RL against make-on-demand catalogs lifts KRAS molecule hits from 10% to 25.5%
bravo_abad · x · 2026-11-09
Neeser et al. train a flow-matching model that generates 3D molecules for protein binding pockets with RL, using a concrete reward: exact membership in the Enamine REAL make-on-demand catalog, instead of vague "synthetic accessibility" scores. For KRAS, unique catalog matches rose from 10.0% to 25.5% of candidates; longer training increased repetition, so an earlier checkpoint yielded more distinct testable compounds. Chart shows DNFT-REAL-S/FREE-S variants well above baseline while DPO variants lag. Caveat: catalog membership doesn't guarantee synthesizability.
More from Research
- PostTrainBench v1.2: Gemini 4 Argon Scores 40.4% (#5), Muse Wonder 24.2% (#17) — maksym_andr · 2026-11-09
- SPDMX releases 254K-song open multitrack music dataset synthesized from public-domain scores — umpedronosapato · 2026-11-09
- MathArena adds October BrokenArXiv: GPT-6.1 Sol leads but drifts down monthly, Opus 5.5 rock solid — teortaxesTex · 2026-11-09
- AI Proves the Aaronson-Ambainis Conjecture, Another Longstanding Quantum Theory Open Problem Falls — ns123abc · 2026-11-09
- Why gradients rapidly increase near the end of training: weight decay meets LR decay, says Defazio — aaron_defazio · 2026-11-09
- Perplexity launches Frontier Research group to openly advance models, data, and algorithms — andrewgwils · 2026-11-09