Casting Protein Structure Predictors as Energy-Based Models for Binder Design and Scoring

Abstract

Protein binder design has been transformed by hallucination-based methods that optimize structure prediction confidence metrics, such as the interface predicted TM-score (ipTM), via backpropagation. However, these metrics are imperfect proxies for binding affinity and do not reflect the statistical likelihood of a binder–target complex under the learned distribution. In this work, we propose a principled alternative: an energy-based framework that directly extracts the statistical likelihood of a predicted binder–target complex from a structure predictor’s internal confidence distributions. Building on the Joint Energy-based Modeling (JEM) framework, we introduce pTMEnergy, a statistical energy function over structures that is derived from predicted inter-residue error distributions. We incorporate pTMEnergy into BindEnergyCraft (BECraft), a hallucination-based binder design pipeline that maintains the same optimization framework as BindCraft but replaces ipTM with our energy-based objective. Across a diverse panel of challenging protein targets, BECraft achieves higher in silico success rates compared to BindCraft, RFDiffusion, and ESM3. Beyond design, we evaluate pTMEnergy as an unsupervised scoring function for retrospective virtual screening tasks. Without any task-specific supervision or retraining, pTMEnergy consistently outperforms baseline methods across both protein–protein and protein–RNA interaction benchmarks. Our results demonstrate that confidence-derived energy functions offer a powerful and generalizable signal for binder design and scoring.