acceptodds
Under review as a conference paper at ICLR 2027

Explainability from Training with Applications to TCR-Epitope Prediction

Abstract

Deep learning models have achieved strong performance in artificial intelligence for science, yet their black-box nature limits our understanding of how they learn scientific tasks. Existing methods for interpretability provide limited insight into how models organize evidence and evolve during learning. We introduce explainability from training (EFT), a model-agnostic paradigm that traces model interpretation during training to explain why models rely on specific features and how they organize these features as predictive evidence. We apply EFT to four state-of-the-art T cell receptor (TCR)-epitope prediction models, TCR-SRIM, TULIP, MixTCRpred, and NetTCR-2.2, spanning post-hoc and interpret-by-design approaches as well as transformers and CNNs. To investigate how structural information affects model explanations, we introduce a benchmark, TCR-XAI2, containing 388 unique experimentally resolved TCR-epitope structures, complemented by structures predicted using AlphaFold3, Boltz-2, TCRModel2, tFold-TCR, and OpenFold3. Using \ModelName[abbr] with TCR-XAI2, we demonstrate that (1) CNN and transformer models exhibit distinct learning trajectories; (2) TCR and evidence can conflict during learning, limiting the benefits of jointly modeling both chains, while MHC information mitigates this; and (3) real versus predicted structural data for TCR-epitope prediction exhibits distinct TCR and peptide feature preferences as well as differing trajectories of model certainty.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
TCR, XAI, Explainability, TCR-Epitope, T Cell

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_CTwEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Explainability From Training (EFT), which tracks feature importance across training checkpoints and clusters similar trajectories to summarize how model explanations evolve. It applies EFT to four TCR-epitope prediction models and introduces TCR-XAI2, which contains 388 experimentally resolved structures, 619 complex instances, and corresponding predicted structures. Through trajectory analysis and binding region hit rate, the authors investigate architectural differences, the role of MHC information in coordinating TCR α\alpha/β\beta-chain evidence, and the effects of structural regularization.

Critical strengths
  1. Tracking explanations throughout training provides a useful perspective on changes that final-checkpoint analysis can miss.
  2. TCR-XAI2 provides a useful structural reference for evaluating residue-level explanations and comparing sources of structural supervision.
Critical weaknesses
  1. The evidence for α\alpha/β\beta - chain learning conflict is indirect. Different or opposing BRHR trajectories do not establish that learning from one chain interfaces with learning from the other. In particular, the correlation of approximately 0.06 without MHC information in Fig. 5 indicates littele linear association, rather than antagonism. The proposed mechanism is plausible, but the evidence supports a narrower descriptive conclusion. Can the authors provide evidence beyond trajectory correlations that distinguishes interference from independent changes in structural alignment?
  2. The architectural conclusions are confounded by other differences. Models use different training datasets, input modalities, and attribution methods (Appendix A.5). Appendix A.7 also shows that changing the attribution method changes the explanations. Consequently, the reported CNN-Transformer differences cannot be confidently attributed to architecture, weakening the generality of a central finding. Which headline conclusions remain consistent across attribution methods, and are there comparisons controlling for training data and input modalities?
Questions for authors

Please see the questions included in Critical Weakness 1 and 2.

5 comments

Thank you for your constructive suggestions. Below are our responses to weaknesses described in the review.

W1. Direct evidence for the conflict between the TCR chains

To provide the direct evidence for the TCR conflicts, we have added an extra dedicated experiment in Section 3.4. We train an encoder-only BERT model with separate encoders for the peptide, CDR3a, and CDR3b, whose concatenated representations are passed through a linear layer to predict the binding probability. The model is optimized with a binding prediction loss and a masked language modeling (MLM) loss on TCRvdb, a recently published and experimentally verified dataset. We compare three strategies: (1) "Vanilla", joint training of the full model throughout; (2) "Pair TCR", training only the CDR3a and CDR3b encoders with the MLM loss for the first 50 epochs before resuming joint training; and (3) "Pair TCR & Frozen", which uses the same first stage but then freezes the TCR encoders and trains only the peptide encoder. Even when the peptide branch is frozen, the BRHR and attention mass trajectories of the alpha and beta chains are already inconsistent. Once the peptide joins training, the attention mass of the two chains becomes substantially more inconsistent. When we instead freeze the TCR encoders to suppress this inconsistency, the variance of the peptide BRHR increases sharply, suggesting that the pressure shifts from TCR encoder learning to peptide encoder learning.

