• Sources: primary
  • Summary: Pyrite is a pairwise benchmark of 1,500 protein-ligand pose pairs on which expert chemists with 3D viewers and physics scores perform near chance. An agent that reasons directly over a structured text representation reaches up to 67%, and one that writes code to build features for logistic regression reaches up to 74%. GRPO training lets open-weight models match or beat closed frontier models on the task.
  • Why it matters: Teams using cofolding models get a benchmark and a code-generating agent pattern for picking among candidate poses.
  • Background: Cofolding models such as Boltz and Chai predict a protein and ligand structure together and can sample many candidate binding poses. Choosing the correct pose is a separate problem, since physics-based scores and visual inspection often disagree. GRPO is a reinforcement learning method that scores groups of sampled answers against each other instead of using a learned value function.