scPerturbClock: Genome-scale Causal Discovery of Aging Regulators with Machine Learning

Machine Learning Framework for Translating Correlational Aging Signatures into Druggable Interventions

Andrew Ding

Massachusetts Institute of Technology, Department of Biological Engineering

The Ragon Institute of Mass General Brigham, MIT, and Harvard

The Broad Institute of MIT and Harvard

1. Overview

Scientific rationale and importance: Molecular aging clocks provide a sensitive readout of biological aging and have helped validate specific interventions, yet translating molecular signatures into actionable targets remains a central challenge [1]. While multi-omics studies have produced extensive catalogs of age-associated genes and pathways, these datasets are largely observational, limiting causal interpretation and target prioritization.

This project proposes a scalable framework that couples Perturb-seq, which enables systematic genetic interventions, with cell-type-specific single-cell RNA-sequencing (scRNA-seq) aging clocks, which quantify shifts in aging-associated transcriptomic programs [2,3]. This integration converts correlational aging signatures into causal rankings and mechanistic hypotheses, creating a practical route from discovery to druggable intervention strategies.

Funding/facilities/ethics approvals: High performance computing infrastructure is available through MIT Office of Research Computing and Data via the Engaging cluster. We have access to state-of-the-art experimental facilities, lab space, and equipment at the Ragon Institute where the proposed research will be conducted. No additional equipment or ethics approvals are required for this phase.

Estimated timeline: 6 months. Months 1-2: dataset selection and harmonization; aging score computation; quality control and confounder sensitivity checks. Months 3-4: perturbation ranking by rejuvenation effect; mechanistic mapping of top ranked perturbations to transcription factor (TF) modules and pathways. Months 5-6: translate mechanisms into candidate ligand/drug hypotheses; in silico validation and, if feasible, one minimal in vitro test; write and share results.

Data sharing commitment: All code, models, processed results, data will be released in an open-access repository with clear documentation. Findings will be available as a preprint. Results will be shared regardless of outcome.

2. Introduction

Background

Single-cell transcriptomic aging clocks developed on scRNA-seq data provide a quantitative method of summarizing complex, often cell-type-specific transcriptional changes into a single readout of “aging state”. Recent work has shown that cell-type-specific clocks can measure aging and report the effects of selected interventions in vivo with remarkable accuracy [3,4]. However, validating interventions in vivo remains slow and resource-intensive, which limits the aging field’s ability to systematically screen the large space of candidate regulators and especially combinations of perturbations. As a result, scRNA-seq clocks are poised to be a high-throughput, integrative readout for intervention effects, but the interventions themselves have not yet been evaluated at scale.

Perturb-seq provides a complementary interventional framework in which genetic perturbations can be applied at scale and profiled using single-cell transcriptomes. This enables direct tests of whether disrupting a specific gene shifts cellular programs toward a rejuvenated or pro-aged transcriptional phenotype [2]. Combining Perturb-seq with cell-type-specific aging clocks creates an opportunity to move beyond correlation by quantifying the causal impact of perturbations on aging-associated cell states, thereby generating a prioritized list of rejuvenating regulators and a foundation for more efficient follow-up experimentation.

Innovation

This project introduces a unified framework that transforms cell-type-specific aging clocks from descriptive biomarkers into a scalable causal discovery readout for perturbation biology.

The central innovation is to quantify the causal impact of genetic perturbations on aging-associated cellular programs by projecting Perturb-seq profiles onto a standardized, cell-type-specific aging axis. This produces a ranked set of candidate rejuvenating regulators with uncertainty and reproducibility criteria, enabling systematic prioritization that is difficult to achieve with observational ‘omics’ alone.

A second innovation is a translation layer that converts genetic hits into intervention-relevant mechanisms. Rather than stopping at a gene list, we will map rejuvenating perturbations onto mediating transcription factor modules and pathway-level explanations, yielding mechanistic hypotheses that are easier to interpret across contexts and connect to druggable biology (receptors, ligands, small-molecule modulators, etc.) [5–8].

