acceptodds
Under review as a conference paper at ICLR 2027

HATIS: Hierarchical Atomistic Graph Transfer and Interface-Aware Stable Learning for TCR–pMHC Recognition under Distribution Shifts

Abstract

Candidate screening for T-cell immunotherapy requires accurate prediction of T cell receptor (TCR) recognition of peptide–MHC (pMHC) complexes. Such screening often involves peptides and TCRs not represented in the training data, requiring models to generalize under shifts in peptide and TCR distributions. Despite recent advances, generalization across peptide distributions remains challenging, particularly when test peptides differ substantially from those used for training. A key limitation underlying this challenge is the limited coverage of TCR–pMHC interactions in training data, which restricts the learning of transferable recognition patterns. Models may also exploit spurious correlations arising from the frequent co-occurrence of molecular patterns across peptides, MHC molecules, and TCRs in training data, learning predictive shortcuts that fail under distribution shifts. We therefore propose HATIS, a framework integrating hierarchical atomistic graph transfer and interface-aware stable learning. To alleviate limited triplet supervision, we progressively transfer molecular representations and interaction knowledge across monomer, pairwise, and triadic levels to support TCR–pMHC recognition. To reduce reliance on spurious interaction correlations, we introduce interface-aware stable learning at both pairwise and triadic levels, using the molecular composition of interaction clusters to guide sample reweighting and interface-gated pooling. For systematic evaluation, we construct a benchmark covering peptide-only, TCR-only, and joint peptide–TCR distribution shifts, complemented by evaluation on an external test set. HATIS improves AUPR over the strongest baseline by 12.86–13.98 percentage points under peptide-only and joint shifts and by 4.65 points on the external test set, while remaining competitive under TCR-only shifts.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
TCR–pMHC binding prediction, distribution shifts, graph representation learning, transfer learning, stable learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Q1u2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies generalization of TCR–pMHC binding prediction under distribution shifts (unseen peptides, unseen TCRs, and jointly unseen peptide–TCR pairs). The proposed framework, HATIS, consists of two components: Hierarchical Atomistic Graph Transfer and Interface-Aware Stable Learning (IASL). The authors construct a unified benchmark with cluster-level hold-outs (MMseqs2, 0.4 identity, 0.7 coverage) covering peptide-only, TCR-only, and joint shifts, plus an external test set (TCRCube). HATIS reportedly improves AUPR over the strongest baseline by 12.86–13.98 percentage points under peptide-only and joint shifts, by 4.65 points on the external test set, and remains competitive under TCR-only shifts.

Critical strengths
  1. Well-motivated problem and a benchmark of independent value. Unifying peptide-only, TCR-only, and joint shifts in one task with cluster-level hold-outs (0.4 sequence identity) plus an external test set with experimentally supplied labels is more systematic than the protocols of TChar, TITAN, EPACT, or PISTE. The benchmark itself is a contribution the community can build on even independent of the method.
  2. A sensible and empirically supported strategy for alleviating triplet-data scarcity. Lower-order supervision is orders of magnitude richer than triplet data (1.35M peptide monomers, 341K pMHC binding pairs, 170K affinity records), and the ablations show that removing all pretraining degrades performance substantially (external AUPR 26.39 → 17.15), with pMHC transfer being the most critical stage. The narrative and the data agree here.
  3. IASL is a non-trivial adaptation of stable learning. The observation that not all inter-cluster dependencies are spurious. And the use of molecular composition to differentiate dependence penalties is a natural but previously unexplored way to inject domain structure into StableNet/StableGNN-style reweighting. The interface gate is inexpensive and remains active at inference.
  4. Rich analytical experiments. The representation-geometry probes (neighborhood gap, silhouette, centroid and kNN transfer AUROC) mechanistically support the claim that peptide shifts are harder and that HATIS preserves more transferable binding-class structure. Appendix F.1 partially controls for data-access fairness (EPACT + binding), and sensitivity analyses over λ_w and K are provided.
  5. Good reproducibility. Anonymized code and the benchmark are submitted; data sources, hyperparameters, software/hardware environment, and proofs are documented in detail.
Critical weaknesses

1."Atomistic" and "interface-aware" framing overstates what the model does, which is a central concern for this domain. The inputs are 2D covalent graphs built from sequences by RDKit; there are no 3D coordinates, no PDB/AlphaFold structures, and no geometric information of any kind (Appendix E.2: "Structure supervision is disabled"). A covalent graph is essentially an enriched sequence representation and carries no spatial information beyond the sequence. More importantly, the "interface score" is merely the balance of DiffPool assignment mass across the two molecules, and it is never validated against actual contact atoms/residues in solved or predicted TCR–pMHC complexes. The field's core bottleneck is precisely the lack of ternary structural data; this work sidesteps rather than addresses it, and no structure-aware method is included among the baselines. The title and framing risk misleading readers about the nature of the contribution.

