Epitope-Only Strategy in Immuno-Design

• Epitope-only strategy is a computational paradigm that focuses on the epitope to reduce immune recognition problems to local, sequence-based features.
• It enables targeted binder redesign and generative modeling by conditioning on epitope sequences while keeping the broader molecular context fixed or indirectly represented.
• Its effectiveness varies by application, showing strong results in nanobody design and peptide selection, yet limitations in predicting conformational B-cell epitopes.

Epitope-only strategy denotes a family of computational and design paradigms in which the epitope is treated as the primary object of optimization, conditioning, prediction, or selection, while broader molecular context is fixed, omitted, or only indirectly represented. In recent literature, the term is used in several non-equivalent senses: local epitope-targeted optimization with a fixed binder scaffold, inference conditioned only on an epitope sequence, peptide-only vaccine or presentation screening, and, in a more cautionary sense, an “epitope-as-class” shortcut in which models encode epitope identity rather than interaction rules (Hu et al., 11 Jun 2026, Zhang et al., 9 Sep 2025, Weber et al., 2021, Wan et al., 16 Dec 2025). The common thread is a reduction of the immune recognition problem to epitope-centered variables, but the consequences of that reduction differ sharply across nanobody design, TCR modeling, B-cell epitope mapping, and antigen-presentation prediction.

1. Semantic scope and conceptual variants

In structure-guided binder design, epitope-only usually means that the desired binding patch is fixed by the user and all optimization pressure is concentrated on bringing a constrained scaffold into productive contact with that patch. In EasyNano, this is formalized by fixing the nanobody framework sequence, the antigen sequence and structure, and the framework pose, while optimizing only Chothia-defined CDR residues toward a user-specified epitope set $\mathcal{E}$ (Hu et al., 11 Jun 2026). In the discontinuous-fragment binder-design literature, the same phrase denotes a more aggressive target reduction: only the discontinuous surface residues surrounding the binding site are retained, and the full target domain is discarded under a “local-first” hypothesis for protein folding neural networks (Deng et al., 29 Sep 2025).

In generative immunoreceptor design, epitope-only refers to conditioning at inference time on the epitope and nothing else. LSMTCR defines this regime explicitly as generation from the amino-acid sequence of a peptide epitope, optionally with MHC or pMHC context encoded in the epitope string, with no example TCRs for that epitope and no template receptor at generation time; the modeled object is $p(\text{TCR}\mid\text{epitope})$ (Zhang et al., 9 Sep 2025). In sequence-only epitope generation, epiGPTope uses only primary amino-acid sequences of linear epitopes, bypassing geometric frameworks, antigen structures, antibody structures, MHC context, and handcrafted features (Manrique et al., 3 Sep 2025).

In predictive immunology, the phrase can identify either a modeling regime or a failure mode. TITAN distinguishes an explicit “epitope-as-class / epitope-only strategy,” in which each epitope is treated as an independent label and unseen epitopes are in principle unresolvable, from generic bimodal models that encode both TCR and epitope (Weber et al., 2021). By contrast, in calibrated TCR–pMHC prediction under epitope shift, the relevant epitope-only concern is not omission of the receptor but reliable behavior when the main source of distribution shift is the peptide itself (Bekov et al., 14 Apr 2026).

In vaccine and antigen-presentation pipelines, epitope-only typically means operating on short peptides rather than full antigens. This includes linear B-cell epitope selection for peptide-based vaccines, peptide-level MHC-II binding or eluted-ligand prediction, and active-learning systems that screen 9-mer epitopes against a fixed receptor without modeling the entire pathogen protein (Ghoshal et al., 2021, Wan et al., 16 Dec 2025, Brisebois et al., 27 Jun 2026). This suggests that “epitope-only strategy” is best understood not as one method, but as a recurring reduction principle whose validity depends on what information can safely be fixed or ignored.

