Paper Overview
- Field: Machine Learning
- Authors: Megha Srivastava, Michael H. Lim, Allison M. Keys, Pranav Murugan, Nate Gruver
- Released: 2026-11-10
- arXiv: 2611.14511
- PyriteReasoner: directly compares poses through text-based reasoning
- PyriteFeaturizer: generates executable code to construct features for logistic regression, with the option to fuse features produced by LLM sub-calls
- The two agents achieve accuracies of 67% and 74%, surpassing human experts on the benchmark.
- LLM-guided feature discovery generally outperforms direct reasoning over structured representations.
- GRPO training improves open-weight models, enabling them to match or exceed closed-source frontier models on pose ranking.
Abstract
Predicting how a ligand binds to its target is central to small-molecule drug discovery. Modern cofolding models can generate diverse candidate poses, but reliably ranking these candidates remains challenging. The authors introduce Pyrite, a pairwise pose-ranking benchmark comprising 1,500 pose pairs across diverse protein-ligand complexes. Expert chemists equipped with 3D visualization tools and physics-based scores perform near chance, highlighting the difficulty of interpreting available chemical signals.
To investigate automated ranking, they introduce PyriteRep, a structured representation of molecular geometry, protein-ligand interactions, and physicochemical scores, and develop two LLM-based agents:
Key Findings
Significance
The results establish Pyrite as a challenging test of pose-ranking capability and highlight LLM-guided feature discovery as a promising direction for understanding and combining the outputs of computational chemistry tools.
---
*Auto-collected on 2026-11-12*