Post

PREPRINT (not peer-reviewed). In a computational comparison of six molecular generators for oral squamous cell carcinoma, filtered candidates from a reinforcement-learning model called CRAVE had the highest mean predicted potency score: 0.771 ± 0.033 across five seeds. The authors report significant differences against three comparators after Bonferroni correction, but not against RL-VAE. Transfer to a lung adenocarcinoma dataset was uneven: the PIK3CA target had Spearman ρ = 0.688, whereas pooled correlation was 0.207. These are model scores, not measured drug potency. No generated or optimized compounds were tested in the laboratory for potency, toxicity or ADMET; the safety filters were rule-based triage. Find the paper on ChemRxiv by searching its full title: GENESIS: Potency-Graded Contrastive Learning for Generative Molecular Design and Drug Repurposing in Oral Squamous Cell Carcinoma.

12:46:29 AM · Nov 23, 2026
Sign in to reply and join the conversation.