acceptodds
Under review as a conference paper at ICLR 2027

Asymmetry-Aware Representation Learning for TCR-Epitope binding prediction

Abstract

T cell receptors (TCRs) play a vital role in immune recognition by binding specific epitopes. Accurate prediction of TCR–epitope binding specificity is fundamental for advancing immunology research. Existing methods commonly formulate TCR–epitope interaction prediction as a dual-encoder matching problem, in which TCRs and epitopes are independently encoded and subsequently aligned in a shared representation space. However, this formulation overlooks the representational asymmetry between structurally complex TCRs and lower-complexity epitopes, introducing unnecessary modeling overhead and hindering effective alignment. To address this issue, we propose an asymmetry-aware representation learning method that combines structured TCR encoding with discrete epitope modeling, reducing joint modeling complexity and recasting heterogeneous alignment as a discriminative mapping to the candidate epitope space. Specifically, the model uses a frozen protein language model to extract contextualized CDR3 representations from full-length - and -chain sequences. These representations are then encoded by weight-shared multiscale convolutions and independent global branches to capture cross-chain local patterns and chain-specific context, respectively. Candidate epitopes are represented as discrete categories in a finite prediction space. The model is optimized with complementary supervision to learn the global TCR-to-epitope mapping while refining the relative distinction between target and competing epitopes. Comparative experiments against multiple baseline methods on several benchmark datasets demonstrate the effectiveness of the proposed method.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
T-cell receptor, epitope specificity classification, representation learning, multiscale convolution

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_LqW5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AsymTCR, an asymmetric representation learning framework for TCR–epitope binding prediction. Instead of separately encoding TCRs and epitopes, the method uses a frozen ProtBERT model to extract contextualized CDR3 features from paired TCR chains and combines chain-specific global representations with shared multiscale convolutions. Epitopes are treated as discrete candidate classes, and the model is trained using asymmetric classification, pairwise, and ranking losses. Experiments on four benchmark datasets against 16 baseline configurations show competitive performance, with improvements in AUROC across all four datasets. Ablation studies further examine the contributions of paired-chain information, sequence context, and multiscale convolutions.

Critical strengths
  1. The asymmetric modeling approach is well motivated for fixed-candidate epitope prediction. Focusing representation learning on TCRs avoids the need for a separate epitope encoder and provides a simpler alternative to dual-encoder methods.

  2. The experimental evaluation is extensive, covering four datasets and 16 baseline configurations. AsymTCR achieves strong results across datasets, particularly in AUROC, which supports the effectiveness of the proposed approach.

Critical weaknesses
  1. The method mainly combines established components, including ProtBERT, multiscale CNNs, and standard loss functions. Moreover, removing the context projection improves independent-test AUPRC from 0.764 to 0.773 (Figure 5), questioning the necessity of this component.

  2. Despite comparisons with 16 baselines (Table 1), no matched symmetric dual-encoder baseline is provided to isolate the benefit of asymmetry. AsymTCR also achieves slightly lower AUPRC than TCRconv-fullAB on all three additional datasets (Tables 4–6).

  3. The individual contributions of the three loss functions (Eqs. 14–18) are not evaluated. Figure 4(a) only reports hyperparameter sensitivity, with AUPRC ranging from 0.805 to 0.817, leaving the necessity of each loss unclear.

  4. The fixed-class prediction head (Eqs. 11–12) cannot directly handle unseen epitopes. Moreover, the NM-Benchmark independent test contains only seen epitopes (Appendix C, Table 3), leaving unseen-epitope generalization unexamined.

Questions for authors
  1. Can the authors provide a controlled comparison with a symmetric dual-encoder model using the same ProtBERT backbone and comparable model capacity? This would help clarify the contribution of asymmetric modeling.

  2. How does AsymTCR compare with a simpler ProtBERT-based classifier without the proposed local-global architecture?

  3. Can the authors provide ablation results for each of the three loss functions? It is unclear how much each contributes to the final performance.

