BConformeR: A Conformer Based on Mutual Sampling for Unified Prediction of Continuous and Discontinuous Antibody Binding Sites

TL;DR

BConformeR tackles the challenge of predicting both linear and conformational B-cell epitopes by integrating local CNN-based features with global Transformer dependencies in a Conformer framework. It introduces feature coupling units and a calibrated per-residue score computed via GAMs, enabling antigen-aware, residue-level predictions. The model, trained separately on AlphaFold-predicted and experimentally determined structures, outperforms state-of-the-art baselines across MCC, ROC-AUC, PR-AUC, and F1, particularly for conformational epitopes. This hybrid architecture advances computational epitope prediction with practical implications for vaccine design, diagnostics, and antibody engineering.

Abstract

Accurate prediction of antibody-binding sites (epitopes) on antigens is crucial for vaccine design, immunodiagnostics, therapeutic antibody development, antibody engineering, research into autoimmune and allergic diseases, and advancing our understanding of immune responses. Despite in silico methods that have been proposed to predict both linear (continuous) and conformational (discontinuous) epitopes, they consistently underperform in predicting conformational epitopes. In this work, we propose Conformer-based models trained separately on AlphaFold-predicted structures and experimentally determined structures, leveraging convolutional neural networks (CNNs) to extract local features and Transformers to capture long-range dependencies within antigen sequences. Ablation studies demonstrate that CNN enhances the prediction of linear epitopes, and the Transformer module improves the prediction of conformational epitopes. Experimental results show that our model outperforms existing baselines in terms of MCC, ROC-AUC, PR-AUC, and F1 scores on both linear and conformational epitopes.

Figures (5)

• Figure 1: Overview of epitope classification and BConformeR prediction. (A) Epitopes, i.e., Antibody-binding sites, include conformational fragments (n=2,632) and linear regions (n=2,388). Data were derived from 1,432 antigen records. (B) Epitope prediction workflow. Antigen inputs are processed to predict linear and conformational epitopes, which enables applications in vaccine design and diagnostic development.
• Figure 2: BConformeR overview. (A) Preparation pipeline. (Right) Preparation of Alphafold-predicted structures and epitope extraction. (Left) Antigen embeddings. (B) Architecture overview. Stem module applies convolution and max-pooling to the ESM Cambrian (ESMC) embeddings; The Conv block and Trans block are based on ResNet bottleneck and Vision Transformer modules respectively; FCU-Up and FCU-Down connect the Conv and Trans blocks; The classifier linearly fuses features from both branches.
• Figure 3: Improved performance of BConformeR. (A–B) GAMs for estimating $\mu$ and $\text{std}$. (C–D) Comparison between raw and calibrated scores on the SARS-CoV-2 test set. (E–F) ROC-AUC and PR-AUC; BConformeR trained on AlphaFold-predicted structures and evaluated on the blind test set of 24 antigens. (G) PR-AUC; BConformeR trained on experimentally solved structures and evaluated on the SARS-CoV-2 test set.
• Figure 4: Structural visualization of epitopes predicted by BConformeR vs. actual epitopes. (A-C) Comparison of predicted and actual epitopes of 7KDB. Prediction is fully covered by the ground truth. (D-F) Comparison of predicted and actual epitopes of 7DFB. Prediction partially overlaps with the ground truth. (G-I) Comparison of predicted and actual epitopes of 7C88. Real epitopes are fully covered by the prediction.
• Figure 5: Linear regions and discontinuous fragments. (Left) The blue structure represents antigen 7c88, and the light green and light orange correspond to part of the antibody light chain and heavy chain, respectively; Linear regions are marked with red rectangles, and discontinuous fragments with green. (Right) Residues highlighted with numbers indicate epitopes.