acceptodds
Under review as a conference paper at ICLR 2027

EMBER: Predicting TCR–pMHC Binding Without a Docked Complex

13 views

Abstract

We propose **EMBER**, a structure-based framework for TCR–pMHC binding prediction that eliminates the need for a predicted docked complex, and with it the placement error that dominates such predictions. Traditional structure-based predictors utilize the co-folded complex structure of the receptor and the antigen and interpret binding from its contact surface. Consequently, every downstream feature inherits any deviation the structure-prediction tool introduced. Methodologically, EMBER folds the TCR and pMHC independently, encodes them as SE(3)-invariant residue graphs, and compares their learned representations without constructing a shared coordinate frame. A relative pose is therefore not an input of the model, so placement error cannot propagate into the prediction. Computationally, EMBER preprocesses monomer folds where a complex-based predictor needs operations. Experimentally, under epitope-disjoint evaluation with six retrained baselines, EMBER is first on every metric under all four folding tools, by**0.9 to 14.4%** in micro-AUROC, **18.5 to 25.3%** in micro-AUPRC and **10.6 to 24.1%** in MCC. Its ranking is also up to **85%** more stable across folding tools.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
TCR--pMHC, Protein, Binding, Assembly-induced Interface Corruption

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SvysOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

EMBER predicts TCR–pMHC binding from independently folded components without requiring a docked complex. It separately encodes their structural and sequence features, then uses cross-attention between CDR3β and peptide residues to predict binding. The authors introduce assembly-induced interface corruption (AIC) to describe how docking errors can distort the information passed to downstream models. Their approach removes dependence on the relative pose and allows component structures to be reused across candidate pairs.

Validation includes structural analysis of 49 crystal complexes and binding prediction on 45,909 pairs using five epitope-disjoint folds, six retrained baselines, and four structure-prediction tools. The authors report improved prediction performance, greater consistency across structure sources, and lower preprocessing costs. However, the experiments do not isolate whether avoiding docking error causes the improvement, particularly because the complex-input ablation also changes component conformations and graph processing. The binding evaluation also mixes assay-derived and generated negatives, while the exact-peptide split does not establish separation of similar peptides or related receptors. These limitations leave the proposed explanation for the gains less convincing.

Critical strengths

1.Reusing component structures offers a clear computational advantage. EMBER predicts each TCR and pMHC structure once and reuses them across candidate pairs, avoiding a separate complex prediction for every pairing. This matters for screening many receptors against many targets, where structure generation can become expensive. I consider this the paper’s strongest contribution because the practical benefit remains meaningful even if the explanation based on AIC needs further evidence.

2.The reported improvements persist across several structure sources. The authors evaluate four structure-prediction tools and retrain the baselines on the same data splits. EMBER improves the average binding-prediction results across these settings and produces more consistent rankings across structure sources. This gives some confidence that its performance is not dependent on one particular folding pipeline and supports the usefulness of the approach within the evaluated benchmark.

Critical weaknesses

1.The experiments do not show that avoiding AIC explains the improvement. The 49 crystal complexes demonstrate errors in reconstructing known binding interfaces, but do not establish how those errors affect binder–non-binder discrimination. The complex-input ablation also changes component conformations, node features, and cross-molecular message passing together. The missing control is to co-fold the complex, extract its components, and process them separately through the original EMBER architecture. A pose-only perturbation or correction would then test whether docking error actually causes prediction errors. This is central to my assessment because AIC is the paper’s main explanation for its advantage, yet the current experiments cannot distinguish that explanation from other architectural or input differences.

2.The binding benchmark leaves important questions about specificity and generalization unresolved. The negatives mix assay-derived non-binders with generated mismatches, without provenance allowing separate evaluation. Approximately 58% of negative rows come from receptors never observed in positive pairs, raising the possibility that receptor-background differences help distinguish the labels. Meanwhile, the split holds out exact peptide identities but reports neither sequence-similarity filtering nor an audit of receptor overlap across folds. These observations do not prove that EMBER exploits shortcuts, but the paper needs controls that exclude them. A receptor-only baseline, similarity-aware evaluation, and an independent assay-grounded test with positive and negative partners for the same receptors would be particularly informative. Without these, I cannot tell how much the reported gain reflects learning TCR–pMHC specificity that generalizes beyond the curated dataset.

Questions for authors
  1. Do you have results from co-folding the complex, extracting its components, and processing them separately through the unchanged EMBER architecture? The reported complex ablation also changes features and graph connectivity. This missing comparison would help determine whether independent folding matters, while a pose-only intervention would establish whether docking error actually causes the downstream performance loss. Either result could materially change my assessment of the main contribution.

  2. Were any peptide- or receptor-similarity filters applied before constructing the epitope-disjoint folds? Please quantify exact receptor overlap and the nearest-training-sequence similarities of test peptides and CDR3β sequences. This would clarify whether the reported results mainly reflect transfer among related examples or generalization to substantially different receptors and peptides.

  3. How well does a TCR-only model perform under the same protocol? Given that approximately 58% of negative rows involve receptors never observed in positive pairs, this control would help determine how much performance is achievable without identifying the correct pMHC partner. If available, results comparing measured positive and negative partners within the same receptor would provide stronger evidence that EMBER learns pair specificity.

4 comments

We thank the reviewer for naming the control our experiments were missing and for saying why it is central to the assessment. Both concerns are actionable as written, and we take them in turn.


Reviewer's Comment: The missing control is to co-fold the complex, extract its components, and process them separately….// Do you have results from co-folding the complex, extracting its components, and processing them separately through the unchanged EMBER architecture ….

Response: We ran both variations and present the results in the revised paper. For the split control of Table 4, the ESMFold2 co-folded complex is cut into its TCR and pMHC halves, each half is featurized alone, and the two are encoded by the unchanged EMBER architecture, which is the arm described. It keeps the bound conformation and the complex-derived features, and removes only the joint graph and the cross-molecular message passing.