W2. Controlling for confounds in the architecture comparison

For the interpretation method, Appendix A.8 compares TULIP-MHC interpreted with QCAI and with AttnLRP, and MixTCRpred interpreted with GradCAM and with AttnLRP. The explanations differ in minor respects, as each interpretation method targets different aspects of the model by design, but the overall pattern is unchanged. We also include an interpret-by-design model, TCR-SRIM, whose explanations are faithful to its predictions by construction.

For the dataset, we additionally add extra experiments to train all models on TCRvdb so that the comparison uses common data with reduced label noise. As shown in Appendix A.11, the resulting explanations are consistent with those obtained on the original training datasets.

Thank you for the additional experiments, which adequately address my second concern regarding the architecture comparisons. My remaining concern is that divergent BRHR and attention trajectories do not directly establish learning interference between the TCR chains, and increased peptide BRHR variability after freezing the TCR encoders does not necessarily demonstrate a transfer of learning pressure. Given the limited time for further experiments, I suggest tempering these claims throughout the manuscript and presenting these mechanisms as possible interpretations rather than established conclusions.

Thank you for your response. We agree with your suggestions and have tempered the corresponding claims in the revised manuscript, with the changes highlighted in purple.

Please let us know if you have any further suggestions or concerns.

Reviewer_oTB3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Explainability From Training (EFT), which tracks feature attribution across multiple training checkpoints and clusters features with similar importance trajectories to analyze how feature importance changes during training. The authors apply EFT to four TCR–epitope prediction models and introduce TCR-XAI2, an extended benchmark containing experimentally determined and predicted structures. Their analysis examines differences between CNNs and Transformers, potential conflicts between TCR α- and β-chain information, and how MHC information and structural regularization affect the learning process.

Critical strengths
  1. The research problem is meaningful: how feature importance changes during training, beyond what final-checkpoint explanations can show.
  2. EFT supports both post-hoc and interpret-by-design methods without changing the underlying prediction models.
  3. The experiments cover four models, three random seeds, and different explainers. TCR-XAI2 also provides a useful benchmark with structural annotations.
  4. The paper is well organized, and the overview figure and trajectory plots help readers understand the framework and follow changes during training.
Critical weaknesses
  1. Tracking attribution during training is not new. Both Lapuschkin et al. [1] and Voita et al. [2] studied this idea. EFT mainly adds trajectory clustering and structured summarization, and the paper should make this distinction clear.
  2. The added value of trajectory clustering is unclear. The paper does not show what clustering adds over grouped attribution curves or feature-by-checkpoint heatmaps. The main trends in Fig. 5, for example, do not depend on clustering. A comparison using the same attributions would clarify its benefits for insight or readability. Sensitivity checks would also help assess robustness to clustering settings.
  3. The comparison does not isolate the effect of architecture. The models differ in training data, training settings, and explainers (Appendix A.5). Appendix A.7 further shows that changing the explainer changes EFT results for the same model. Without controlling for these factors, the observed differences cannot be attributed to architecture alone and do not establish general differences between CNNs and Transformers.
  4. The evidence for α–β learning conflict is insufficient. Eq. 8 uses argmax to select the dominant cluster, so dominance can switch even when both groups become more important. BRHR measures structural alignment rather than each chain’s predictive contribution. These observations do not establish learning interference. The authors should clearly define “learning conflict” and provide targeted behavioral validation, or narrow the claim.

