BConformeR

BConformeR predicts which residues on an antigen an antibody will bind. The task splits awkwardly in two: linear epitopes are contiguous stretches of sequence, while conformational epitopes are residues scattered along the chain that fold into spatial proximity. Predictors built for one shape tend to fail on the other, and the field's persistent weakness has been conformational epitopes. Existing tools such as BepiPred-3.0, SEMA-1D 2.0, SEPPA-3.0, and DiscoTope-3.0 each address part of the problem, but none handles both regimes well from a single backbone.

The model, from groups at the University of Hong Kong, the Hong Kong Generative AI Research and Development Center at HKUST, and BayVax Biotech, answers with a dual-branch Conformer: convolutional blocks capture the local residue motifs that define linear epitopes, Transformer blocks capture the long-range couplings that define discontinuous ones, and feature coupling units exchange information between the two branches at every stage rather than merging them only at the end. Both branches read per-residue embeddings from ESM Cambrian, so the antigen representation is inherited from a pretrained protein language model rather than learned from the small epitope corpora available.

#Key Features

• Unified linear and conformational prediction: One trained classifier scores both epitope types, with an ablation confirming the division of labor — convolutions carry linear performance and self-attention carries discontinuous performance.
• Bidirectional feature coupling: FCU-Down projects convolutional features into the Transformer space and FCU-Up maps Transformer features back, so local and global representations inform each other throughout the network instead of at a single fusion point.
• Language model input features: 1,152-dimensional per-residue ESM-C embeddings feed a shared convolutional stem, removing the need to learn amino acid representations from the limited supply of solved antigen-antibody complexes.
• Length-calibrated scores: Raw residue scores are z-normalized per antigen using two generalized additive models — one predicting mean score from antigen length, the other predicting standard deviation from mean score — which cuts false positives on long antigens.
• Separate solved and predicted-structure models: Two checkpoints are trained, one on experimentally determined complexes and one on AlphaFold-modeled complexes, matching the structural evidence a user actually has.

#Technical Details

BConformeR is a 12-stage dual-branch network of about 289M parameters. The CNN branch uses ResNet-style bottlenecks widening from 144 to 576 channels; the Transformer branch uses ViT-style blocks with 24-head self-attention at width 1,536 and no positional embeddings, since the convolutional branch already encodes position, and no CLS token, since the task is residue-level. Epitope residues are defined by a heavy-atom minimum distance below 4 Å; linear stretches with at least three contacting residues and gaps of up to three residues are labeled regions, and the rest are treated as discontinuous fragments. The solved structure model trains on 3,674 non-redundant antigen-antibody complexes from AACDB; the predicted-structure model starts from 1,300 complexes in AbEpiTope-1.0, models each 30 times with AlphaFold-2.3 through ColabFold, keeps the top-ranked prediction, discards pLDDT below 0.82, and ends with 1,180 complexes split 1,080 for training and 100 for validation. On a blind test set of 24 antigens drawn from SEMA-1D 2.0, BConformeR reaches ROC-AUC 0.777 and PR-AUC 0.313 against 0.755 and 0.243 for DiscoTope-3.0, with linear F1 0.349, discontinuous F1 0.262, and MCC 0.296 versus 0.303, 0.231, and 0.227. On a SARS-CoV-2 test set built from CoV-AbDab, the solved-structure model reaches PR-AUC 0.268 against DiscoTope-3.0's 0.233, with linear F1 0.405, discontinuous F1 0.135, and MCC 0.350. Ablations on the validation set show a pure ResNet1D-152 leading on linear F1 (0.462) but collapsing on discontinuous F1 (0.272), while the full model holds 0.458 and 0.330.

#Applications

Epitope prediction sits upstream of vaccine antigen design, therapeutic antibody discovery, immunodiagnostic assay development, and antibody engineering, all of which otherwise depend on crystallography or cryo-EM to map an interface. BConformeR takes an antigen sequence with a solved or predicted structure and returns per-residue scores, so the practical use is prioritizing which surface regions to target with an immunogen or to probe experimentally. Users working from AlphaFold models rather than crystal structures have a checkpoint trained for exactly that input distribution, which matters because most antigens of interest lack a solved antibody complex.

#Impact

BConformeR's contribution is architectural: it demonstrates that the convolution-plus- attention hybrid familiar from speech recognition transfers to epitope prediction, and that coupling the branches beats either alone on the metric each is naturally good at. The reported gains over DiscoTope-3.0, the strongest prior baseline, are consistent but modest in absolute terms — discontinuous F1 remains below 0.3 on the blind set, underscoring how far conformational epitope prediction remains from solved. The work is a preprint. The code link given in the paper points at an anonymized review repository that has since expired, though a public implementation carrying the model and training code has since been published separately; it declares no license and ships no trained weights, so the reported results cannot be reproduced from a released checkpoint.