Inputmacro-AUROCmacro-AUPRCmicro-AUROCmicro-AUPRC[email protected]MCC*
Independently folded monomers0.7590.5840.6700.3720.2730.304
Halves of the co-folded complex0.7390.5740.6560.3460.2360.276

The arm loses on all six metrics, by 0.020 of macro-AUROC and 0.026 of micro-AUPRC, so the drop is present once the halves carry the bound conformation. For the pose-only perturbation of Table 5, the docked arm reads the 49 crystal complexes at the crystal relative pose, and then with the receptor translated rigidly by 3 Å and by 8 Å in a fixed random direction. Nothing else changes and every target is a binder, so the correct response is no drop.

Receptor displacementΔ logit over targetsΔ logit over target-checkpoint pairs
3 Å-0.137-0.164
8 Å-0.363-0.440

Table 11 puts the magnitudes in context. The median TCR docking RMSD is 5.87 to 8.09 Å for five of the six tools, and 22% to 51% of targets sit beyond 8 Å. Boltz-2 is the apparent exception at 1.50 Å, and Table 12 separates it, because 39 of the 49 targets precede its training cutoff and on the ten it cannot have memorized its median rises to 6.55 Å with no near-native prediction. The limit of this evidence is that all 49 targets are binders, so Table 5 reports a score shift and not a change in discrimination. What the two controls establish together is that a pose-reading score moves under a pose change alone, at displacements the tools produce, and that the loss is already present when the two halves are processed separately.

Reviewer's Comment: the split holds out exact peptide identities but reports neither sequence-similarity filtering nor an audit of receptor overlap…. // Were any peptide- or receptor-similarity filters applied before constructing the epitope-disjoint folds ….

Response: Appendix F.1 now sets out the composition of STAG's negative samples and the filtering STAG applies to them before release. It passes the dextramer block through ITRAP, excludes the alleles it finds non-specific, and prunes on CDR3alpha and CDR3beta kernel similarity at 0.90 for non-binders and 0.95 for binders. We added no further similarity filter before the split, the folds hold out exact epitope identity, and both audits have been added. Table 15 counts receptor reuse, where a test receptor is reused when it also appears in one of the three training folds.

FoldOf test receptorsOf test rowsOf binder rowsShared (receptor, epitope) pairs
00.5790.6090.3530
10.6610.6840.7550
20.6780.7060.7920
30.7160.7510.8040
40.6790.7060.7150
Mean0.6630.6910.6840

Receptors are reused and no (receptor, epitope) pair is shared with training in any fold. The split is on epitopes by design, and Sec. 5.1 now states the reason, that a screen ranks receptors against one epitope. The same receptor is then a candidate for the next epitope. Table 16 gives the Levenshtein distance from each distinct test sequence to the nearest sequence of the three training folds, pooled over folds, with the last five columns giving the share of test sequences at that distance.

SequenceDistinctMedianMean0123>=4
Test epitope34343.770.0000.2590.1110.0610.569
Test CDR3beta36,12600.820.6600.0690.1300.0910.049

The two axes are not symmetric: no test epitope occurs in training and 0.569 sit at distance 4 or more, while 0.660 of the test CDR3beta sequences occur verbatim in training. For a similarity-aware evaluation, Appendix F.6 reruns the screen on the test epitopes at distance 4 or more. Under ESMFold2 EMBER reaches an enrichment of 2.264 at ten candidates against 2.158 for STAG-LLM and 2.100 for the strongest sequence baseline, the remaining baselines fall between 1.982 and 2.100, and EMBER's lead is maintained at distance.


Reviewer's Comment: How well does a TCR-only model perform under the same protocol….

Response: This control is now Appendix F.7, Table 28, on the rows of Table 6.

Scorermacro-AUROCmacro-AUPRCmicro-AUROCmicro-AUPRC[email protected]MCC*
EMBER0.7590.5840.6700.3720.2730.304
No peptide0.7070.5150.6300.3000.2350.270
Receptor identity only0.4560.1990.6220.2030.0000.210

No peptide is EMBER retrained and tested with one fixed pMHC on every row, so it never sees the peptide's specificity. Receptor identity only is not trained at all. It assigns a row to class one when its TCR is a binder in some row of the three training folds, and scores the class by its positive rate on the validation fold. The receptor background does carry the pooled signal the 58% figure predicts, and the lookup reaches micro-AUROC 0.622 when all epitopes of a fold are pooled. Within an epitope its macro-AUROC is 0.456, below chance. Its [email protected] of 0.000 does not say the lookup cannot discriminate, only that a fixed cut at 0.5 does not suit a score that takes two values, and the reading we would point to is its MCC* of 0.210 against our 0.304. The shortcut therefore does not account for the whole of the model's discrimination. We cannot assemble measured positive and negative partners for the same receptor from this release, and the closest the paper comes is the within-epitope reading of Table 3. We agree that a second sampling strategy would strengthen the evidence, and regenerating its inputs is what we could not complete during the discussion period. The three baselines that take the predicted complex as their input would each need a new complex folded for every new pair. Table 7 puts the 45,909 pairs at 1,602 GPU-hours of inference for TCRmodel2, and the 38,538 non-binding rows are 84% of that. Generating the structures for a new negative set is close to 2,500 GPU-hours on an A100-SXM4-80GB, which is over three months and not likely be done during the rebuttal period.

Reviewer's Comment: an independent assay-grounded test with positive and negative partners for the same receptors would be particularly informative.

Response: We thank the reviewer for asking for an assay-grounded test. Our earlier replies said that regenerating the inputs for a second negative set over the curated collection would cost time beyond the discussion period. We therefore ran the comparison on a smaller collection, TCRvdb (Messemaker et al., 2026), with ESMFold2, the tool the paper's ablations use because it leads every interface and placement metric on the leakage-controlled targets of Table 12. TCRvdb holds 1,779 pairs over five epitopes, and its TCR-pMHC pairs were re-tested experimentally under a standardized readout, so both its binders and its non-binders are determined by experiments. Every row here carries one assay provenance, so the mixture of assay-derived and generated non-binders that Appendix F.1 could not separate does not arise. We evaluate EMBER and other baseline methods under the same protocol as in the paper to guarantee no leakage. A fold here is a single epitope, so a fold's score and the validation signal that selects the checkpoint both rest on one peptide, and the numbers carry more spread than the curated set gives. Each cell is the mean over the five folds, with the standard deviation over seeds in brackets.