Reference [1] Lapuschkin et al. (2019). Unmasking Clever Hans Predictors and Assessing What Machines Really Learn. Nature Communications. https://arxiv.org/pdf/1902.10178 [2] Voita, E., Sennrich, R., and Titov, I. (2021). Analyzing the Source and Target Contributions to Predictions in Neural Machine Translation. ACL-IJCNLP 2021. https://aclanthology.org/2021.acl-long.91/

Questions for authors

Nothing in particular.

2 comments

Thank you for your constructive suggestions. Below are our responses to weaknesses in the review.

W1 & W2. Novelty and value of tracking the entire training process using EFT

Prior work that tracks interpretations during training follows only a few pre-defined attribution statistics or hand-selected input regions. It does not track the importance trajectories of all input features, which show when and how a model comes to rely on each one. (We have added this discussion and related citations at Line 067.) EFT is not only a combination of existing feature attribution techniques. It provides a systematic way to organize the evidence by how feature importance changes along the training trajectory, rather than clustering attributions directly or tracking a set of pre-selected features.

Compared with inspecting only the final checkpoints, EFT recovers the model's learning trajectory. Rather than a static heatmap, it shows how and why a model comes to rely on a feature, which helps diagnose whether that evidence is appropriate. For example, EFT reveals a conflict between the learning of the TCR alpha and beta chains, and this observation leads to a simple training strategy that mitigates the conflict and improves performance.

W3. Controlling for confounds in the architecture comparison

For the interpretation method, Appendix A.8 compares TULIP-MHC interpreted with QCAI and with AttnLRP, and MixTCRpred interpreted with GradCAM and with AttnLRP. The explanations differ in some respects, as each interpretation method targets different aspects of the model by design, but the overall pattern is unchanged. We also include an interpret-by-design model, TCR-SRIM, whose explanations are faithful to its predictions by construction.

For the dataset, we additionally train all models on TCRvdb so that the comparison uses common data with reduced label noise. As shown in Appendix A.11, the resulting explanations are consistent with those obtained on the original training datasets.

W4. Direct evidence for the conflict between the TCR alpha and beta chains

To provide the direct evidence for the TCR conflicts, we have added an extra dedicated experiment in Section 3.4. We train an encoder-only BERT model with separate encoders for the peptide, CDR3a, and CDR3b, whose concatenated representations are passed through a linear layer to predict the binding probability. The model is optimized with a binding prediction loss and a masked language modeling (MLM) loss on TCRvdb, a recently published and experimentally verified dataset. We compare three strategies: (1) "Vanilla", joint training of the full model throughout; (2) "Pair TCR", training only the CDR3a and CDR3b encoders with the MLM loss for the first 50 epochs before resuming joint training; and (3) "Pair TCR & Frozen", which uses the same first stage but then freezes the TCR encoders and trains only the peptide encoder.

Even when the peptide branch is frozen, the BRHR and attention mass trajectories of the alpha and beta chains are already inconsistent. Once the peptide joins training, the attention mass of the two chains becomes substantially more inconsistent. When we instead freeze the TCR encoders to suppress this inconsistency, the variance of the peptide BRHR increases sharply, suggesting that the pressure shifts from TCR encoder learning to peptide encoder learning.

Thank you for your constructive suggestions. We believe our responses have addressed your questions and concerns. As the rebuttal deadline is approaching, we would like to kindly ask whether you have any further questions or concerns. We would be happy to discuss them and provide further clarification.

Reviewer_kdtHOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a method called Explainability From Training (EFT), which essentially tracks feature importance throughout training and clusters features based on how their importance evolves over time. This allows us to track how a model changes its reliance on different features during training. They apply this methodology to the TCR-epitope prediction task in computational biology, evaluating it on several models and using different feature-importance techniques. They analyze how feature importance changes over time and how different models rely on different types of features. Moreover, to support these analyses, they introduce TCR-XAI2, an extension of TCR-XAI containing additional TCR-epitope structures as well as structures predicted by state-of-the-art models.

