Automatic Knowledge Base for Immunological Aging: Grant Proposal

1. Overview

Scientific rationale and importance: Aging biology is a complex, multi‑factorial process driven by genetic, cellular and environmental factors. The Gene Ontology (GO) provides a structured vocabulary describing molecular functions, cellular components and biological processes. It contains over 47,000 terms and more than one million experimentally supported annotations [1]. However, GO is not specific to aging, and manually curated ontologies cannot keep pace with the accelerating literature. We propose to develop an AI agent system that automatically builds an aging biology knowledge base akin to GO. The system will integrate multiple biomedical databases such as GO, GenAge, DrugBank, Cell Ontology and the HALD aging knowledge graph. We will leverage a large language model (LLM) agent to mine PubMed literature for relationships between genes, cytokines, immune cells and processes involved in immunosenescence and inflammaging. Recent work shows that LLMs can accurately annotate cell types (>80–90% accuracy) and gene sets (>80% agreement with expert annotations) [2], suggesting they can contribute to ontology curation. Automating aging ontology construction will accelerate discovery of new relationships, enable cross‑species comparisons, and provide a living resource that updates as new knowledge emerges. Equally important, the agentic AI approach that we develop in this project can be generalized to other domains (e.g. cancer, developmental biology, dementia, etc.) to build more informative knowledge graphs.

Facilities, supervision and approvals: The project will be conducted in the laboratory of Prof. Eric Sun at the MIT Department of Biological Engineering and the Ragon Institute (Massachusetts General Hospital). Ethics approvals for data handling and cell experiments will be obtained.

Estimated timeline:

Month 0-3: implement data connectors for GO, GenAge, DrugBank, Cell Ontology and PubMed; build a skeleton graph of aging‑related entities.

Month 3–6: develop the LLM‑based extraction pipeline with retrieval‑augmented generation (RAG) and extract candidate relations. We will benchmark against curated knowledge and refine prompts.

Month 6–9: evaluate and validate the knowledge base using knowledge‑graph perturbation experiments and cross‑validation; perform wet‑lab experiments to test selected novel predictions.

Month 9–12: prepare manuscripts and release data and code.

Total project duration: 12 months.

Data sharing commitment: All code, extracted knowledge and derived ontologies will be shared openly. Wet‑lab data will be shared through appropriate repositories. Results will be published open access.

2. Introduction

Background and theoretical rationale: Immune aging is characterized by chronic, low‑grade inflammation and decline of adaptive and innate immune functions (immunosenescence). Inflammaging results from persistent activation of innate immune pathways and is marked by elevated cytokines such as interleukin‑6 (IL‑6), tumour necrosis factor‑α (TNF‑α) and IL‑1β[3]. These cytokines drive age‑related diseases—high IL‑6, TNF and IL‑1β levels correlate with cardiovascular, neurodegenerative and metabolic disorders[4]. Adaptive immunity declines; naive T cells decrease and memory T cells accumulate[5], B‑cell subsets shift toward pro‑inflammatory double‑negative populations[6], and macrophages exhibit reduced antigen presentation but increased baseline secretion of inflammatory mediators[7]. These changes are influenced by genetic factors and environmental triggers such as mitochondrial dysfunction, cellular senescence and dysbiosis[8].

Status quo and limitations: Current knowledge about aging biology is dispersed across curated databases and the primary literature. The Gene Ontology (GO) offers a hierarchical vocabulary (molecular function, biological process, cellular component) but is not tailored for aging and requires manual curation [1]. GenAge provides a curated list of human aging genes and candidate genes from model organisms[9], but lacks detailed relationships among genes, cells and processes. DrugBank contains ~4,900 drug entries with detailed pharmacological and molecular information[10]; linking this to aging processes is laborious. The Cell Ontology contains >2,700 cell‑type classes[11], and the HALD knowledge graph mined 339,918 PubMed articles to create 12,227 entities and 115,522 relations related to aging and longevity[12], but HALD is static and cannot update automatically. Furthermore, existing AI‑driven biomedical agents have been benchmarked primarily on question‑answering tasks; new benchmarks such as BioKGBench highlight that evaluating agents’ ability to verify knowledge graph facts remains largely unexplored[13]. BioKGBench decomposes the evaluation into knowledge‑graph question answering and scientific claim verification, finding that state‑of‑the‑art agents fail without adaptation[14]. This underscores the need for dynamic, evidence‑grounded knowledge graphs and robust evaluation.

Research questions and hypotheses: We ask whether a large‑language‑model agent, equipped with retrieval from curated databases and literature, can automatically construct a high‑quality ontology for immunological aging. H1: The AI‑driven system will recover known relationships among aging genes, cytokines, immune cells and processes with precision and recall comparable to curated resources. H2: The system will discover novel associations and clusters that are supported by literature and can be validated experimentally. H3: Perturbation tests (edge shuffling, node label scrambling) will reveal that the knowledge graph captures meaningful biology beyond random chance. H4: Wet‑lab experiments on selected predictions (e.g., drug–cytokine interactions) will confirm the biological relevance of the inferred relationships.

