Structure-Fused Immunogenic B-Epitope Predictor

Predicting which residues of an antigen a B-cell receptor will bind is the starting point for subunit vaccine and therapeutic antibody design, and protein language models have already moved the field forward: BepiPred-3.0 and SEMA 2.0 both built on evolutionary-scale embeddings and beat the previous generation of feature-engineered predictors. Precision, however, remains too low for practical prioritization, and language-model embeddings alone say little about whether a residue is actually exposed on the antigen surface.

This work, from Rakshanda Sajeed, Swatantra Pradhan, Rajgopal Srinivasan, and Sadhna Rana at TCS Research and posted to bioRxiv in August 2025, asks a narrow, well-posed question: given frozen ESM-2 embeddings, does adding explicit geometric structural features improve epitope prediction, and does how the two feature types are combined matter? The answer to both is yes. A deep transformation of the embedding followed by late concatenation with structural descriptors consistently beats naive early concatenation, and both beat an embeddings-only baseline — for conformational and linear epitopes alike.

A secondary finding concerns immunogenicity rather than antigenicity. Attention analysis of vanilla ESM-2 weights indicates the model already encodes some form of T-B reciprocity, the observed tendency of B-cell epitopes near T-cell epitopes to be more immunogenic, offering a way to filter high-scoring epitopes for likely immune dominance.

#Key Features

• Late fusion beats early fusion: Transforming the 1,280-dimensional embedding through four network layers before concatenating structural features outperforms simply stacking the two feature sets at the input, which ignores that they live in different spaces.
• Geometric structural descriptors: Contact number, protrusion index, and half-sphere exposure encode how buried or surface-protruding each residue is, computed from antibody-free antigen structures.
• Linear and conformational coverage: Separate models are trained for both epitope types, showing that structural features help linear epitope prediction too — a setting where most tools use sequence features only.
• Leakage-controlled evaluation: The independent test set is filtered at 25% sequence identity against the training sets of every competing predictor, not just its own.
• T-B reciprocity signal: Statistical analysis of ESM-2 attention links high-scoring B-cell epitopes to T-cell epitope proximity, a filter for immunogenic rather than merely antigenic candidates.

#Technical Details

Conformational training data derives from the IEDB-3D full-BCR assay set, filtered to antigen-antibody crystal structures better than 3 Å and antigens of at least 60 residues, then clustered at 50% identity with MMseqs2, yielding 901 complexes. Antigen sequences were folded with ESMFold to produce apo structures, avoiding leakage from antibody-bound conformations. Evaluation uses the SAbDab-derived Dset_anti set, reduced to 45 complexes after 25% identity filtering against all comparator training sets. The linear epitope track uses 438 IEDB antigens split 339/27/72. ESM-2 650M serves as a frozen feature extractor producing 1,280-dimensional per-residue vectors, classified by a four-layer network trained with nested five-fold cross-validation, dropout, and early stopping.

The best conformational model — late fusion of ESM-2 with contact number — reaches AUROC 0.813 and AUPRC 0.445 on the independent set, against 0.755 and 0.180 for DiscoTope-3.0. The embeddings-only baseline reaches AUROC 0.798. The best linear model, late fusion with contact number and protrusion index, reaches AUROC 0.78 versus 0.49 for BepiPred-3.0 and 0.56 for Epidope on the same data. Substituting the structure-aware SaProt embeddings for ESM-2 lowers AUROC to 0.655, and adding explicit structural features improves SaProt too. The T-B reciprocity test compares predicted scores for antigens where all B-cell epitopes are highly attended by T-cell epitopes against those where none are: p = 0.03 for conformational epitopes across 58 antigens, p = 0.007 for linear across 66. The preprint is released under CC BY-NC-ND; no code, weights, model card, or data card accompany it.

#Applications

Vaccine designers and antibody engineers can use the approach to rank candidate epitopes on a target antigen before committing to peptide synthesis or experimental mapping. Because the models require only sequence plus a predicted structure, they apply to antigens with no solved antibody complex. The T-B reciprocity filter is the more distinctive offering: among residues the model scores highly, it flags those most likely to provoke a strong antibody response rather than merely be recognized, which is the property that matters for immunogen selection.

#Impact

The contribution here is methodological rather than architectural. It establishes on a leakage-controlled benchmark that structural features still add real signal on top of large protein language model embeddings, that the fusion mechanism is itself a meaningful design choice, and that a structure-aware language model does not automatically subsume hand-computed geometric descriptors. The T-B reciprocity analysis matters for immunoinformatics because it suggests immunogenicity signal can be pulled from a general-purpose protein model without dedicated T-cell epitope prediction. It sits alongside other language-model epitope predictors such as the patch-centric conformational predictor; as a preprint with no released implementation, its uptake depends on independent reimplementation.