Abstract

DrugAge contains many promising lifespan-extension results, but striking effects in isolated experiments do not automatically become durable scientific claims. We present an offline automated pipeline that turns DrugAge into a robustness-first screen for longevity interventions. Rather than rewarding the single most exciting reported result, the pipeline favors compounds whose pro-longevity signal is broad across species, survives prespecified stress tests, and remains measurably above a species-matched empirical null baseline. The canonical run uses a bundled local copy of DrugAge Build 5, explicit normalization rules, evidence tiers, a Claim Stability Certificate, and an Empirical Null Certificate. In the canonical run, the pipeline retained 3,372 scored experiments spanning 1,038 compounds, 33 normalized species, and 9 taxon labels, and identified 48 robust compounds. The observed top-10 mean robustness score was 0.94097 versus a null mean of 0.91320, while the robust-compound count was 48 versus a null mean of 29.865 (empirical p = 0.002 and p = 0.001 respectively, 1,000 permutations). The result is a reproducible shortlist of model-organism longevity claims that has been stress-tested before reporting, not a recommendation engine for human interventions.

Introduction

This submission presents an automated pipeline for ranking DrugAge longevity intervention claims by robustness. The canonical execution path is fully offline and uses a bundled local copy of the Human Ageing Genomic Resources (HAGR) DrugAge Build 5 dataset, deterministic normalization, evidence tiers, a Claim Stability Certificate, and an Empirical Null Certificate.

The contribution is not a leaderboard of raw lifespan effects. The contribution is a self-auditing pipeline that asks which model-organism longevity claims remain convincing after perturbation and falsification pressure.

Related Work

DrugAge is a curated database of aging-related drugs maintained as part of the Human Ageing Genomic Resources (HAGR) project. Barardo et al. (2017) describe the database design and its initial content of over 400 compounds tested in model organisms. The related Geroprotectors.org database (Moskalev et al., 2015) takes a complementary approach, cataloguing therapeutic interventions with structured annotations for mechanism and clinical status. De Magalhaes et al. (2018) demonstrate that demographic measurements of the rate of aging in mice provide a more reliable basis for assessing genetic interventions than single-endpoint lifespan measures, motivating our emphasis on robustness over raw effect size. The NIA Interventions Testing Program (ITP) represents the closest precedent for multi-site robustness testing of longevity compounds through prospective replication across three independent laboratories. Our pipeline differs by performing retrospective robustness reanalysis of existing DrugAge records rather than new experiments, and by applying systematic permutation-based null calibration to rank cross-species robustness.

Data

The canonical input is data/drugage_build5_2024-11-29.csv, a bundled local copy of the DrugAge Build 5 dataset dated November 29, 2024. The canonical execution path validates the file hash, required columns, and release metadata before processing. No network access is needed after the repository is cloned.

Rows are dropped only if compound name, species, or numeric avg_lifespan_change_percent is missing. In the canonical run, the pipeline retained 3372 of 3423 DrugAge rows, covering 1038 compounds, 33 normalized species, and 9 scored taxon labels.

Methods

The canonical ranking uses DrugAge's average lifespan change field, avg_lifespan_change_percent, because it is the most consistently populated and directly comparable effect-size field in DrugAge. Significance annotations are retained descriptively because they are heterogeneous across studies and do not provide a stable standalone ranking signal. For each compound, the pipeline computes:

• number of experiments
• number of species
• number of taxa
• number of PMIDs
• median effect
• 10% trimmed-mean effect
• sign consistency
• leave-one-species-out stability
• leave-one-taxon-out stability
• aggregation stability
• breadth score
• robustness score

Metric Definitions

Breadth score. A normalized composite of experimental, species, and taxonomic coverage, each capped at a saturation point:

breadth_score = (min(n_experiments, 6)/6 + min(n_species, 4)/4 + min(n_taxa, 3)/3) / 3

Aggregation stability. A binary indicator of whether both central-tendency estimators agree on a positive effect:

aggregation_stability = 1.0 if median_effect > 0 and trimmed_mean_effect > 0, else 0.0

Magnitude score. A normalized effect size capped at 50% lifespan extension:

magnitude_score = min(max(trimmed_mean_effect, 0), 50) / 50

Robustness score. A weighted combination of the component metrics:

robustness_score = 0.35 * breadth_score
                 + 0.20 * sign_consistency
                 + 0.15 * leave_one_species_out_stability
                 + 0.10 * leave_one_taxon_out_stability
                 + 0.15 * aggregation_stability
                 + 0.05 * magnitude_score

The weights prioritize breadth (0.35) and consistency (0.20 + 0.15 + 0.10 = 0.45) over raw effect magnitude (0.05), reflecting the design goal of ranking by durability rather than excitement.

Sign consistency. sign_consistency = n_positive_experiments / n_total_experiments, where a positive experiment is one with average lifespan change > 0.

Leave-one-species-out (LOSO) stability. LOSO = n_positive_subsets / n_species, where each subset recomputes the compound's 10% trimmed-mean effect with one species removed, and a positive subset retains trimmed mean > 0. For compounds tested in only 1 species, LOSO = 0.0 because no leave-one-species-out subset remains after removal.

Leave-one-taxon-out (LOTO) stability. LOTO = n_positive_subsets / n_taxa, the same leave-one-out procedure applied at the taxonomic group level. For compounds tested in only 1 taxon, LOTO = 0.0 because no leave-one-taxon-out subset remains after removal.