2. Epitope-only as local structural optimization

EasyNano provides the clearest explicit formalization of an epitope-only strategy in binder redesign. The user specifies an epitope as a residue-index set $\mathcal{E}$, only the nanobody CDR residue logits $\mathbf{L}\in\mathbb{R}^{n_{\text{CDR}}\times 20}$ are optimized, and the nanobody framework sequence remains fixed as one-hot throughout optimization (Hu et al., 11 Jun 2026). The central targeting term is

$$\mathcal{L}_\text{epitope} = \frac{1}{|\mathcal{C}|}\sum_{i\in\mathcal{C}} \mathrm{ELU}\!\left(\min_{j\in\mathcal{E}}\bar d_{ij}-8.0\right),$$

where $\bar d_{ij}$ is the expected residue–residue distance under the differentiable ESMFold2-Fast distogram. This term penalizes CDR residues whose expected distance to every epitope residue exceeds 8 Å, but stops pushing once at least one epitope residue is within that range. The remainder of the objective stabilizes intra-chain packing, inter-chain contacts, global regularity, amino-acid usage, and, critically, a structure prior derived from full ESMFold2 to prevent pose drift. The fixed-framework, fixed-antigen, fixed-pose formulation is the essence of epitope-only in this setting: only the sequence-encoded micro-geometry of the CDR loops is allowed to move (Hu et al., 11 Jun 2026).

This formulation is not only conceptual but operationally effective. EasyNano runs in approximately 10–20 minutes on a high-end personal workstation, uses Adam with learning rate 0.05 for 60 steps, and depends strongly on a wild-type logit bias $\beta$, with $\beta\approx 2.0$ reported as optimal (Hu et al., 11 Jun 2026). Across six target-framework pairs, it improves ipTM by up to $+0.559$, from 0.143 to 0.702 for Ty1/RBD, achieves a 4.6-fold improvement from 0.117 to 0.538 in a de novo AQP4 case, and remains 5.7 sigma above the random mean for Ty1; Kabsch cross-validation against crystal structures indicates that framework pose is preserved while CDR–epitope geometry improves (Hu et al., 11 Jun 2026).

A related but more radical formulation appears in “Discontinuous Epitope Fragments as Sufficient Target Templates for Efficient Binder Design,” where the target is reduced to only the discontinuous surface residues surrounding the binding site (Deng et al., 29 Sep 2025). Residues are retained if they lie within a distance cutoff from known functional hot spots, and the remaining fragments are presented to AlphaFold2-Multimer as a disconnected but spatially coherent target patch. The paper argues that PFNNs behave in a “local-first” manner, prioritizing short-range interactions while displaying limited sensitivity to global foldability. Under that hypothesis, the epitope-only strategy improves in silico success rates by up to 80% and reduces average time per successful design by up to forty-fold, particularly on structurally large targets such as ClpP and ALS3 (Deng et al., 29 Sep 2025). This suggests that, for local interface design, the epitope can sometimes serve not merely as a target specification but as a sufficient structural template.

3. Epitope-only as the sole conditioning signal in generative models

LSMTCR generalizes epitope-only from local optimization to de novo sequence generation. Its inference-time objective is

$$p(\text{TCR}\mid\text{epitope}) = p(\text{CDR3}_\alpha,\text{CDR3}_\beta,V_\alpha,J_\alpha,V_\beta,J_\beta\mid\text{epitope}),$$

with the epitope sequence as the sole conditioning input to the CDR3 generators (Zhang et al., 9 Sep 2025). The architecture separates specificity from immunogenetic constraints: a diffusion-enhanced BERT encoder learns time-conditioned epitope representations; conditional GPT decoders generate CDR3$p(\text{TCR}\mid\text{epitope})$0 and transferred CDR3$p(\text{TCR}\mid\text{epitope})$1 under cross-modal attention to the epitope embedding; a gene-aware Transformer predicts V/J usage and assembles full-length TRA and TRB chains. This factorization allows generation of full-length, gene-contextualized $p(\text{TCR}\mid\text{epitope})$2 receptors from epitope input alone, without grafting or mutating a template TCR (Zhang et al., 9 Sep 2025).

