging. Bioinformatics. 2009;25(7):875-881. doi:10.1093/bioinformatics/btp073
Sansone, S.A. et al. Toward interoperable bioscience data. Nat. Genet. 44, 121-126 (2012).
López-Otín, C., Blasco, M.A., Partridge, L., Serrano, M. & Kroemer, G. Hallmarks of aging: an expanding universe. Cell 186, 243-278 (2023).
Aramillo Irizar, P. et al. Transcriptomic alterations during ageing reflect the shift from cancer to degenerative diseases in the elderly. Nat. Commun. 9, 327 (2018).
Palmer, D. et al. Ageing transcriptome meta-analysis reveals similarities and differences between key mammalian tissues. Aging 13, 3313-3341 (2021).
Voisin, S. et al. Exercise is associated with younger methylome and transcriptome profiles in human skeletal muscle. Aging Cell 23, e13859 (2024).
Winterhalter, P.R. et al. The circadian rhythm: a key variable in aging? Aging Cell 23, e14268 (2024).
Barrett, T. et al. NCBI GEO: archive for functional genomics data sets, update. Nucleic Acids Res. 41, D991-D995 (2013).
Brazma, A. et al. Minimum information about a microarray experiment (MIAME): toward standards for microarray data. Nat. Genet. 29, 365-371 (2001).
Lee, P. et al. Benefits, limits, and risks of GPT-4 as an AI chatbot for medicine. N. Engl. J. Med. 388, 1233-1239 (2023).
Jiang, L.Y. et al. Health system-scale language models are all-purpose prediction engines. Nature 619, 357-362 (2023).
Churchill, A. et al. GPT models for text annotation: an empirical exploration in public policy research. Policy Stud. J. (2025).
Anderson, W. et al. Unlocking the capabilities of large language models for accelerating drug development. Clin. Pharmacol. Ther. 116, 38-41 (2024).
Cao, H. et al. Genome-wide profiles of DNA damage represent highly accurate predictors of mammalian age. Aging Cell 23, e14122 (2024).
Tacutu, R. et al. Human Ageing Genomic Resources: new and updated databases. Nucleic Acids Res. 46, D1083-D1090 (2018).
Aging Atlas Consortium. Aging Atlas: a multi-omics database for aging biology. Nucleic Acids Res. 49, D825-D830 (2021).
Avelar, R.A. et al. A multidimensional systems biology analysis of cellular senescence in aging and disease. Genome Biol. 21, 91 (2020).
Lopes, I. et al. Six drivers of aging identified among genes differentially expressed with age. bioRxiv 2024.08.02.606402 (2024).
Law, C.W., Chen, Y., Shi, W. & Smyth, G.K. voom: precision weights unlock linear model analysis tools for RNA-seq read counts. Genome Biol. 15, R29 (2014).
Love, M.I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550 (2014).
Ritchie, M.E. et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 43, e47 (2015).
Tseng, G.C., Ghosh, D. & Feingold, E. Comprehensive literature review and statistical considerations for microarray meta-analysis. Nucleic Acids Res. 40, 3785-3799 (2012).
Smyth, G.K. Linear models and empirical Bayes methods for assessing differential expression in microarray experiments. Stat. Appl. Genet. Mol. Biol. 3, Article 3 (2004).
Hedges, L.V. & Olkin, I. Statistical Methods for Meta-Analysis. Academic Press (1985).
Viechtbauer, W. Conducting meta-analyses in R with the metafor package. J. Stat. Softw. 36, 1-48 (2010).

AI Disclosure

5f:T864e,GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome 1. Overview Scientific Rationale: A search for aging-related terms ("aging", "senescence", "ageing", "lifespan", "longevity") on the NCBI Gene Expression Omnibus returns over 30,000 dataset entries, with approximately 10,000 involving human samples (NCBI E-utilities query, February 2026). Yet no study has systematically mined this resource at scale. Existing aging meta-analyses are limited to manually curated subsets of 25 to 50 datasets (1,2), leaving the vast majority of public data unexploited. Two compounding bottlenecks explain why. First, GEO metadata is notoriously inconsistent: sample ages are buried in free-text descriptions, tissue annotations use non-standard vocabularies, and experimental conditions are logged in idiosyncratic formats (3). Second, and critically, keyword-based GEO searches massively overcount relevant datasets: a study that merely reports upregulation of a senescence-associated pathway (e.g., p16/CDKN2A) in its results section will appear in a "senescence" search even if the study has nothing to do with aging or cellular senescence. This false-positive problem means the true number of directly aging-relevant transcriptomic datasets is likely a fraction of the raw search total, but no systematic triage has ever been performed. This project deploys an LLM-powered curation pipeline, built on Anthropic's Claude 4.5 and above model family, to systematically extract, standardize, and quality-control metadata across all candidate GEO entries, intelligently distinguishing genuinely aging-focused datasets from incidental keyword matches. This enables the first mega meta-analysis of human aging transcriptomes at unprecedented scale. The resulting resource, a curated, machine-readable atlas of aging gene expression, will be released as an open community tool. Timeline: 6 months Month 1-2: LLM pipeline development, GEO-wide metadata extraction, and validation. Month 3-4: Expression data download, harmonization, batch correction, and quality control. Month 5: Meta-analysis and statistical modeling (Random-effects meta-analysis, subgroup analyses, and meta-regression.) Month 6: Cross-validation, interactive visualization, manuscript preparation, and submission Data Sharing: All curated metadata, processed expression matrices, analysis code, and the LLM curation pipeline will be deposited in open repositories (GitHub under MIT license, Zenodo under CC-BY-4.0). The curated metadata catalogue itself will be released as a standalone resource for community reanalysis. An interactive Shiny web application will be deployed for the community to explore meta-analytic results, query gene-level aging signatures across tissues and conditions, and download subsets of interest. The study protocol will be preregistered on ResearchHub prior to analysis. The final manuscript will be published as a ResearchHub Registered Report with regular progress updates shared via the Updates tab throughout the project. 2. Introduction Aging is the dominant risk factor for cardiovascular disease, neurodegeneration, cancer, and metabolic dysfunction, yet the molecular programs that drive age-dependent decline remain incompletely mapped (4,5). Over the past two decades, transcriptomic profiling has emerged as the primary tool for characterizing aging at the molecular level. Individual studies have identified age-associated changes in inflammation, mitochondrial function, proteostasis, and DNA damage response pathways. However, these findings exhibit surprisingly limited overlap across studies, a problem that has persisted despite the growing volume of available data (1,6). Two landmark meta-analyses illustrate both the promise and the ceiling of current approaches. Shokhirev and Johnson reprocessed approximately 6,000 raw RNA-seq samples from the Sequence Read Archive, ultimately retaining 3,060 high-quality human samples after stringent filtering, and achieved remarkable predictive accuracy for transcriptomic age (R² = 0.96) (1). Voisin et al. assembled 3,176 human skeletal muscle samples across methylome and transcriptome datasets to demonstrate that exercise training shifts molecular profiles toward a younger state (7). Both studies relied on extensive manual curation and were restricted to specific tissues or contexts. A recent preprint used only 25 GEO datasets across multiple species to identify six causal aging driver genes, explicitly noting that their small dataset count was a limitation and calling for expanded meta-analyses as more data become available (19). Meanwhile, Palmer et al. demonstrated that transcriptomic aging signatures show minimal overlap across organs, raising the fundamental question of whether truly universal aging markers exist or whether the apparent inconsistency is an artifact of small, heterogeneous datasets (6). The core problem is scale, and specifically the curatorial bottleneck that prevents achieving it. GEO currently hosts over 200,000 series entries comprising more than 6 million individual samples (9). A broad keyword search for aging-related terms returns over 30,000 dataset entries, of which approximately 10,000 involve human samples. However, these raw counts are profoundly misleading. GEO's full-text search indexes not only study titles and designs but also result descriptions, pathway annotations, and associated publications. A cancer genomics study that incidentally reports upregulation of CDKN2A or other senescence-associated genes will appear in a "senescence" search despite having no relevance to aging biology whatsoever. Similarly, studies examining age as a covariate rather than a primary variable, or studies using "aging" only in the context of sample storage or reagent shelf life, inflate the counts. The true number of datasets whose primary focus is aging, cellular senescence, DNA damage in the context of aging, or lifespan modulation is likely an order of magnitude smaller than the raw search suggests, but no systematic triage has been performed, because manual screening of 10,000+ entries is prohibitively labor-intensive. This is precisely the problem that LLM-powered curation pipelines can solve. Large language models excel at reading unstructured text descriptions, understanding experimental intent from context, and making nuanced classification decisions that go beyond keyword matching (11,12). The Critical Path Institute has demonstrated that LLMs can perform arbitrary data cleaning