Binder motif prediction from receptor structure alone, mapping 14 functional-group types across a protein surface as reusable interaction profiles.
No providers recorded yet. Browse all providers
Most binder design starts by copying. A designer takes the interaction motifs of a known ligand or natural partner — the hydrogen-bond donor reaching a particular backbone carbonyl, the aromatic ring stacking in a hydrophobic pocket — and builds a molecule that reproduces them. The method works, and it fails exactly where it is most needed: on a receptor with no characterised binder, or on a surface patch never targeted, there is nothing to copy.
MotifGen removes that dependency by predicting the motifs directly. Given only a receptor structure, it produces a profile across the protein surface indicating which chemical groups would be favourable at each point, expressed both as 14 specific functional-group types and as six coarser interaction property classes. The result is a map of what the surface wants rather than of what has previously bound to it, so it applies to novel targets and unexplored sites. Because the profiles are spatial fields over chemical categories, they are directly readable by a medicinal chemist, and they serve as pretrained embeddings that downstream models can consume with little task-specific data.
The network was developed by Hahnbeom Park's group at the Korea Institute of Science and Technology with Chaok Seok's group at Seoul National University, and appeared in the Journal of Cheminformatics in 2025.
MotifGen places a virtual node at each candidate motif point and passes the surrounding receptor atoms through ten SE(3)-transformer layers, after which linear layers convert the virtual node embedding into predictions. Node features combine one-hot amino acid type, Rosetta generic atom type, solvent accessibility, atomic charge, and distance to the virtual node; edges carry distance encodings, bond information, and displacement vectors. During training the grid points sit on exact ligand atom coordinates; at inference they are laid out on a 1.5 Å voxel grid restricted to solvent-accessible points.
Training data came from 68,918 PDB entries resolved better than 3.0 Å as of September 2020, clustered into 40,543 unique interfaces yielding 271,870 distinct binding motifs, with a motif retained when detaching the partner chain changed Rosetta energy by more than 2.0 kcal/mol. Training ran in two stages over 217 epochs on four NVIDIA A5000 GPUs in roughly three days, with adversarial negatives added at a 1:1 ratio in the second round. On peptide motif prediction, F1 sat around 0.3 for top-quartile predictions while at least half the true motifs were recovered for over 90% of targets — a recall-oriented profile suited to proposing candidates rather than pinpointing them.
Two downstream applications demonstrate reuse. MotifPepScore combines motif agreement with AlphaFold 2 predicted structures and their pLDDT and PAE confidence values to discriminate peptide binders, cross-validated over 1,783 binders and 17,830 non-binders across 18 receptor domains. MotifSite predicts small-molecule binding sites, trained on 748 complexes from PDBbind 2019 and reaching 82.7% success on the COACH420 benchmark under a 4.0 Å distance-to-centre criterion, against 80.0% for P2Rank.
The primary use is opening a design campaign on a receptor that offers no template: the motif profile nominates where to place chemistry before any scaffold exists, which suits protein-protein interaction surfaces, allosteric pockets, and newly characterised targets. MotifPepScore filters peptide candidates without a large target-specific training set, and MotifSite gives a structure-based first pass at where a small molecule could bind. Because the profiles are embeddings, a group with a small in-house assay set can fit a light head on top rather than train from scratch.
MotifGen's argument is that the motif — not the whole ligand and not the whole pocket — is the right intermediate representation for binder discovery, and that it can be predicted from structure alone. Framing the output as a reusable embedding rather than a final answer is what lets one pretrained network serve both peptide and small-molecule problems, and the small-molecule result shows the representation adds value against a strong geometric baseline. The evidence base is still early: two case studies, modest absolute F1 on motif prediction itself, and a repository that ships weights but carries no licence file, leaving reuse terms unsettled.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.