Critical strengths
  • The idea of using interpretability techniques to better understand models and extract useful insights to expand our knowledge of what they have learned (especially in XAI for science) has long been a goal, but it has been difficult to put into practice or to derive meaningful insights from it. In this case, the authors apply XAI techniques and reframe them within their method to extract useful information during model training for a real-world application in biology, which is indeed interesting and potentially impactful.

  • The paper is well written and easy to understand and follow. At times, it contains too many details and/or excessive formalization, but these are simple enough not to disrupt the reader's flow.

  • The extension of TCR-XAI, while including structures predicted by state-of-the-art models is valuable.

Critical weaknesses

My biggest concern is whether what is being measured really corresponds to what the authors intend to measure, and how robust these findings are, given a few inconsistencies and possible confounding factors.

  • Assuming the feature-importance method is accurate, EFT shows when the model starts relying more on a feature. However, this does not explain why (which I consider a more important question), nor does it tell us whether the model has actually learned to use that feature.

  • The method relies on feature-importance estimates, which, as we know from the literature, can be unreliable. Consequently, EFT can only be as reliable as the feature-importance method it uses. You apply different attribution techniques to different models, making it difficult to determine whether the differences observed across models are driven by the models themselves or by the attribution methods. How robust are your findings to changing the attribution method for a given model?

  • The rationale behind several methodological choices is not discussed. How are the thresholds, σ\sigma, and other hyperparameters selected? Moreover, all notions of being "active" are defined relative to the feature's own importance values. For instance, what happens if a feature's importance remains constant throughout training? Would it always be considered active, even if its importance is extremely low?

Questions for authors

See above.

  • Figure 1 is not very informative; it is dense and complex. I suggest abstracting away some of the details and making it more informative and intuitively understandable.

  • How did you choose DTD_T, and how do you determine whether it is sufficiently large and diverse? Would different subsets lead to different results?

  • Figure 3 is not referenced anywhere in the text.

3 comments

Thank you for your constructive suggestions. Below are our responses to weaknesses and questions in the review.

W1. Interpretation reliability

Our response has two parts. First, EFT also supports interpret-by-design models such as TCR-SRIM, whose evidence is faithful to its predictions by construction, so the concern does not apply there; it applies only to black-box models analyzed with post-hoc methods. Second, as discussed in Appendix A.8, the explanations of a single model do differ across post-hoc methods, but the differences are minor. For MixTCRpred, AttnLRP and GradCAM differ mainly in the CDR1a and CDR1b clusters, while the general pattern, including the conflict between the TCR chains, is unchanged. This is corroborated by the attention mass trajectories in Appendix A.7.

W2. Does EFT explain "why" a feature is needed?

A static interpretation of the final checkpoint shows the evidence the model uses to produce its prediction, but that final state is formed during training under the loss. A model is not optimized for semantic or causal correctness; it is optimized to minimize the loss, and this signal is reflected in how feature importance changes. Because the features in turn determine the predictions that enter the loss, the final static interpretation is itself determined by the intermediate changes in feature importance. EFT therefore recovers the search path through feature space that the loss induces, which is what allows it to address why a model comes to rely on a feature rather than only when. For example, EFT indicates that a TULIP without MHC input comes to rely on one TCR chain as a proxy for MHC-dependent information, a hypothesis we then test by adding MHC information, which reduces the conflict: for TULIP, the correlation between the alpha and beta attention mass trajectories rises from r=0.1r=0.1 to r=0.91r=0.91 (Appendix A.7).

W3. Hyperparameter selection & ablation studies

We provide ablation studies in Appendix A.13 and choose the combination that yields sufficient sparsity while retaining enough features to remain informative. Moreover, EFT organizes the change in feature importance along training rather than a single static importance value, so a feature whose importance always remains constant and very low is filtered out.

Q2. Explanation subset selection

The tracking set is drawn from the TCR-XAI2 benchmark, which allows us to compute the quantitative BRHR metric and to assign residue-level structural labels. Its size is therefore limited by the number of available experimentally resolved structures.