MethodAUROCAUPRC[email protected]MCC*
STAG-LLM0.5891 (0.0072)0.5609 (0.0154)0.0618 (0.0512)0.2279 (0.0191)
SageTCR0.5726 (0.0189)0.5632 (0.0083)0.0924 (0.0160)0.2045 (0.0107)
Abeer0.4472 (0.0028)0.4665 (0.0066)-0.0565 (0.0136)0.1211 (0.0216)
TRAP0.5435 (0.0123)0.5313 (0.0075)0.0192 (0.0196)0.1960 (0.0265)
NetTCR-2.20.5519 (0.0188)0.5335 (0.0108)0.0428 (0.0318)0.1921 (0.0338)
ERGO-II0.5074 (0.0336)0.5095 (0.0240)0.0403 (0.0365)0.1861 (0.0484)
EMBER0.5931 (0.0104)0.5708 (0.0077)0.0931 (0.0249)0.2341 (0.0256)

From the results above, EMBER takes the lead compared with the other baselines. The drop in the absolute level of the metrics appears in all the baselines as well, and it is caused by the high variation that comes with a limited number of peptide classes. However, EMBER keeps its lead on a second dataset with different negative samples compared with the STAG dataset.


Reference

Messemaker, M., Kwee, B. P. Y., Moravec, Ž., de Paauw, S., Urbanus, J., Álvarez-Salmoral, D., Geerligs, J., Xie, R., Voogd, R., Morris, B., Guislain, A., Winkler, Y., Richterich, C. A., Mussmann, M., Bakker, M., Steinmetz, M., Iras, M., Marcus, E., Teuwen, J., Perrakis, A., Beijersbergen, R. L., Scheper, W., & Schumacher, T. N. (2026). Functional evaluation of TCR-pMHC pairs at scale allows in silico TCR reactivity prediction. Immunity, 59(10), 2775-2788.e9. https://doi.org/10.1016/j.immuni.2026.09.007

AuthorsOpenReview ↗

Follow-up on Rebuttal

Dear reviewers,

As the rebuttal deadline is approaching, I would like to kindly check whether our responses have addressed your concerns and questions. If there is anything that is still unclear or could benefit from further clarification, please let us know, and we will be happy to elaborate.

Thank you very much for your time and effort in reviewing our work.

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

This paper proposes EMBER, a structure-based framework for TCR–pMHC binding prediction without relying on predicted docked complexes. Motivated by assembly-induced interface corruption (AIC), the method independently encodes TCR and pMHC structures using equivariant GNNs and combines their representations through cross-attention. The authors evaluate the method across multiple folding models and report improvements in binding prediction, robustness, and computational efficiency.

Critical strengths

The paper addresses an important and well-motivated problem. The systematic analysis of AIC across multiple folding models provides useful insights into the limitations of existing complex-based binding prediction methods. The proposed approach offers an interesting alternative to existing methods. Instead of attempting to correct inaccurate docking predictions, EMBER avoids relying on predicted interfaces altogether. I think this is a reasonable and potentially useful direction, particularly when reliable complex structures are difficult to obtain. The experiments are solid and reasonably comprehensive, including multiple retrained baselines, epitope-disjoint splits, different folding tools, and ablation studies. The consistent improvements in macro-level metrics and AUPRC are encouraging, and the gains in ranking stability and computational efficiency further support the practical value of the approach.

Critical weaknesses

I think the methodological novelty is somewhat limited. The proposed architecture largely combines established components, including equivariant GNNs and cross-attention. While the motivation for avoiding unreliable docked complexes is interesting, the method itself is relatively straightforward. The theoretical pose-invariance result also follows naturally from this design. I would be interested in understanding whether the authors see broader implications of this approach beyond TCR–pMHC binding prediction. Although the results are generally positive, some of the performance improvements are relatively modest. For example, micro-AUROC improves from 0.659 to 0.670 over STAG-LLM with ESMFold2. The gains in AUPRC and macro-level metrics are more substantial, but I am not entirely sure how much they would translate into practical improvements in TCR screening. Given the relatively incremental methodology, stronger evidence of practical impact would make the contribution more convincing. I also think the ablation results raise questions about the importance of structural information. Removing structure only reduces micro-AUPRC from 0.372 to 0.349, suggesting that much of the predictive capability is retained without it. While this does not necessarily undermine the proposed approach, it would be useful to better understand how much of the improvement comes from the structural representations rather than the sequence embeddings and interaction architecture.

Overall, I think this is a solid piece of work addressing an important problem, and the empirical analysis is quite thorough. The AIC analysis is interesting, and I appreciate the practical advantages of the proposed approach. My main concerns are the limited methodological novelty and whether the reported improvements are significant enough to establish a strong contribution. I currently lean slightly negative, but I am open to discussion and would consider raising my score if the authors can provide more convincing evidence to address these points.

Questions for authors

Could the authors provide a more direct evaluation of practical candidate prioritization, such as precision@k, recall@k, or hit enrichment under fixed screening budgets? I think this would help clarify the practical significance of the reported improvements.

Could the authors further analyze when and why structural information helps? In particular, are there certain epitopes or challenging cases where the independent structural representations provide substantially greater benefits over the structure-free variant?

How sensitive are the results to the construction of negative examples? Since the dataset includes synthetically constructed nonbinding pairs, it would be useful to understand whether similar improvements hold under different negative sampling strategies or on experimentally measured nonbinders.

4 comments

We thank the reviewer for pressing on whether the reported gains translate into practice, and for three questions that make the test concrete. We take the weaknesses first and then the questions.


Reviewer's Comment: I think the methodological novelty is somewhat limited….