Impact statement: If successful, this project will deliver the first automated, extensible ontology for aging biology focused on immunology. It will enable researchers to query relationships between genes, cell types, processes and interventions, support enrichment analyses, and accelerate hypothesis generation. Even if the AI system identifies few novel associations, the curated knowledge base will unify disparate data sources and reduce manual curation efforts, benefiting the aging research community.

3. Methods

Participants/Sample

This study does not involve human or animal participants. Instead, the “sample” consists of data sources and biological entities. Inclusion criteria: (i) human genes and proteins listed in GenAge[9] and associated GO terms; (ii) cytokines and immune cell types implicated in inflammaging and immunosenescence (e.g., IL‑6, TNF‑α, IL‑1β, CD4⁺ T cell, macrophage); (iii) drugs relevant to aging and immune modulation obtained from DrugBank[10]; (iv) cell types from the Cell Ontology[11]; (v) literature abstracts retrieved from PubMed for queries such as “aging AND cytokine”, “immunosenescence”, and each entity. Exclusion criteria: non‑human data unless conserved across species, non‑immune pathways unrelated to aging, and retracted publications.

Materials and Procedures

Data acquisition. Using Python, we will write connectors to download and parse: (1) GO basic OBO file via GOATOOLS to extract relevant terms and relationships (is‑a, part‑of); (2) GenAge gene list as a TSV file; (3) DrugBank XML to retrieve drug names, targets and mechanisms; (4) Cell Ontology OBO to obtain immune cell definitions and hierarchies; (5) PubMed abstracts via NCBI E‑Utilities (ESearch and EFetch). A local cache will store retrieved abstracts.

Knowledge representation. A graph data model will represent entities (genes, proteins, cytokines, cell types, biological processes, drugs) as nodes with attributes (ID, name, definition, source) and relations (e.g., has_function, participates_in, expressed_in, targets, involves_cell, is_a, part_of). We will implement the graph using NetworkX and provide export to RDF/OWL for interoperability. GO terms will seed the structure—e.g., “cytokine activity” (GO:0005125) is a molecular function term that will link to cytokine genes and drugs.

LLM‑assisted extraction. We will use a state‑of‑the‑art LLM (GPT‑4 or a domain‑specific model like BioGPT) with retrieval‑augmented generation. For each entity pair, we will assemble a prompt containing definitions from GO or GenAge and top‑k PubMed sentences (obtained via embedding‑based similarity search) and ask the LLM to extract triples in the form ⟨subject, relation, object⟩. Example prompt: “From the provided abstract snippets, identify relations involving IL‑6, macrophages and aging. Represent them as subject‑predicate‑object triples.” Extracted statements will be mapped to predefined relation types. To reduce hallucinations, we will ask the LLM to cite supporting sentences and will automatically verify that extracted relations are present in the source text (string matching). Only validated facts will be added to the graph.

Graph integration and reasoning: We will merge curated relations from GO, GenAge, DrugBank and the Cell Ontology with LLM‑extracted relations. A reasoning module will infer higher‑level links—for example, combining drug→target and target→process to infer drug→process interactions. The system will also support complex queries (e.g., “Which drugs modulate IL‑6 mediated inflammaging?”) by traversing the graph.

Planned Analyses

We will evaluate our system using multiple approaches:

Comparison to curated resources: A held‑out set of relationships from GO and GenAge will be used as ground truth. For each relation type, we will compute precision, recall and F1‑score of the AI‑extracted relations. In line with biomedical NLP benchmarks, we will use metrics such as exact match accuracy and F1. Domain‑tuned models like BioBERT have been shown to outperform general models on biomedical tasks; we will evaluate whether retrieval‑augmented GPT‑4 matches or exceeds their performance.

Knowledge graph perturbation: To assess whether the graph encodes meaningful structure, we will perform edge perturbation and node‑label shuffling. We will randomly reassign relation labels or shuffle edges and compute metrics such as link prediction accuracy or embedding similarity. A significant drop in performance compared to the original graph will indicate that the extracted relations capture real biological patterns (hypothesis H3).

Query‑answer evaluation: Inspired by BioKGBench, we will formulate knowledge‑graph question answering tasks (“Which cytokines produced by macrophages increase with age?”). The system’s answers will be scored against curated answers; we will measure accuracy and evaluate retrieval of supporting evidence. We will also perform scientific claim verification tasks, asking the system to decide whether a proposed relation is supported by literature[14].

Case studies and wet‑lab validation: We will identify novel predicted relations that are biologically plausible but not well documented. For example, if the system suggests that drug metformin modulates IL‑1β secretion in macrophages through AMPK, we will test this in vitro using human monocyte‑derived macrophages at the Ragon Institute. Cytokine levels (IL‑1β, IL‑6, TNF‑α) will be measured by ELISA after drug treatment. Positive results would support hypothesis H4.

