CARTnet treats known epitopes as templates: internal AUPRC rises from 0.225 to 0.403, and it also leads on an external set

CARTnet extends B-cell epitope prediction from relying on sequence or structural features alone to accepting a few known epitope residues as prior input and using them to find additional epitope residues on the same antigen. BepiPred-3 predicts linear and conformational epitopes from sequence using protein language model representations, and DiscoTope-3 predicts conformational epitopes from inverse-folding structure representations; neither accepts user-supplied known epitopes as guidance. On the 61-target internal test set, the no-template CARTnet reached an AUPRC of 0.243, versus 0.220 for DiscoTope-3, 0.198 for BepiPred-3 and 0.140 for BepiPred-2; on the 52-target external test set CARTnet reached 0.267, versus 0.192, 0.215 and 0.111 respectively.

Template-guided prediction lets the model complete an epitope from existing experimental evidence rather than merely repeating the supplied residues. The template is a per-residue 0/1 vector where 1 marks a known epitope site, and it directly initializes the prediction-logit vector; without a template that vector is set to zero and the model falls back to de novo prediction. On the internal test set, AUPRC was 0.225 without a template and rose to 0.403 after supplying a single template epitope residue, with diminishing and eventually saturating gains as more templates were added; two simple rule-based ways of folding template information into DiscoTope-3 scores performed substantially worse than CARTnet.

Template information can transfer across viruses in the same genus, using epitopes characterized on a related virus to guide prediction on a target virus. Using Ross River virus epitopes (PDB 6VYV) as a template mapped by sequence alignment onto Chikungunya and Mayaro viruses, the transferred templates recovered 61.1% and 83.3% of their experimental epitopes respectively. After adding the transferred template, recall rose from 0.389 to 1.000 for Chikungunya virus (6W09) and from 0.833 to 1.000 for Mayaro virus (6W1C); Prec@Top-10 rose from 0.200 to 1.000 and from 0.200 to 0.900 respectively.

The evaluation covers a 61-target internal test set and a 52-target external test set, the latter drawn from SAbDab entries released after January 1, 2025 and stripped of sequences exceeding 20% identity to training data; an epitope is defined as an antigen residue with at least one heavy atom within 6 Å of an antibody residue. The authors state that broader evaluation across additional antigen families is needed to determine how template usefulness depends on sequence similarity, structural similarity and epitope conservation.

The authors report that the template gain diminishes as more templates are added and eventually saturates, which they attribute to the finite number of identifiable epitope residues; the perturbation experiment shows that spatial deviations up to roughly 18 Å still beat the no-template baseline, while larger deviations lose the template benefit and can fall below it. The authors also note that epitope annotation is inherently incomplete, because experimentally resolved antigen–antibody complexes cover only a subset of possible antibodies, antigen conformations and accessible binding sites.