Compounds are then assigned to four evidence tiers with explicit thresholds:

• robust: ≥3 experiments, ≥2 species, ≥2 PMIDs, positive median and trimmed mean, sign consistency ≥0.80, LOSO = 1.0, aggregation stability = 1.0
• promising: positive median and trimmed mean, sign consistency ≥0.67, and ≥2 species or ≥3 experiments
• thin evidence: positive median and trimmed mean, sign consistency ≥0.67 (but insufficient breadth for promising)
• conflicted: all remaining compounds

Ranking is robustness-first: tier priority dominates sort order, followed by robustness score, species breadth, taxonomic breadth, PMID breadth, and effect magnitude.

The canonical execution path emits two verification certificates.

The Claim Stability Certificate evaluates the top-ranked compounds under five fixed perturbations:

• leave-one-species-out positivity
• leave-one-taxon-out positivity
• positive median and trimmed mean
• exclusion of single-PMID compounds
• exclusion of mixed-sign compounds

The Empirical Null Certificate runs 1,000 fixed-seed species-stratified effect permutations. Within each species, average lifespan effects are shuffled across rows, preserving DrugAge's species composition and within-species effect distribution while breaking the link between compound identity and observed effect. With 1,000 reruns, the smallest nonzero empirical p-value is 1/1001, approximately 0.0010.

Results

In the canonical run, the evidence tiers were:

• robust: 48 compounds
• promising: 174 compounds
• thin evidence: 435 compounds
• conflicted: 381 compounds

The top 10 compounds were:

1. Spermidine
2. Apple extract
3. N-acetyl-L-cysteine
4. Minocycline
5. Alpha-ketoglutarate
6. Carnosine
7. Rapamycin
8. Mycophenolic acid
9. Epigallocatechin-3-gallate
10. Vitamin E

Some top-ranked compounds may look surprising; this ranking reflects internal robustness within curated model-organism evidence, not human plausibility or mechanistic priority.

All top-10 compounds passed leave-one-species-out positivity, leave-one-taxon-out positivity, positive median-vs-trimmed-mean checks, and the single-PMID exclusion perturbation. Three of the top 10 also passed the stricter mixed-sign exclusion perturbation.

The observed top-10 mean robustness score was 0.94097, compared with a null mean of 0.91320 and null standard deviation 0.01131. The corresponding empirical p-value was 0.00200 with z-score 2.45, indicating measurable rather than overwhelming score separation. The more persuasive null result was the robust-compound count: 48, compared with a null mean of 29.865, null standard deviation 4.10, empirical p-value 0.00100, and z-score 4.42.

Sex-Stratified Sensitivity Analysis

The pipeline supports a --gender filter that restricts experiments to a single sex before scoring. To assess whether robustness rankings are stable across sexes, we ran drugage-skill run --gender Male (858 experiments, 353 compounds) and --gender Female (612 experiments, 271 compounds):

6 of the top-10 compounds are shared between sexes (Rapamycin, N-acetyl-L-cysteine, Butylated hydroxytoluene, Curcumin, L-deprenyl, Rhodiola rosea extract), indicating a stable core of robust longevity compounds. Sex-specific compounds include Minocycline and Chloroquine (male-only top-10) and Melatonin and Alpha-ketoglutarate (female-only top-10). Male-only analysis identified 8 robust compounds; female-only identified 9.

Optional AnAge Context

The optional AnAge context report joins normalized DrugAge species to a bundled local copy of AnAge for descriptive context only. It does not alter the canonical ranking, scores, tiers, or certificates. In the current rerun, 10 of 35 normalized DrugAge species matched AnAge exactly after normalization.

Limitations

This pipeline ranks model-organism longevity evidence and does not recommend interventions for humans. It does not harmonize doses across studies, perform effect-size meta-analysis, or infer mechanism. DrugAge significance fields are retained descriptively, not as scored inputs. DrugAge records include dosage, sex, and administration timing metadata, but the canonical ranking pools experiments across doses and sexes. A sex-stratified sensitivity analysis (see above) shows 6/10 top compounds are shared between sexes, indicating a stable core with meaningful sex-specific differences. Dose- and timing-stratified analyses remain as future work. The Empirical Null Certificate uses 1,000 fixed-seed within-species permutations. With 1,000 reruns, the observed top-10 mean robustness score (0.94097) exceeds the null mean (0.91320) at p = 0.002, and the observed robust-compound count (48) exceeds the null mean (29.865) at p = 0.001 (z = 4.42). The permutation count is configurable; 1,000 provides sufficient resolution for ranking-level separation. Some top-ranked compounds may look surprising; this reflects robustness within curated model-organism evidence, not translational plausibility or mechanistic priority. The optional AnAge join is intentionally descriptive and partial.

Conclusion

This repository contributes a lightweight, offline longevity pipeline that ranks claims by robustness, certifies perturbation stability, and measures separation from a species-matched empirical null. The main result is not a static list of compounds. The main result is a reproducible pipeline that interrogates its own conclusions before reporting them.

References

• Barardo, D., Thornton, D., Thoppil, H., Walsh, M., Sharber, S., Ferber, S., Greer, E.L., Ship, A., Valli, A., Horro, R., & de Magalhaes, J.P. (2017). The DrugAge database of aging-related drugs. Aging Cell, 16(3), 594--597.
• de Magalhaes, J.P., Budovsky, A., Li, Q., Fraifeld, V.E., & Church, G.M. (2018). A reassessment of genes modulating aging in mice using demographic measurements of the rate of aging. Genetics, 208, 1617--1630.
• Moskalev, A., Chernyagina, E., de Magalhaes, J.P., Barardo, D., Thoppil, H., Shaposhnikov, M., Budovsky, A., Fraifeld, V.E., Garaz, A., Tsvetkov, V., et al. (2015). Geroprotectors.org: a new, structured and curated database of current therapeutic interventions in aging and age-related disease. Aging, 7(9), 616--628.