Why we built it: Molecule generators can produce huge numbers of structures, but many outputs are difficult to synthesize, lack aging-specific context, or report precise-looking predictions without useful uncertainty.
What we tried: - Target × tissue × hallmark organization for geroscience - Reaction-first generation with proposed synthesis routes - CPU-only execution rather than GPU infrastructure - Prediction intervals alongside model scores - Comparison against simple baselines
Current results: - 0.945 binding ROC-AUC - approximately 20x enrichment at the top 1% - 90.3% empirical interval coverage - lifespan-ranking correlation of 0.145 on 627 DrugAge compounds
Important limitations: The work is retrospective, the lifespan signal is modest, and none of the generated candidates should be treated as experimentally validated drugs. The preprint is not peer reviewed.
I would particularly value criticism of the validation design, reaction-first generation, and uncertainty calibration.