  1. The theoretical analysis provides almost no generalization guarantee. Theorem 1's recombination distribution Q_comp is an analytical construction that, by the authors' own statement, need not coincide with the target distribution; the comparison J_T ≤ J_unif follows trivially from 0 ≤ T ≤ 1, and the "strictly smaller" claim is an immediate consequence of the definitions. The result is motivational at best; presenting it as "theoretical analysis" inflates its weight.

  2. Ablation results contradict the design narrative. (a) Removing pTCR transfer improves peptide-only and joint-shift performance (AUPR 46.99 vs 46.07; 46.53 vs 44.54), one branch of the three-stage design is a net negative on th headline settings, yet the Full model retains it, and only a qualitative explanation is offered. (b) Removing IASL entirely (II+III) also improves TCR-only AUPR over Full (66.79 vs 66.24). (c) Under TCR-only shifts, HATIS trails EPACT by 5.08 AUPR points, so "remaining competitive" is an optimistic characterization.

  3. Residual fairness concerns in baseline comparison. In the main table, HATIS exclusively uses large auxiliary corpora (monomer data, affinity, pTCR binding); the Appendix F.1 control augments only EPACT. UniPMT (Nat. Mach. Intell. 2025) is cited in the related work but not included as a baseline, and deepAntigen (Nat. Commun. 2025), an atom-level graph method and one of this paper's data sources, is also missing. In addition, the pTCR corpus covers only 676 peptides (vs 124K TCRs), an extreme imbalance that may itself explain the negative transfer, but no diagnosis is provided.

  4. Internal negatives are randomly recombined "unobserved combinations," not experimentally confirmed non-binders. With 5:1 recombination that explicitly does not preserve pMHC pairing, some negatives are likely false; the external set mitigates this for one setting, but absolute metric values on the three internal settings should be interpreted with caution.

  5. No unseen-MHC evaluation. The triplet training pool contains only 33 MHCs and 321 peptides, peptide diversity in training is itself extremely low. The practically important scenario of novel/rare patient alleles is not evaluated, even though the MHC sequence representation would in principle support it.

Questions for authors
  1. Structural validation: Have you checked whether the learned cross-molecular (top-k) edges or high-interface-score clusters correspond to actual contact atoms/residues in solved TCR–pMHC complexes (PDB) or AlphaFold-predicted structures? Appendix E.2 mentions "Structure supervision is disabled," implying the code supports it. What happens when it is enabled, and why is it not used?
  2. Interpretability of the interface gate: Do interface scores correlate with known functional regions (e.g., CDR loops, pMHC contact zones)? Could you provide 1–2 case studies to justify the "interface-aware" naming?
  3. Model selection: Given that pTCR transfer is a net negative on the headline settings, why is the Full model the default? What is the basis for reporting Full rather than w/o pTCR transfer in the main table? Were all configurations (including K = 8) selected on the validation split or the test split? Please state the selection protocol explicitly to rule out test-set peeking.
  4. Disentangling IASL's gains: Could you add an ablation with uniform penalties (reducing to StableGNN-style reweighting) plus the interface gate, to separate how much of the gain comes from composition weighting T, from reweighting itself, and from the gate?
  5. Negative-sample quality: What is the estimated false-negative rate in the recombined negatives? Have you tried calibration or fine-tuning using the experimentally validated negative labels in the external set?
  6. Unseen MHC: Why is there no unseen-MHC evaluation? How does the model perform on novel alleles given that the training pool contains only 33 MHCs?
  7. Missing baselines: Why are UniPMT and deepAntigen not compared against? The latter is an atom-level graph method and one of your data sources, so the comparison should be nearly free.
  8. Effect of training coverage: With only 321 training peptides, how much of the peptide-only-shift improvement compensates for the tiny training peptide pool versus reflecting genuinely transferable mechanisms? Could you report performance stratified by training-peptide frequency?
9 comments
AuthorsOpenReview ↗

Response to Reviewer Q1u2 (1/8)

We thank the reviewer for the constructive comments.