Together, these innovations deliver near-term outputs (ranked aging perturbations, mechanistic modules, prioritized intervention hypotheses) and a reusable framework for causal, perturbation-driven aging discovery across contexts, enabling researchers to iteratively measure, model, and manipulate aging.

Research Questions and Hypotheses

H1: Cell-type-specific scRNA-seq aging clocks can measure reproducible directional shifts in transcriptomic aging signatures from Perturb-seq genetic perturbations.

H2: Genetic perturbations which cause directional changes in transcriptomic aging signatures map to specific, druggable, upstream mediating TF modules and pathways.

H3: Pharmacological targeting of a rejuvenating and pro-aging pathway nominated from genetic perturbations will show expected directional shifts in transcriptome projected on cell-type-specific aging axis when profiled with RNA-seq.

Impact Statement

Even if no strongly rejuvenating perturbations are found in a given dataset, the project will produce a reusable framework for causal evaluation of perturbations using aging readouts, along with a set of negative results that clarify which perturbations do not shift aging state in the tested context.

3. Methods

Participants/Sample

Target sample: This study will use publicly available single-cell Perturb-seq datasets as the primary data source, paired with external scRNA-seq aging reference datasets used to train or calibrate cell-type-specific aging clocks. The unit of analysis is the single cell, grouped by perturbation condition (gene perturbation vs matched controls) within each Perturb-seq experiment.

Inclusion/exclusion criteria: This study will prioritize Perturb-seq datasets that 1) include non-targeting controls, 2) provide clear perturbation annotations, 3) contain sufficient cells per perturbation condition to estimate effects with uncertainty, and 4) provide metadata to support basic quality control and confounder adjustment (such as batch/replicate identifiers).

Sample size justification and power analysis: Sample size will be determined by the size of selected Perturb-seq datasets. We will justify adequacy by reporting per-perturbation cell counts after QC, estimating uncertainty for each perturbation effect, and using resampling (bootstrap or permutation) to confirm that top-ranked effects are stable.

Materials and Procedures

Data sources & selection criteria: Public Perturb-seq datasets will be selected using predefined inclusion criteria, prioritizing screens that match the cellular context of available cell-type-specific scRNA-seq aging clocks.

Aging score computation: Per-cell aging scores will be computed by applying a cell-type-specific aging clock to Perturb-seq expression profiles, enabling estimation of within-screen perturbation effects as directional shifts in aging state relative to matched controls. Shift aging score will be computed as ΔAgingScore = AgingScore_perturbed-AgingScore_control.

Mechanistic mapping: Top rejuvenating and pro-aging perturbations will then be translated into mechanistic hypotheses by mapping their transcriptional signatures to mediating TF programs and upstream pathways using known regulons and prior knowledge networks [5,6,9].

In vitro validation: One druggable node each from a pathway upstream of a top ranked rejuvenating and pro-aging perturbation will be selected for minimal in vitro validation in a matching cell line using the same scRNA-seq readout and aging-score framework.

Planned Analyses

Primary outcome: The perturbation effect on aging state, quantified as a within-screen shift in aging score relative to matched non-targeting controls (ΔAgingScore) for each perturbation. Perturbations will be ranked by effect size, uncertainty, and reproducibility to prioritize robust rejuvenating and pro-aging shifts.

Perturbation shift aging state: ΔAgingScore will be estimated within each dataset using regression-based comparisons of perturbation versus control, with uncertainty from model-based standard errors or resampling. Multiple testing will be controlled using false discovery rate correction and prioritization will emphasize stability across replicates. Experimental followup will be done to validate the top rejuvenating and pro-aging hits and rule confounders such as stress response

Mechanistic interpretation: For top perturbations, transcriptional responses will be summarized as inferred TF activity changes using known regulons and mapped to pathway-level explanations using prior knowledge networks with parsimony constraints to favor minimal causal mechanisms. Convergence will be assessed by testing whether rejuvenating/pro-aging perturbations consistently map to a compact set of shared TF modules and pathways across robustness checks (for example, resampling and repeated model fitting), yielding mechanism-level prioritization.

