Loading…
Viral protein annotation model predicting ten residue-level classes from sequence alone: topology, glycosylation, cleavage sites and disorder.
No providers recorded yet. Browse all providers
A vaccine designer holding a newly sequenced viral glycoprotein needs several answers at once: where the signal peptide stops, which stretch crosses the membrane, which asparagines carry glycans, where furin cuts the precursor into mature subunits, and which cysteines pair. Each question has its own predictor — DeepTMHMM, NetNGlyc, ProP, AIUPred, DeepCoil — each built for cellular proteins, each with its own dependencies and formats, all stitched into a pipeline first. ViralMap replaces that pipeline with one forward pass over the primary sequence, trained on eukaryotic viral proteins, whose rapid evolution puts them far from the host proteomes most annotation tools learned from.
The reformulation is to treat annotation as multi-label residue classification. For every position, the model emits ten independent probabilities spanning topology and localization (signal peptide, transmembrane, cytoplasmic, extracellular), post-translational modifications (N-glycosylation, furin cleavage, chain cleavage, disulfide-bond cysteine), and structural features (coiled coil, disordered region). Predicting them jointly keeps overlapping annotations coherent, since an extracellular domain and the glycosylation sites inside it come from one representation.
ViralMap comes from the ADAPT lab at Houston Methodist with collaborators at Rice University, funded in part by CEPI's Immunogen Design for Disease X program and published in Journal of Virology.
The base model pairs the pretrained 33-layer, 650M-parameter ESM-2 protein language model with a two-layer fully connected head (1,280-dimensional residue embeddings, ReLU, dropout 0.3) emitting ten sigmoid outputs per position; all but the final four transformer layers stay frozen. Training data came from UniProt taxonomy 10239 minus phage: 4,090,737 proteins filtered on annotation score and a 100–1,024 residue window, clustered with MMseqs2 at 60% identity, then reduced by representative selection to 8,238 proteins (3,131,029 residues) split into five viral-family-aware folds. Post-processing applies the four HMMs to the six region classes and F2-optimized thresholds — deliberately recall-favoring, since a missed site costs more in antigen design than a spurious one — to the four site classes.
ViralMap reaches residue-level PR-AUC ≥ 0.75 for 7 of the 10 classes, led by disordered regions (0.942) and chain cleavage (0.895). Against NetNGlyc on N-glycosylation it improves both precision (0.648 vs 0.270) and recall (0.913 vs 0.827); against AIUPred on disorder it trades recall (0.894 vs 0.993) for far higher precision (0.832 vs 0.388); against DeepTMHMM it gains transmembrane recall (0.868 vs 0.695) at slightly lower precision (0.884 vs 0.941). Coiled coils remain the hardest class for both ViralMap and DeepCoil.
The intended user is an antigen engineer deciding what to change in a viral surface protein: stabilizing an ectodomain, deleting a transmembrane anchor, removing an epitope-shielding glycan, mutating a protease site. ViralMap takes a FASTA file and returns a per-residue prediction table, a region-and-site summary, and an interactive visualization. Four held-out case studies show the profile in practice. On the SARS-CoV-2 spike it recovered all 22 reference glycosylation sites, the 30 disulfide-bonded cysteines, both heptad-repeat coiled coils, and the S1/S2 and S2′ cleavage sites; on the Lassa virus glycoprotein complex it correctly declined to call the retained stable signal peptide a cleaved one. HIV-1 gp160 and the Nipah virus fusion protein complete the set.
ViralMap argues for specializing an annotation stack rather than scaling it: a compact head over a general protein language model, trained on 8,238 carefully chosen viral proteins, matches or beats single-task tools across eight benchmarked classes and covers two more — chain cleavage and disulfide-bond state — that no comparable sequence-only tool addressed. The authors state its limits plainly. Sparse UniProt annotations make the reported precision a lower bound, and cross-fold variance is high for rare classes like furin cleavage. The model predicts which cysteines bond, not which pair with which, and it missed the spike protein's short disordered regions despite strong global disorder performance. The release is inference-only: checkpoints, HMM parameters, and the curated dataset are public, but no training code accompanies them.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.