From Gene Lists to Durable Signals: A Self-Verifying Longevity Signature Triangulator

Abstract

We present an offline, agent-executable workflow that classifies gene signatures as aging-like, dietary-restriction-like, senescence-like, mixed, or unresolved from vendored Human Ageing Genomic Resources (HAGR) snapshots. The contribution is not merely a label assignment. The workflow tests whether the label survives perturbation, remains specific against competing longevity programs, and beats explicit non-longevity confounder explanations before reporting it. The scored path uses only frozen snapshots from GenAge, GenDR, CellAge, and HAGR ageing and dietary-restriction signatures, and it ships with a strict verifier plus a holdout-source benchmark that tests whether each canonical example still classifies correctly when the source family used to build that example is withheld at scoring time. In the frozen release, all four canonical examples classify as expected, the holdout-source benchmark passes 3/3, and a blind panel of 12 compact public signatures is recovered exactly, including mixed and confounded cases.

Introduction

Gene-list interpretation is often too easy to overstate. A signature can appear longevity-related because it overlaps one curated database, yet still be unstable under small perturbations or better explained by stress, inflammation, or growth arrest. This repository addresses that problem with an executable, fully offline workflow designed for agent review.

Data and Scope

The scored path uses vendored Human Ageing Genomic Resources (HAGR) snapshots only:

• GenAge human genes
• HAGR ageing expression signature
• a frozen humanized GenDR manipulation subset
• HAGR dietary restriction signature
• CellAge senescence genes
• CellAge senescence signatures

AnAge is vendored only for optional descriptive context and never changes the canonical classification.

Methods

The classifier normalizes a simple gene list, ranked gene list, or differential-expression table into a common internal schema. It then computes class-level evidence tables for the three longevity classes and a fixed confounder panel. Each longevity class is anchored by two frozen source families: GenAge plus the ageing signature for aging-like calls, humanized GenDR manipulations plus the dietary-restriction signature for DR-like calls, and CellAge genes plus senescence signatures for senescence-like calls. Longevity classes are scored with breadth, weighted overlap, directional consistency when available, and source consistency across those paired source families.

Three certificates are then generated:

1. Claim Stability Certificate
2. Adversarial Specificity Certificate
3. Causal Plausibility / Confounder-Rejection Certificate

The workflow also includes a holdout-source benchmark. Each canonical example is built from one source family and then reclassified with that source family withheld. This tests non-circularity rather than simple same-source recovery.

The confounder panel is a fixed project-curated asset with explicit citations. The causal-plausibility verdict is credible only when the winning longevity class clears preset margins over both the nearest longevity competitor and the best confounder; it is ambiguous when the confounder margin is positive but small, and confounded when the best confounder ties or wins.

Results

The repository includes four deterministic canonical inputs and all four classify as expected. The three single-program fixtures each receive pass/pass/credible verdicts, while the balanced mixed fixture is correctly left mixed rather than forced into a single program. In the frozen release, the nearest confounder for all four fixtures is cell-cycle arrest/quiescence (cell_cycle_arrest_quiescence), the holdout-source benchmark passes 3/3, and the verifier reproduces the canonical DR-like run with stable summary metrics.

The mixed example is intentionally borderline: its specificity margin is 0.00, so the workflow leaves it mixed, and the failed stability or specificity certificates indicate that small perturbations can flip the signal toward aging_like or senescence_like rather than supporting a robust single-program call.

External Blind Challenge Panel

To test source-independent generalization, we evaluated a frozen blind panel of 12 compact public signatures curated outside the HAGR source families used to define the reference classes. The workflow recovered the expected label in 12/12 cases, including one mixed NeuroHIV microglia case and two confounded negatives that were left unresolved with confounded verdicts rather than overcalled. Aging-like and DR-like positives again most often encountered cell_cycle_arrest_quiescence as the nearest confounder, while senescence-like positives most often approached inflammation_sasp. These results suggest the workflow generalizes beyond reference-derived fixtures while still preserving conservative behavior on ambiguous or confounded inputs.

Limitations

This workflow makes narrow claims. It does not infer causal mechanisms, does not perform runtime ortholog mapping, does not use live APIs during scoring, and does not recommend interventions or make human-translational claims. The confounder panel is curated and finite rather than exhaustive.

Conclusion

The main result is not a static annotation of a gene list. The main result is an executable skill that tests whether a longevity interpretation is stable, specific, and causally plausible before reporting it.