and knowledge extraction tasks dynamically through prompt engineering, achieving results that would otherwise require weeks of manual curation (14). By deploying an LLM pipeline to read every GEO entry's title, summary, experimental design, and sample annotations, we can distinguish genuinely aging-focused datasets from incidental keyword matches while simultaneously extracting and standardizing the metadata needed for downstream integration. This project combines these two threads, the unmet need for large-scale human aging meta-analysis and the emerging capability of LLMs for intelligent data triage and curation, into a unified pipeline we call GEO-Age. The pipeline, built on Anthropic's Claude 4.5 and above model family, will systematically process all candidate GEO entries, classify each for aging relevance, extract and standardize metadata into a machine-readable schema, and feed the curated catalogue into a multi-level random-effects meta-analysis. The scope is restricted to human samples, including primary tissues, primary cell cultures, and cell lines (both normal and cancer-derived), to maximize translational relevance and minimize the confounds introduced by cross-species ortholog mapping. Cancer-derived cell lines are included specifically for the senescence and DNA damage analyses (RQ2, RQ5), where therapy-induced and oncogene-induced senescence are biologically relevant phenomena, but are excluded from the primary chronological aging meta-analysis (RQ1, RQ3) to avoid confounding by baseline transcriptomic aberrations (telomerase reactivation, p53 loss-of-function, aneuploidy). A particular focus will be the relationship between DNA damage responses and chronological aging. Genomic instability is a canonical hallmark of aging, and recent work has shown that genome-wide DNA damage profiles can predict mammalian age with accuracy exceeding transcriptome-based models (15). Hundreds of GEO datasets examine DNA damage induction (via irradiation, oxidative stress, genotoxic agents, or replicative senescence) in human cell lines, yet no systematic comparison has tested whether these acute damage signatures recapitulate the molecular program of chronological aging or represent a distinct biological response. The scale of our meta-analysis is uniquely positioned to resolve this question. Research Questions. RQ1. What molecular signatures of aging are robust across tissues, cell types, and experimental contexts? RQ2. Which aging-associated changes are specific to particular biological systems or perturbations? RQ3. Can meta-analysis identify high-confidence aging biomarkers that replicate across independent studies? RQ4. What factors (technical vs. biological) explain heterogeneity in reported aging signatures? RQ5. To what extent do DNA damage and genotoxic stress signatures overlap with, or diverge from, chronological aging programs? This is a discovery-driven study. Rather than testing a fixed set of hypotheses, the project is designed to generate a comprehensive descriptive atlas of human aging transcriptomics at unprecedented scale, from which specific mechanistic hypotheses can be derived and subsequently tested by the community. The analysis plan (Section 3.3) specifies pre-registered analytical frameworks and decision rules, but the primary output is a curated resource and empirical characterization rather than hypothesis confirmation. Impact: This project delivers three high-value outputs to the aging research community regardless of any specific finding. First, the curated metadata catalogue, a structured, machine-readable annotation of every aging-relevant human GEO dataset, constitutes a standalone resource that will accelerate future meta-analyses across any aging subfield. Second, the GEO-Age LLM pipeline establishes a generalizable methodology for LLM-powered data curation that can be applied to any domain with heterogeneous public data. Third, the mega meta-analysis itself will either identify genuinely universal aging markers (advancing biomarker development) or definitively demonstrate that human aging transcriptomes are context-dependent (redirecting the field toward tissue-specific intervention strategies). The interactive Shiny server will ensure these outputs remain accessible and explorable by the broader community well beyond the funding period. 3. Methods Participants/Sample Target population. All publicly available high-throughput transcriptomic datasets in GEO involving human samples that measure age-dependent molecular changes, cellular senescence, DNA damage responses, or lifespan-modulating interventions in primary tissues, primary cell cultures, or cell lines (normal and cancer-derived). Inclusion criteria. (i) Human samples (primary tissues, primary cells, or established cell lines). (ii) Study's primary focus involves aging, cellular senescence, DNA damage response, or age-modulating intervention, as determined by the GEO-Age pipeline's relevance classification, not merely keyword co-occurrence. (iii) At least 3 biological replicates per experimental condition. (iv) Raw or processed expression data available for download. (v) Sufficient metadata to assign age/condition groups (after AI-assisted extraction). Exclusion criteria. (i) Non-human samples. (ii) Datasets where aging/senescence terms appear only incidentally (e.g., pathway annotations in results, age as a minor covariate) rather than as the study's primary focus. (iii) Fewer than 3 biological replicates per group. (iv) Metadata insufficient for age/condition assignment even after AI extraction and manual review of the original publication. (v) Studies of age-associated diseases (e.g., Alzheimer's, cancer prognosis) without age-matched healthy controls or explicit aging comparisons. (vi) Datasets flagged for known technical failures in GEO. (vii) Single-cell RNA-seq datasets, which require distinct analytical frameworks (pseudobulk aggregation, cell-type deconvolution) that are outside the scope of this bulk transcriptomics meta-analysis; extension to scRNA-seq is planned as future work pending proof of concept. Expected sample size. Approximately 10,000 human GEO entries match aging-related keyword searches. Based on the high rate of false-positive keyword matches inherent to GEO's full-text indexing, we estimate that 1,000 to 2,500 of these represent genuinely aging-focused studies after AI-assisted relevance filtering. After applying the remaining inclusion/exclusion criteria (replication, data availability), we conservatively estimate 500 to 1,500 high-quality transcriptomic datasets representing 20,000 to 60,000 individual samples. For meta-analysis, a minimum of 20 independent studies per subgroup analysis is required to achieve stable random-effects estimates; we expect to exceed this threshold for all major tissue categories. 3.2 Materials and Procedures Phase 1: LLM Pipeline for Metadata Extraction and Relevance Classification (Months 1 to 2) The GEO-Age pipeline is a multi-step LLM classification and extraction system built on Anthropic's Claude 4.5 and above model family that performs two critical functions: (a) relevance classification, meaning determining whether a GEO entry is genuinely focused on aging/senescence versus incidentally matching keywords, and (b) structured metadata extraction for entries classified as relevant. Step 1: Retrieval. Use the NCBI E-utilities API (esearch + efetch) and the GEOquery R package to programmatically download all human GEO Series (GSE) records matching aging-related search terms ("aging", "ageing", "senescence", "cellular senescence", "lifespan", "longevity", "gerontology", "DNA damage aging", "replicative senescence", "oxidative stress aging"). For each entry, retrieve: title, summary, overall design, sample characteristics (GSM-level), platform annotation, publication DOI, and supplementary file links. Store raw metadata in a structured JSON schema. Step 2: AI-Powered Relevance Classification. For each GEO entry, submit the concatenated metadata fields to the Claude API with a classification prompt that instructs the model to assign one of four relevance categories: Primary aging study: aging, senescence, or DNA damage in the context of aging is the central research question. Secondary aging study: age is a major covariate or secondary variable, with sufficient aging data for extraction. Incidental mention: aging/senescence terms appear only in pathway annotations, background literature, or results interpretation, but the study does not investigate aging. Ambiguous: insufficient information to classify; flagged for manual review. Only entries classified as "primary" or "secondary" proceed to metadata extraction. "Incidental mention" entries are excluded. "Ambiguous" entries are flagged for manual review by the research assistant, with the original publication abstract retrieved via DOI to aid classification. Step 3: Structured Metadata Extraction. For entries passing relevance classification, submit to the Claude API with a structured extraction prompt specifying the following target fields: Tissue/cell type (mapped to Cell Ontology for cell types, Uberon for tissues, or CLO for cell lines; with explicit annotation of whether the sample is primary tissue, primary cells, or a cell line, and for cell lines whether normal or cancer-derived) Age or time point (normalized to years for humans) Sex (male, female, mixed, not reported) Experimental condition (natural aging, accelerated aging/progeroid model, senescence induction, DNA damage treatment, lifespan intervention, control) Intervention details (agent, dose, duration where applicable) Sample size (number