Carbohydrate-binding prediction for proteins, calling both whole-protein binding and the specific binding-site residues from sequence and structure.
No providers recorded yet. Browse all providers
How much of a proteome touches sugar? The question is hard to answer experimentally, because settling it would require testing every protein against a large glycan library. Current catalogs label only 1.5 to 5% of proteins as carbohydrate-binding, yet 50 to 70% of proteins are glycosylated — a mismatch suggesting that carbohydrate recognition is far more widespread than the annotated lectin repertoire implies. PiCAP and CAPSIF2, from the GrayLab at Johns Hopkins, attack that gap computationally.
The two models answer complementary questions about the same input structure. PiCAP (Protein interaction of CArbohydrate Predictor) makes a whole-protein call: does this protein bind a carbohydrate non-covalently? CAPSIF2 (CArbohydrate-Protein Site IdentiFier 2) localizes the interaction, labeling which residues form the binding site. They share a training corpus, an equivariant graph backbone, and a single inference entry point, and are distributed together as one package with fixed weights applicable to any new structure.
Preprinted in March 2025 and published in PNAS, the work's headline result comes from running the pair across three complete proteomes, where PiCAP calls 35 to 40% of proteins carbohydrate-binding — an order of magnitude above prior estimates.
Both networks are built from equivariant graph convolutional layers operating on C-beta coordinates with radial basis function edge features encoding inter-residue distance and orientation, and both take per-residue ESM2 embeddings as node inputs. PiCAP uses 12 residual layers at 128 embedding dimensions with a neighborhood that widens with depth — 10 nearest neighbors in the first three layers, then 20, 40, and 60 — followed by adaptive pooling to a fixed length of 150 and a convolutional-plus-dense classification head, totaling 1,798,895 parameters. CAPSIF2 keeps 12 residual layers at 128 dimensions with a fixed 16-neighbor window and a two-layer dense decoder emitting a per-residue binding likelihood, totaling 1,600,387 parameters.
The NoCAP dataset holds 30,849 unique structures: 9,608 with bound carbohydrates drawn from DIONYSUS, UniLectin, SAbDab, and the original CAPSIF set, against 21,242 curated non-binders. Sequences were clustered at 60% identity with MMseqs before an 80/5/15 split into 24,957 training, 1,479 validation, and 4,413 test structures. PiCAP reaches 89.6% balanced accuracy on the held-out set (96.3% true positive rate, 82.8% true negative rate), with perfect rejection of ribosomal proteins but weaker discrimination on designed proteins (60.8%) and antibodies (56.2%). CAPSIF2 scores 0.573 Dice and 0.574 MCC on the larger, more sequence-diverse DR benchmark, ahead of PeSTo-Carbs at 0.493 and CAPSIF:V at 0.226; on the smaller TS-90 set CAPSIF2 reaches 0.616 Dice while PeSTo-Carbs leads at 0.638. Applied to whole proteomes, PiCAP called 1,677 of 4,339 E. coli proteins (39%), 8,177 of 21,304 mouse proteins (38%), and 7,029 of 20,067 human proteins (35%) as carbohydrate binders, agreeing with 225 of 225 known mouse lectins and 229 of 230 human lectins from LectomeXplore.
Glycobiologists gain a way to nominate candidate carbohydrate-binding proteins before committing to glycan array or pull-down experiments, and to prioritize which residues to mutate when testing a proposed interaction. 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. Because the pipeline consumes predicted structures, it scales to organisms with no experimental structural coverage.
The substantive claim here is biological rather than architectural: if a third or more of a proteome engages carbohydrates, glycan recognition is a general property of protein surfaces rather than a specialty of a small lectin family. The authors are explicit that this first pass is noisy and requires experimental follow-up, and they identify concrete sources of error — the negative set is likely contaminated with undiscovered binders, crystallographic complexes over-represent high-affinity interactions relative to physiological avidity-driven binding, and performance degrades on antibodies and designed proteins whose hypervariable sequences are poorly represented in language model pretraining. Weights for both models are distributed 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.