Response: We thank the reviewer for asking where the contribution sits, and we would place it on the problem formulation. We identify and measure a failure mode the literature had not characterized, assembly-induced interface corruption, and Table 11 shows it is not a corner case. On 49 crystal complexes the median receptor placement error is 5.87 to 8.09 Å for five of the six co-folding tools, and 22% to 51% of targets are mis-docked beyond 8 Å. Structure-based predictors are always given that complex, and Table 5 shows the consequence, since displacing the receptor alone moves the score of a model that reads the pose. EMBER keeps the structure of both molecules in full and avoids reading their relative pose. The same reasoning shapes the contact head, where the peptide is pooled on a canonical frame of nine slots that pins the class-I anchors, the free N-terminus, P2 and P-omega, to fixed slots and places surplus residues centrally, so an 8-mer and an 11-mer present the same anchor in the same slot. Our experiments attribute the gain to this constraint and not to an unrelated redesign, the split arm of Table 4, the pose perturbation of Table 5 and the ablation ladder of Table 6 among them, and the gain itself is first on every metric under all four folding tools, with [email protected] ahead of the strongest baseline by 51% to 75%. The construction should carry over beyond TCR-pMHC. In general, monomer folding is more reliable than complex co-folding, so any downstream task that consumes the assembly inherits the same problem, and the argument is strongest where the interface is hard to place, for small, shallow or transient interfaces and for assemblies with few templates. We expect this method could extend to similar pair-binding problems, e.g. protein-peptide and antibody-antigen binding. Removing the relative pose also makes the next step affordable. A conformational ensemble multiplies the preprocessing cost by the number of samples, and our side of that multiplication is O(N+M) monomer folds where a complex-based method would pay O(NM).


Reviewer's Comment: Although the results are generally positive, some of the performance improvements are relatively modest….

Response: We thank the reviewer for pressing on practical significance. A whole-list metric averages over the entire candidate list, which is not where a screen reads, so we added the evaluation that reads the budget a screen actually runs on. The new Table 3 reports it. Under ESMFold2 EMBER reaches precision@10 of 0.361, recall@10 of 0.350 and enrichment at ten candidates of 2.353, against 0.322, 0.327 and 2.100 for STAG-LLM. EMBER is first on all six screening metrics under every folding tool, and its enrichment at ten candidates leads the strongest baseline of each tool, sequence baselines included, by 3.5% to 11.3%. In Table 1 our model also leads [email protected] by 51% to 75%, and ranking stability under a change of folding tool is Kendall's W of 0.840 against 0.454 to 0.666, a separation of 0.174 against a seed-to-seed variation of 0.008 to 0.023. Taken together, these results establish EMBER's lead under deployment conditions at a fixed budget.


Reviewer's Comment: I also think the ablation results raise questions about the importance of structural information….

Response: Table 6 now carries a no-language-model arm over three seeds, at 0.650 macro-AUROC, 0.514 macro-AUPRC, 0.587 micro-AUROC, 0.285 micro-AUPRC, 0.189 [email protected] and 0.213 MCC*. That is the largest single loss in the ladder, 0.109 macro-AUROC and 0.087 micro-AUPRC, against 0.006 and 0.023 for removing the structure. The share of the margin over the strongest baseline attributable to the structure is 2.3% on macro-AUROC against 11.1% on micro-AUPRC and 13.6% on [email protected], so what it buys is the head of the list and the fixed threshold. Appendix F.7 adds the control that removes the pair, where EMBER retrained without sight of the peptide reaches 0.707 macro-AUROC and 0.300 micro-AUPRC against 0.759 and 0.372.

Reviewer's Comment: Could the authors provide a more direct evaluation of practical candidate prioritization….

Response: We added this evaluation as Table 3 for two folding tools, Appendix F.5 for all four and Appendix F.6 for the distant-epitope case. Within each epitope of a test fold the candidates are ranked by the method's score. We report precision, recall and enrichment at ten candidates and at a tenth of the pool, averaged over the epitopes of a fold with at least twenty candidates and one binder. EF@k divides precision at the same k by the pool prevalence, so a random ranking gives 1. Every cell is a mean over five folds and three seeds, with the standard deviation over seeds in brackets.

ESMFold2prec@10recall@10EF@10prec@10%recall@10%EF@10%
EMBER0.361 (0.010)0.350 (0.004)2.353 (0.061)0.340 (0.004)0.250 (0.003)2.285 (0.034)
STAG-LLM0.322 (0.008)0.327 (0.009)2.100 (0.052)0.310 (0.010)0.227 (0.005)2.075 (0.059)
ERGO-II0.332 (0.021)0.324 (0.028)2.162 (0.134)0.306 (0.031)0.224 (0.022)2.039 (0.198)
NetTCR-2.20.321 (0.013)0.303 (0.002)2.093 (0.085)0.307 (0.012)0.224 (0.008)2.052 (0.075)
Abeer0.272 (0.014)0.257 (0.004)1.773 (0.103)0.265 (0.005)0.196 (0.005)1.768 (0.063)
SageTCR0.264 (0.003)0.253 (0.011)1.690 (0.015)0.233 (0.012)0.169 (0.008)1.538 (0.071)
TRAP0.196 (0.005)0.260 (0.017)1.302 (0.028)0.229 (0.005)0.168 (0.005)1.544 (0.043)

Reviewer's Comment: Could the authors further analyze when and why structural information helps….

Response: We evaluate our method's lead on the distant-epitope stratum, which is a hard case for every baseline and for our model, and Appendix F.6 reports the screen on it. We keep only the test epitopes at Levenshtein distance at least four from every epitope of the three training folds, which removes transfer from a near-identical training epitope.