W1 & Q1–Q2: Atomistic representation, interface scores, and structural supervision

Meaning of “atomistic.” We acknowledge that “atomistic” can suggest three-dimensional modeling in some contexts. Here, it refers to the resolution of our representation and computation: we construct atom–bond graphs from sequences, explicitly encoding atom attributes, chemical bonds, and intra-residue topology. These graphs contain no three-dimensional coordinates. This terminology has precedent: deepAntigen (Nat. Commun. 2025) [1] describes atomic-level modeling and includes a sequence-based prediction branch built on atom graphs. We therefore consider atom-level modeling an appropriate description, while agreeing that its distinction from geometric modeling should be explicit.

Meaning and interpretation of “interface-aware” (Q2). The interface score measures the balance of DiffPool assignment mass from the two interacting components within a cluster; at the triplet stage, these are pMHC and TCR. It describes cross-component mixing, rather than physical contact probability. Composition balance guides readout weighting, while differences between cluster compositions modulate HSIC regularization. The motivation is to reduce training-specific co-occurrence dependencies between interaction patterns with different molecular compositions; larger composition differences therefore receive stronger decorrelation penalties.

This is the intended meaning of “interface-aware” in our aggregation and stable-learning design. We do not claim that high-scoring clusters recover physical interfaces or identify functional regions such as CDR loops.

Why structure supervision is disabled (Q1). We agree that the scarcity of experimental ternary structures is an important bottleneck. The code includes an optional structure-supervision module to support future fine-tuning with predicted or experimentally determined crystal structures. This module is disabled in the model reported in this paper, as stated in Appendix E.2. Incorporating such structural supervision into the sequence-input binding prediction framework is a direction for future work.

Application scope and structure-aware comparisons. Our target applications include personalized neoantigen screening and antigen-specific TCR screening: identifying potentially recognized combinations from candidate peptides, patient MHCs, and TCR repertoires. These applications typically provide sequences before corresponding complex structures are available. Some datasets contain only TCR CDR3 sequences, without complete variable regions or paired chains. Predicting a complete TCR–pMHC structure from such records generally requires additional sequence information or reconstruction of the missing regions.

Our comparisons therefore focus on binding prediction without complex-structure inputs. Effectively incorporating experimental or predicted geometry, and converting it into reliable improvements in triplet binding prediction, remains an important direction for future work.

References

[1] Que, J., et al. Identifying T cell antigen at the atomic level with graph convolutional network. Nat. Commun. 2025, 16, 5171. DOI: 10.1038/s41467-025-60461-6.

AuthorsOpenReview ↗

Response to Reviewer Q1u2 (2/8)

W2: Scope of the analysis

We agree that JT≤JunifJ_T\leq J_{\mathrm{unif}} and its strictness conditions follow directly from the definitions. However, the ordering of the bounds and the validity of the weighted risk bound are separate questions. The proof first controls the risk difference through ∑k<ℓqkℓ2Hkℓ\sum_{k<\ell}q_{k\ell}^{2}H_{k\ell}, then uses qkℓ2≤Tkℓq_{k\ell}^{2}\leq T_{k\ell} under the composition-guided recombination mechanism to obtain the bound involving JTJ_T. Thus, arbitrary weights in [0,1][0,1] cannot simply replace TT and retain the same guarantee. The analysis therefore links composition-weighted dependence to a specified risk difference, beyond comparing two penalty values. We acknowledge that this mechanism is an analytical construction and does not provide a guarantee for the actual target distribution. To reflect this scope, we have removed the “Theorem 1” heading in the revision and presented the result under “A supporting analysis of composition-guided recombination”, with practical benefits assessed through empirical comparisons.

AuthorsOpenReview ↗

Response to Reviewer Q1u2 (3/8)

W3 & Q3: Ablation trade-offs and the choice of the Full model

pTCR transfer and the Full model. Given limited triplet supervision, we use both pMHC and pTCR data to provide lower-order interaction knowledge and report this complete hierarchical framework as the Full configuration. Table 2(A) shows that the benefit of pTCR transfer depends on the evaluation setting: removing it improves peptide-only and joint-shift AUPR by 0.92 and 1.99 percentage points, respectively, but lowers external AUPR from 26.39% to 24.51%. Thus, the branch has a design rationale in providing additional interaction supervision, and its value is supported by the external result. Reporting Full evaluates the complete framework across settings. We also report the variant without pTCR transfer to make this trade-off explicit.