of biological replicates per condition) Sequencing platform (microarray model or RNA-seq chemistry) Batch structure (single batch, multi-batch, unknown) The prompt enforces JSON output conforming to a predefined schema with constrained value sets for categorical fields. When the model cannot confidently extract a field, it returns "uncertain" with a free-text explanation, flagging the entry for manual review. Step 4: Validation. Three validation layers ensure accuracy, targeting ≥95% field-level agreement with human annotation and aiming for ≥99% on critical fields (relevance classification, tissue type, age): (a) Automated consistency checks: Cross-validate extracted organism against the GEO platform annotation. Verify that reported ages fall within plausible human ranges. Flag entries where the model's output contradicts structured GEO fields. (b) Independent LLM re-classification: For entries flagged as uncertain or where automated checks detect inconsistencies, re-submit to a separate API call with the original publication abstract (retrieved via DOI) appended to the prompt, enabling the model to resolve ambiguities from richer context. Because each API call operates without memory of prior classifications, this constitutes a functionally independent re-classification rather than self-correction. Concordance rates between first-pass and re-classification will be reported as an empirical measure of extraction reliability. (c) Human spot-check: A random 10% sample of extracted entries will be manually validated against the original publications by a trained research assistant with graduate-level experience in aging biology and GEO data structures. If field-level accuracy falls below 95% on any target field, the extraction prompt will be iteratively refined and the affected batch re-processed. All ambiguous entries flagged by the pipeline will be manually reviewed regardless of the spot-check results. Cohen's kappa will be computed between LLM and human annotations for the four-category relevance classification to quantify inter-rater reliability. Step 5: Ontology Mapping and Standardization. All extracted terms are mapped to controlled vocabularies using the EMBL-EBI Ontology Lookup Service API: tissue terms to Uberon, cell types to Cell Ontology, diseases to Disease Ontology, experimental interventions to the Experimental Factor Ontology (EFO). Unmapped terms are flagged for manual resolution. Phase 2: Expression Data Harmonization (Months 3 to 4) Step 1: Data Download. Download raw or normalized expression matrices for all included datasets. For microarray data, download CEL files where available; otherwise, use pre-processed series matrices. For RNA-seq, download count matrices from GEO Supplementary Files or re-quantify from FASTQ via the Recount3 resource where available. Step 2: Preprocessing and Normalization. Apply platform-appropriate normalization. For Affymetrix microarrays, apply RMA normalization via the oligo R package. For two-color arrays, apply loess within-array normalization followed by quantile between-array normalization via limma. For RNA-seq count data, apply the limma-voom transformation pipeline (20), which converts raw counts to log-counts-per-million with observation-level precision weights. We use limma-voom rather than DESeq2 as the primary RNA-seq framework because DESeq2's negative binomial GLM becomes computationally prohibitive beyond approximately 10,000 samples (21), whereas limma-voom scales efficiently to mega-analysis (20,22). Probe-to-gene mapping uses manufacturer annotation files with ambiguous probes removed. Gene identifiers are harmonized to HGNC symbols with Ensembl gene IDs as a secondary namespace. Step 3: Quality Control. For each dataset: (a) Remove outlier samples with inter-sample correlation < 0.8 (Pearson r on global expression profile). (b) Assess batch effects via principal component analysis; flag datasets where the first principal component separates by technical batch rather than biological condition. (c) Apply ComBat-seq for within-study batch correction where batch information is available. Step 4: Per-Dataset Differential Expression (Stage 1 of Two-Stage Meta-Analysis). Each dataset is analyzed independently to extract gene-level effect sizes (23). For microarray datasets, fit gene-wise linear models using limma's empirical Bayes moderated t-statistics (24). For RNA-seq datasets, apply the limma-voom pipeline: filter lowly expressed genes (requiring at least 10 counts in a minimum number of samples), apply the voom transformation, and fit gene-wise weighted linear models with empirical Bayes moderation (20). For studies with discrete age groups (young vs. old), the model is a simple contrast between groups. For studies with continuous age, fit a linear regression of expression on age. Where metadata permits, include sex, batch, and tissue subregion as covariates in the per-dataset model to reduce confounding. For each gene, extract the log2 fold-change (logFC) and its standard error. To combine continuous and categorical contrasts in a single meta-analysis, all effect sizes are converted to a common scale: the standardized mean difference (Hedges' g), computed from the logFC divided by the pooled within-group standard deviation, with a small-sample correction factor (25). For discrete age-group comparisons, this conversion is direct. For continuous age designs (regression of expression on age), the regression slope is converted to an equivalent standardized mean difference using the age range of each study to define the contrast (youngest versus oldest quartile), with the residual standard deviation of the regression as the denominator. Where the age range or residual SD cannot be recovered from GEO metadata, the dataset is flagged for manual extraction from the original publication. A sensitivity analysis will compare meta-analytic results when restricted to discrete-group studies only versus including converted continuous designs, to assess robustness to conversion assumptions. Phase 3: Random-Effects Meta-Analysis (Month 5) Stage 2: Combining Effect Sizes Across Datasets. For each gene, combine the per-dataset effect sizes using a random-effects meta-analysis with REML estimation of between-study variance via the metafor R package (26). REML is preferred over the DerSimonian-Laird moment estimator, which underestimates between-study variance under substantial heterogeneity (23,26). For each gene, compute: (a) the summary effect size (weighted mean Hedges' g) and its 95% confidence interval, (b) Cochran's Q statistic and I² for heterogeneity, and (c) the prediction interval. Apply Benjamini-Hochberg FDR correction across all tested genes (q < 0.05). Computational Strategy. Both stages are embarrassingly parallel. The per-dataset pipeline is orchestrated via Snakemake on the IISc HPC cluster (or equivalent GCP instances). Stage 2 parallelizes trivially across genes via BiocParallel. The computational bottleneck is I/O (downloading and parsing thousands of GEO files), not model fitting. Analysis for RQ1 and RQ3: conserved aging signatures and biomarker identification. Using the genome-wide REML random-effects results from Stage 2, restricted to primary tissues, primary cell cultures, and normal cell lines (excluding cancer-derived cell lines to avoid confounding by oncogenic transcriptomic aberrations), identify genes with consistent direction of effect and q < 0.05 across a high proportion of datasets as candidate conserved aging markers. Compare against the Human Ageing Genomic Resources (GenAge, CellAge) and the Aging Atlas (16,17,18). Assess replication by leave-one-study-out cross-validation and by testing whether top markers replicate in held-out datasets not included in the meta-analysis. Analysis for RQ2: Tissue-specific and perturbation-specific signatures. Stratify datasets by sample type: (a) primary tissues by category (immune/hematopoietic, epithelial, neural, muscle, adipose, other), (b) primary cell cultures, (c) normal cell lines, (d) cancer cell lines. Perform separate meta-analyses within each stratum. Compute pairwise Spearman correlations of gene-level effect sizes between strata. Apply hierarchical clustering on the correlation matrix to identify natural groupings of aging signatures. Perform pathway enrichment (KEGG, Reactome, GO) within each cluster to characterize biological themes. Analysis for RQ4: Sources of heterogeneity. Perform subgroup meta-analyses stratified by: sequencing platform (microarray vs. RNA-seq), tissue category, sample size per group (above/below median), and publication year (pre-2015 vs. post-2015). Report I² within each stratum to assess whether heterogeneity is driven by technical or biological factors. Use leave-one-study-out sensitivity analysis to identify influential datasets. Analysis for RQ5: DNA damage versus chronological aging. Separate datasets into two groups: (a) chronological aging studies (natural aging across timepoints) and (b) DNA damage/genotoxic stress studies (irradiation, oxidative stress, replicative senescence), including cancer-derived cell lines undergoing therapy-induced or DNA damage-induced senescence. Perform meta-analysis within each group. To compare the two signatures, apply three complementary approaches: (i) rank-rank hypergeometric overlap (RRHO) analysis, which provides a threshold-free visualization of concordance and discordance across the full ranked gene lists without arbitrary significance cutoffs; (ii) Spearman ρ restricted to genes reaching nominal significance (p < 0.05) in at least one of the two categories, to avoid inflating the correlation