Q1 & Q3. Figure revisions

Reviewers noted that Figure 1 is too complex and that Figure 3 is never referenced in the text. We have simplified Figure 1 (see Figure 1) and now reference Figure 3 in Section 3.2 (Line 358) as part of case study.

I thank the authors for the additional clarifications, and most of my concerns are now resolved. I will increase my score to 3.

Thank you for your responses and for raising the score. We sincerely appreciate your time and constructive feedback.

Reviewer_zw6GOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces EFT, which is an interpretability technique that applies feature-attribution methods to every checkpoint during training and combines them into trajectories. They apply this method to 4 TCR-epitope prediction models, and also create a benchmark of ~400 structures with results on them. They provide evidence of different models using different ways of learning and encoding information, and provide mechanistic insights into how the models function.

Critical strengths
  • I like how they model explainability as a trajectory instead of insights about a static, final model. It shows the change in how features contribunte to the final prediction.
  • Since the feature attribution methods are diverse, this method seems to be method agnostic, which means it can be applied to various types of prediction models, regardless of architecutre or framework.
  • It's great that some insights align with findings that the biology community know from prior literature, showing that models can learn useful models of underlying biological phenomena.
Critical weaknesses
  • Novelty is a bit underwhelming since it's a combination of existing feature attribution techniques, but on different checkpoints. While it's nice to track the development of features through the checkpoints, it seems more like pipelining rather than a new method.
  • There is no contribution of a new architecture with the insights gained from the individual models that were analysed. There is no practical utility of the suggestions provided in the paper either.
  • There doesn't seem to be a null model provided where intentionally running EFT on labels or annotations with poor signal (by shuffling or scrambling them, or randomly initialising the models) show weaker scores. The lower bounds of explainability isn't shown either.
Questions for authors
  • I'm curious what EFT explains that running the same explainer on the final checkpoint plus one early checkpoint wouldn’t? What’s the marginal value of the full trajectory instead of just the final 1-2 checkpoints? Won't the initial checkpoints just be noisy or erratic?

  • The models used here don't really generalise to new, unseen epitopes. Does running EFT distinguish the ones that generalise from the ones that don’t OR does it explain all of them equally well?

  • The authors use 3d structure as ground-truth explanation. But should a binding predictor’s feature attribution match the ground truth contact map since binding depends mostly on energetics and dynamics, which the static structure doesn’t really capture?

2 comments

Thank you for your constructive suggestions. Below are our responses to weaknesses and questions from the review.

W1. Novelty of EFT

EFT is not merely a combination of existing feature attribution techniques. It provides a systematic way to organize the evidence by how feature importance changes along the training trajectory, rather than clustering attributions directly or tracking a set of pre-selected features.

W2 & Q2. Contribution to model improvement and generalization to unseen epitopes

We provide insights for model design in Appendix A.14. We also add a new section showing that freezing the TCR encoders during training, as suggested by EFT, reduces the conflict between the TCR chains. On a subset of IMMREP23 with unseen epitopes, this improves ROC-AUC@FPR$<0.1by1.5percentagepoints(0.1 by 1.5 percentage points (p=0.017$).

W3. Null model explanation

We add an extra experiment as Appendix A.12 to train a null model on a dataset with 50% of the labels flipped and compare its explanation with that of the model trained on the clean dataset. The patterns differ clearly: under noisy labels, the model expends more capacity fitting a single interaction and leaves less room for the remaining ones. EFT distinguishes the two cases.

Q1. Value of tracking the entire training process

Compared with inspecting only the final checkpoints, EFT recovers the model's learning trajectory. Rather than a static heatmap, it shows how and why a model comes to rely on a feature, which helps diagnose whether that evidence is appropriate. For example, EFT reveals a conflict between the learning of the TCR alpha and beta chains, and this observation leads to a simple training strategy that mitigates the conflict and improves performance, as discussed in W2. Moreover, because we track the change in feature importance rather than its value at a single checkpoint, even early-stage explanations carry usable signal rather than pure noise.