IASL. The ablations already reported in Table 2(B) remove IASL from both pairwise and triplet training to assess its overall contribution. Table 2(B) shows a 0.55-point increase in TCR-only AUPR, alongside decreases of 0.42, 2.09, and 4.29 points under peptide-only shift, joint shift, and external testing, respectively. These results support IASL's contribution primarily to generalization involving unseen peptides, especially in the joint and external settings, while showing that its benefits vary with the type of shift.

TCR-only performance. HATIS achieves 92.34% AUROC, close to EPACT’s 93.14%, while its AUPR is 5.08 percentage points lower. Our description of “remaining competitive” reflects the comparable AUROC, rather than superiority across both metrics. We acknowledge EPACT’s stronger AUPR in this setting. HATIS’s main performance advantages are under peptide-only and joint shifts and on the external test set.

Configuration choice (Q3). The Full architecture and the default K=8K=8 were fixed in advance, rather than selected based on test-set performance. Here, KK denotes the final number of DiffPool clusters. Checkpoints were selected based on validation loss.

Reviewer_u9vkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses generalization of TCR–pMHC recognition models to peptides and TCRs absent from training. It proposes HATIS, combining two components. Hierarchical atomistic graph transfer builds atom-level molecular graphs, pretrains per-molecule encoders with masked side-chain prediction and contrastive learning, trains separate pMHC and peptide–TCR interaction models via a cross-attention module that induces cross-molecular edges, then composes the triadic graph from the frozen pMHC interaction graph and peptide-conditioned TCR node states. Interface-aware stable learning describes each DiffPool cluster by the proportion of its assignment mass from each component, uses pairwise differences in those proportions to weight an HSIC dependence penalty for sample reweighting, and applies a learned gate that upweights compositionally balanced clusters during pooling. The authors construct a benchmark with cluster-level hold-outs on peptide and TCR axes, plus an external test set, and report AUPR gains of 12.86 and 13.98 points under peptide-only and joint shifts and 4.65 points externally, with fourth-of-six performance under TCR-only shift.

Critical strengths
  • The benchmark construction is more careful than is typical for this task, applying MMseqs2 cluster-level hold-outs to both molecular axes rather than peptide-only clustering or exact entity exclusion, and complementing internal splits with an externally sourced test set.
  • Data access is disclosed honestly. Tables 4 and 5 state exactly which corpora each method uses, and Appendix F.1 is a real attempt to control for it, retraining EPACT with the pMHC binding data HATIS receives. Most submissions in this area do not run that control.
  • Limitations are stated rather than buried. The authors note that triplet negatives do not preserve pMHC pairing, that negatives are unobserved combinations rather than confirmed non-binders, that no MHC hold-out is imposed, and that Theorem 1’s reference distribution need not coincide with any target environment.
  • The representation probes in Appendices E.4 and F.2 go beyond headline metrics and are defined precisely enough to reproduce.
  • The TCR-only result is reported without adjustment in Table 1, where HATIS places fourth of six.
Critical weaknesses
  • The encoder, pooling, HSIC criterion, multi-task weighting, and pretraining recipe are all adopted unmodified from prior work. Two elements are original: preserving the pairwise pMHC interaction graph as a subgraph of the triadic graph (Eq. 4), and weighting the HSIC penalty by cluster composition difference. The first is one equation; the second supplies half the title but is never ablated against uniform weighting. Per Table 2(A), the first carries the entire result and the second does not measurably contribute, inverting the paper’s emphasis.
  • Chance AUPR on the external set is 16.99%. Three of five baselines fall at or below it, and the ranking inverts against TCR-only, with NetTCR-2.0 going from last to first and ERGO-II from second to below chance. Table 6 shows PISTE’s representation outperforming its own head by 7.17 AUROC points on external alone, suggesting misconfiguration rather than genuine weakness. The 4.65-point external margin is therefore a margin over a baseline set that appears mis-trained, and no diagnosis is offered.
  • The architecture defines a per-cluster interface score and induces explicit cross-molecular edges, yet neither is ever compared against physical reality: no check of interface scores against contact residues in solved TCR–pMHC structures, no check of induced peptide–MHC edges against known anchor positions for the common alleles in the data. No per-peptide breakdown is given, though Table 3 shows 90,169 training positives spread over 321 distinct peptides, so gains concentrated on a few immunodominant epitopes would be indistinguishable from broad improvement in the aggregate figures.
  • The composition vector is one number per cluster, used twice (as T and as the interface score), and at the triadic stage it cannot separate peptide from MHC. Theorem 1’s conclusion follows from T∈[0,1]T \in [0,1] alone, so any unit-interval weighting gives it; it compares two bounds under a distribution built to separate the pairs T down-weights. Proposition 1 covers only the favorable perturbation case, and the gate is near-uniform at initialization, with its learned slope never reported.
Questions for authors

Please see weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer u9vk (1/4)

We thank the reviewer for the constructive comments.