through shared null effects across thousands of uninformative genes; and (iii) pathway-level comparison using gene set enrichment analysis (GSEA) against KEGG and Reactome, testing whether shared and divergent pathways align with predictions from the hallmarks of aging framework (4). Sensitivity analyses. (i) Restrict to studies with n ≥ 5 per group. (ii) Restrict to primary tissues only (exclude cell lines). (iii) Stratify by publication year (pre-2015 microarray era vs. post-2015 RNA-seq era). (iv) Funnel plots and Egger's test for publication bias. (v) Restrict to discrete age-group designs only (excluding continuous-age regression designs) to assess robustness to the Hedges' g conversion procedure. Phase 4: Validation and Deliverables (Month 6) Cross-reference validation. Compare meta-analytic aging gene sets against: (a) the GenAge database of aging-related genes (16), (b) the CellAge database of cellular senescence genes (18), (c) the Aging Atlas multi-omics resource (17), and (d) recently published independent aging cohorts not included in the meta-analysis. Quantify overlap using hypergeometric enrichment tests. Interactive Shiny server. Develop and deploy an interactive Shiny web application enabling users to: query the curated metadata catalogue; visualize gene-level meta-analytic results across tissues, cell types, and conditions; explore tissue-specific and perturbation-specific aging signatures; and download subsets of interest for their own analyses. The server will be deployed via shinyapps.io or a containerized instance and maintained as a community resource. Open resource release. Deposit the following on Zenodo with DOI: (a) complete curated metadata catalogue (JSON + CSV), (b) processed expression matrices with harmonized annotations, (c) full meta-analytic results for all genes across all subgroup analyses, (d) the GEO-Age LLM pipeline code and prompt library (GitHub, MIT license), (e) Docker container for full pipeline reproducibility. 3.3 Ethics and Data Management Ethics. This study exclusively reanalyzes publicly available, de-identified datasets from GEO. No new human subjects research, animal experimentation, or data collection involving identifiable individuals is conducted. All source datasets comply with the original studies' ethics approvals and GEO's data sharing policies. Data Management. All analysis code will be version-controlled on GitHub with MIT license from project initiation. The curated metadata catalogue and processed datasets will be deposited on Zenodo with a persistent DOI upon completion. Analysis pipelines will be containerized using Docker for full reproducibility. The final manuscript will be published as a ResearchHub Registered Report. The study protocol will be preregistered on ResearchHub prior to initiating the meta-analysis phase, per the RFP requirements. 4. Budget Request: $10,000 USD (in ResearchCoin) Item Cost Justification Cloud computing (GCP) $3,000 Data storage, bulk expression data processing, meta-analysis computation across thousands of datasets LLM API costs (Claude 4.5 Sonnet + API) $2,000 Relevance classification and metadata extraction for 10,000 GEO entries with multi-step prompting ($0.20/dataset including validation passes) Research assistant (100 hrs × $25/hr) $2,500 Graduate-level research assistant with aging biology expertise for manual validation of 10% metadata sample, review of ambiguous entries flagged by pipeline, literature cross-referencing Interactive Shiny server (development + deployment) $1,000 Web application development and hosting for community access to curated atlas and meta-analytic results ResearchHub Registered Report publication $1,000 Open-access publication of final manuscript as ResearchHub Registered Report Contingency $500 Additional compute for expanded analyses, unforeseen API cost overruns Budget is flexible and will be adjusted based on Phase 1 results. If metadata extraction proves more efficient than estimated, surplus funds will be allocated to expanding the analysis to additional sample types, scRNA-Seq or deeper pathway-level analyses. 6. References Shokhirev MN, Johnson AA. Modeling the human aging transcriptome across tissues, health status, and sex. Aging Cell. 2021;20(1):e13280. doi:10.1111/acel.13280 de Magalhães JP, Curado J, Church GM. Meta-analysis of age-related gene expression profiles identifies common signatures of aging. Bioinformatics. 2009;25(7):875-881. doi:10.1093/bioinformatics/btp073 Sansone, S.A. et al. Toward interoperable bioscience data. Nat. Genet. 44, 121-126 (2012). López-Otín, C., Blasco, M.A., Partridge, L., Serrano, M. & Kroemer, G. Hallmarks of aging: an expanding universe. Cell 186, 243-278 (2023). Aramillo Irizar, P. et al. Transcriptomic alterations during ageing reflect the shift from cancer to degenerative diseases in the elderly. Nat. Commun. 9, 327 (2018). Palmer, D. et al. Ageing transcriptome meta-analysis reveals similarities and differences between key mammalian tissues. Aging 13, 3313-3341 (2021). Voisin, S. et al. Exercise is associated with younger methylome and transcriptome profiles in human skeletal muscle. Aging Cell 23, e13859 (2024). Winterhalter, P.R. et al. The circadian rhythm: a key variable in aging? Aging Cell 23, e14268 (2024). Barrett, T. et al. NCBI GEO: archive for functional genomics data sets, update. Nucleic Acids Res. 41, D991-D995 (2013). Brazma, A. et al. Minimum information about a microarray experiment (MIAME): toward standards for microarray data. Nat. Genet. 29, 365-371 (2001). Lee, P. et al. Benefits, limits, and risks of GPT-4 as an AI chatbot for medicine. N. Engl. J. Med. 388, 1233-1239 (2023). Jiang, L.Y. et al. Health system-scale language models are all-purpose prediction engines. Nature 619, 357-362 (2023). Churchill, A. et al. GPT models for text annotation: an empirical exploration in public policy research. Policy Stud. J. (2025). Anderson, W. et al. Unlocking the capabilities of large language models for accelerating drug development. Clin. Pharmacol. Ther. 116, 38-41 (2024). Cao, H. et al. Genome-wide profiles of DNA damage represent highly accurate predictors of mammalian age. Aging Cell 23, e14122 (2024). Tacutu, R. et al. Human Ageing Genomic Resources: new and updated databases. Nucleic Acids Res. 46, D1083-D1090 (2018). Aging Atlas Consortium. Aging Atlas: a multi-omics database for aging biology. Nucleic Acids Res. 49, D825-D830 (2021). Avelar, R.A. et al. A multidimensional systems biology analysis of cellular senescence in aging and disease. Genome Biol. 21, 91 (2020). Lopes, I. et al. Six drivers of aging identified among genes differentially expressed with age. bioRxiv 2024.08.02.606402 (2024). Law, C.W., Chen, Y., Shi, W. & Smyth, G.K. voom: precision weights unlock linear model analysis tools for RNA-seq read counts. Genome Biol. 15, R29 (2014). Love, M.I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550 (2014). Ritchie, M.E. et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 43, e47 (2015). Tseng, G.C., Ghosh, D. & Feingold, E. Comprehensive literature review and statistical considerations for microarray meta-analysis. Nucleic Acids Res. 40, 3785-3799 (2012). Smyth, G.K. Linear models and empirical Bayes methods for assessing differential expression in microarray experiments. Stat. Appl. Genet. Mol. Biol. 3, Article 3 (2004). Hedges, L.V. & Olkin, I. Statistical Methods for Meta-Analysis. Academic Press (1985). Viechtbauer, W. Conducting meta-analyses in R with the metafor package. J. Stat. Softw. 36, 1-48 (2010). AI Disclosure Gemini 3 Pro Nano Banana tool was used to generate the graphical abstract based on the proposal given as input. Claude Opus 4.6 was used to draft the initial proposal. The idea, methodology and workplan are completely original ideas of the author. The llm was used for initial draft after which the author then thoroughly vetted and heavily edited the proposal for accuracy and flow of thought and added crucial details. All references have also been cross-validated.22:["$","$L56",null,{"children":["$","$L4",null,{"topBanner":["$","$L57",null,{"work":{"id":5927,"type":"$undefined","contentType":"preregistration","title":"GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome","slug":"multi-omics-meta-analysis-of-aging-trajectories-using-ai-powered-data-integration","createdDate":"2026-02-15T13:12:26.759468Z","createdByUserId":74374,"updatedDate":"2026-02-22T19:41:32.725940Z","publishedDate":"2026-02-15T13:12:26.759468Z","authors":[{"authorProfile":{"id":5146850,"fullName":"Amruth Bhat","firstName":"Amruth","lastName":"Bhat","profileImage":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/06/01/blob_QMU7UYj","headline":"PhD Student at Indian Institute of Science","profileUrl":"/author/5146850","user":{"id":0,"email":"","firstName":"","lastName":"","fullName":"Unknown User","isVerified":false,"authorProfile":"$undefined","balance":0,"lockedBalance":0,"promotionalBalance":0,"fundingCredits":0,"balances":"$undefined","totalRsc":0,"hasCompletedOnboarding":false,"createdDate":"$undefined","moderator":false,"editorOfHubs":"$undefined","isModerator":false,"isFunder":false,"referralCode":"$undefined","authProvider":"$undefined","isStakingOptedIn":false,"balanceHistory":0,"raw":74374},"description":"PhD researcher mapping biological aging across India. Immune aging, multiomics, systems biology, ML & mathematical modeling to decode how and why Indian populations age differently. Building tools for mechanistic insight at scale.","createdDate":"2024-10-23T10:45:05.775749Z","education":[{"id":2876,"name":"Indian Institute of