ESMFold2, distance >= 4prec@10recall@10EF@10prec@10%recall@10%EF@10%
EMBER0.348 (0.004)0.336 (0.016)2.264 (0.023)0.299 (0.014)0.221 (0.011)2.018 (0.105)
STAG-LLM0.332 (0.010)0.318 (0.011)2.158 (0.069)0.286 (0.005)0.211 (0.004)1.925 (0.045)
ERGO-II0.323 (0.018)0.299 (0.037)2.100 (0.125)0.284 (0.023)0.209 (0.017)1.901 (0.150)
NetTCR-2.20.305 (0.010)0.271 (0.008)1.982 (0.051)0.260 (0.010)0.191 (0.005)1.744 (0.048)
Abeer0.277 (0.011)0.244 (0.010)1.793 (0.094)0.256 (0.001)0.188 (0.006)1.704 (0.062)
SageTCR0.265 (0.006)0.230 (0.015)1.685 (0.036)0.213 (0.015)0.154 (0.011)1.399 (0.093)
TRAP0.172 (0.010)0.234 (0.014)1.136 (0.063)0.189 (0.017)0.141 (0.015)1.278 (0.129)

EMBER stays first on all six metrics under every folding tool, and the two structure-free baselines stay behind it on all six. The advantage therefore holds where the test epitope is far from everything seen in training, and it is not an artifact of test epitopes which are near-duplicates of training ones. By metric we can be more specific about what the structure buys, since removing it costs 2.3% of the macro-AUROC margin against 11.1% of micro-AUPRC and 13.6% of [email protected], so what it improves is the metrics read at a fixed threshold and the yield of a screen.


Reviewer's Comment: How sensitive are the results to the construction of negative examples….

Response: Our negative samples are the negative samples of STAG, and Appendix F.1 now reports what STAG filters before release. It passes the dextramer block through ITRAP, excludes the alleles it finds non-specific, and prunes the whole set on CDR3alpha and CDR3beta kernel similarity at 0.90 for non-binders and 0.95 for binders. Every method in the paper is trained and evaluated on the same rows, so the comparison between methods does not directly depend on the sampling. An assay-only evaluation is not constructible from this release, because as Appendix F.1 states the file carries no provenance column, so the assay-derived non-binders cannot be separated from the swapped pairs. We agree that a second sampling strategy would strengthen the evidence, and regenerating its inputs is what we could not complete during the discussion period. The three baselines that take the predicted complex as their input would each need a new complex folded for every new pair. Table 7 puts the 45,909 pairs at 1,602 GPU-hours of inference for TCRmodel2, and the 38,538 non-binding rows are 84% of that. Generating the structures for a new negative set is close to 2,500 GPU-hours on an A100-SXM4-80GB, which is over 3 GPU-months.

Reviewer's Comment: whether similar improvements hold under different negative sampling strategies or on experimentally measured nonbinders.

Response: We thank the reviewer for mentioning the negative sampling strategy. Our earlier replies said that regenerating the inputs for a second negative set over the curated collection would cost time beyond the discussion period. We therefore ran the comparison on a smaller collection, TCRvdb (Messemaker et al., 2026), with ESMFold2, the tool the paper's ablations use because it leads every interface and placement metric on the leakage-controlled targets of Table 12. TCRvdb holds 1,779 pairs over five epitopes, and its TCR-pMHC pairs were re-tested experimentally under a standardized readout, so both its binders and its non-binders are determined by experiments. This is the second of the two cases the reviwer names, measured non-binders instead of a synthetic negative sampling, and it is the more trustworthy one. We evaluate EMBER and other baseline methods under the same protocol as in the paper to guarantee no leakage. A fold here is a single epitope, so a fold's score and the validation signal that selects the checkpoint both rest on one peptide, and the numbers carry more spread than the curated set gives. Each cell is the mean over the five folds, with the standard deviation over seeds in brackets.

MethodAUROCAUPRC[email protected]MCC*
STAG-LLM0.5891 (0.0072)0.5609 (0.0154)0.0618 (0.0512)0.2279 (0.0191)
SageTCR0.5726 (0.0189)0.5632 (0.0083)0.0924 (0.0160)0.2045 (0.0107)
Abeer0.4472 (0.0028)0.4665 (0.0066)-0.0565 (0.0136)0.1211 (0.0216)
TRAP0.5435 (0.0123)0.5313 (0.0075)0.0192 (0.0196)0.1960 (0.0265)
NetTCR-2.20.5519 (0.0188)0.5335 (0.0108)0.0428 (0.0318)0.1921 (0.0338)
ERGO-II0.5074 (0.0336)0.5095 (0.0240)0.0403 (0.0365)0.1861 (0.0484)
EMBER0.5931 (0.0104)0.5708 (0.0077)0.0931 (0.0249)0.2341 (0.0256)

From the results above, EMBER takes the lead compared with the other baselines. The drop in the absolute level of the metrics appears in all the baselines as well, and it is caused by the high variation that comes with a limited number of peptide classes. However, EMBER keeps its lead on a second dataset with different negative samples compared with the STAG dataset.


Reference

Messemaker, M., Kwee, B. P. Y., Moravec, Ž., de Paauw, S., Urbanus, J., Álvarez-Salmoral, D., Geerligs, J., Xie, R., Voogd, R., Morris, B., Guislain, A., Winkler, Y., Richterich, C. A., Mussmann, M., Bakker, M., Steinmetz, M., Iras, M., Marcus, E., Teuwen, J., Perrakis, A., Beijersbergen, R. L., Scheper, W., & Schumacher, T. N. (2026). Functional evaluation of TCR-pMHC pairs at scale allows in silico TCR reactivity prediction. Immunity, 59(10), 2775-2788.e9. https://doi.org/10.1016/j.immuni.2026.09.007

Dear reviewer,

Thank you for your positive scoring. If you have any further questions, we are willing to discuss.

Thank you very much for your time and effort in reviewing our work.

Reviewer_rWP2OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes EMBER, a structure-based framework for TCR–pMHC binding prediction that independently encodes TCR and pMHC structures without constructing a docked complex. This design avoids dependence on predicted relative poses and enables reuse of independently predicted structures. Evaluation across four folding tools under epitope-disjoint splits shows consistent improvements over six retrained baselines.

