ConforFlux: Particle-Guided Trunk Repulsion for Diverse Protein Conformations

Abstract

Deep-learning protein structure predictors achieve near-experimental accuracy on individual folds, yet their default inference samples concentrate around a single dominant conformation. We introduce ConforFlux, an inference-time procedure for Boltz-2 that couples multiple structure-prediction trajectories through a pair-wise C-RMSD repulsion gradient on the trunks single and pair embeddings. Because the trunk conditions every block of the diffusion module, this update propagates to every subsequent denoising step. On four conformational-change categories, ConforFlux improves the per-state success rate over default Boltz-2 by 3-17 percentage points. On twelve transporter pairs with at least one alternate-state reference released after the Boltz-2 cutoff, ConforFlux raises the 2[A] success rate from 33% to 75%. Under an extended bandwidth sweep, ConforFlux samples reach the inward, occluded, and outward states of the human dopamine transporter alternating-access cycle, while default samples cluster between them.