The reported behavior is explicitly epitope-conditioned rather than merely repertoire-like. Across GLIPH, TEP, MIRA, McPAS, and a curated dataset, LSMTCR achieves higher predicted binding than baselines on most datasets, more faithfully recovers positional and length grammars, and yields temperature-tunable diversity (Zhang et al., 9 Sep 2025). For full-length $p(\text{TCR}\mid\text{epitope})$3 assembly, paired co-modeling with epitope attains higher pTM/ipTM than single-chain settings, and de novo full-length sequences preserve k-mer spectra while maintaining low edit distances to references (Zhang et al., 9 Sep 2025). The approach therefore treats epitope-only not as a shortcut around receptor modeling, but as a generative conditioning regime.

An even stricter sequence-only variant appears in epiGPTope, which fine-tunes ProtGPT2 on 504,611 unique human linear epitope sequences and then samples directly in peptide sequence space (Manrique et al., 3 Sep 2025). The model generates 192,222 distinct synthetic epitope sequences and is validated not by structure but by epitope-like statistics: generated sequences reproduce the dominant 7–9 residue length peak, terminal positional biases, low cysteine frequency, and near-zero pairwise mutual information when sample size is sufficient (Manrique et al., 3 Sep 2025). In the same framework, downstream classifiers operate only on peptide embeddings to distinguish bacterial or viral origin, with fine-tuned ProtBERT and ProtGPT2 classifiers reaching, for example, F1 $p(\text{TCR}\mid\text{epitope})$4 and PR AUC $p(\text{TCR}\mid\text{epitope})$5 on bacterial MHC, and F1 $p(\text{TCR}\mid\text{epitope})$6 and PR AUC $p(\text{TCR}\mid\text{epitope})$7 on viral MHC (Manrique et al., 3 Sep 2025). Here epitope-only means that the object being designed is the epitope itself rather than a receptor against it.

These two uses differ in directionality. LSMTCR uses the epitope as a generative condition for receptors, whereas epiGPTope treats the epitope as the direct generative object. The common feature is that all non-epitope context is deferred to downstream filtering, scoring, or experimental validation.

4. Epitope-only shortcuts, generalization, and reliability in TCR prediction

TITAN shows that epitope-only can also arise as an unintended shortcut. The formal task is a binary interaction map $p(\text{TCR}\mid\text{epitope})$8, but the paper distinguishes generic bimodal models from categorical “epitope-as-class / epitope-only” approaches that simply learn TCR patterns associated with epitope IDs (Weber et al., 2021). TITAN itself encodes TCRs and epitopes jointly, either at the amino-acid level or with SMILES-based atom-level epitope representations, yet its attention analysis shows that sparse epitope coverage encourages an implicit class-like treatment of epitopes. In the AA CDR3 setting, the residues with attention $p(\text{TCR}\mid\text{epitope})$9 still uniquely identify 185 of 192 epitopes, and epitope attention is almost constant across contexts, effectively turning the epitope branch into a label hash (Weber et al., 2021). That shortcut is useful on seen epitopes but harmful under strict unseen-epitope evaluation: K-NN achieves ROC-AUC $\mathcal{E}$0 on the TCR split but only $\mathcal{E}$1 on the strict split, while TITAN’s best strict-split ROC-AUC remains only $\mathcal{E}$2 despite richer epitope modeling (Weber et al., 2021).

Calibrated abstention work reframes the same problem as reliability under epitope shift rather than model architecture alone. A dual-encoder TCR–pMHC model with temperature scaling and conformal abstention reaches AUROC 0.813 and ECE 0.043 on an epitope-held-out protocol, reducing ECE by 69.7% relative to an uncalibrated baseline (Bekov et al., 14 Apr 2026). At 80% coverage, the selective model reduces error rate from 18.7% to 10.9% (Bekov et al., 14 Apr 2026). In this setting, an epitope-focused strategy is not to ignore the receptor, but to require calibrated confidence or abstention whenever the peptide lies outside the support of the training epitopes.