Q3. 3D structure rationale

BRHR, which is based on 3D structure, is defined from interatomic contact distances, which provide a general proxy for interactions at the binding interface, since a distance criterion is agnostic to the specific type of interaction. It does not account for conformational variation or dynamics, as the available structural data remain too limited to support this. To corroborate this choice, we add the attention mass trajectories in Appendix A.7. Each chain shows the same pattern as in the BRHR trajectories, with a stronger signal. For NetTCR-2.2, MixTCRpred, and TCR-SRIM, the two TCR chains are negatively correlated (r<−0.55r < -0.55). For TULIP, adding MHC information raises the correlation of the attention mass trajectories from r=0.1r=0.1 to r=0.91r=0.91. These results reproduce the pattern seen in the BRHR trajectories and support the use of BRHR, which additionally relates the explanations to structural information.

Thank you for your constructive suggestions. We believe our responses have addressed your questions and concerns. As the rebuttal deadline is approaching, we would like to kindly ask whether you have any further questions or concerns. We would be happy to discuss them and provide further clarification.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Consolidated Rebuttal

We thank the reviewers for their constructive and detailed reviews. Below we respond to the questions and weaknesses raised. We first summarize the concerns shared across reviewers and provide a consolidated response together with an updated manuscript, in which the updated content is marked in blue. Individual responses to each reviewer follow.

1. Novelty and value of tracking the entire training process using EFT

Reviewers asked about the novelty of EFT, how it differs from existing methods, and what is gained by tracking all checkpoints rather than only the last one or two.

Prior work that tracks interpretations during training follows only a few pre-defined attribution statistics or hand-selected input regions. It does not track the importance trajectories of all input features, which show when and how a model comes to rely on each one. (We have added this discussion at Line 067.) EFT is not only a combination of existing feature attribution techniques. It provides a systematic way to organize the evidence by how feature importance changes along the training trajectory, rather than clustering attributions directly or tracking a set of pre-selected features.

Compared with inspecting only the final checkpoints, EFT recovers the model's learning trajectory. Rather than a static heatmap, it shows how and why a model comes to rely on a feature, which helps diagnose whether that evidence is appropriate. For example, EFT reveals a conflict between the learning of the TCR alpha and beta chains, and this observation leads to a simple training strategy that mitigates the conflict and improves performance. Please see Responses 2 and 3 for details.

2. Direct evidence for the conflict between the TCR alpha and beta chains

Reviewers noted that the evidence for a conflict between the TCR alpha and beta chains is not sufficient.

To provide the direct evidence for the TCR conflicts, we have added an extra dedicated experiment in Section 3.4. We train an encoder-only BERT model with separate encoders for the peptide, CDR3a, and CDR3b, whose concatenated representations are passed through a linear layer to predict the binding probability. The model is optimized with a binding prediction loss and a masked language modeling (MLM) loss on TCRvdb, a recently published and experimentally verified dataset. We compare three strategies: (1) "Vanilla", joint training of the full model throughout; (2) "Pair TCR", training only the CDR3a and CDR3b encoders with the MLM loss for the first 50 epochs before resuming joint training; and (3) "Pair TCR & Frozen", which uses the same first stage but then freezes the TCR encoders and trains only the peptide encoder.

Even when the peptide branch is frozen, the BRHR and attention mass trajectories of the alpha and beta chains are already inconsistent. Once the peptide joins training, the attention mass of the two chains becomes substantially more inconsistent. When we instead freeze the TCR encoders to suppress this inconsistency, the variance of the peptide BRHR increases sharply, suggesting that the pressure shifts from TCR encoder learning to peptide encoder learning.

3. Contribution to model improvement and generalization to unseen epitopes

Reviewers asked how EFT can be used to improve a model and its generalization to unseen epitopes. Rather than an architecture-level change, EFT points to a simpler intervention: the "Pair TCR & Frozen" training strategy.

