acceptodds
Under review as a conference paper at ICLR 2027

HIFT-TCR: Interface-Structured TCR Library Screening under Epitope-Family Shift

Abstract

Emerging pathogen variants and tumor neoantigens motivate screening characterized T-cell receptor (TCR) libraries against new peptide families. We introduce the Hierarchical Interface Field Transformer for T-cell Receptor screening (HIFT-TCR), a structure-based graph Transformer for modeled interfaces between TCRs and peptide–major histocompatibility complexes (pMHCs). It combines typed local interactions, molecular hierarchy, and separate TCR–peptide and TCR–MHC compatibility. Five-fold cross-validation on the STAG pan-peptide benchmark holds out complete peptide components defined at 75% sequence identity. Across 25 matched seed–fold runs, HIFT-TCR achieves 0.5874 AUROC, 0.2679 AUPRC, and 0.5112 standardized partial AUROC at a false-positive rate of 0.1, leading six same-split sequence, language-model, and structure baselines across all three mean metrics. Performance is higher without an exact paired-CDR3 training match. Retrained ablations support the combined objective, interface topology, and hierarchy, while interventions identify TCR residues and TCR–peptide edges as consistent support for positive predictions. These findings support interface-structured receptor prioritization for experimental binding validation under peptide-family shift.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
TCR–pMHC binding prediction, structure-based modeling, graph Transformer, hierarchical representation learning, molecular interfaces, epitope-family shift

Reviews

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

This paper proposes a structure-based graphTransformer for modeled interfaces betweenTCRs and peptide–major histocompatibility complexes. It combines typed local interactions,molecular hierarchy, and separate TCR–peptide and TCR–MHC compatibility. The experimental results are quite abundant and have verified its effectiveness.

Critical strengths
  1. This paper proposes a hierarchical GNN architecture combining edge-gated local message passing with role-aware multi-hop attention for TCR-pMHC binding prediction.
  2. A role-context broadcast mechanism is designed that fuses residue-level representations with role-level and complex-level context.
  3. The comparison results show that the proposed method has achieved an improvement in effectiveness.
Critical weaknesses
  1. Introduction and related work lack computer-oriented focus and clear problem positioning.
  2. Key thresholds and design choices lack justification, sensitivity analysis, and ablation.
  3. Baseline selection lacks recent comparison methods.
  4. Reported performance is low and not interpretable without prevalence and baselines.
Questions for authors
  1. The introductions are all background information on the medical field. However, this is a computer-oriented conference. The author should at least describe the development of this field or why AI is used, or summarize the work and problems of cross-research. Additionally, the structure of this part is not perfect. It is recommended to refer to published articles.
  2. The summary of related work is not obvious, making it difficult to accurately obtain the starting point of this work and the problems to be solved.
  3. The author used 8 Å as the trimming threshold for the interface residues of TCR-pMHC, and 5 Å as the reconstruction target for contact loss. However, the selection of these two thresholds was not given any basis. 8 Å is commonly used in protein interface prediction, but the amino acid side chains length distribution and the range of hydrogen bonds/salt bridges in the TCR-pMHC interface are not exactly the same as those in general protein interfaces. It is recommended to add a threshold sensitivity analysis (such as the impact of 6/7/8/9/10 Å on the final AUPRC); or provide a theoretical explanation.
  4. The restoration mechanism in the clipping rules may introduce distribution bias. The restoration mechanism will cause inconsistent graph density distributions for different complexes. Some complexes are sparse graphs strictly cut according to distance, while others are almost completely retained dense graphs. This will make the subsequent graph neural networks face very different graph structures on different samples. Will it not affect the recognition results? Is it more reasonable to use other mechanisms, such as retaining the top-k nearest residues, with k adapting to the size of the complex?
  5. What is the specific role of the edge gating in Equation 3? If a visualization or statistics can be provided to explain the distribution differences of the gating values on "peptide-MHC edge" and "TCR-peptide edge", it will significantly improve the interpretability of the method.
  6. The odd-even division of multi-hop attention is hardcoded. It is recommended to at least conduct an ablation study: let all heads look two hops, or let the heads learn the hop number preference on their own, and compare the effect of the hardcoded division.
  7. Equation 8 uses role bias tables to add a role-specific vector to each residue, and Equation 14 concatenates the role context to the residue representation. Both are injecting role information and have functional overlap. It is recommended to conduct an ablation study.
  8. The ERGO-IILSTM and ERGO-IIAE models are not introduced in the report and have not appeared in related work. Why is this method compared? Also, in 2026, are there updated related algorithms? The author should compare them.
  9. The reported prediction performance is low. AUROC is 0.5874, only slightly higher than the random baseline (0.5); AUPRC is 0.2679, but in the case of unknown positive sample proportion in the test set, it is impossible to determine whether this value has practical discriminative ability.
Reviewer_N21NOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents HIFT-TCR for screening binding candidates from characterized T-cell receptor (TCR) libraries against unseen peptide sequence-identity families. It represents modeled TCR–peptide–MHC complexes as residue-level graphs and combines edge-gated message passing, role-aware interface-shell attention, hierarchical context aggregation. On peptide-component-disjoint splits of the STAG benchmark, HIFT-TCR achieves an AUROC of 0.5874, an AUPRC of 0.2679, and a standardized [email protected] of 0.5112, outperforming the reported baselines in mean performance.

Critical strengths
  1. Well-motivated evaluation setting. Screening characterized TCR libraries against unseen peptide families has clear practical relevance. The authors recognize that random splits may overestimate generalization due to peptide similarity and adopt peptide-component-disjoint splits.

  2. Detailed architectural specification. The paper describes graph construction, node and edge features, attention, hierarchical aggregation, and training objectives in sufficient detail to support technical assessment.

