Pyrite improves ranking of protein ligand binding poses with AI assistance
Pyrite: Building and Benchmarking Protein-Ligand Pose Ranking Agents
Machine Learning
Summary
Predicting how small molecules called ligands attach to proteins is important for drug discovery, but it is hard to tell which predicted shapes are best. The authors created a benchmark named Pyrite to test how well people and computers rank these predicted poses. Expert chemists perform about as well as random guessing, showing the challenge. Using large language models to analyze and generate features from molecular data, the authors built automated agents that outperform experts in ranking accuracy. This work highlights new ways to combine AI tools to better understand molecular interactions.
What this means in practice
- •For drug discovery teams: Improve prediction and ranking of how small molecules bind proteins to support more effective drug design.
- •For computational chemists: Use AI-guided feature extraction to interpret complex protein-ligand interaction data and enhance pose scoring.
Authors
Megha Srivastava, Michael H. Lim, Allison M. Keys, Pranav Murugan, Nate Gruver
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. We 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, we introduce PyriteRep, a structured representation of molecular geometry, protein-ligand interactions, and physicochemical scores, and develop two LLM-based agents: PyriteReasoner, which directly compares poses through text-based reasoning, and PyriteFeaturizer, which generates executable code to construct features for logistic regression, optionally incorporating features derived from LLM subcalls. These agents achieve up to 67% and 74% accuracy, respectively, exceeding observed expert performance. LLM-guided feature discovery generally outperforms direct reasoning, while Group Relative Policy Optimization (GRPO) improves open-weight models, enabling them to match or surpass closed-source frontier models in our evaluations. These results establish Pyrite as a challenging test of pose-ranking ability and highlight LLM-guided feature discovery as a promising approach to interpret and combine the outputs of computational chemistry tools.