CARTnet Extends B-cell Epitope Prediction Beyond Single-Feature Models
Predicting B-cell epitopes remains one of the most stubborn problems in antigen design and antibody-oriented immunology. The biological target is deceptively simple: identify which amino acid residues on an antigen are likely to be recognized by antibodies. In practice, the signal is thin, experimentally annotated data are limited, and the structural rules of antibody–antigen recognition are only partly mapped. CARTnet enters that space as a contact-aware deep learning framework built to combine multimodal features with template-derived epitope information, rather than relying on any single representation of the antigen surface.
The model was evaluated against established prediction tools on both internal and external test datasets, and it outperformed the comparison methods across most metrics. On the internal test set, CARTnet reached an AUPRC of 0.243, compared with 0.220 for DiscoTope-3, 0.140 for BepiPred-2, and 0.198 for BepiPred-3. On the external test set, CARTnet again led the field with an AUPRC of 0.267, versus 0.192, 0.111, and 0.215 for those same comparators. In a domain where small performance gains can determine whether a candidate epitope is carried forward or discarded, those differences are not cosmetic.
Why Contact Awareness Changes the Prediction Problem
Antibody recognition is not governed only by linear sequence signals. Surface shape, residue proximity, and local structural context all influence whether an epitope is accessible and immunologically relevant. CARTnet’s contact-aware attention mechanism is designed around that fact. Instead of treating residues as isolated tokens, the framework learns from interactions among neighboring positions and structural contacts, a strategy that better reflects how antibodies physically engage antigen surfaces.
This matters because many false positives in epitope prediction arise when a residue looks plausible in sequence space but sits in the wrong structural environment. A solvent-exposed patch can still fail as an epitope if the broader contact map makes it inaccessible or conformationally unsuitable. By integrating multimodal features, CARTnet is trying to recover the layered biology of the antigen surface rather than compressing it into a single score.
The model’s performance on an external test set is especially relevant. A method that performs well only inside the dataset that trained it may simply be memorizing assay-specific patterns or annotation artifacts. CARTnet’s stronger external AUPRC suggests the framework is capturing signals that survive dataset shift, which is a central requirement for real-world translational use in vaccine design, therapeutic antibody discovery, and antigen engineering.
Template-Guided Prediction Raises the Signal When Prior Knowledge Exists
CARTnet also supports template-guided prediction, where limited prior epitope information is used to refine the search for additional epitope residues. In the internal test set, introducing a single template residue increased the AUPRC of the template-based variant from 0.225 to 0.403. That is a substantial jump for a problem in which usable prior information is often sparse.
A single residue can change the geometry of the entire inference problem. In antibody biology, partial epitope knowledge can anchor the model to a local structural neighborhood that is more likely to be immunologically active, allowing the network to expand from one known residue to adjacent residues that share exposure, proximity, or contact patterns. This mirrors how experimental immunologists often work in practice: one validated site is not the finish line, but it can guide a much sharper search for the surrounding epitope patch.
CARTnet’s template-guided architecture is therefore not just a convenience feature. It creates a route for combining weak prior evidence with machine inference, which is often the most realistic configuration in infectious disease research, where only fragments of epitope annotation may exist for a pathogen or viral lineage.
Cross-Genus Transfer Suggests a Broader Immunologic Signal
The study also showed that epitope information from closely related viruses can be transferred across the genus Alphavirus to improve prediction accuracy. This is an important biological observation because it implies that CARTnet is not simply learning organism-specific quirks. It is able to exploit conserved immunologic and structural patterns across related viral antigens, where homologous regions may present similar surface chemistry and spatial arrangements.
Cross-genus transfer is one of the more attractive features of a contact-aware model. Viral families often preserve enough structural logic for antibody recognition to remain partly predictable, even when sequence identity has drifted. A model that can use that relatedness may be particularly useful in outbreak settings, where novel viral proteins must be evaluated before exhaustive wet-lab mapping is available. The practical benefit is straightforward: a better first-pass ranking of candidate epitopes can prioritize experiments and reduce wasted laboratory cycles.
What the Performance Numbers Mean for Antigen Design
The AUPRC values reported here are the clearest quantitative measure of CARTnet’s advantage. Precision-recall metrics are especially meaningful in epitope prediction because positive residues are typically rare relative to the full protein surface. In that setting, a higher AUPRC reflects better ranking of true epitopes above the large background of non-epitopic residues, which is usually more informative than raw accuracy.
CARTnet’s gains over DiscoTope-3 and the BepiPred series suggest that multimodal contact-aware learning is capturing patterns that older methods miss. Those older tools remain useful references, but their weaker performance on both test sets points to the limits of approaches that lean more heavily on conventional feature engineering or less context-rich inference. When the underlying biological question depends on three-dimensional exposure and local contact structure, a model that explicitly encodes those relationships has a clear advantage.
Even so, the problem remains constrained by the quality and breadth of experimental annotations. Sparse ground truth limits how far any model can generalize, and the field still depends on iterative cycles of computational prediction and laboratory validation. CARTnet improves the ranking problem, but it does not replace the need for empirical epitope mapping, binding assays, and structural confirmation.
A Preprint With Immediate Utility and Open Validation Needs
This work appears as a bioRxiv preprint, so the findings have not yet undergone peer review. The absence of peer review does not erase the signal in the data, but it does place the results in a provisional category. Independent validation, broader benchmarking, and testing on additional antigen classes will determine whether the reported gains persist outside the current datasets.
For now, the contribution is clear: CARTnet offers a contact-aware framework that improves de novo epitope prediction and becomes stronger when even minimal template information is available. Its strongest result may be conceptual rather than numerical. The model treats epitope recognition as a structural and contextual problem, which is much closer to antibody biology than residue-by-residue scoring alone. The next question is whether that advantage holds across larger, more heterogeneous antigen collections, where structural diversity and annotation noise are both higher than in the current evaluation sets.