Science","year":{"label":"2024","value":"2024"},"major":"Bioengineering","degree":{"label":"Master's Degree","value":"Master's Degree"},"summary":"Bioengineering Master's Degree '24, Indian Institute of Science","is_public":false,"university":{"id":2876,"name":"Indian Institute of Science"}},{"id":9316,"name":"Visvesvaraya Technological University","year":{"label":"2022","value":"2022"},"major":"Biotechnology","degree":{"label":"Bachelor of Engineering","value":"BE"},"summary":"Biotechnology BE '22, Visvesvaraya Technological University","is_public":false,"university":{"id":9316,"city":"","name":"Visvesvaraya Technological University","state":"","country":"India","updated_date":"2020-11-24T21:44:27.631606Z"}},{"id":2876,"name":"Indian Institute of Science","year":{"label":"2030","value":"2030"},"major":"Bioengineering","degree":{"label":"Doctor of Philosophy","value":"PhD"},"summary":"Bioengineering PhD '30, Indian Institute of Science","is_public":true,"university":{"id":2876,"name":"Indian Institute of Science"}}],"twitter":"https://x.com/__adb__7","facebook":"$undefined","linkedin":"https://www.linkedin.com/in/amruth-bhat/","googleScholar":"https://scholar.google.com/citations?user=kpg2O-QAAAAJ","orcidId":"https://orcid.org/0000-0002-4075-1939","isOrcidConnected":true,"isClaimed":true,"hIndex":1,"i10Index":"$undefined","userId":"$undefined","isVerified":true,"isHubEditor":false,"editorOfHubs":[],"raw":{"id":5146850,"is_public":true,"is_removed":false,"is_removed_date":null,"user":74374,"first_name":"Amruth","last_name":"Bhat","created_date":"2024-10-23T10:45:05.775749Z","updated_date":"2026-02-19T06:14:54.746437Z","description":"PhD researcher mapping biological aging across India. Immune aging, multiomics, systems biology, ML & mathematical modeling to decode how and why Indian populations age differently. Building tools for mechanistic insight at scale.","h_index":1,"i10_index":0,"profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/06/01/blob_QMU7UYj","author_score":0,"university":null,"orcid_id":"https://orcid.org/0000-0002-4075-1939","openalex_ids":["https://openalex.org/A5082829184"],"education":"$22:props:children:props:topBanner:props:work:authors:0:authorProfile:education","headline":"PhD Student at Indian Institute of Science","facebook":null,"twitter":"https://x.com/__adb__7","linkedin":"https://www.linkedin.com/in/amruth-bhat/","google_scholar":"https://scholar.google.com/citations?user=kpg2O-QAAAAJ","claimed":true,"country_code":null,"created_source":"RESEARCHHUB","last_full_fetch_from_openalex":null,"two_year_mean_citedness":0,"merged_with_author":null,"added_as_editor_date":null,"is_hub_editor_of":[],"is_hub_editor":false,"num_posts":1,"is_orcid_connected":true,"reputation":102,"reputation_v2":{"hub":{"id":231,"name":"Molecular Biology","slug":"molecular-biology-1"},"score":4,"percentile":0.001,"bins":[[0,1000],[1000,10000],[10000,100000],[100000,1000000]]},"reputation_list":[{"hub":{"id":231,"name":"Molecular Biology","slug":"molecular-biology-1"},"score":4,"percentile":0.001,"bins":[[0,1000],[1000,10000],[10000,100000],[100000,1000000]]}],"suspended_status":{"probable_spammer":false,"is_suspended":false},"total_score":null,"wallet":null,"is_verified":true}},"isCorresponding":false,"position":"middle","raw":"$22:props:children:props:topBanner:props:work:authors:0:authorProfile:raw"}],"abstract":"$58","doi":"10.55277/researchhub.hvwscfmu","journal":"$undefined","topics":[{"id":594084,"name":"AI","slug":"ai"},{"id":665770,"name":"Data Analysis with R","slug":"data-analysis-with-r"},{"id":669794,"name":"Database","slug":"database-2"},{"id":670130,"name":"Aging","slug":"aging-2"}],"formats":[],"license":"$undefined","pdfCopyrightAllowsDisplay":true,"figures":[],"versions":[],"metrics":{"votes":0,"adjustedScore":"$undefined","comments":0,"saves":0,"reviewScore":4,"reviews":1,"earned":0,"views":0},"unifiedDocumentId":9011527,"registeredReportId":null,"postType":"PREREGISTRATION","fundraise":"$undefined","note":{"id":22704,"access":"WORKSPACE","organization":{"id":72595,"memberCount":1,"userPermission":"$undefined","createdDate":"2024-10-23T10:45:05.836503Z","coverImage":null,"description":"","name":"Amruth Bhat's Org","slug":"amruth-bhats-org"},"createdDate":"2026-02-15T11:35:34.550386Z","updatedDate":"2026-02-16T21:41:09.249004Z","title":"GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome","isRemoved":false,"post":{"id":5927,"slug":"multi-omics-meta-analysis-of-aging-trajectories-using-ai-powered-data-integration","contentType":"preregistration","documentType":"PREREGISTRATION","moderationStatus":"APPROVED","fundraise":{"id":288,"amountRaised":{"usd":4.986445502324,"rsc":77.132},"goalAmount":{"usd":10000,"rsc":154683.33097002984},"contributors":{"numContributors":3,"topContributors":[{"id":50877,"authorProfile":{"id":984288,"fullName":"Suramya Asthana","firstName":"Suramya","lastName":"Asthana","profileImage":"https://storage.prod.researchhub.com/uploads/author_profile_images/2024/11/17/blob_lMGDGID","headline":"Early Career Fellow | Aging Biology & Longevity","profileUrl":"/author/984288","user":{"id":0,"email":"","firstName":"","lastName":"","fullName":"Unknown User","isVerified":false,"authorProfile":"$undefined","balance":0,"lockedBalance":0,"promotionalBalance":0,"fundingCredits":0,"balances":"$undefined","totalRsc":0,"hasCompletedOnboarding":false,"createdDate":"$undefined","moderator":false,"editorOfHubs":"$undefined","isModerator":false,"isFunder":false,"referralCode":"$undefined","authProvider":"$undefined","isStakingOptedIn":false,"balanceHistory":0,"raw":50877},"description":"I'm a Cell Biologist with a specialization in Aging Biology and Cellular Senescence. I completed my PhD from Indian Institute of Science, Bengaluru India. I'm an Early Career Fellow in the Longevity Research Group at IISc.","createdDate":"2024-01-08T04:17:03.868581Z","education":[{"id":2876,"name":"Indian Institute of Science","year":{"label":"2023","value":"2023"},"major":"Aging Biology","degree":{"label":"Doctor of Philosophy","value":"PhD"},"summary":"Aging Biology PhD '23, Indian Institute of Science","is_public":true,"university":{"id":2876,"name":"Indian Institute of Science"}},{"id":3351,"name":"Jaypee Institute of Information Technology","year":{"label":"2016","value":"2016"},"major":"Biotechnology","degree":{"label":"Master of Technology","value":"MTech"},"summary":"Biotechnology MTech '16, Jaypee Institute of Information Technology","is_public":false}],"twitter":"https://twitter.com/suramyaasthana","facebook":"$undefined","linkedin":"https://www.linkedin.com/in/suramyaa/","googleScholar":"https://scholar.google.com/citations?user=HWI44igAAAAJ&hl=en","orcidId":"https://orcid.org/0009-0007-4148-467X","isOrcidConnected":false,"isClaimed":true,"hIndex":2,"i10Index":2,"userId":"$undefined","isVerified":true,"isHubEditor":false,"editorOfHubs":[],"raw":{"id":984288,"is_verified":true,"is_public":true,"is_removed":false,"is_removed_date":null,"first_name":"Suramya","last_name":"Asthana","created_date":"2024-01-08T04:17:03.868581Z","updated_date":"2026-07-14T09:41:53.990618Z","description":"I'm a Cell Biologist with a specialization in Aging Biology and Cellular Senescence. I completed my PhD from Indian Institute of Science, Bengaluru India. I'm an Early Career Fellow in the Longevity Research Group at IISc.","h_index":2,"i10_index":2,"profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2024/11/17/blob_lMGDGID","author_score":0,"orcid_id":"https://orcid.org/0009-0007-4148-467X","openalex_ids":["https://openalex.org/A5023886114"],"education":"$22:props:children:props:topBanner:props:work:note:post:fundraise:contributors:topContributors:0:authorProfile:education","headline":"Early Career Fellow | Aging Biology & Longevity","facebook":null,"twitter":"https://twitter.com/suramyaasthana","linkedin":"https://www.linkedin.com/in/suramyaa/","google_scholar":"https://scholar.google.com/citations?user=HWI44igAAAAJ&hl=en","claimed":true,"country_code":null,"created_source":"RESEARCHHUB","last_full_fetch_from_openalex":null,"two_year_mean_citedness":0.75,"user":50877,"university":null,"merged_with_author":null}},"totalContribution":{"rsc":54.494,"usd":0},"contributions":[{"amount":54.494,"currency":"RSC","date":"2026-02-20T10:21:01.867276Z"}]},{"id":109841,"authorProfile":{"id":7136539,"fullName":"Zhen-Yu YIN","firstName":"Zhen-Yu","lastName":"YIN","profileImage":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/01/11/blob_WZNkWbV","headline":"Artificial Intelligence | Computer Science","profileUrl":"/author/7136539","user":{"id":0,"email":"","firstName":"","lastName":"","fullName":"Unknown User","isVerified":false,"authorProfile":"$undefined","balance":0,"lockedBalance":0,"promotionalBalance":0,"fundingCredits":0,"balances":"$undefined","totalRsc":0,"hasCompletedOnboarding":false,"createdDate":"$undefined","moderator":false,"editorOfHubs":"$undefined","isModerator":false,"isFunder":false,"referralCode":"$undefined","authProvider":"$undefined","isStakingOptedIn":false,"balanceHistory":0,"raw":109841},"description":"My research focuses on machine learning tools that help humans make better decisions, mainly interpretable machine learning and its applications.","createdDate":"2025-01-11T02:37:16.465858Z","education":[],"twitter":"$undefined","facebook":"$undefined","linkedin":"$undefined","googleScholar":"$undefined","orcidId":"$undefined","isOrcidConnected":false,"isClaimed":true,"hIndex":2,"i10Index":"$undefined","userId":"$undefined","isVerified":true,"isHubEditor":false,"editorOfHubs":[],"raw":{"id":7136539,"is_verified":true,"is_public":true,"is_removed":false,"is_removed_date":null,"first_name":"Zhen-Yu","last_name":"YIN","created_date":"2025-01-11T02:37:16.465858Z","updated_date":"2026-01-21T13:01:03.321052Z","description":"My