DockFormer

DockFormer predicts the bound structure and the binding affinity of a protein-small molecule complex in a single forward pass. It was developed by Ben Shor and Dina Schneidman-Duhovny in the Department of Computer Science at the Hebrew University of Jerusalem and first posted as a preprint in November 2024, with a revised version in April 2025.

The method sits deliberately between two families. Classical and deep-learning docking tools hold the receptor backbone rigid and lose accuracy when only an apo or predicted structure is available. Co-folding systems such as AlphaFold 3 and Boltz-1 build the complex from scratch and are correspondingly expensive. DockFormer adapts the AlphaFold 2 architecture but replaces multiple sequence alignments and templates with a distogram of an input receptor structure — either an experimental apo conformation or a predicted one. The receptor is never held rigid: the distogram acts only as a soft structural constraint, so the model still co-folds the complex while skipping the protein-folding problem it was handed a solution to.

Removing the MSA stack lets the trunk shrink from AlphaFold 2's 48 Evoformer layers to 8 Pairformer-style blocks. A second head predicts affinity from the same pair representation, making DockFormer a multi-task model that learns docking and affinity together — to the authors' knowledge, the first adaptation of the AlphaFold 2 architecture to both tasks at once.

#Key Features

• Fully flexible receptor: No part of the receptor is fixed during prediction. In 97% of PoseBusters cases the receptor moves less than 1 Å Cα RMSD, but where it does move substantially, the output is usually closer to the experimental structure than the input was.
• Joint structure and affinity prediction: An affinity module reads the ligand-atom-to-residue block of the pair representation and predicts affinity as a 30-bin classification over pKd values from 0 to 15. A second variant instead reads a dedicated classification token.
• MSA-free and fast: With no alignment search and only 8 trunk blocks, the median prediction time on PoseBusters is 2.67 seconds, which matters most in screening, where one receptor structure is amortized across many ligands.
• Three released checkpoints: DockFormer-PDBBind, DockFormer-Screen, and DockFormer-PLINDER are trained on different data splits and released together, so users can pick the one matching their task.
• Affinity without a correct pose: The model frequently predicts affinity accurately in cases where its structural prediction fails, indicating the pair representation captures interaction properties independently of pose accuracy.

#Technical Details

Inputs are embedded as per-token single representations (amino acid type for protein tokens; atom type, charge, chirality, and bond type for ligand atoms) plus a pair representation carrying intra-protein Cα-Cα distances binned into 15 bins from 3.25 Å to 20.75 Å and intra-ligand distances in 10 bins from 0.75 Å to 9.75 Å. Eight Pairformer-like blocks with three recycling iterations feed the AlphaFold 2 structure module; ligand frame rotations are discarded and only translations are used as atom positions. The implementation builds on OpenFold.

DockFormer-PDBBind was trained on 18,101 PDBBind 2020 complexes after removing 479 entries similar to CASF-2016. DockFormer-Screen fine-tunes it on 12,374 BindingDB affinity measurements spanning just 119 proteins, with structures predicted by Boltz-1. DockFormer-PLINDER was trained on 138,402 samples across 6,979 PLINDER clusters, then fine-tuned on the 36,594 affinity-labeled entries. On CASF-2016, the model reaches a Pearson correlation of 0.803 (interface head) or 0.817 (token head) and a mean Kendall's tau of 0.60, which DockFormer-Screen raises to 0.62. Structure success rate (ligand RMSD < 2 Å) is 69% on CASF-2016, 20% on PoseBusters, and 15% on PLINDER's strictly separated test set, where affinity Pearson reaches 0.753 against a best competing value of 0.569.

#Applications

The combination of second-scale inference, no MSA dependency, and a ranking-tuned checkpoint targets structure-based virtual screening: docking a large ligand library against one target, then ranking hits by predicted affinity. In the CASP16 Pharmaceutical Protein-Ligand Challenge, DockFormer-Screen matched top-scoring groups on Chymase and Autotaxin while running fully automatically, where most participants used templates and manual refinement.

#Impact

DockFormer is an argument that a much smaller model suffices when protein folding is factored out of the docking problem — it beats the AlphaFold 2-derived Umol on PoseBusters (20% versus 18%) with a fraction of the depth and no alignments. Its structural accuracy remains below diffusion-based co-folding models, which the authors attribute to treating side chains implicitly, and the work is a preprint that has not been peer reviewed. Code, training scripts, and all three checkpoints are Apache-2.0 licensed; the checkpoints themselves are distributed through the project's Hugging Face Space, which also runs a Gradio app that takes a protein sequence, an apo PDB file, and a ligand SMILES string and returns the docked complex plus predicted affinities.