Machine Learned Interatomic Forces as Inference-Time Physical Guidance for All-Atom Diffusion Peptide Design

Abstract

Diffusion-based biomolecular generative models can efficiently explore peptidereceptor sequence-structure space, but the learned generative distribution does not explicitly represent the underlying atomistic potential-energy landscape. As a result, generated complexes may contain locally unfavorable atomic arrangements even when they remain plausible under the learned structural prior. Here, we develop an inferencetime physical guidance framework that couples machine learned interatomic potential (MLIP)-derived forces to the all-atom reverse diffusion trajectory of RFDiffusion3 for receptor-conditioned peptide design. Rather than applying energy minimization after generation, atomistic physical information is introduced before the joint sequencestructure trajectory termination. MLIP-derived forces therefore provide atomic coordinate corrections that are propagated through subsequent denoising steps, allowing physical information to influence not only atomic geometry but subsequent residue selection. The framework requires neither retraining nor modification of the pretrained diffusion model and uses late-stage force guidance, adaptive step-size control, and geometry-preserving projection to retain the underlying generative prior. Across eight receptor systems for peptide designs, MLIP guidance reduced peptidereceptor steric clashes by 24.7% and intrapeptide clashes by 37.4%, while promoting more hydrogen-bond-compatible interface geometries. Guided structures shifted toward more favorable energetic regions in both the guiding MACE potential and an independent MM/GBSA evaluation, with 82.4% of retained structures exhibiting more favorable MM/GBSA energetic estimates. These improvements were achieved with a median peptide backbone deviation of only 0.051 Å. Importantly, 95.0% of generated peptide sequences were preserved, whereas 5.0% underwent residue substitutions, demonstrating that physical guidance can feed back into sequence-structure generation without globally overriding the learned RFDiffusion3 distribution. This distinguishes inference-time physical guidance from post-generation relaxation, which optimizes only the coordinates at fixed chemical identity. More broadly, the results establish a general strategy for coupling pretrained biomolecular generators with differentiable atomistic energy models, enabling physical information to participate directly in molecular generation rather than serving only as a post hoc filter or refinement step.