W1: Novelty and the contribution of IASL

Novelty. We use established building blocks, but our contribution lies in a task-specific hierarchical learning framework with four aspects:

  • Monomer pretraining. We remove side chains and predict amino-acid identities from the remaining atomic context, using differently masked views of the same molecule as contrastive positives. This adapts the pretraining objectives to learning atom-level representations of the molecular components.

  • Pairwise interaction modeling and transfer. We build atom-level peptide–MHC and peptide–TCR interaction graphs through cross-molecular attention, then transfer both sources of interaction knowledge into the triplet model. Preserving the pMHC graph in Eq. (4) is one part of this process; Table 2(A) evaluates the broader hierarchical training and transfer pipeline, not that equation alone.

  • More comprehensive use of available data. Our pretraining recipe combines monomer data, pMHC affinity supervision, and binding labels for both pMHC and pTCR interactions to exploit complementary information across training stages. Compared with EPACT's original pretraining scheme, which also uses lower-order data, we additionally incorporate pMHC binding and pTCR binding data. This broader supervision supports the transfer of lower-order interaction knowledge to triplet recognition.

  • Interface-aware stable learning. We adapt stable learning to cross-molecular interaction graphs by using cluster composition to distinguish which dependencies to penalize. Composition differences determine the strength of HSIC penalties during sample reweighting, while composition balance guides gated pooling. This introduces a molecular-composition prior into both dependence regularization and representation aggregation, rather than treating all cluster pairs uniformly.

IASL. In addition to the stage-wise ablations already reported in Table 2(B), we conducted an additional comparison against Plain HSIC (uniform cluster weighting), reported in Table R1. Together, these results show that IASL makes a measurable contribution. Removing IASL from both pairwise and triplet training reduces AUPR by 0.42, 2.09, and 4.29 percentage points under peptide-only shift, joint shift, and external testing, respectively. The additional comparison shows smaller gains under peptide-only and joint shifts, with mixed results externally. Thus, IASL provides a complementary benefit to hierarchical transfer, although its contribution varies across settings.

Table R1. Full IASL versus Plain HSIC. Metrics are AUROC / AUPR (%).

Test setFull IASLPlain HSIC
TCR-only shift92.335 / 66.24492.334 / 66.106
Peptide-only shift79.535 / 46.07579.392 / 45.899
Joint shift79.514 / 44.54479.159 / 43.771
External test62.507 / 26.39461.809 / 26.638
AuthorsOpenReview ↗

Response to Reviewer u9vk (2/4)

W2: External baseline performance and the PISTE representation–head gap

Implementation checks. We believe that the low absolute performance and ranking changes on the external set primarily reflect differences in generalization under a challenging distribution shift. All baselines were adapted from their official implementations to our evaluation setting. We rechecked data processing, checkpoint loading, and inference, and found no configuration errors in the code paths examined.

External-set difficulty and ranking changes. TCRcube (Cell Syst. 2026) reports a similar pattern on the same public pan-peptide test set (Fig. 2B in the TCRcube paper): among three general-purpose models evaluated using their official checkpoints, only TULIP clearly outperforms chance, while PanPep and TABR-BERT perform near or below chance, despite partial coverage of the test epitopes in their training data. This provides independent evidence of the difficulty of this test set. In our experiments, the baselines exceed chance on the other splits and generally perform better under TCR-only shift than under peptide-only or joint shift, indicating that they learn predictive signals whose transfer becomes more difficult when peptides are unseen. Ranking changes also have precedent in the same TCRcube study: its structure-based scoring function performs substantially worse than learned models in the known-peptide setting, but approaches the best TCRCube variant on the public pan-peptide test set. We therefore interpret NetTCR-2.0's relative rise as reflecting differences in how strongly models degrade under distribution shift.

PISTE's representation–head gap. The nearest-neighbor metric TAUROCT_{\mathrm{AUROC}} complements the centroid-based metric CAUROCC_{\mathrm{AUROC}}: they assess transfer through local neighborhoods and global class-centroid structure, respectively, with different geometric assumptions and sensitivities. PISTE's classification head outperforms its centroid probe on all four splits. Across the other model–split comparisons, classification heads also generally outperform or closely match the nearest-neighbor probe. For PISTE on the external set, the stronger nearest-neighbor result suggests that its representation retains binding information accessible through similar training examples, while the decision rule learned by its classification head transfers less effectively. Thus, the representation can contain predictive information that the head does not fully exploit.

