Loading…
Residue-level carbohydrate binding-site identification on protein structures, using equivariant graph convolutions over ESM2 per-residue embeddings.
No providers recorded yet. Browse all providers
CAPSIF2 (CArbohydrate-Protein Site IdentiFier 2) labels which residues of a protein structure form a carbohydrate binding site. It comes from the GrayLab at Johns Hopkins, was preprinted in March 2025, and was published in PNAS alongside PiCAP, the whole-protein classifier that decides whether a protein binds carbohydrate at all. The two share a training corpus, an equivariant graph backbone, and a single inference entry point, and ship as one package with fixed weights applicable to any new structure.
The division of labor matters because the two questions have different answers and different error modes. A protein-level call is enough to prioritize a proteome; it is not enough to design an experiment. Localizing the interaction to specific residues is what turns a prediction into a mutagenesis plan or a docking restraint, and it is the harder of the two problems — carbohydrate binding sites are typically shallow, solvent-exposed, and lack the deep pockets that make small-molecule sites easy to find geometrically.
As the numeral suggests, this is a successor: it replaces CAPSIF:V, an earlier model from the same lab, and more than doubles its accuracy on the harder of the two benchmarks used here.
The network uses equivariant graph convolutional layers over C-beta coordinates, with radial basis function edge features encoding inter-residue distance and orientation, and takes per-residue ESM2 embeddings as node inputs. Twelve residual layers at 128 embedding dimensions with a fixed 16-neighbor window feed a two-layer dense decoder emitting a per-residue binding likelihood, totaling 1,600,387 parameters.
Training used the NoCAP dataset — 30,849 unique structures, 9,608 of them with bound carbohydrates drawn from DIONYSUS, UniLectin, SAbDab, and the original CAPSIF set — with sequences clustered at 60% identity by MMseqs before an 80/5/15 split. On the larger, more sequence-diverse DR benchmark the model scores 0.573 Dice and 0.574 MCC, ahead of PeSTo-Carbs at 0.493 and CAPSIF:V at 0.226. The ordering reverses on the smaller TS-90 set, where it reaches 0.616 Dice while PeSTo-Carbs leads at 0.638 — a gap the authors attribute to that benchmark's narrower sequence diversity.
The residue-level output is directly usable as a restraint or hotspot definition for protein-carbohydrate docking, complementing structure-side tools such as GlycanGT that model the glycan itself. For experimentalists it prioritizes which residues to mutate when testing a proposed interaction, converting a predicted binder into a specific, falsifiable hypothesis. In a screening workflow it runs second, mapping sites on the proteins PiCAP has already flagged, and a single command runs both models over a directory of structures.
Site-level carbohydrate prediction has lagged behind small-molecule and nucleic-acid site prediction, largely because the sites are shallow and the training data thin, and the jump from 0.226 to 0.573 Dice over the lab's own predecessor is the clearest evidence here that language model embeddings plus equivariant geometry close part of that gap. The split benchmark result is worth reading honestly: leading on the diverse set while trailing on the narrow one suggests the model generalizes better than it specializes. The authors also flag degraded performance on antibodies and designed proteins, whose hypervariable sequences are poorly represented in language model pretraining. Weights download openly from the lab's server and the code is MIT-licensed, though the convenience Docker image bundles PyRosetta, which requires a commercial license for non-academic use.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.