GlycoWeave Predicts Bound Sugar Chains with Chemical Constraints

Glycans, branched chains of sugar units, coat cell surfaces and decorate many secreted proteins, and the proteins that recognise them take part in signalling, immune responses and host-pathogen contact ^[1], Main. Structures of those recognition events are scarce. The authors count 6,715 of 167,616 non-redundant protein sequences in the RCSB structure database (4.0%) as associated with a resolved protein-bound glycan, and the available examples lean toward short glycans, common sugars and recurrent linkages. Of 77,562 chemically permitted disaccharide linkage types they enumerate, only 312 appear in the database and 134 appear at least ten times ^[1], Main. General co-folding models represent carbohydrates as standard chemical components rather than through glycan-specific constraints, and earlier glycan-focused evaluations cited by the authors reported wrong anomeric and epimeric configurations and strained ring shapes ^[1], Main. The preprint reviewed here, GlycoWeave, is built around that mix of intricate chemistry and thin data.

The finding

GlycoWeave is a diffusion model. It takes a glycan sequence plus a protein or pocket structure, starts from a disordered glycan, and denoises it step by step toward a bound pose ^[1], Figs. 1G and 2D.

Sampling only the twists a sugar chain can make

Instead of moving every atom freely, the model updates four kinds of freedom: global translation, global rotation, torsions around the glycosidic bonds that join sugars, and the remaining non-glycosidic torsions. Bond geometry, ring closure, stereochemistry, anomeric configuration and modification state stay under constraint, so the output is the molecule that was requested ^[1], Main. Starting geometries come from a builder that assigns bond lengths, angles and dihedrals according to each sugar's local context, using distributions measured in experimentally resolved glycans. In the Man9GlcNAc2 glycan from PDB 8U48, for example, the 11 sugars fall into nine distinct local contexts ^[1], Fig. 2B. The vocabulary covers 139 stereochemically defined monosaccharide forms and four chemical modifications.

Computed structures added to scarce experimental ones

A tensor-product message-passing network exchanges information among protein residues, glycan residues and glycan atoms, so a contact can be read in light of the sugar's identity, its linkage and its place in the larger chain ^[1], Fig. 1F. Training combines 17,006 experimental complexes with about 6 million complexes refined by the GLYCAM06j carbohydrate force field and about 4 million optimised by a machine-learned potential the authors describe as quantum-mechanics-informed. One shared network feeds source-specific output heads, so the three data types can teach shared chemistry without being forced into one distribution ^[1], Fig. 2E.

The evidence

The test set is ProGlyDB-RecentSet, 698 protein-glycan complexes assembled by the authors. Comparators include the carbohydrate docking program Vina-Carb, HADDOCK3 and other docking tools, plus the co-folding models AlphaFold 3 ^[2] and Boltz-2 ^[3]. The docking tools and GlycoWeave receive a protein structure, whereas the co-folding models predict the protein and glycan together, a difference the authors mark in Fig. 3. The version 1 PDF refers to Methods and supplementary notes for the benchmark split and the quality thresholds but does not include them, so those definitions are not available to check.

Pose accuracy against docking and co-folding

At the acceptable-quality level, 79.4% of GlycoWeave's top-ranked poses succeeded, against 17.9% for AlphaFold 3 and 17.8% for Vina-Carb. Quality tiers combine ring placement, scaffold orientation and recovery of native contacts. Under the strictest tier, 71.9% of complexes had a high-quality GlycoWeave pose among 25 candidates, against 19.2% for AlphaFold 3 and 16.9% for Vina-Carb. The authors report sampling at 0.34 seconds per pose, roughly 17 times faster than AlphaFold 3 by their account ^[1], Main.

Chemical validity and recovered contacts

Every one of the 698 top-ranked GlycoWeave predictions passed the authors' checks for glycan identity, internal geometry and distance from the protein. Vina-Carb passed 99.7%. AlphaFold 3 passed 41.0% and Boltz-2 33.7%, after the authors standardised glycan topology so the outputs could be analysed ^[1], Fig. 3A. On the 497 complexes that every method could be evaluated on, GlycoWeave recovered 70% of native hydrogen bonds and 86% of native salt bridges, against 40% and 9% for AlphaFold 3 and 33% and 39% for Vina-Carb. Its mean fraction of native contacts was 0.75, against 0.51 for AlphaFold 3 and 0.50 for Vina-Carb ^[1], Fig. 3H.

The sampled glycosidic torsion angles were also compared with experimental structures. Across six representative linkages, predicted distributions reproduced the dominant populations. In a cross-comparison of 12 related linkage types, 9 predicted distributions were closest to the experimental distribution of their own linkage ^[1], Fig. 3E and 3F. For four linkage types absent from the experimental training set, the observed conformations fell within regions GlycoWeave sampled with high probability ^[1], Supplementary Figs. 31-34.

Pockets with low similarity to the training set

The continuous-score analyses use the 497-complex set that every method could be evaluated on. Splitting its pockets by structural similarity to the training set shows where the extra data matters. In the lowest-similarity subset, GlycoWeave's average continuous accuracy score (CARVE-Score) was 0.75. AlphaFold 3 and Boltz-2 fell from 0.45 and 0.39 in the highest-similarity subset to 0.13 and 0.09 in the lowest ^[1], Fig. 3I. The same score for GlycoWeave rose from 0.40 with experimental training data alone to 0.75 after adding GLYCAM-refined complexes, and to 0.78 with the machine-learned-potential set as well ^[1], Robustness across structural and glycan complexity; the ablation itself is shown in Supplementary Fig. 19. Across glycan sizes the average stayed above 0.8, and among oligosaccharides of three or more sugars it was 0.84 for linear and 0.78 for branched chains ^[1], Fig. 3.

Future directions

The 25-candidate result and the single-pose result answer different questions, and the version 1 main text does not describe how a user would choose among 25 poses without the experimental answer. A proposed check is to rank GlycoWeave's candidates with a confidence score and report how often the top-ranked pose is the high-quality one, using the same 698 complexes. The numbers of samples drawn from AlphaFold 3 and the docking tools for their 25-candidate rows are not given in the main text, so a matched-budget comparison would settle whether the 71.9% versus 19.2% gap holds when every method gets the same number of attempts.

The refinement ablation raises a second question. Adding about 6 million force-field-refined complexes moved the low-similarity score from 0.40 to 0.75, while adding about 4 million machine-learned-potential complexes moved it only to 0.78. Because the sets were added in sequence, the paper does not separate what each contributes. Training on the machine-learned-potential set alone, and on the force-field set alone, then scoring the same low-similarity pockets, would show whether the second set adds anything beyond the first. That is a proposal; the paper reports the combined result.

The linkage tests cover four absent linkage types, out of the large space of chemically permitted ones the authors enumerate. Newly deposited structures containing rare linkages are the natural next test, and the open measurement is whether the torsion-angle distributions GlycoWeave samples for those linkages fall on the experimental distributions in the same way the six representative linkages did.