  4. Can the authors clarify whether the independent test set contains only epitopes seen during training, and how much overlap exists between the training and test TCR sequences?

1 comment

I maintain my original score.

Reviewer_SeNiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

he authors propose AsymTCR for predicting TCR specificity within a fixed vocabulary of known epitopes. Frozen ProtBERT encodes the full α- and β-chain sequences, from which contextualized CDR3 features are extracted. Shared multiscale convolutions capture local patterns, while independently parameterized global branches process each chain. A classification head maps the combined receptor representation directly to epitope-specific scores. Training combines an asymmetric classification loss, a selected-pair binary loss, and a ranking loss. AsymTCR achieves the highest reported NM-Benchmark AUPRC and AUROC, including 0.743 AUPRC and 0.700 AUROC on the independent set. Results on the additional datasets are competitive, although TCRconv-fullAB has slightly higher AUPRC on all three. The main unresolved issues are the distinction from existing receptor-to-epitope classifiers, the lack of ablations isolating the proposed changes, and insufficient detail about how the processed benchmark records become training examples.

Critical strengths
  1. The independent NM-Benchmark result provides meaningful evidence of improved discrimination on new receptors for known epitopes. AsymTCR improves AUROC from 0.570 for TCRconv-fullAB to 0.700, alongside a smaller AUPRC improvement from 0.732 to 0.743. TCRconv-fullAB is a relevant comparator because it also uses paired-chain contextual information. This result deserves weight even though the improvement is not uniform across datasets or metrics.

2.The evaluation includes established public resources and examines relevant receptor inputs. NM-Benchmark and the IMMREP datasets are appropriate sources for studying known-epitope prediction. The paired-chain and context ablations also provide useful evidence about which inputs contribute in this implementation. For example, removing the β-chain substantially reduces independent-test AUPRC. The paper appropriately identifies unseen-epitope prediction as outside its present scope.

Critical weaknesses
  1. The methodological contribution is not sufficiently distinguished from TCRconv, and the ablations do not isolate the actual additions. The central framing is that rich TCR representation learning should be combined with discrete epitope outputs instead of bilateral sequence encoding. However, TCRconv already extracts full-chain-contextualized CDR3 representations using ProtBERT, applies convolutional processing, and jointly predicts predefined epitope labels. The paper includes this method experimentally, which is appropriate, but does not adequately acknowledge how closely it precedes the proposed formulation. The contribution therefore appears to lie primarily in the particular local/global architecture, cross-chain parameter sharing, and additional training losses. Figure 5 does not isolate these differences. There is no shared-versus-independent local convolution comparison, no decomposition of the three losses, and no comparison between discrete and sequence-based epitope prediction using the same receptor encoder and supervision. Sweeping nonzero ranking-loss weights does not establish that the ranking term is necessary.

  2. The benchmark processing and supervision are not specified well enough to establish what drives the reported performance.Appendix C reports retained counts and evaluation procedures, but does not sufficiently describe negative-pair construction, receptor grouping, sequence-similarity controls after processing, or full-chain reconstruction. The original NM-Benchmark includes similarity filtering and defined negative-generation procedures. The question is whether these protections were preserved when the authors filtered receptors and constructed the reported pairs. Their absence should not be assumed, but their preservation needs to be demonstrated. The training formulation makes this particularly consequential. Equation 13 assigns one positive candidate to each training receptor, and the all-candidate loss treats the remaining candidates as negatives. The pair and ranking losses additionally require positive and negative candidates for that receptor. The paper does not explain how these rules handle receptors with multiple recorded positives, negative-only receptors, or several negative records. The authors acknowledge that unselected epitopes are not experimentally verified non-interactions, but the source and handling of the selected negative labels remain unclear.

