BridgeDock: Taming Generative Co-Folding Prior for Molecular Docking with Diffusion Bridge

| Source: arXiv AI

Tags: BridgeDock, AlphaFold 3, molecular docking, diffusion bridge, co-folding, NeurIPS 2026

BridgeDock, accepted at NeurIPS 2026, adapts pretrained co-folding models like AlphaFold 3 to flexible docking by modeling the apo-to-holo transition as a diffusion bridge, beating co-folding baselines even with few denoising steps.

Details

Co-folding models such as AlphaFold 3 predict protein-ligand complexes accurately, but they start from sequence-derived inputs and do not use apo structures as priors. That widens the search space and weakens robustness in hard or few-step settings. BridgeDock treats the move from apo receptor to holo complex as a diffusion bridge, so generation starts from the initial protein-ligand state rather than Gaussian noise. An alignment mechanism maps bridge states onto the pretrained model's original denoising schedule so existing denoisers can be reused. On standard flexible docking benchmarks, the authors report consistent gains over strong co-folding baselines, efficient inference, and good results with very few denoising steps.