research focuses on machine learning tools that help humans make better decisions, mainly interpretable machine learning and its applications.","h_index":2,"i10_index":0,"profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/01/11/blob_WZNkWbV","author_score":0,"orcid_id":"","openalex_ids":["https://openalex.org/A5010261718","https://openalex.org/A5041310778"],"education":"$22:props:children:props:topBanner:props:work:note:post:fundraise:contributors:topContributors:1:authorProfile:education","headline":"Artificial Intelligence | Computer Science","facebook":null,"twitter":"","linkedin":"","google_scholar":"","claimed":true,"country_code":null,"created_source":"RESEARCHHUB","last_full_fetch_from_openalex":null,"two_year_mean_citedness":3,"user":109841,"university":null,"merged_with_author":null}},"totalContribution":{"rsc":11.475,"usd":0},"contributions":[{"amount":11.475,"currency":"RSC","date":"2026-03-10T11:52:07.109347Z"}]},{"id":184629,"authorProfile":{"id":8385365,"fullName":"Qi-hang Guo","firstName":"Qi-hang","lastName":"Guo","profileImage":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/12/20/blob_261d2PH","headline":"Trustworthy AI & Information Systems Research","profileUrl":"/author/8385365","user":{"id":0,"email":"","firstName":"","lastName":"","fullName":"Unknown User","isVerified":false,"authorProfile":"$undefined","balance":0,"lockedBalance":0,"promotionalBalance":0,"fundingCredits":0,"balances":"$undefined","totalRsc":0,"hasCompletedOnboarding":false,"createdDate":"$undefined","moderator":false,"editorOfHubs":"$undefined","isModerator":false,"isFunder":false,"referralCode":"$undefined","authProvider":"$undefined","isStakingOptedIn":false,"balanceHistory":0,"raw":184629},"description":"I am a researcher working at the intersection of Trustworthy AI and Information Systems. I serve as a reviewer for leading international journals such as ACM Computing Surveys, IEEE TNNLS, IEEE TKDE, and ACM TKDD.","createdDate":"2025-12-20T01:58:09.452226Z","education":[],"twitter":"$undefined","facebook":"$undefined","linkedin":"$undefined","googleScholar":"https://scholar.google.com/citations?hl=zh-CN&view_op=list_works&gmla=AIqSsVvLfzhl5UNKDCn8TpToVpI5UD2nQoI4zim2ITnk_qE-GEZY9GXVu4Rh6yBadODo9Pp4DvXXDSfavWsEbnd9uEc&user=tmXSCUMAAAAJ","orcidId":"https://orcid.org/0009-0003-4053-3880","isOrcidConnected":false,"isClaimed":true,"hIndex":"$undefined","i10Index":"$undefined","userId":"$undefined","isVerified":true,"isHubEditor":false,"editorOfHubs":[],"raw":{"id":8385365,"is_verified":true,"is_public":true,"is_removed":false,"is_removed_date":null,"first_name":"Qi-hang","last_name":"Guo","created_date":"2025-12-20T01:58:09.452226Z","updated_date":"2026-05-25T04:20:46.222245Z","description":"I am a researcher working at the intersection of Trustworthy AI and Information Systems. I serve as a reviewer for leading international journals such as ACM Computing Surveys, IEEE TNNLS, IEEE TKDE, and ACM TKDD.","h_index":0,"i10_index":0,"profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/12/20/blob_261d2PH","author_score":0,"orcid_id":"https://orcid.org/0009-0003-4053-3880","openalex_ids":[],"education":"$22:props:children:props:topBanner:props:work:note:post:fundraise:contributors:topContributors:2:authorProfile:education","headline":"Trustworthy AI & Information Systems Research","facebook":null,"twitter":"","linkedin":"","google_scholar":"https://scholar.google.com/citations?hl=zh-CN&view_op=list_works&gmla=AIqSsVvLfzhl5UNKDCn8TpToVpI5UD2nQoI4zim2ITnk_qE-GEZY9GXVu4Rh6yBadODo9Pp4DvXXDSfavWsEbnd9uEc&user=tmXSCUMAAAAJ","claimed":true,"country_code":null,"created_source":"RESEARCHHUB","last_full_fetch_from_openalex":null,"two_year_mean_citedness":0,"user":184629,"university":null,"merged_with_author":null}},"totalContribution":{"rsc":11.163,"usd":0},"contributions":[{"amount":11.163,"currency":"RSC","date":"2026-03-10T21:55:17.466429Z"}]}]},"createdBy":{"id":0,"email":"","firstName":"","lastName":"","fullName":"Unknown User","isVerified":false,"authorProfile":"$undefined","balance":0,"lockedBalance":0,"totalRsc":0,"hasCompletedOnboarding":false,"moderator":false,"raw":null,"isModerator":false,"isFunder":false,"authProvider":"$undefined","isStakingOptedIn":false,"balanceHistory":0},"status":"OPEN","goalCurrency":"USD","startDate":"2026-02-15T13:12:26.813630Z","endDate":"2026-03-17T13:12:26.813475Z","createdDate":"","updatedDate":"","postId":"$undefined","postTitle":"$undefined","postSlug":"$undefined","postImage":null,"reviewMetrics":"$undefined","nonprofit":"$undefined","raw":{"id":288,"amount_raised":{"usd":4.986445502324,"rsc":77.132},"goal_amount":{"usd":10000,"rsc":154683.33097002984},"contributors":{"total":3,"top":[{"id":50877,"author_profile":"$22:props:children:props:topBanner:props:work:note:post:fundraise:contributors:topContributors:0:authorProfile:raw","is_funder":false,"rsc_earned":null,"balances":null,"benefits_expire_on":null,"editor_of":[],"is_verified":true,"last_login":"2026-04-12T17:02:50.947775Z","is_superuser":false,"username":"asthana40@gmail.com","first_name":"Suramya","last_name":"A","email":"asthana40@gmail.com","is_staff":false,"is_active":true,"date_joined":"2024-01-08T04:17:02.886863Z","is_public":true,"is_removed":false,"is_removed_date":null,"agreed_to_terms":false,"clicked_on_balance_date":"2026-08-03T15:27:56.093064Z","country_code":null,"created_date":"2024-01-08T04:17:03.816260Z","has_seen_first_coin_modal":true,"has_seen_orcid_connect_modal":true,"has_completed_onboarding":true,"is_suspended":false,"moderator":false,"probable_spammer":false,"referral_code":"cdc4b9bb-4cb5-45fa-9b77-d95f32a21c1d","reputation":128,"should_display_rsc_balance_home":true,"is_staking_opted_in":true,"staking_opted_in_date":"2026-04-21T04:15:27.042804Z","spam_updated_date":"2025-06-16T19:25:08.333977Z","suspended_updated_date":null,"updated_date":"2025-06-24T20:55:24.629836Z","upload_tutorial_complete":false,"is_official_account":false,"invited_by":null,"groups":[],"user_permissions":[],"total_contribution":{"rsc":54.494,"rsc_usd_snapshot":5.000044001832,"usd":0},"contributions":[{"amount":54.494,"currency":"RSC","date":"2026-02-20T10:21:01.867276Z"}]},{"id":109841,"author_profile":"$22:props:children:props:topBanner:props:work:note:post:fundraise:contributors:topContributors:1:authorProfile:raw","is_funder":false,"rsc_earned":null,"balances":null,"benefits_expire_on":null,"editor_of":[],"is_verified":true,"last_login":"2026-02-06T08:59:50.411899Z","is_superuser":false,"username":"211110701107@stu.just.edu.cn","first_name":"Z.","last_name":"Yin","email":"211110701107@stu.just.edu.cn","is_staff":false,"is_active":false,"date_joined":"2025-01-11T02:37:15.869006Z","is_public":true,"is_removed":false,"is_removed_date":null,"agreed_to_terms":false,"clicked_on_balance_date":"2026-04-25T02:49:22.815489Z","country_code":null,"created_date":"2025-01-11T02:37:16.409716Z","has_seen_first_coin_modal":true,"has_seen_orcid_connect_modal":false,"has_completed_onboarding":true,"is_suspended":true,"moderator":false,"probable_spammer":true,"referral_code":"2a7de26a-2150-4a41-9ea6-7d87288738f7","reputation":116,"should_display_rsc_balance_home":false,"is_staking_opted_in":true,"staking_opted_in_date":"2026-04-11T03:02:39.681455Z","spam_updated_date":"2026-04-28T21:47:11.519766Z","suspended_updated_date":"2026-04-28T21:47:11.547081Z","updated_date":"2026-04-28T21:47:39.741454Z","upload_tutorial_complete":false,"is_official_account":false,"invited_by":null,"groups":[],"user_permissions":[],"total_contribution":{"rsc":11.475,"rsc_usd_snapshot":0.9999622610325,"usd":0},"contributions":[{"amount":11.475,"currency":"RSC","date":"2026-03-10T11:52:07.109347Z"}]},{"id":184629,"author_profile":"$22:props:children:props:topBanner:props:work:note:post:fundraise:contributors:topContributors:2:authorProfile:raw","is_funder":false,"rsc_earned":null,"balances":null,"benefits_expire_on":null,"editor_of":[],"is_verified":true,"last_login":"2026-07-20T01:39:53.521683Z","is_superuser":false,"username":"just_gqh@163.com","first_name":"Qi-hang","last_name":"Guo","email":"just_gqh@163.com","is_staff":false,"is_active":true,"date_joined":"2025-12-20T01:58:09.039109Z","is_public":true,"is_removed":false,"is_removed_date":null,"agreed_to_terms":false,"clicked_on_balance_date":"2026-07-20T01:39:59.977739Z","country_code":null,"created_date":"2025-12-20T01:58:09.418363Z","has_seen_first_coin_modal":false,"has_seen_orcid_connect_modal":false,"has_completed_onboarding":true,"is_suspended":false,"moderator":false,"probable_spammer":false,"referral_code":"36ce4128-15d3-4c69-932d-1cdeeafdc94b","reputation":100,"should_display_rsc_balance_home":true,"is_staking_opted_in":true,"staking_opted_in_date":"2026-04-16T15:44:08.328272Z","spam_updated_date":"2026-04-28T21:47:28.011901Z","suspended_updated_date":"2026-04-28T21:47:28.049968Z","updated_date":"2026-05-02T02:05:25.602177Z","upload_tutorial_complete":false,"is_official_account":false,"invited_by":null,"groups":[],"user_permissions":[],"total_contribution":{"rsc":11.163,"rsc_usd_snapshot":0.9999666150690001,"usd":0},"contributions":[{"amount":11.163,"currency":"RSC","date":"2026-03-10T21:55:17.466429Z"}]}]},"status":"OPEN","goal_currency":"USD","start_date":"2026-02-15T13:12:26.813630Z","end_date":"2026-03-17T13:12:26.813475Z"}},"grant":"$undefined","doi":"10.55277/researchhub.hvwscfmu","topics":[{"id":594084,"name":"AI","slug":"$undefined","imageUrl":"$undefined","description":"$undefined","namespace":"$undefined","category":"$undefined","discussionCount":"$undefined","paperCount":"$undefined","subscriberCount":"$undefined","isLocked":"$undefined","isRemoved":"$undefined","isUsedForRep":"$undefined","raw":{"id":594084,"name":"AI"}},{"id":665770,"name":"Data