PhysDock

Deep-learning docking models have closed much of the accuracy gap with experiment, but they routinely produce poses that are geometrically close to the truth and physically wrong: inverted stereocenters, distorted bond geometry, steric clashes with the receptor, and hydrogen bonds or π-stacking contacts placed on the wrong atoms. Most also treat the protein as rigid, which forecloses the induced-fit rearrangements that let a single pocket accommodate chemically distinct scaffolds.

PhysDock, from ShanghaiTech University with collaborators at Lingang Laboratory, the Australian National University, and Cellverse, attacks both problems inside the diffusion process itself. It is an all-atom denoising diffusion model that generates protein and ligand coordinates jointly, so the pocket can make the subtle side-chain adjustments the authors call "precision-flexibility." Physical priors — pocket residues and the specific residues that form non-covalent contacts — enter as diffusion conditioning, and a two-phase physics guidance step projects each partially denoised ligand back into a physically realizable conformational space at every denoising step.

The result is a docking model that plays in the same league as the large co-folding systems — AlphaFold 3 and Chai-1 — while running several times faster, and that pulls clearly ahead of them when a binding pocket is known.

#Key Features

• Two-phase physics guidance: At high noise levels the model matches the denoised ligand against a pre-computed conformer library and replaces it; at low noise it runs a few steps of MMFF force-field relaxation, correcting chirality and steric errors before they propagate.
• Physical priors as conditioning: Binding-pocket residues and residues involved in hydrogen bonds, hydrophobic contacts, salt bridges, and π-interactions are fed to the conditioning module, with probabilistic dropout during training so the model works with partial or absent prior information.
• Iterative sampling with clustering-based ranking: Multiple rounds of conditioning and sampling, each with MSA resampling, feed a K-means clustering step over PAL-RMSD; the top-5 representatives outrank AlphaFold 3's learned confidence score as a selection criterion.
• Streamlined architecture for screening: A downscaled PairFormer trunk with no recycling and fewer sampling steps gives 2–4x faster inference than compiled AlphaFold 3 and 4–16x faster than compile-free Chai-1.

#Technical Details

PhysDock is built from a DiffusionConditioning module and a DiffusionModule and is trained on Plinder, a curated protein-ligand dataset. Conditioning inputs include the ligand molecular graph, protein sequence, backbone Cβ distogram, multiple sequence alignments, and the physical priors, encoded into atom-level and token-level representations. The DiffusionModule is a dual-scale diffusion transformer: a full-atom transformer without masking attends over geometrically neighboring atom pairs, with token-level blocks interleaved to capture residue-scale interactions.

Evaluation uses protein-aligned ligand RMSD (PAL-RMSD) across three redocking benchmarks — PoseBusters (428 complexes), DeepDockingDare (425 complexes, 38% membrane proteins), and the authors' PhiBench set (206 complexes deposited June–December 2024). Without pocket information, top-5 success at PAL-RMSD < 2 Å is 65.2%, 53.5%, and 61.2% respectively. With pocket information it reaches 95.3%, 89.1%, and 83.0%, against 90.2% for AlphaFold 3's non-open-source docking mode and 81.2% for Chai-1 on PoseBusters; under the stricter PB-valid and PAL-RMSD < 2 Å criterion PhysDock scores 90.9%, 87.0%, and 77.7%. On a 50-case time-split cross-docking set built from 25 proteins each resolved with two different ligands, it exceeds the next-best model by at least 15% under both criteria. Averaged across benchmarks it recovers 50.3% of hydrogen bonds and 62.8% of hydrophobic contacts identified by PLIP.

#Applications

The intended setting is structure-based drug design, where a pocket is usually known and the question is which of many candidate molecules binds and how. Two case studies exercise this directly: PhysDock reproduces the experimentally measured selectivity of 40 compounds between the closely related CB1 and CB2 cannabinoid receptors starting from cryo-EM structures, and a virtual-screening pipeline against NTRK3 kinase separates most high-affinity candidates, including compounds in preclinical and clinical development, from weak binders.

#Impact

Released under an MIT license with pretrained weights on Zenodo, PhysDock is one of the few open co-folding-class docking models whose pocket-conditioned accuracy exceeds the reported figures for closed docking modes of larger systems, and its physics-guidance mechanism is a transferable idea rather than a model-specific trick. The caveats are real: the released model handles multiple protein chains but only a single small molecule per prediction; the benchmark gains depend heavily on supplying binding-pocket priors, with the pocket-free numbers only matching rather than beating AlphaFold 3; and the work remains a preprint that has not completed peer review.