Critical strengths
  1. The motivation is clear and practically relevant: removing dependence on predicted docking poses mitigates error propagation from incorrectly constructed interfaces.
  2. The evaluation covers epitope-disjoint splits, multiple folding tools, retrained baselines, and multiple random seeds, alongside cross-tool ranking stability.
  3. Reusing independently predicted structures reduces structural preprocessing costs, making the approach attractive for large-scale screening.
Critical weaknesses
  1. Binding-related conformational changes remain insufficiently validated. A single independently predicted structure may not capture binding-associated rearrangements, particularly in flexible CDR3 loops and peptide regions. Learned coordinate updates do not necessarily recover these changes. Evaluation stratified by the magnitude of experimentally observed conformational changes would strengthen the claims.
  2. The trade-off between avoiding pose errors and discarding interface information remains unclear. Improvements over predicted-complex baselines do not establish that explicit intermolecular geometry is dispensable. Comparisons using experimental complexes or high-quality docking poses would help distinguish unreliable structural inputs from limitations of interface-based modeling.
  3. The contribution of structural information requires clearer isolation. Removing structure reduces micro-AUPRC from 0.372 to 0.349 and micro-AUROC from 0.670 to 0.649, retaining much of the predictive performance. Further controls are needed to separate gains from structural geometry, protein language model embeddings, and architectural choices.
  4. Negative sampling and threshold selection complicate interpretation. Sampled non-binders may contain unverified negatives or introduce sampling biases. Moreover, test-optimized MCC* is an optimistic reference rather than a deployable result. Validation-selected thresholds and experimentally verified or challenging negatives would provide stronger evidence.
Questions for authors

None

4 comments

We thank the reviewer for reading the ablation table closely and for tying each concern to a specific number in it. We take the four weaknesses in the order they are raised.


Reviewer's Comment: Binding-related conformational changes remain insufficiently validated….

Response: Because all 49 crystal targets are binders and no discrimination metric can be defined on that set, we measure instead what a geometry error of a known magnitude does to a score that reads the complex. Table 5 takes the docked arm of Table 6, reads the 49 crystal complexes at the crystal relative pose, and then moves the receptor rigidly by 3 Å and by 8 Å.

Receptor displacementΔ logit over targetsΔ logit over target-checkpoint pairs
3 Å-0.137-0.164
8 Å-0.363-0.440

These magnitudes are the ones the folding tools produce on the same targets. Table 11 gives a median TCR docking RMSD from 1.50 Å to 8.09 Å across the six tools, and for the five other than Boltz-2 between 22% and 51% of targets sit beyond 8 Å, so the 3 to 8 Å window falls inside the error distribution. Boltz-2 is the apparent exception and Table 12 separates it, because 39 of the 49 targets were released before its training cutoff of 2023-06-01. Its median interface error is 0.68 Å inside that window against 1.72 Å outside it, and on the ten targets it cannot have memorized its median TCR docking RMSD is 6.55 Å with no near-native prediction and last place on contact Jaccard. On the loops our claim does not extend to recovering the bound conformation. EMBER removes the relative pose from the score, and the relative pose is the quantity Table 11 shows to be unreliable, but the internal conformation of each monomer still enters the score, so an unbound CDR3 loop that rearranges on binding remains a limitation of our input. Table 11 also gives median CDR3beta deviations of 2.56 to 4.27 Å and CDR3alpha deviations of 3.96 to 5.44 Å for those five tools, which is the loop error a model reading the loops inside a predicted complex inherits on top of the pose error. A single static monomer cannot represent the change either, and our response to Reviewer XEZJ describes the conformational-ensemble direction we see for it.


Reviewer's Comment: The trade-off between avoiding pose errors and discarding interface information remains unclear….

Response: Two new controls separate the two causes. Table 4 cuts the co-folded complex into its TCR and pMHC halves and feeds the halves to the unchanged architecture, which keeps the bound conformation and the complex-derived features and removes only the joint graph.

Inputmacro-AUROCmacro-AUPRCmicro-AUROCmicro-AUPRC[email protected]MCC*
Independently folded monomers0.7590.5840.6700.3720.2730.304
Halves of the co-folded complex0.7390.5740.6560.3460.2360.276
Full docked complex (Table 6)0.7450.5780.6540.3460.2320.274

The split arm lands within 0.006 of the full docked arm on every metric, so the loss against independently folded monomers is carried by the bound conformation and not by the joint graph. Table 5 is the comparison on experimental complexes, reading crystal complexes at the crystal relative pose and showing that displacing the receptor alone moves the score of a model that reads the pose. Figure 5 reads the question from the other side. Every method gains from the least to the most confident band, and the two sequence baselines see no structure, so their gain measures what the pair alone is worth. The three complex consumers do not exceed that level, averaging 16% against 17%, while EMBER reaches 56% without reading the complex, so the metric is not saturated. These results do not establish that intermolecular geometry is dispensable in principle, and we do not make that claim. Our claim is that the predicted intermolecular geometry available for this task does not improve the prediction, and Tables 4, 5 and 11 support it.

Reviewer's Comment: The contribution of structural information requires clearer isolation….

Response: We agree that the previous ablation left the embedding uncontrolled. Table 6 now carries a no-language-model arm over three seeds, which drops the ESM-C embedding and keeps the geometry and the residue identities.

Armmacro-AUROCmacro-AUPRCmicro-AUROCmicro-AUPRC[email protected]MCC*
EMBER (control)0.7590.5840.6700.3720.2730.304
Remove structure0.7530.5700.6490.3490.2360.283
No language model0.6500.5140.5870.2850.1890.213

Dropping the embedding costs 0.109 macro-AUROC and 0.087 micro-AUPRC, against 0.006 and 0.023 for removing the structure, so the embedding carries most of what the score reads, which Sec. 5.2 now states. The contribution of the structure is smaller and it is uneven. As a share of the margin over the strongest baseline, removing it costs 2.3% of the macro-AUROC margin, 11.1% of the micro-AUPRC margin and 13.6% of [email protected], so what it buys is the head of the ranked list and a usable fixed threshold. For the architectural term, Appendix F.7 reports EMBER retrained and tested with one fixed pMHC on every row, so that it never sees the peptide, and it reaches 0.707 macro-AUROC and 0.300 micro-AUPRC against 0.759 and 0.372 for the full model.