Analysis with R","slug":"$undefined","imageUrl":"$undefined","description":"$undefined","namespace":"$undefined","category":"$undefined","discussionCount":"$undefined","paperCount":"$undefined","subscriberCount":"$undefined","isLocked":"$undefined","isRemoved":"$undefined","isUsedForRep":"$undefined","raw":{"id":665770,"name":"Data Analysis with R"}},{"id":669794,"name":"Database","slug":"$undefined","imageUrl":"$undefined","description":"$undefined","namespace":"$undefined","category":"$undefined","discussionCount":"$undefined","paperCount":"$undefined","subscriberCount":"$undefined","isLocked":"$undefined","isRemoved":"$undefined","isUsedForRep":"$undefined","raw":{"id":669794,"name":"Database"}},{"id":670130,"name":"Aging","slug":"$undefined","imageUrl":"$undefined","description":"$undefined","namespace":"$undefined","category":"$undefined","discussionCount":"$undefined","paperCount":"$undefined","subscriberCount":"$undefined","isLocked":"$undefined","isRemoved":"$undefined","isUsedForRep":"$undefined","raw":{"id":670130,"name":"Aging"}}],"authors":[{"authorId":5146850,"userId":74374,"name":"Amruth Bhat","raw":{"id":5146850,"first_name":"Amruth","last_name":"Bhat","user":74374}}],"contacts":"$undefined","image":"https://storage.prod.researchhub.com/uploads/posts/users/74374/2433ddc7-a198-4c63-97a8-36997dcde34e/geoage-graphical-abstractv2.png","raw":{"id":5927,"status":"APPROVED","authors":["$22:props:children:props:topBanner:props:work:note:post:authors:0:raw"],"hubs":["$22:props:children:props:topBanner:props:work:note:post:topics:0:raw","$22:props:children:props:topBanner:props:work:note:post:topics:1:raw","$22:props:children:props:topBanner:props:work:note:post:topics:2:raw","$22:props:children:props:topBanner:props:work:note:post:topics:3:raw"],"unified_document":{"fundraise":"$22:props:children:props:topBanner:props:work:note:post:fundraise:raw","grant":null},"image_url":"https://storage.prod.researchhub.com/uploads/posts/users/74374/2433ddc7-a198-4c63-97a8-36997dcde34e/geoage-graphical-abstractv2.png","document_type":"PREREGISTRATION","slug":"multi-omics-meta-analysis-of-aging-trajectories-using-ai-powered-data-integration","doi":"10.55277/researchhub.hvwscfmu"}},"documentType":null,"proposalId":null,"registeredReportPrefill":null,"image":null,"previewImage":null,"topics":"$undefined","authors":"$undefined","raw":{"id":22704,"access":"WORKSPACE","latest_version":{"id":345724,"src":null,"created_date":"2026-02-16T21:55:56.671255Z","plain_text":"$59","json":"$5a","note":22704},"organization":{"id":72595,"member_count":1,"user_permission":null,"created_date":"2024-10-23T10:45:05.836503Z","updated_date":"2024-10-23T10:45:05.836516Z","cover_image":null,"description":"","name":"Amruth Bhat's Org","note_created":false,"org_type":"PERSONAL","slug":"amruth-bhats-org","user":74374},"post":"$22:props:children:props:topBanner:props:work:note:post:raw","registered_report_prefill":null,"unified_document":{"id":9011421,"is_removed":false},"created_date":"2026-02-15T11:35:34.550386Z","updated_date":"2026-02-16T21:41:09.249004Z","document_type":null,"title":"GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome","created_by":74374},"content":null,"versionId":345724,"versionDate":"2026-02-16T21:55:56.671255Z","plainText":"$5b","contentJson":"$5c"},"previewContent":"$5d","contentUrl":"https://storage.prod.researchhub.com/uploads/post_discussion/2026/02/19/RH-POST-PREREGISTRATION-USER-74374.txt","tips":[],"peerReviews":[{"id":16297,"createdBy":{"id":60659,"authorProfile":{"id":993725,"fullName":"Tahir Ahmed","profileImage":"https://storage.prod.researchhub.com/uploads/author_profile_images/2024/04/26/blob_m68nRIQ","isVerified":false}},"score":4,"createdDate":"2026-03-01T17:08:35.246322Z","contentPreview":"$undefined","isAssessed":false},{"id":16294,"createdBy":{"id":184629,"authorProfile":{"id":8385365,"fullName":"Qi-hang Guo","profileImage":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/12/20/blob_261d2PH","isVerified":false}},"score":4,"createdDate":"2026-03-01T11:38:39.192184Z","contentPreview":"$undefined","isAssessed":true}],"image":"https://storage.prod.researchhub.com/uploads/posts/users/74374/2433ddc7-a198-4c63-97a8-36997dcde34e/geoage-graphical-abstractv2.png","enrichments":[],"moderationStatus":"APPROVED","raw":{"score":0,"user_flag":null,"authors":["$22:props:children:props:topBanner:props:work:authors:0:authorProfile:raw"],"id":5927,"created_by":{"id":74374,"author_profile":{"id":5146850,"is_public":true,"is_removed":false,"is_removed_date":null,"user":74374,"first_name":"Amruth","last_name":"Bhat","created_date":"2024-10-23T10:45:05.775749Z","updated_date":"2026-02-19T06:14:54.746437Z","description":"PhD researcher mapping biological aging across India. Immune aging, multiomics, systems biology, ML & mathematical modeling to decode how and why Indian populations age differently. Building tools for mechanistic insight at scale.","h_index":1,"i10_index":0,"profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/06/01/blob_QMU7UYj","author_score":0,"university":null,"orcid_id":"https://orcid.org/0000-0002-4075-1939","openalex_ids":["https://openalex.org/A5082829184"],"education":[{"id":2876,"name":"Indian Institute of Science","year":{"label":"2024","value":"2024"},"major":"Bioengineering","degree":{"label":"Master's Degree","value":"Master's Degree"},"summary":"Bioengineering Master's Degree '24, Indian Institute of Science","is_public":false,"university":{"id":2876,"name":"Indian Institute of Science"}},{"id":9316,"name":"Visvesvaraya Technological University","year":{"label":"2022","value":"2022"},"major":"Biotechnology","degree":{"label":"Bachelor of Engineering","value":"BE"},"summary":"Biotechnology BE '22, Visvesvaraya Technological University","is_public":false,"university":{"id":9316,"city":"","name":"Visvesvaraya Technological University","state":"","country":"India","updated_date":"2020-11-24T21:44:27.631606Z"}},{"id":2876,"name":"Indian Institute of Science","year":{"label":"2030","value":"2030"},"major":"Bioengineering","degree":{"label":"Doctor of Philosophy","value":"PhD"},"summary":"Bioengineering PhD '30, Indian Institute of Science","is_public":true,"university":{"id":2876,"name":"Indian Institute of Science"}}],"headline":"PhD Student at Indian Institute of Science","facebook":null,"twitter":"https://x.com/__adb__7","linkedin":"https://www.linkedin.com/in/amruth-bhat/","google_scholar":"https://scholar.google.com/citations?user=kpg2O-QAAAAJ","claimed":true,"country_code":null,"created_source":"RESEARCHHUB","last_full_fetch_from_openalex":null,"two_year_mean_citedness":0,"merged_with_author":null,"added_as_editor_date":null,"is_hub_editor_of":[],"is_hub_editor":false,"num_posts":1,"is_orcid_connected":true,"reputation":102,"reputation_v2":{"hub":{"id":231,"name":"Molecular Biology","slug":"molecular-biology-1"},"score":4,"percentile":0.001,"bins":[[0,1000],[1000,10000],[10000,100000],[100000,1000000]]},"reputation_list":[{"hub":{"id":231,"name":"Molecular Biology","slug":"molecular-biology-1"},"score":4,"percentile":0.001,"bins":[[0,1000],[1000,10000],[10000,100000],[100000,1000000]]}],"suspended_status":{"probable_spammer":false,"is_suspended":false},"total_score":null,"wallet":null,"is_verified":true},"balance":null,"balances":null,"created_date":"2024-10-23T10:45:05.745449Z","has_seen_first_coin_modal":true,"has_seen_orcid_connect_modal":false,"is_funder":false,"is_staking_opted_in":true,"is_suspended":false,"probable_spammer":false,"is_verified":true,"moderator":false,"reputation":102,"updated_date":"2025-04-01T18:19:22.426533Z","upload_tutorial_complete":false,"hub_rep":null,"time_rep":null},"created_date":"2026-02-15T13:12:26.759468Z","discussion_count":0,"document_type":"PREREGISTRATION","doi":"10.55277/researchhub.hvwscfmu","editor_type":"CK_EDITOR","full_markdown":"$5e","grants":[{"id":18,"short_title":"Aging and Longevity","status":"CLOSED","organization":"ResearchHub Foundation","amount":"10000.00","currency":"USD","post_id":5056,"image_url":"https://storage.prod.researchhub.com/uploads/posts/users/39602/50043f15-5552-40ef-8ed8-4c81c50c0468/chatgpt-image-jan-16-2026-05_10_49-pm.png","title":"Request for Proposals - Aging and Longevity","applicant_count":16,"application_visibility":"OPTIONAL","proposal":{"unified_document_id":9011527,"ai_peer_review":null}}],"hubs":[{"editor_permission_groups":[],"hub_image":null,"id":594084,"is_locked":false,"is_removed":false,"name":"AI","slug":"ai","namespace":"journal","is_used_for_rep":false},{"editor_permission_groups":[],"hub_image":null,"id":665770,"is_locked":false,"is_removed":false,"name":"Data Analysis with