Sony AI's Hakken predicted undiscovered biology links — independent lab confirmed 2 of 3
alex_verem · x · 2026-10-28
Sony AI's Hakken (arXiv:2609.04494) predicts scientific relationships before anyone publishes them. Pointed at aging-related genes, it generated 1.5M hypotheses; biologists picked 3 and an independent lab confirmed 2: a TP53–BAMBI link and a RAF1–TNF link, neither previously reported.
- It treats the literature as a knowledge graph of 7.1M links among 255K genes, drugs and diseases, and predicts missing edges, showing the reasoning chain behind each prediction.
- Backtesting on papers up to 2010, it anticipated 70% of links discovered in the next two years and 2/3 over a decade (random baseline 25%); results held for 1990/2000 cutoffs.
- Built around a 7B open model, trained in days on two machines, with an open-source GitHub release. Some flagged-but-never-seen links may be discoveries yet to be made.
More from Research
- Eric Topol highlights immunome paper combining multi-assay data with AI to model immune aging — EricTopol · 2026-10-28
- Google Research's 'sleep framework' lets LLMs keep learning, retaining up to 10M tokens — behrouz_ali · 2026-10-28
- CATCH: ~200 examples of adversarial training make LLMs robustly honest across benchmarks — irinarish · 2026-10-28
- Andrew Critch's dubious credences: P≠NP at 80%, Riemann hypothesis at just 60% — AndrewCritchPhD · 2026-10-28
- AI-driven Indian biotech PopVax, backed by $20M from Vitalik's Balvi fund, enters human trials with fridge-stable broad-spectrum Covid vaccine — chaitjo · 2026-10-28
- Cognition launches FrontierCode 2.0, a coding benchmark testing judgment beyond the spec — DevinAI · 2026-10-28