A molecular coin toss stumps expert chemists. AI-generated features score 74%.

Five experienced chemists struggled to choose which of two plausible molecular poses better matched an experimental structure. A system that had AI write feature-extraction code, then trained a simple classifier, reached 74% accuracy on a curated test set.

Stanford and Genesis Molecular AI built Pyrite from 1,500 pose pairs. Generating possible protein–drug arrangements is increasingly practical; choosing the right one remains difficult. Experts had 3D tools and physical scores. On the newer curated split, their accuracy was 45% counting ties as wrong, or 57% counting only decisive answers.

PyriteFeaturizer combined molecular geometry and interaction clues. A reinforcement-trained Qwen3.8-27B version scored 74% on 265 curated pairs and 71% across the full 484-pair test split. A conventional statistical baseline scored 64% and 60%, respectively.

The preprint points toward AI helping researchers discover useful decision rules. It tests candidate poses from a particular generation pipeline, not chemistry expertise overall or drug effectiveness. Whether the gains reduce real-world drug-development trial and error remains unproven.

Source published: Research team / arXiv, November 11, 2026, 02:56:32 Beijing time / November 10, 2026, 1:56:32 PM EST.