Reviewer's Comment: Negative sampling and threshold selection complicate interpretation….

Response: We agree on the threshold, and the deployable number the paper reports is [email protected], which the conclusion gives as a 51% to 75% lead over the strongest baseline. MCC* is described as an oracle in the caption of Table 18 and in Appendix F.4, and Table 18 also reports MCC@t, fitted on the four folds that are not being scored, so it reads neither the fixed cut nor the scored fold.

tFold-TCRESMFold2TCRmodel2Boltz-2
EMBER, [email protected]0.2510.2730.2670.278
Strongest baseline, [email protected]0.1660.1720.1530.169
EMBER, MCC@t0.2330.2570.2090.263
Strongest baseline, MCC@t0.1700.2240.1690.183

The new Table 3 is read at a fixed budget and not at a threshold, so the practical comparison does not depend on a cut at all. We read MCC mainly as a measure of discrimination, and EMBER is first under the fixed cut and under the fitted one alike. On the negatives, we use the STAG release without resampling, and Appendix F.1 now reports the filtering STAG applies before release. The dextramer block passes ITRAP, which removes false positives from 10x panels, the alleles STAG finds non-specific are excluded, and the whole set is then pruned on CDR3alpha and CDR3beta kernel similarity with no non-binder above 0.90 and no binder above 0.95. We agree that a second negative sampling strategy would make the evidence stronger, and its inputs are what we could not produce in time. The three baselines that take the predicted complex as their input would each need a new complex folded for every new pair, and we were not able to complete that synthesis within the discussion period.

Reviewer's Comment: experimentally verified or challenging negatives would provide stronger evidence.

Response: We thank the reviewer for mentioning the negative sampling strategy. Our earlier replies said that regenerating the inputs for a second negative set over the curated collection would cost time beyond the discussion period. We therefore ran the comparison on a smaller collection, TCRvdb (Messemaker et al., 2026), with ESMFold2, the tool the paper's ablations use because it leads every interface and placement metric on the leakage-controlled targets of Table 12. TCRvdb holds 1,779 pairs over five epitopes, and its TCR-pMHC pairs were re-tested experimentally under a standardized readout, so both its binders and its non-binders are determined by experiments. These non-binders are measured and not sampled, and we report [email protected] beside MCC* so the comparison can be read at a fixed cut and not only at the optimistic reference. We evaluate EMBER and other baseline methods under the same protocol as in the paper to guarantee no leakage. A fold here is a single epitope, so a fold's score and the validation signal that selects the checkpoint both rest on one peptide, and the numbers carry more spread than the curated set gives. Each cell is the mean over the five folds, with the standard deviation over seeds in brackets.

MethodAUROCAUPRC[email protected]MCC*
STAG-LLM0.5891 (0.0072)0.5609 (0.0154)0.0618 (0.0512)0.2279 (0.0191)
SageTCR0.5726 (0.0189)0.5632 (0.0083)0.0924 (0.0160)0.2045 (0.0107)
Abeer0.4472 (0.0028)0.4665 (0.0066)-0.0565 (0.0136)0.1211 (0.0216)
TRAP0.5435 (0.0123)0.5313 (0.0075)0.0192 (0.0196)0.1960 (0.0265)
NetTCR-2.20.5519 (0.0188)0.5335 (0.0108)0.0428 (0.0318)0.1921 (0.0338)
ERGO-II0.5074 (0.0336)0.5095 (0.0240)0.0403 (0.0365)0.1861 (0.0484)
EMBER0.5931 (0.0104)0.5708 (0.0077)0.0931 (0.0249)0.2341 (0.0256)

From the results above, EMBER takes the lead compared with the other baselines. The drop in the absolute level of the metrics appears in all the baselines as well, and it is caused by the high variation that comes with a limited number of peptide classes. However, EMBER keeps its lead on a second dataset with different negative samples compared with the STAG dataset.


Reference

Messemaker, M., Kwee, B. P. Y., Moravec, Ž., de Paauw, S., Urbanus, J., Álvarez-Salmoral, D., Geerligs, J., Xie, R., Voogd, R., Morris, B., Guislain, A., Winkler, Y., Richterich, C. A., Mussmann, M., Bakker, M., Steinmetz, M., Iras, M., Marcus, E., Teuwen, J., Perrakis, A., Beijersbergen, R. L., Scheper, W., & Schumacher, T. N. (2026). Functional evaluation of TCR-pMHC pairs at scale allows in silico TCR reactivity prediction. Immunity, 59(10), 2775-2788.e9. https://doi.org/10.1016/j.immuni.2026.09.007

Dear reviewer,

Thank you for your positive scoring. If you have any further questions, we are willing to discuss.

Thank you very much for your time and effort in reviewing our work.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Rebuttal/Revision Response

We sincerely thank the reviewers for their feedback and valuable insights, which have improved the empirical validation, the controls, and the clarity of our work.

This paper introduces EMBER, a pose-independent framework for TCR-pMHC binding prediction that scores two independently folded monomers and never forms their relative pose. We summarize the key contributions and the revisions made during the rebuttal.


Key Contributions

Assembly-induced interface corruption (AIC): we identify a previously uncharacterized failure mode in structure-based TCR-pMHC prediction, in which pose errors corrupt the residues and contacts exposed to downstream predictors. We measure it across six co-folding tools on 49 held-out crystal complexes (Sec. 3, Table 11).

Pose-free structure-based prediction: we propose EMBER, which folds the TCR and the pMHC independently as unbound monomers. Each molecule is encoded as an SE(3)-invariant residue graph with a shared encoder, and binding is scored by bidirectional attention over the two representations. The score is invariant to independent rigid transforms of either monomer (Sec. 4, Definition 4.1, Proposition 1).

