GroupBind

GroupBind is a molecular docking framework that abandons the assumption every docking method before it made: that each protein-ligand pair should be posed in isolation. Ligands that bind the same pocket tend to adopt similar poses, because a pocket presents a fixed arrangement of hydrogen bond donors, acceptors, and hydrophobic contacts. GroupBind turns that biochemical regularity into an architectural prior, docking a group of co-binding ligands simultaneously so that each one informs the others. It was developed by researchers at the University of Illinois Urbana-Champaign, Tsinghua University, Helixon Research, and four other institutions, and published at ICLR 2025.

The practical framing matters. GroupBind is not restricted to cases where a chemist happens to have several ligands of interest queued up. At inference it retrieves companion ligands from a database of known binders for the target protein and uses them purely as auxiliary conditioning — their 3D binding structures are never supplied, only their 2D molecular graphs. When no such ligands exist for a target, the model gracefully degrades to standard single-ligand docking.

The group modules are not a standalone predictor. They are designed to slot into an existing pairwise-representation docking network, and the paper demonstrates this by integrating them into a diffusion-based docking model that treats pose prediction as generative sampling over ligand translation, rotation, and torsion angles.

#Key Features

• Group-conditioned docking: A query ligand is docked alongside up to four other ligands known to bind the same pocket, with the group assembled by UniProt identifier and capped by agglomerative clustering on Tanimoto similarity.
• No auxiliary structures required: Companion ligands contribute 2D graphs and generated conformers only. Their crystallographic poses are never used at inference, so the method applies to any target with a binding record.
• Cross-ligand message passing: Atoms in different ligands exchange messages over a graph built from maximum common substructure matching, with unmatched atoms connected to neighbors within 4 Å. Edges are SO(3)-equivariant and independent of diffusion time.
• Triangle attention across ligands: Borrowing the geometric consistency mechanism from AlphaFold 2 but applying it between ligands rather than within one, the module enforces that if an atom in one ligand contacts a given residue, the corresponding atom in a similar ligand should contact it too.
• Drop-in for pairwise docking networks: The interaction layer and triangle attention operate on pair embeddings, so they transfer to any docking architecture that maintains protein-ligand pair representations.

#Technical Details

Training and evaluation use PDBBind v2020 with the standard time-based split: roughly 17,000 complexes deposited in 2018 or earlier for training and validation, and 363 complexes from 2019 with no ligand overlap for testing. Grouping the training set yields 7,323 groups covering 17,649 complexes. Proteins within a group are aligned by the Kabsch algorithm over the longest common amino acid subsequence. The diffusion backbone follows the established formulation of docking as score matching on the product manifold of translation, rotation, and torsion, with SO(3) irreducible representations implemented in e3nn; an auxiliary distance-prediction loss supplies additional geometric supervision to the pair embeddings during training.

In the fully blind setting, where pockets come from P2Rank predictions rather than crystal structures, GroupBind improves on the diffusion baseline it extends across nearly every metric. Top-1 poses under 2 Å RMSD rise to 33.2% from 32.4%, and under 5 Å to 66.2% from 59.3%. The gain is largest on top-5: 48.8% of predictions fall under 2 Å and 80.1% under 5 Å, against 41.3% and 75.6% for the baseline. Supplying the native reference pocket instead pushes top-1 accuracy under 2 Å to 36.6%. Ablations confirm that both the cross-ligand message passing and the triangle attention contribute independently, and that combining structurally dissimilar rather than near-identical companion ligands helps most.

#Applications

The setting GroupBind fits is structure-based lead optimization, where a medicinal chemistry program has already accumulated a series of binders against a target and needs accurate poses for new analogues. Every previously characterized binder becomes usable evidence rather than a separate, discarded docking run. Virtual screening campaigns against well-studied targets benefit for the same reason, and the paper notes that where no ligands are recorded for a protein, binders of homologous proteins can serve as the group instead.

#Impact

GroupBind is the first deep learning docking method to encode the multi-ligand pose consistency prior end to end, converting an observation long exploited by physics-based docking heuristics into a learned architecture. Its clearest limitation is the flip side of its premise: performance improvements concentrate on targets with existing binding records, and novel or orphan pockets fall back to single-ligand behavior. Neither code nor trained weights have been released, and the paper states no license, so independent reproduction currently requires reimplementing the group modules on top of a diffusion docking backbone.