CWFBind

CWFBind is an end-to-end blind docking model that predicts where a small-molecule ligand binds on a protein and what conformation it adopts there, without being told the pocket in advance. It was developed by Liyan Jia and Chuan-Xian Ren at Sun Yat-sen University with Hong Yan at City University of Hong Kong, and released as a preprint in August 2025.

Most deep-learning docking methods encode proteins and ligands as sequences or graphs and then extract features with language models or graph networks. That representation carries chemistry and topology but discards much of the local geometry of a binding cavity — how sharply the surface curves, where it closes into a pocket. CWFBind's central idea is to make that geometry explicit: it computes local curvature descriptors based on Ollivier-Ricci curvature, a discrete notion of curvature defined on graph edges through optimal transport, and adds them to the node features of both the protein and the ligand graph.

The design targets the accuracy-versus-speed tradeoff that separates the two families of learned docking methods. Generative samplers explore many candidate poses and rank them with a confidence model, which is accurate but slow; direct regression methods predict coordinates in a single pass, which is fast but usually less accurate. CWFBind is a regression method that recovers much of the sampling methods' accuracy while staying in the sub-second regime. Unlike protein-protein docking models such as DFMDock, it addresses small-molecule ligands.

#Key Features

• Curvature-aware featurization: Local curvature descriptors derived from Ollivier-Ricci curvature enrich the geometric representation of protein and ligand graphs, complementing chemical, sequence, and structural features.
• Degree-aware message passing: The equivariant layers weight messages from spatially adjacent atoms by node degree, suppressing noise from irrelevant contacts and sharpening the distinction between strong and weak interactions.
• Dynamic pocket radius: An MLP predicts a ligand-specific pocket radius rather than using a fixed cutoff, and a class-balanced focal loss addresses the severe imbalance between pocket and non-pocket residues.
• Single-pocket prediction: The model commits to one high-confidence pocket per protein instead of enumerating candidates, which keeps inference cheap and the prediction interpretable.
• Sub-second inference: A single forward pass takes 0.09 seconds per complex, with no diffusion sampling and no external pocket-detection module.

#Technical Details

The encoder combines the local curvature features with evolutionary sequence embeddings from ESM-2 and chemical and topological descriptors from TorchDrug. A pocket predictor built from one FABind layer plus a residue classifier selects the binding region and its radius; the docking module then applies four FABind layers followed by eight coordinate-refinement iterations. Training uses AdamW with a learning rate of 5e-5 and a combined pocket and docking loss.

Training and evaluation use PDBbind v2020 (19,443 protein-ligand complexes) under the same splits and preprocessing as TankBind and FABind. On the blind flexible self-docking benchmark, CWFBind places 38.3% of predictions within 2 Å ligand RMSD and 71.9% within 5 Å, with 66.0% and 85.8% of centroid distances under the same thresholds; mean ligand RMSD and centroid distance improve 8.5% and 27.0% over FABind. On a held-out subset of 144 proteins absent from training and validation, it reaches 63.9% under 5 Å ligand RMSD. Ablation shows the curvature features carry the most weight: removing them drops sub-2 Å accuracy from 38.3% to 30.0%. Inference is 71× faster than FABind+ with 40 sampled conformations and 690× faster than DiffDock, against roughly 146 seconds or more for classical tools such as GLIDE, Vina, SMINA, and GNINA.

#Applications

Blind docking at sub-second cost is aimed at virtual screening, where a pose and a pocket must be produced for hundreds of thousands of candidate ligands against targets whose binding sites may not be annotated. Structural biologists and medicinal chemists can also use it for hypothesis generation on individual targets — locating a plausible cavity and the key residues lining it before committing to more expensive physics-based refinement or experiment.

#Impact

CWFBind is a concrete demonstration that discrete curvature, long used as a descriptor in topological analyses of protein-ligand affinity, transfers usefully into end-to-end docking networks as an inexpensive geometric prior. Its results remain below sampling methods such as FABind+ on the strict sub-2 Å criterion, a gap the authors attribute to FABind+ optimizing a confidence model toward exactly that threshold, and the work is a preprint that has not yet been peer reviewed. No code or trained weights accompany the paper, so the reported numbers cannot currently be reproduced or the model applied directly by other groups.