Critical weaknesses
  1. The architecture appears over-engineered. HIFT-TCR combines many components, but several yield marginal gains in the reported ablations. For example, removing the shell-distance bias changes AUROC by only 0.0002, and removing the interface gate changes it by 0.0018. The paper does not establish that these differences are statistically significant. The necessity of the additional complexity therefore remains unclear.

  2. Potential data leakage through structural templates. The three-dimensional TCR–peptide–MHC structures are generated using structural modeling pipelines that rely on existing templates. Although the evaluation holds out peptide sequence-identity components, the paper does not clarify whether the templates used to model test complexes are highly similar to structures associated with training examples. Such template-level overlap could introduce structural information about held-out complexes and compromise the intended evaluation of generalization to unseen peptide families.

Questions for authors
  1. Proximity reconstruction with target information in the input. Why does the proximity objective improve binding prediction when the input edge representation already contains the target-bearing 5 Å indicator? Could the authors report an ablation that masks or removes this feature from the decoder input to clarify the source of the gains?

  2. Sensitivity to structural uncertainty. How sensitive is HIFT-TCR to the quality of modeled TCR–pMHC structures? Could the authors evaluate prediction stability across independently generated structures or different structural quality levels?

  3. Node and edge feature ablations. Could the authors ablate node features, including amino-acid identity, physicochemical descriptors, and molecular-role encodings, as well as edge features, including distance-derived and physicochemical features? These experiments would clarify which input information is essential and whether a simpler representation could achieve comparable performance.

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

The authors propose HIFT-TCR for screening characterized TCR libraries against new peptide families. The model operates on residue graphs constructed from modeled TCR–peptide–MHC class I complexes. It combines typed local messages, role-aware attention over one- and two-hop neighborhoods, molecular-role and whole-complex pooling, and separate TCR–peptide and TCR–MHC compatibility terms. Training combines binding classification, edge-proximity self-reconstruction, and a global pairwise ranking loss. Peptides are grouped into 249 sequence components at 75% normalized Levenshtein identity. Across five folds and five seeds, HIFT-TCR achieves 0.5874 AUROC, 0.2679 AUPRC, and 0.5112 standardized partial AUROC at FPR ≤0.1, exceeding all six reported baselines on the evaluation set. Retrained ablations support contributions from the training objective, interface connectivity, and hierarchy. However, the evaluation does not directly establish receptor ranking within individual new pMHC queries, and the contribution of the more distinctive attention and gating choices remains unclear.

Critical strengths
  1. The peptide-family split is explicit and appropriate for the stated generalization question. The authors hold out complete sequence components and report that none of 47,096 cross-fold peptide comparisons reaches 75% identity; the maximum is 0.7273. All methods use the same frozen partitions. Allowing receptor recurrence is reasonable for screening an already characterized library, and the separate familiarity analysis makes this choice transparent. This is substantially more informative than a random interaction split.

  2. The overall improvement is supported by matched runs and meaningful retrained controls. Relative to STAG, HIFT-TCR improves mean AUROC by 0.0752 and AUPRC by 0.0428. The paper includes five seeds per fold, component removals, individual loss ablations, and attention-support comparisons that keep the surrounding model fixed. Removing interface edges or hierarchy reduces performance across all three metrics. The authors also correctly distinguish frozen-model sensitivity from biological causality and describe the proximity objective as self-reconstruction rather than independent structural supervision.

Critical weaknesses
  1. The reported evaluation does not establish the receptor-screening benefit claimed by the paper. The intended use is to select receptors for a particular new peptide–MHC target. Table 1 instead computes metrics across complexes from many different targets within each fold. A model can improve this pooled ranking by distinguishing peptide, MHC, or source groups with different positive prevalences, without substantially improving receptor ordering for any one target. Holding out peptide families does not eliminate this distinction. The training objective reinforces its relevance: Appendix A explicitly states that the ranking loss compares positives and negatives globally within a minibatch, rather than within the same peptide or pMHC query.

2.The experiments support the integrated recipe more strongly than the distinctive architectural explanation. However, the largest AUROC decrease comes from replacing the combined objective with BCE alone, at 0.0288. Removing hierarchy costs 0.0145, whereas removing the interface gate costs only 0.0018 and removing shell-distance bias costs 0.0002. The attention comparison is similarly qualified. One-hop RISA achieves the same rounded AUPRC as the proposed one/two-hop version, differs by only 0.0021 AUROC, and has slightly higher partial AUROC. These results leave the preference for complementary first- and second-shell attention uncertain. Paired differences and uncertainty are needed; the reported marginal standard deviations neither establish nor rule out a reproducible benefit.

Questions for authors
  1. Does HIFT-TCR improve receptor ranking within a fixed, previously unseen pMHC? Using the existing held-out predictions, can you report per-pMHC average precision, enrichment relative to each query’s prevalence, and precision or recall at a stated screening budget. Compare HIFT-TCR and STAG on identical candidate sets, report the distribution across queries, and state how many queries contain enough positives and negatives for evaluation. Use unique receptor candidates within each query and report results separately for observed versus swapped negatives where provenance is available. This would directly test the proposed application without requiring a new experimental study. If the intended application instead pools candidates across many different targets, that selection setting should be made explicit.

  2. Which dense-attention result is correct? Figure 2 reports an AUROC decrease of 0.0028 for dense attention, implying 0.5846. Table 3 reports dense Transformer AUROC of 0.5762, a decrease of 0.0112. If these are different configurations, please specify their differences; otherwise. Can you also provide paired differences for RISA-2 versus RISA-1, dense attention, and the no-gate variant. Uncertainty should account for the shared peptide components and repeated seeds. The 25 runs provide useful replication, but they are not 25 independent biological test cohorts.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.