Because the TCR-epitope task mainly requires generalization to unseen epitopes, the model should learn stable features for the TCR and more diverse representations for the peptide. EFT shows that the alpha and beta chains carry conflicting information, which can degrade feature quality. "Pair TCR & Frozen" reduces this conflict and encourages the model to learn more diverse information from the peptide. We compare the three strategies on the HLA-A*01 and HLA-A*02 samples of the IMMREP23 benchmark. Across three random seeds, they achieve ROC-AUC@FPR$\le$0.1 of 0.550±0.0060.550 \pm 0.006, 0.550±0.0060.550 \pm 0.006, and 0.565±0.0030.565 \pm 0.003, respectively, where "Pair TCR & Frozen" yields a significant improvement of 1.5% (p=0.017p=0.017). These results indicate that EFT can identify defects in model learning that can then be addressed to improve generalization. We also provide more model improvement suggestion in Appendix A.14.

AuthorsOpenReview ↗

Consolidated Rebuttal (continue)

[continue]

4. Controlling for confounds in the architecture comparison

Reviewers noted that our conclusions about architectural differences are drawn from models trained on different datasets and interpreted with different methods. We address both confounds.

For the interpretation method, Appendix A.8 compares TULIP-MHC interpreted with QCAI and with AttnLRP, and MixTCRpred interpreted with GradCAM and with AttnLRP. The explanations differ in some respects, as each interpretation method targets different aspects of the model by design, but the overall pattern is unchanged. We also include an interpret-by-design model, TCR-SRIM, whose explanations are faithful to its predictions by construction.

For the dataset, we additionally add an experiment to train all models on TCRvdb so that the comparison uses common data with reduced label noise. As shown in Appendix A.11, the resulting explanations are consistent with those obtained on the original training datasets.

5. Interpretation reliability and the null-model explanation

Reviewers asked about the reliability of the interpretations and how explanations differ for a model trained on a null dataset.

As described in Response 4, EFT recovers similar patterns for the same model across datasets and interpretation methods. To validate this further, we add a null-model comparison in Appendix A.12. We randomly flip the labels of 50% of the samples for each peptide, which reduces the ROC-AUC of NetTCR-2.2 from 0.8 to near chance, and then apply EFT to the resulting model. The model focuses mainly on the CDR3a-peptide interaction, whereas the CDR3b-peptide interaction emerges much later than when the model is trained on the clean dataset. More notably, although the model still attends to CDR1b, this attention shifts to the end of training and never becomes dominant. This suggests that, under noisy labels, the model expends more capacity fitting a single interaction and leaves less room for the remaining ones.

These results show that EFT can distinguish models trained on high- and low-quality data. They also indicate that the noise in the original training set is not severe enough to make the labels effectively random, since the model still learns signal from it.

6. Hyperparameter ablation study and subset selection

Reviewers asked how we choose the subset used for explanation and how we select the hyperparameters.

The tracking set is drawn from the TCR-XAI2 benchmark, which allows us to compute the quantitative BRHR metric and to assign residue-level structural labels. Its size is therefore limited by the number of available experimentally resolved structures. For the hyperparameters, we add extra ablation studies in Appendix A.13. We select the combination that yields sufficient sparsity while retaining enough features to remain informative.

7. Rationale for the BRHR trajectories

Reviewers asked why BRHR is an appropriate basis for evaluating EFT and what the attention mass trajectories look like.

BRHR is defined from interatomic contact distances, which provide a general proxy for interactions at the binding interface, since a distance criterion is agnostic to the specific type of interaction. It does not account for conformational variation or dynamics, as the available structural data remain too limited to support this. To corroborate this choice, we add the attention mass trajectories in Appendix A.7. Each chain shows the same pattern as in the BRHR trajectories, with a stronger signal. For NetTCR-2.2, MixTCRpred, and TCR-SRIM, the two TCR chains are negatively correlated (r<−0.55r<-0.55). For TULIP, adding MHC information raises the correlation of the attention mass trajectories from r=0.1r=0.1 to r=0.91r=0.91. These results reproduce the pattern seen in the BRHR trajectories and support the use of BRHR, which additionally relates the explanations to structural information.