TCR-SRIM sharpens the distinction between epitope-centered and pairwise models. It is explicitly not epitope-only: it models CDR3$\mathcal{E}$3, CDR3$\mathcal{E}$4, and peptide jointly using PLM embeddings, cross-attention fusion, and interpretable contact prototypes regularized by residue-level structural contacts (Li et al., 29 Jun 2026). On unseen epitopes, it achieves state-of-the-art predictive performance; with real structures for regularization, the top-100 unseen-epitope ROC-AUC reaches 0.953 with an ESM2-8M backbone, and interpretation quality is quantified by BRHR values such as 0.855 for peptide$\mathcal{E}$5CDR3$\mathcal{E}$6 and 0.996 for CDR3$\mathcal{E}$7$\mathcal{E}$8peptide with ProteinBERT (Li et al., 29 Jun 2026). The same study shows that AlphaFold3, TCRModel2, and tFold-TCR structures yield competitive performance but less accurate interaction patterns and reduced binding-site diversity than experimentally resolved complexes (Li et al., 29 Jun 2026). This suggests that epitope-only abstractions lose precisely the chain-specific residue–residue geometry that drives both interpretability and unseen-epitope generalization.

5. Antigen-side epitope-only mapping and the shift toward partner awareness

Antigen-side epitope prediction is the most literal epitope-only regime: given only the antigen, predict which residues are likely to belong to an antibody-binding site. Neural message passing for joint paratope–epitope prediction argues that paratopes and epitopes require asymmetric treatment and, in its antigen-only baselines, already shows a meaningful gap between antibody-agnostic and antibody-aware models: Epi-GCN, which uses only the antigen graph, reaches AUC-ROC $\mathcal{E}$9 and AUC-PR $\mathbf{L}\in\mathbb{R}^{n_{\text{CDR}}\times 20}$0, whereas joint Epi-EPMP reaches $\mathbf{L}\in\mathbb{R}^{n_{\text{CDR}}\times 20}$1 and $\mathbf{L}\in\mathbb{R}^{n_{\text{CDR}}\times 20}$2 (Vecchio et al., 2021). Sequence CNNs do not help, and more expressive antigen-only MPNNs underperform the simpler GCN, reinforcing the idea that epitopes are structurally dispersed and context-dependent (Vecchio et al., 2021).

MIPE extends antigen-side modeling by combining sequence and structure modalities with multi-modal contrastive learning and interaction informativeness estimation. In single epitope prediction, MIPE reaches AUC 0.852 and AUPR 0.504, while MIPE with AlphaFold2 structures retains AUC 0.842 and AUPR 0.450 (Wang et al., 2024). The same model performs markedly better in joint paratope–epitope prediction than prior joint methods, indicating that antigen-only inference can benefit from training that has seen explicit antibody interaction patterns (Wang et al., 2024).

SurfBind moves the antigen-side perspective from residue graphs to the molecular surface itself. It is surface-centric and can operate in a partner-agnostic regime or in a binder-aware regime via cross-attention from antigen patches to antibody features (Wu et al., 22 Jun 2026). On SAbDab benchmarks, SurfBind achieves AUC-ROC 81.62, AUC-PR 30.57, balanced accuracy 76.98, and F1 42.95, outperforming sequence-based, backbone-based, and earlier surface-based baselines (Wu et al., 22 Jun 2026). Because labels, loss, and hierarchy are all defined on the antigen surface, SurfBind exemplifies an epitope-centric strategy that remains compatible with optional antibody conditioning.

EpiFormer, by contrast, shows how far performance can improve once epitopes are treated as inherently antibody-specific. Its interleaved cross-attention within GNN encoding layers yields AUC 0.924, AUPRC 0.493, F1 0.482, and MCC 0.464 on the AsEP epitope-ratio split, improving over the previous best method by over 40% in F1 score (Ahmed et al., 2 Jun 2026). The learned cross-attention gates favor antigen-to-antibody information flow, and the model’s preference for geometric over evolutionary features aligns with the finding that epitope residues are not evolutionarily conserved (Ahmed et al., 2 Jun 2026). Taken together, these studies suggest that antigen-only epitope prediction remains useful for coarse localization, but the strongest current results arise when epitopes are modeled as relational properties of antigen–antibody pairs rather than intrinsic properties of antigen residues alone.