R","slug":"data-analysis-with-r","namespace":null,"is_used_for_rep":false},{"editor_permission_groups":[],"hub_image":null,"id":669794,"is_locked":false,"is_removed":false,"name":"Database","slug":"database-2","namespace":"journal","is_used_for_rep":false},{"editor_permission_groups":[],"hub_image":null,"id":670130,"is_locked":false,"is_removed":false,"name":"Aging","slug":"aging-2","namespace":"journal","is_used_for_rep":false}],"image_url":"https://storage.prod.researchhub.com/uploads/posts/users/74374/2433ddc7-a198-4c63-97a8-36997dcde34e/geoage-graphical-abstractv2.png","is_latest_version":null,"is_removed":false,"is_root_version":true,"note":"$22:props:children:props:topBanner:props:work:note:raw","peer_reviews":[{"id":16297,"created_by":{"id":60659,"author_profile":{"id":993725,"first_name":"Tahir","last_name":"Ahmed","profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2024/04/26/blob_m68nRIQ"},"first_name":"Tahir","last_name":"Ahmed"},"created_date":"2026-03-01T17:08:35.246322Z","updated_date":"2026-03-01T17:08:35.246334Z","score":4,"is_assessed":false},{"id":16294,"created_by":{"id":184629,"author_profile":{"id":8385365,"first_name":"Qi-hang","last_name":"Guo","profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/12/20/blob_261d2PH"},"first_name":"Qi-hang","last_name":"Guo"},"created_date":"2026-03-01T11:38:39.192184Z","updated_date":"2026-03-01T11:38:39.192195Z","score":4,"is_assessed":true}],"post_src":"https://storage.prod.researchhub.com/uploads/post_discussion/2026/02/19/RH-POST-PREREGISTRATION-USER-74374.txt","preview_img":null,"registered_report_id":null,"renderable_text":"$5f","reviewed_by":null,"reviewed_date":null,"slug":"multi-omics-meta-analysis-of-aging-trajectories-using-ai-powered-data-integration","status":"APPROVED","title":"GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome","unified_document_id":9011527,"unified_document":{"id":9011527,"documents":[{"id":5927,"title":"GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome","slug":"multi-omics-meta-analysis-of-aging-trajectories-using-ai-powered-data-integration"}],"created_by":{"id":74374,"author_profile":{"id":5146850,"first_name":"Amruth","last_name":"Bhat","profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/06/01/blob_QMU7UYj"}},"reviews":{"avg":4,"count":1},"is_public":true,"is_removed":false,"document_type":"PREREGISTRATION"},"version_number":1,"updated_date":"2026-02-22T19:41:32.725940Z","bounty_type":null},"aiPeerReview":null,"linkedGrant":{"id":18,"postId":5056,"title":"Request for Proposals - Aging and Longevity","shortTitle":"Aging and Longevity","organization":"ResearchHub Foundation","fundingAmount":10000,"currency":"USD","status":"CLOSED","applicationVisibility":"OPTIONAL","imageUrl":"https://storage.prod.researchhub.com/uploads/posts/users/39602/50043f15-5552-40ef-8ed8-4c81c50c0468/chatgpt-image-jan-16-2026-05_10_49-pm.png","applicantCount":16},"isPublic":true},"metadata":{"id":9011527,"score":0,"topics":[{"id":665770,"name":"Data Analysis with R","slug":"data-analysis-with-r","imageUrl":"$undefined","description":"$undefined","namespace":null,"category":"$undefined","discussionCount":"$undefined","paperCount":"$undefined","subscriberCount":"$undefined","isLocked":"$undefined","isRemoved":"$undefined","isUsedForRep":false,"raw":{"id":665770,"name":"Data Analysis with R","namespace":null,"slug":"data-analysis-with-r","created_date":"2025-10-21T04:01:05.903235Z","is_used_for_rep":false}}],"metrics":{"votes":0,"comments":0,"saves":0,"reviewScore":4,"conversationComments":0,"reviewComments":2,"bountyComments":1},"fundraising":{"id":288,"amountRaised":{"usd":4.986445502324,"rsc":77.132},"goalAmount":{"usd":10000,"rsc":154683.33097002984},"contributors":{"numContributors":3,"topContributors":[{"id":50877,"authorProfile":{"id":984288,"fullName":"Suramya Asthana","firstName":"Suramya","lastName":"Asthana","profileImage":"https://storage.prod.researchhub.com/uploads/author_profile_images/2024/11/17/blob_lMGDGID","headline":"","profileUrl":"/author/984288","user":"$undefined","description":"$undefined","createdDate":"$undefined","education":"$undefined","twitter":"$undefined","facebook":"$undefined","linkedin":"$undefined","googleScholar":"$undefined","orcidId":"$undefined","isOrcidConnected":false,"isClaimed":false,"hIndex":"$undefined","i10Index":"$undefined","userId":"$undefined","isVerified":false,"isHubEditor":false,"editorOfHubs":[],"raw":{"id":984288,"first_name":"Suramya","last_name":"Asthana","profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2024/11/17/blob_lMGDGID"}},"totalContribution":{"rsc":54.494,"usd":0},"contributions":[{"amount":54.494,"currency":"RSC","date":"2026-02-20T10:21:01.867276Z"}]},{"id":109841,"authorProfile":{"id":7136539,"fullName":"Zhen-Yu YIN","firstName":"Zhen-Yu","lastName":"YIN","profileImage":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/01/11/blob_WZNkWbV","headline":"","profileUrl":"/author/7136539","user":"$undefined","description":"$undefined","createdDate":"$undefined","education":"$undefined","twitter":"$undefined","facebook":"$undefined","linkedin":"$undefined","googleScholar":"$undefined","orcidId":"$undefined","isOrcidConnected":false,"isClaimed":false,"hIndex":"$undefined","i10Index":"$undefined","userId":"$undefined","isVerified":false,"isHubEditor":false,"editorOfHubs":[],"raw":{"id":7136539,"first_name":"Zhen-Yu","last_name":"YIN","profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/01/11/blob_WZNkWbV"}},"totalContribution":{"rsc":11.475,"usd":0},"contributions":[{"amount":11.475,"currency":"RSC","date":"2026-03-10T11:52:07.109347Z"}]},{"id":184629,"authorProfile":{"id":8385365,"fullName":"Qi-hang Guo","firstName":"Qi-hang","lastName":"Guo","profileImage":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/12/20/blob_261d2PH","headline":"","profileUrl":"/author/8385365","user":"$undefined","description":"$undefined","createdDate":"$undefined","education":"$undefined","twitter":"$undefined","facebook":"$undefined","linkedin":"$undefined","googleScholar":"$undefined","orcidId":"$undefined","isOrcidConnected":false,"isClaimed":false,"hIndex":"$undefined","i10Index":"$undefined","userId":"$undefined","isVerified":false,"isHubEditor":false,"editorOfHubs":[],"raw":{"id":8385365,"first_name":"Qi-hang","last_name":"Guo","profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/12/20/blob_261d2PH"}},"totalContribution":{"rsc":11.163,"usd":0},"contributions":[{"amount":11.163,"currency":"RSC","date":"2026-03-10T21:55:17.466429Z"}]}]},"createdBy":{"id":0,"email":"","firstName":"","lastName":"","fullName":"Unknown User","isVerified":false,"authorProfile":"$undefined","balance":0,"lockedBalance":0,"totalRsc":0,"hasCompletedOnboarding":false,"moderator":false,"raw":null,"isModerator":false,"isFunder":false,"authProvider":"$undefined","isStakingOptedIn":false,"balanceHistory":0},"status":"OPEN","goalCurrency":"USD","startDate":"2026-02-15T13:12:26.813630Z","endDate":"2026-03-17T13:12:26.813475Z","createdDate":"","updatedDate":"","postId":"$undefined","postTitle":"$undefined","postSlug":"$undefined","postImage":null,"reviewMetrics":"$undefined","nonprofit":"$undefined","raw":{"id":288,"amount_raised":{"usd":4.986445502324,"rsc":77.132},"goal_amount":{"usd":10000,"rsc":154683.33097002984},"contributors":{"total":3,"top":[{"id":50877,"author_profile":"$22:props:children:props:topBanner:props:metadata:fundraising:contributors:topContributors:0:authorProfile:raw","first_name":"Suramya","last_name":"A","total_contribution":{"rsc":54.494,"rsc_usd_snapshot":5.000044001832,"usd":0},"contributions":[{"amount":54.494,"currency":"RSC","date":"2026-02-20T10:21:01.867276Z"}]},{"id":109841,"author_profile":"$22:props:children:props:topBanner:props:metadata:fundraising:contributors:topContributors:1:authorProfile:raw","first_name":"Z.","last_name":"Yin","total_contribution":{"rsc":11.475,"rsc_usd_snapshot":0.9999622610325,"usd":0},"contributions":[{"amount":11.475,"currency":"RSC","date":"2026-03-10T11:52:07.109347Z"}]},{"id":184629,"author_profile":"$22:props:children:props:topBanner:props:metadata:fundraising:contributors:topContributors:2:authorProfile:raw","first_name":"Qi-hang","last_name":"Guo","total_contribution":{"rsc":11.163,"rsc_usd_snapshot":0.9999666150690001,"usd":0},"contributions":[{"amount":11.163,"currency":"RSC","date":"2026-03-10T21:55:17.466429Z"}]}]},"status":"OPEN","goal_currency":"USD","start_date":"2026-02-15T13:12:26.813630Z","end_date":"2026-03-17T13:12:26.813475Z"}},"bounties":[{"id":6980,"amount":"1525.0000000000","status":"CLOSED","expirationDate":"2026-08-10T23:36:50.807Z","bountyType":"BOUNTY","createdBy":{"id":39602,"email":"","firstName":"","lastName":"","fullName":"Unknown User","isVerified":false,"authorProfile":{"id":973869,"fullName":"ResearchHub Foundation","firstName":"ResearchHub","lastName":"Foundation","profileImage":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/05/28/blob_aBYdPow","headline":"","profileUrl":"/author/973869","user":"$undefined","description":"$undefined","createdDate":"$undefined","education":"$undefined","twitter":"$undefined","facebook":"$undefined","linkedin":"$undefined","googleScholar":"$undefined","orcidId":"$undefined","isOrcidConnected":false,"isClaimed":false,"hIndex":"$undefined","i10Index":"$undefined","userId":"$undefined","isVerified":false,"isHubEditor":false,"editorOfHubs":[],"raw":{"id":973869,"first_name":"ResearchHub","last_name":"Foundation","profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/05/28/blob_aBYdPow"}},"balance":0,"lockedBalance":0,"promotionalBalance":0,"fundingCredits":0,"balances":"$undefined","totalRsc":0,"hasCompletedOnboarding":false,"createdDate":"$undefined","moderator":false,"editorOfHubs":"$undefined","isModerator":false,"isFunder":false,"referralCode":"$undefined","authProvider":"$undefined","isStakingOptedIn":false,"balanceHistory":0,"raw":{"id":39602,"author_profile":"$22:props:children:props:topBanner:props:metadata:bounties:0:createdBy:authorProfile:raw"}},"solutions":[],"contributions":[],"totalAmount":"1525","comment":"$undefined","raw":{"id":6980,"created_by":"$22:props:children:props:topBanner:props:metadata:bounties:0:createdBy:raw","amount":"1525.0000000000","status":"CLOSED"}}],"activeBounties":0,"closedBounties":1},"contentType":"post"}],"rightSidebar":"$L60","children":"$L61"}]}]

GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome

GeoAge: LLM-Powered Metadata Curation and Mega Meta-Analysis of the Human Aging Transcriptome

Peer Reviews

Topics

DOI