Preprocessing complexity: we fold O(N+M) monomers where a complex-based predictor needs O(NM) complex predictions, which cuts preprocessing by up to 19.4x on the same evaluation set (Appendix F.9).

Accuracy and stability across folding tools: we compare against six published baselines retrained on identical data and splits, and EMBER is first on every metric under all four folding tools. It improves MCC at the fixed threshold of 0.5 by 51% to 75% over the strongest baseline. Its ranking is also the least disturbed by a change of folding tool, at Kendall's W of 0.840 against 0.454 to 0.666 (Sec. 5.1, Tables 1 and 2).


Major Revisions

1. Evaluation of practical utility

New Sec. 5.1, Table 3: candidate prioritization under a fixed screening budget. It reports precision, recall and enrichment at ten candidates and at a tenth of the pool, for two folding tools in the main text. reviewer XEZJ

New Appendix F.5: the same evaluation for all four folding tools. reviewer XEZJ

New Appendix F.6: the same evaluation restricted to test epitopes at Levenshtein distance at least four from every training epitope. EMBER stays first on all six metrics under every tool, and its enrichment at ten candidates leads the strongest baseline by 1.3% to 8.2%. reviewer XEZJ reviewer Svys

2. Isolating the pose from everything else the docked arm changes

New Table 4: the split control, in which the co-folded complex is cut into its TCR and pMHC halves and the halves are encoded by the unchanged EMBER architecture. The arm lands within 0.006 of the full docked arm on every metric. reviewer Svys reviewer rWP2

New Table 5: a pose-only perturbation on the 49 crystal complexes, every target a binder. Translating the receptor by 3 Å lowers the docked arm's logit by 0.137 and 8 Å lowers it by 0.363. reviewer Svys reviewer rWP2

Revised Table 6: a no-language-model arm over three seeds, which costs 0.109 macro-AUROC and 0.087 micro-AUPRC against 0.006 and 0.023 for removing the structure. reviewer rWP2 reviewer XEZJ

3. Audits of the split, the negatives and the partner

New Appendix F.1, Table 15: receptor reuse across folds, 0.663 of test receptors and 0.691 of test rows on average, with 0 shared (receptor, epitope) pairs in every fold. reviewer Svys

New Appendix F.1, Table 16: the Levenshtein distance from every distinct test sequence to the nearest training sequence, for test epitopes and test CDR3beta separately. reviewer Svys

Revised Appendix F.1: the filtering STAG applies before release. It passes the dextramer block through ITRAP, excludes the alleles it finds non-specific, and prunes on CDR3alpha and CDR3beta kernel similarity at 0.90 for non-binders and 0.95 for binders. all reviewers

New Appendix F.7, Table 28: two scorers that do not see the partner, EMBER retrained with one fixed pMHC on every row and an untrained receptor-identity lookup. reviewer Svys reviewer rWP2


We once again thank the Area Chair and reviewers for their valuable time and constructive feedback. These revisions strengthen the empirical evidence for the pose-independent design, separate the contribution of the structure from that of the language model, and characterize the receptor reuse and the sequence distance that the epitope-disjoint split leaves. All revisions are in the updated manuscript, with changes highlighted in red.

We thank the reviewers for mentioning the negative sampling strategy. All three reviews ask for evidence from outside the curated collection, in different words, and this comment reports it.

Our earlier replies said that regenerating the inputs for a second negative set over the curated collection would cost time beyond the discussion period. We therefore ran the comparison on a smaller collection, TCRvdb (Messemaker et al., 2026), with ESMFold2, the tool the paper's ablations use because it leads every interface and placement metric on the leakage-controlled targets of Table 12. TCRvdb holds 1,779 pairs over five epitopes, and its TCR-pMHC pairs were re-tested experimentally under a standardized readout, so both its binders and its non-binders are determined by experiments. We evaluate EMBER and other baseline methods under the same protocol as in the paper to guarantee no leakage. A fold here is a single epitope, so a fold's score and the validation signal that selects the checkpoint both rest on one peptide, and the numbers carry more spread than the curated set gives. Each cell is the mean over the five folds, with the standard deviation over seeds in brackets.

MethodAUROCAUPRC[email protected]MCC*
STAG-LLM0.5891 (0.0072)0.5609 (0.0154)0.0618 (0.0512)0.2279 (0.0191)
SageTCR0.5726 (0.0189)0.5632 (0.0083)0.0924 (0.0160)0.2045 (0.0107)
Abeer0.4472 (0.0028)0.4665 (0.0066)-0.0565 (0.0136)0.1211 (0.0216)
TRAP0.5435 (0.0123)0.5313 (0.0075)0.0192 (0.0196)0.1960 (0.0265)
NetTCR-2.20.5519 (0.0188)0.5335 (0.0108)0.0428 (0.0318)0.1921 (0.0338)
ERGO-II0.5074 (0.0336)0.5095 (0.0240)0.0403 (0.0365)0.1861 (0.0484)
EMBER0.5931 (0.0104)0.5708 (0.0077)0.0931 (0.0249)0.2341 (0.0256)

From the results above, EMBER takes the lead compared with the other baselines. The drop in the absolute level of the metrics appears in all the baselines as well, and it is caused by the high variation that comes with a limited number of peptide classes. However, EMBER keeps its lead on a second dataset with different negative samples compared with the STAG dataset.


Reference

Messemaker, M., Kwee, B. P. Y., Moravec, Ž., de Paauw, S., Urbanus, J., Álvarez-Salmoral, D., Geerligs, J., Xie, R., Voogd, R., Morris, B., Guislain, A., Winkler, Y., Richterich, C. A., Mussmann, M., Bakker, M., Steinmetz, M., Iras, M., Marcus, E., Teuwen, J., Perrakis, A., Beijersbergen, R. L., Scheper, W., & Schumacher, T. N. (2026). Functional evaluation of TCR-pMHC pairs at scale allows in silico TCR reactivity prediction. Immunity, 59(10), 2775-2788.e9. https://doi.org/10.1016/j.immuni.2026.09.007

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.