GENESIS: Potency-Graded Contrastive Learning for Generative Molecular Design and Drug Repurposing in Oral Squamous Cell Carcinoma
Year: 2026
Type: preprint
Abstract: Computational drug discovery typically binarizes bioactivity, discarding the graded doseresponse signal needed for lead prioritization. We introduce Potency-Graded Contrastive Learning (PGCL), projecting frozen molecular language-model embeddings into a 128dimensional space via an ordinal contrastive objective, six-tier negative taxonomy, and tier-margin ranking. Across three seeds, PGCL achieves Spearman ρ = 0.555 ± 0.003 for potency ordering and target-switch AUROC = 0.967, outperforming flat-contrastive and binaryclassification ablations and roughly doubling ECFP4 activity-cliff discrimination. Among six generative architectures, our reinforcement-learning method CRAVE (Contrastive Reward-Aligned VAE Exploration) significantly exceeds non-RL baselines on OSCC potency after Bonferroni correction (p = 0.016). The frozen model transfers to lung adenocarcinoma without retraining — strongly along the PI3K/AKT/mTOR axis (PIK3CA ρ=0.688, MTOR ρ=0.614, AKT1 ρ=0.583, EGFR ρ=0.561) but only modestly when pooled across all 16 LUAD-relevant targets (ρ=0.207, n=52,373). An integrated docking, ADMET, network-pharmacology, and leadoptimization pipeline screens 23,308 repurposing candidates, prioritizing 46 Tier-1 candidates, including 13 triple-validated elite compounds, with transparently reported safety pass rates (14.6%) traced to an EGFR/aniline training-corpus skew. The pipeline runs end-to-end on consumer-grade hardware, with code and protocols enabling transfer to other cancers.
Source: ChemRxiv
Author: Saurabh Bundela
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7237225736
