Hi PH 👋 I'm Dinesh, one of two people building this. Every AI drug-discovery tool I looked at had the same two problems: it would hand you a molecule no chemist could actually make, and it would report a benchmark score without telling you when that score stops working. So we built GeroQubit backwards. Instead of generating a structure and reverse-engineering a synthesis, the engine evolves recipes — scaffold + building blocks + named reactions. The molecule is the output of a route that already exists. Nothing gets proposed that can't be built. The part I'd rather show you than hide: our binding score hits ROC-AUC 0.945 on known chemistry — and collapses to 0.62 on genuinely novel molecules. That collapse is universal in this field; almost nobody publishes it. We put it on our benchmark page in the same size type as the good number, because a score you can't trust is worse than no score. Other things that are real: matches the strongest unconstrained generators on the PMO benchmark (0.938, QED oracle) while every candidate stays makeable; the whole search runs on a CPU in seconds; and each efficacy prediction carries a calibrated confidence interval (~90% coverage). It's research-use only — a hypothesis engine for medicinal chemists, not a clinical claim. We have no wet-lab data yet, and I'd rather say that up front. Happy to answer anything, especially the sceptical questions. 🧬

GeroQubit

De-novo longevity drug design you can actually synthesize

De-novo longevity drug design you can actually synthesize