Intervention hypotheses and validation: Mechanistic modules will be translated into upstream, druggable hypotheses (for example, receptors, ligands, or pathway modulators). In silico validation will prioritize independent perturbation datasets that target the same inferred pathway modules to test whether these interventions reproduce the same ΔAgeScore change relative to controls. Minimal in vitro testing will compare treated versus control conditions using the same aging-score framework and a lightweight transcriptional readout in a cell line, focusing on directional consistency.

Reporting: Deliverables will include a list of ranked perturbation effects with uncertainty and robustness flags, a set of stable TF and pathway modules mapped from top ranked perturbations, and a list of intervention hypotheses with in silico or minimal in vitro validation.

Contingency: IF existing aging clocks do not transfer well to the perturbation context, THEN we will train or calibrate a context-matched clock using available aging references. IF ligand or stimulation reference datasets are insufficient for in silico validation, THEN report pathway-level hypotheses and prioritize a minimal, feasible wet validation target instead.

Ethics and Data Management

The primary work uses publicly available de-identified datasets and does not involve human subjects enrollment. Optional minimal in vitro validation, will use established cell models and standard institutional biosafety procedures within available lab infrastructure at Ragon Institute, Broad Institute, and Massachusetts Institute of Technology, without new human or animal subject approvals for this phase. All code, trained models, processed outputs (including ranked perturbation effects, TF modules, and candidate intervention lists), and documentation will be released in a public repository. Results will be shared as a preprint regardless of outcome. All analyses will be designed to be reproducible from deposited processed datasets.

4. Pilot Data (Optional)

We have developed cell-type-specific single-cell RNA clocks using a custom machine learning framework, which can outperform baseline prediction even for rare cell types. These clocks can robustly predict directional shifts under rejuvenating perturbations from external datasets including for heterochronic parabiosis (i.e. young blood factors) and exercise.

5. Budget

Personnel cost: $10,000 (approximately 12% effort)

Total: $10,000

NOTE: Experimental and high-performance computing costs will be supported through other support, including lab start-up funds.

6. References

Sun, E. D., Nagvekar, R., Pogson, A. N. & Brunet, A. Brain aging and rejuvenation at single-cell resolution. Neuron 113, 82–108 (2025).
Dixit, A. et al. Perturb-Seq: Dissecting Molecular Circuits with Scalable Single-Cell RNA Profiling of Pooled Genetic Screens. Cell 167, 1853-1866.e17 (2016).
Buckley, M. T. et al. Cell-type-specific aging clocks to quantify aging and rejuvenation in neurogenic regions of the brain. Nat. Aging 3, 121–137 (2022).
Sun, E. D. et al. Spatial transcriptomic clocks reveal cell proximity effects in brain ageing. Nature 638, 160–171 (2025).
Badia-i-Mompel, P. et al. decoupleR: ensemble of computational methods to infer biological activities from omics data. Bioinforma. Adv. 2, vbac016 (2022).
Türei, D., Korcsmáros, T. & Saez-Rodriguez, J. OmniPath: guidelines and gateway for literature-curated signaling pathway resources. Nat. Methods 13, 966–967 (2016).
Liu, A. et al. From expression footprints to causal pathways: contextualizing large signaling networks with CARNIVAL. Npj Syst. Biol. Appl. 5, 1–10 (2019).
Browaeys, R., Saelens, W. & Saeys, Y. NicheNet: modeling intercellular communication by linking ligands to target genes. Nat. Methods 17, 159–162 (2020).
Müller-Dott, S. et al. Expanding the coverage of regulons from high-confidence prior knowledge for accurate estimation of transcription factor activities. Nucleic Acids Res. 51, 10934–10949 (2023).

scPerturbClock: Genome-scale Causal Discovery of Aging Regulators with Machine Learning

scPerturbClock: Genome-scale Causal Discovery of Aging Regulators with Machine Learning

Funding Progress

Related Funding

Request for Proposals - Aging and Longevity

Nonprofit Recipient

Topics

DOI