BindEnergyCraft: Casting Protein Structure Predictors as Energy-Based Models for Binder Design

TL;DR

This work tackles the misalignment between optimizer signals and true binder–target likelihood by converting structure-prediction confidence into a probabilistic energy function. By deriving pTMEnergy from AlphaFold2-Multimer's pAE outputs under a Joint Energy-based Modeling framework, the authors create a dense, differentiable objective that guides sequence optimization more effectively than ipTM. Integrating pTMEnergy into BindEnergyCraft yields higher in silico design success rates and fewer atomic clashes, and moreover establishes state-of-the-art unsupervised scores for miniprotein and RNA aptamer virtual screening. This principled energy-based approach broadens the design toolkit for biomolecular engineering and holds promise for cross-modal design tasks, albeit with caveats around the calibration of confidence models and the need for experimental validation.

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 do not reflect the statistical likelihood of a binder-target complex under the learned distribution and yield sparse gradients for optimization. In this work, we propose a method to extract such likelihoods from structure predictors by reinterpreting their confidence outputs as an energy-based model (EBM). By leveraging the Joint Energy-based Modeling (JEM) framework, we introduce pTMEnergy, a statistical energy function derived from predicted inter-residue error distributions. We incorporate pTMEnergy into BindEnergyCraft (BECraft), a design pipeline that maintains the same optimization framework as BindCraft but replaces ipTM with our energy-based objective. BECraft outperforms BindCraft, RFDiffusion, and ESM3 across multiple challenging targets, achieving higher in silico binder success rates while reducing structural clashes. Furthermore, pTMEnergy establishes a new state-of-the-art in structure-based virtual screening tasks for miniprotein and RNA aptamer binders.

Figures (4)

• Figure 1: A: BindEnergyCraft (BECraft) optimizes binder sequences by backpropagating pTMEnergy, computed from pAE logits output by AlphaFold2-Multimer. B: Gradients from ipTM are sparse across interface residue pairs, since the maximum over target residue indices zeroes out the gradient from all but one position. pTMEnergy preserves gradients across the interface.
• Figure 2: (A) A VirB8-binding complex designed using ipTM exhibits severe atomic clashes at the interface. (B) The corresponding pTMEnergy-optimized design for VirB8 forms a clash-free interface. (C) An ipTM-optimized binder for ALK with substantial steric overlap. (D) The pTMEnergy-optimized design for ALK forms a well-packed, physically realistic interaction.
• Figure 3: (A–B): Magnitude of maximum gradient error bin for each binder–target residue pair (binder on $y$, target on $x$), normalized to $[0,1]$, for ipTM and pTMEnergy respectively on an IL2Ra design task. (C–D): Frequency with which each target residue ranks in the top-10 by gradient magnitude across all iterations, for ipTM and pTMEnergy respectively.
• Figure 4: Distribution of predicted pTMEnergy scores. Scores are negated so that higher values indicate stronger predicted binding. Red vertical lines indicate scores assigned to true binders.