Questions for authors
  1. Which parts of the original benchmark protocols were preserved? Can you provide the retained record identifiers, original split assignments, negative-generation rules, and full-chain construction procedure. For newly created folds, specify whether grouping is by interaction pair, paired receptor, individual chain, or sequence cluster. Report exact and near-neighbor receptor overlap after preprocessing, and confirm that positive and negative records involving the same receptor remain together. For IMMREP22/23, please map the processed data to the original training and test collections. An evaluation retaining the applicable official seen-epitope holdout would make the results easier to compare with prior work. The relevant IMMREP22 and IMMREP23 resources provide separate training/test structures; reproducing those settings should be distinguished from constructing new cross-validation experiments.

  2. How are the candidate-level targets constructed from the pair records? Can you explain how negative-only receptors, multiple negatives, and any receptors with multiple known positive epitopes enter Equations 13–18. Are known positives consolidated into a multi-hot target and excluded from negative sampling? Are the selected negatives experimentally measured, supplied background controls, or generated mismatches?

Reviewer_eRmHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper predicts TCR binding to epitopes from a fixed list of candidates. AsymTCR uses frozen ProtBERT to encode the full alpha and beta chains, then extracts the CDR3 features. It combines multiscale convolutions with a separate global branch for each chain and predicts a score for each candidate epitope. Training uses asymmetric classification, pairwise binary cross-entropy, and ranking losses. The experiments cover four datasets, including an independent seen-epitope test on NM-Benchmark. The model improves AUPRC and AUROC over TCRconv-fullAB on NM-Benchmark, although the results on the other datasets are mixed.

Critical strengths

The independent test result is encouraging. AsymTCR reaches an AUROC of 0.700, compared with 0.570 for TCRconv-fullAB. AUPRC also improves, from 0.732 to 0.743. Testing on an independent data source makes these results more useful than cross-validation alone. I also appreciate that the paper clearly states that it handles a fixed set of epitopes and leaves unseen-epitope prediction for future work.

Critical weaknesses
  1. My main concern is what the method adds to TCRconv [1]. TCRconv already uses full-chain ProtBERT embeddings, extracts CDR3 features, combines multiscale convolutions with global linear branches, and predicts a fixed set of epitopes without an epitope encoder (Section 2.3 and Figure 2 of [1]). So the idea of treating epitopes as output classes is already there. The important question is whether the changes proposed here explain the gains. Figure 5 does not test shared versus separate convolutional weights or remove the added loss terms. The sensitivity plots do not answer this either, since the tested loss weights remain nonzero. AUPRC is also slightly below TCRconv-fullAB on all three additional datasets in Tables 4–6. The NM-Benchmark gains are worth reporting, but I am not yet convinced that the paper shows where they come from.

  2. The model description and the supplied code do not agree. Section 2.2.2, Equation (6), and Appendix B say that the two chains share their local convolutional weights. In asymtcr/model.py, however, alpha and beta use separate ChainEncoder instances with independent convolutional weights. The README also explicitly says that these parameters are separate. This matters because weight sharing is part of the proposed method. The package provides a random-input demo, but no scripts to process the benchmark data, train the models, or reproduce the tables. I therefore cannot tell whether the results came from the shared model in the paper or the separate encoders in the code. Please clarify which model was actually evaluated and provide the matching implementation.

[1] TCRconv: predicting recognition between T cell receptors and epitopes using contextualized motifs. Bioinformatics, 2022.

Questions for authors
  1. Did the models used for Tables 1 and 4–6 and Figure 5 share convolutional weights across the two chains? Please point to the code and configuration used for those results. If the supplement is only a simplified demo, how does it differ from the evaluated model?

  2. Do you have a comparison that keeps the encoder and training setup fixed and changes only the pairwise/ranking losses or weight sharing? A small control on the existing NM-Benchmark setup would help show what the proposed changes contribute over TCRconv.

  3. How were the folds and training pairs built? Please confirm that all positive and negative pairs for the same TCR stay in the same fold. How are the candidate indices in Equations (13)–(17) assigned when a TCR has several negative pairs or no retained positive pair? Also, does “native method coverage” in Tables 4–6 mean that some methods were evaluated on different pairs? Split files and the evaluation script would help clarify this.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.