6. Selection pipelines, vaccine uses, and the limits of epitope-only reasoning

Epitope-only strategies are especially prominent in peptide selection workflows. In linear B-cell epitope prediction for SARS-CoV-2, a Bayesian neural network with MC-DropWeights operates on engineered peptide features rather than raw sequence and improves confusion-matrix accuracy from 82% to 85%, while also producing aleatoric and epistemic uncertainty estimates that correlate with error (Ghoshal et al., 2021). The paper explicitly connects this to peptide-based or epitope-only vaccine design: select peptides with high predicted epitope probability and low aleatoric uncertainty, then use them as candidate vaccine components (Ghoshal et al., 2021).

BeeTLe provides a more modern sequence-only version of the same logic. It predicts linear B-cell epitopes versus non-epitopes and classifies epitopes by IgA, IgE, and IgM using a BiLSTM+Transformer hybrid with a BLOSUM62-derived amino-acid encoding and logit-adjusted focal objectives for imbalance (Yuan, 2023). On curated IEDB-derived data, BeeTLe reports epitope prediction AUC 85.80%, accuracy 77.29%, and Ig-type balanced accuracy 72.21% (Yuan, 2023). This is an archetypal epitope-only classifier: the peptide is the object of inference, and broader antigen context is absent.

A closely related screening paradigm appears in vaccine epitope selection for PRRS. Here the task is binary prediction of strong versus weak docking-derived binding affinity between 9-mer epitopes and a fixed SLA receptor, using only one-hot encoded peptide sequences as input to models including a small Transformer (Brisebois et al., 27 Jun 2026). Under strict low-data conditions, transformer-based sequence models consistently emerge as the best-performing architecture; with active incremental learning, the optimized configuration reaches a peak accuracy of 86.8% at $\mathbf{L}\in\mathbb{R}^{n_{\text{CDR}}\times 20}$3, and at $\mathbf{L}\in\mathbb{R}^{n_{\text{CDR}}\times 20}$4 it achieves 80.5%, outperforming a standard baseline trained on twice the amount of data (Brisebois et al., 27 Jun 2026). In this formulation, epitope-only means sequence-to-affinity prediction against a fixed receptor, with active learning deciding which epitopes are worth docking next.

The MHC-II multi-scale framework shows both the strength and the ceiling of peptide-only reasoning. In its best peptide-only EL configuration, the model achieves EL accuracy 0.7347, AUC$\mathbf{L}\in\mathbb{R}^{n_{\text{CDR}}\times 20}$5 0.8662, and CR-AUC 0.7349, using peptide and HLA information together with ESM2, MHC structure, joint BA/EL training, and binding-core supervision (Wan et al., 16 Dec 2025). Separately, in the antigen-presentation setting, a peptide-based model trained on EL only reaches CR-AUC 0.6092, while antigen-based models with windowed training reach values from 0.6346 to 0.6649 (Wan et al., 16 Dec 2025). This suggests that antigen context can recover signal that simpler epitope-only models miss, even though a strong peptide-only configuration remains competitive.

The sharpest caution comes from conformational B-cell epitope prediction. A benchmark of nine state-of-the-art webservers on over 250 antibody–antigen structures concludes that all methods achieve very low performances and that some do not perform better than randomly generated patches of surface residues; even consensus strategies are at best only marginally better than random (Cia et al., 2023). The best generic methods, such as DiscoTope2 and BEpro, reach ROC-AUC 0.58, BAC 0.53, and MCC 0.06, while antibody-specific EpiPred has PR-AUC 0.35 but no corresponding gain in MCC or BAC (Cia et al., 2023). The same benchmark shows poor performance on the SARS-CoV-2 spike protein case study (Cia et al., 2023). In encyclopedic terms, this is the central limitation of epitope-only reasoning: it is strongest when the epitope is explicitly specified, linearly encoded, or embedded in a fixed local design problem, and weakest when it is asked to discover generic conformational B-cell epitopes from antigen information alone.

A plausible synthesis is that epitope-only strategies are best viewed as controlled reductions. They are highly effective when local geometry is dominant, when the epitope is the intended design handle, or when peptide screening must be massively scalable. They are much less reliable when immune recognition depends on antibody specificity, antigen conformational ensembles, glycosylation, oligomeric context, or repertoire-dependent interaction patterns. The current literature therefore treats epitope-only not as a universal solution, but as a technically powerful abstraction whose usefulness is tightly conditioned by the biological level at which the problem is posed.