Contingent analysis decisions: If the LLM shows high hallucination rates (>10 % false positives), we will tighten the extraction criteria (e.g., require multiple supporting sentences) or switch to a smaller, domain‑finetuned model. If PubMed retrieval yields insufficient context, we will expand to full‑text articles or additional databases (e.g., HALD). For wet‑lab experiments, if initial results do not confirm predictions, we will refine the selection criteria for candidate relations.

Ethics and Data Management

The project does not involve human subjects; however, wet‑lab experiments will use commercially available human cell lines and will follow institutional biosafety and ethical guidelines. We will also obtain IRB approval or exemption for use of human cell lines at MGH. . Genomic and chemical data are from public databases; no identifiable personal data are used.

Confidentiality and data sharing: We will store all data on secure MIT servers with access controls. The knowledge graph will be released under a Creative Commons license. Scripts and pipelines will be documented and deposited in a public repository. We will adhere to FAIR data principles.

5. Budget

The requested budget totals US$5,000 and covers:

$4000 for 1 full time graduate researcher

$500 for experimental resources

$500 for computational resources

Additional institutional support (computing infrastructure, office space) will be provided by MIT and the Ragon Institute.

6. References

1. Gene Ontology Consortium. The Gene Ontology knowledgebase in 2026. Nucleic Acids Res. 54, D1779–D1792 (2026). doi:10.1093/nar/gkaf1292

2. Crowley, G. et al. Benchmarking cell type and gene set annotation by large language models with AnnDictionary. Nat. Commun. 16, 9511 (2025). doi:10.1038/s41467-025-64511-x

3. Karpuzoglu, E., Holladay, S. D. & Gogal, R. M. F. Jr. Inflammaging: triggers, molecular mechanisms, immunological consequences, sex differences, and cutaneous manifestations. Front. Immunol. 16, 1704203 (2025). doi:10.3389/fimmu.2025.1704203

4. Tylutka, A., Walas, Ł. & Zembroń-Lacny, A. Level of IL-6, TNF, and IL-1β and age-related diseases: a systematic review and meta-analysis. Front. Immunol. 15, 1330386 (2024). doi:10.3389/fimmu.2024.1330386

5. Bektas, A., Schurman, S. H., Sen, R. & Ferrucci, L. Human T cell immunosenescence and inflammation in aging. J. Leukoc. Biol. 102, 977–988 (2017). doi:10.1189/jlb.3RI0716-335R

6. Frasca, D., Diaz, A., Romero, M., Garcia, D. & Blomberg, B. B. B Cell Immunosenescence. Annu. Rev. Cell Dev. Biol. 36, 551–574 (2020). doi:10.1146/annurev-cellbio-011620-034148

7. Oishi, Y. & Manabe, I. Macrophages in age-related chronic inflammatory diseases. npj Aging Mech. Dis. 2, 16018 (2016). doi:10.1038/npjamd.2016.18

8. de Magalhães, J. P. et al. Human Ageing Genomic Resources: updates on key databases in ageing research. Nucleic Acids Res. 52, D900–D908 (2024). doi:10.1093/nar/gkad927

9. Wishart, D. S. et al. DrugBank: a knowledgebase for drugs, drug actions and drug targets. Nucleic Acids Res. 36, D901–D906 (2008). doi:10.1093/nar/gkm958

10. Mungall, C. et al. Uberon Ontology (release 2016-01-26). http://purl.obolibrary.org/obo/uberon/releases/2016-01-26/uberon.owl (2016).

11. Wu, Z. et al. HALD, a human aging and longevity knowledge graph for precision gerontology and geroscience analyses. Sci. Data 10, 851 (2023). doi:10.1038/s41597-023-02781-0

12. Lin, X., Ma, S., Shan, J., Zhang, X., Hu, S. X., Guo, T., Li, S. Z. & Yu, K. BioKGBench: A Knowledge Graph Checking Benchmark of AI Agent for Biomedical Science. arXiv:2407.00466 (2024). doi:10.48550/arXiv.2407.00466

13. Laurent, A. LLM Evaluation for Biotech: A Methodological Guide. IntuitionLabs https://intuitionlabs.ai/articles/llm-evaluation-biotech-guide (2026) (accessed 8 Feb 2026).

14. Zheng, S. et al. PharmKG: a dedicated knowledge graph benchmark for biomedical data mining. Brief. Bioinform. 22, bbaa344 (2021). doi:10.1093/bib/bbaa344

Automatic Knowledge Base for Immunological Aging: Grant Proposal

Automatic Knowledge Base for Immunological Aging: Grant Proposal

Funding Progress

Funders

Peer Reviews

Related Funding

Request for Proposals - AI Tools & AI Agents for Biomedical Research

Nonprofit Recipient

Topics

DOI