8. Figure revisions

Reviewers noted that Figure 1 is too complex and that Figure 3 is never referenced in the text. We have simplified Figure 1 (see Figure 1) and now reference Figure 3 in Section 3.2 (Line 358) as part of case study.

AuthorsOpenReview ↗

Rebuttal Summary for Area Chair

Rebuttal Summary for Area Chair

Dear Area Chair,

Thank you for handling our submission. A brief summary of the review and rebuttal for your reference.

The value of this work

Our experiments and the discussion with reviewers established that interpreting a model during training reveals how its evidence is formed, and where that process goes wrong. We see two contributions:

  1. Interpretability as diagnosis, not only description. EFT identifies a specific failure in how a model organizes evidence, which leads a loop from interpretation to measurable improvement.
  2. A general instrument for TCR-epitope models. EFT is model-agnostic across architectures and explainers. Applied here, it surfaced the alpha-beta possible conflict and motivated a fix that improves performance on unseen epitopes.

We hope this encourages the community to treat interpretability as a tool for improving models, not only for understanding them, and a way to assist TCR-epitope model improving on noise TCR-epitope dataset.

Reviewer Responses

Reviewers kdtH and CTwE engaged with our rebuttal, confirmed their concerns were resolved, and raised their scores [2 →\to 3 and 2 →\to 3]. Reviewers zw6G and oTB3 did not respond during the discussion period.

Reviewer confirmed addressed shared concerns. Three concerns account for nearly all comments across the four reviews. All four reviewers raised at least one, and oTB3's comments fall entirely within them.

1. Novelty of EFT and the value of tracking training

Prior work tracks a few pre-defined attribution statistics or hand-selected regions, not the importance trajectory of every input feature. EFT organizes evidence by how importance evolves during training rather than clustering attributions directly (discussion added at Line 067).

2. Confounds in the architecture comparison

A.8 holds the model fixed and varies the explainer (TULIP-MHC under QCAI vs. AttnLRP; MixTCRpred under GradCAM vs. AttnLRP): details differ by design, the overall pattern does not. A.11 retrains all models on common data (TCRvdb); explanations remain consistent. TCR-SRIM is interpret-by-design and faithful by construction.

3. Evidence for possible conflict between TCR alpha-beta chain

New controlled experiment in Section 3.4: alpha-beta inconsistency is present before the peptide encoder trains, grows sharply once it joins, and when suppressed by freezing the TCR encoders the variance is displaced onto the peptide. Attention-mass trajectories (A.7) show r<−0.55r < −0.55 for NetTCR-2.2, MixTCRpred and TCR-SRIM; for TULIP, adding MHC raises the correlation from r=0.1r = 0.1 to r=0.91r = 0.91.

Each was addressed with experiments and manuscript revisions, and both reviewers who engaged confirmed they were satisfied. Since oTB3's concerns coincide with these, we believe the same evidence applies directly to their review.

Additional questions raised only by zw6G, each addressed in the revision:

  1. Model improvement and unseen epitopes.

A.14 gives design insights derived from EFT. The staged-training experiment above also yields a concrete fix: freezing the TCR encoders reduces inter-chain conflict and improves ROC-AUC@FPR$<0.1$ by 1.5 points (p = 0.017) on unseen epitopes in IMMREP23. The diagnosis EFT provides therefore leads directly to a measurable gain.

  1. Null model.

A.12 trains a model on 50% flipped labels. Its trajectories differ clearly from the clean model: under label noise, capacity concentrates on a single interaction at the expense of the rest. EFT separates the two cases.

  1. Rationale for 3D structure.

BRHR uses interatomic contact distance, an interaction-agnostic proxy for the binding interface. It does not capture conformational dynamics, for which the available structural data remain too limited. The attention-mass trajectories in A.7 reproduce the BRHR pattern with a stronger signal, corroborating the measure.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.