Based on our implementation checks, published evidence on the same test set, and representation analyses, we believe these results reflect differences in generalization under the reported evaluation protocol, and that the baselines are not misconfigured.

AuthorsOpenReview ↗

Response to Reviewer u9vk (3/4)

W3: Interface scores, MHC anchors, and per-peptide performance

Interface score. Our interface score is defined as the balance of assignment mass from the two interacting components within each DiffPool cluster. At the triadic stage, these components are pMHC and TCR. Here, “interface-aware” denotes a composition-based inductive bias: composition balance guides readout weighting, while differences between cluster compositions guide HSIC regularization to reduce training-specific dependencies between representations with different compositions. The score therefore describes cross-component mixing, rather than physical contact probability. This is the meaning of “interface-aware” in our pooling and stable-learning design.

MHC anchor validation. We conducted an additional validation comparing induced peptide–MHC edges with literature-supported anchor positions across 21 allele–peptide-length groups covering 1,884 unique peptide–MHC pairs. We have added this experiment to Appendix F.4 (Figures 6–7) of our revised manuscript. Figure 6 presents the complete positional overview, and Figure 7 shows enlarged examples, with literature references and sample counts provided in the appendix. The anchor references account for both MHC allele and peptide length. We observe partial positional agreement, strongest at C-terminal anchors, while internal-anchor agreement varies across alleles.

Per-peptide breakdown. Test peptides under peptide-only shift, joint shift, and in the external test set do not overlap with the triplet training set. To directly examine the concern about frequent training peptides, we added a per-peptide analysis in the TCR-only shift setting. We computed AUPR separately for each peptide using average precision and averaged these values with equal weight across peptides (macro-AUPR). Table R2 reports the results.

Table R2. Per-peptide macro-AUPR on TCR-only shift. Metrics are percentages.

Peptide subsetNumber of peptidesHATISEPACTPISTE
All test peptides20871.2370.3170.22
Training-seen peptides18570.2072.4372.27
Training-unseen peptides2379.5553.2753.71
≥5 positive and ≥5 negative test examples11570.9971.1569.42

As shown in Table R2, HATIS has a small overall macro-AUPR advantage, with a larger gain on training-unseen peptides and lower performance on training-seen peptides. Thus, the small overall macro-AUPR advantage is driven by training-unseen peptides rather than frequent training epitopes.

AuthorsOpenReview ↗

Response to Reviewer u9vk (4/4)

W4: Composition vectors, theoretical scope, and the learned gate

Stage-specific composition. Our composition vector describes the two sides of the interaction modeled at each stage: peptide versus MHC or peptide versus TCR in the pairwise models, and pMHC versus TCR in the triplet model. This follows the model's hierarchy from pairwise interactions to triplet recognition. Treating pMHC as one component matches the recognition task: MHC presents the peptide, and the TCR recognizes their combined ligand. The triplet composition vector therefore measures how much a cluster spans the pMHC–TCR sides. Peptide–MHC interactions have already been modeled in the pairwise stage and are retained in the transferred graph. Keeping separate peptide, MHC, and TCR proportions would provide a finer description and is a natural extension of this design.

Theorem 1. We agree that JT≤JunifJ_T\le J_{\mathrm{unif}} and its strictness conditions follow from Tkℓ∈[0,1]T_{k\ell}\in[0,1] and nonnegative HSIC. However, the ordering of two bounds and the validity of the weighted bound are separate questions. The proof first establishes

\le B\sqrt{\sum_{k<\ell}q_{k\ell}^{2}H_{k\ell}}+2\epsilon_{\mathrm{app}},$$ then uses $q_{k\ell}^{2}\le T_{k\ell}$ under the stated composition-based recombination mechanism to obtain the bound involving $J_T$. Arbitrary weights in $[0,1]$ cannot simply replace $T$ and guarantee control of the same risk difference. Regarding the direction of recombination, separation probability and composition weight are defined from the angular difference $d$ through $2d/\pi$ and $1-\cos d$, respectively. Thus, pairs with larger composition differences are more likely to be separated, while low-$T$ pairs are more likely to remain together. This is an analytical construction based on the composition prior. The theorem compares risk bounds for the same predictor under the same recombination model. To reflect this conditional scope, we have removed the “Theorem 1” heading in the revision and presented the result under “A supporting analysis of composition-guided recombination”, with the appendix heading updated accordingly. Practical generalization benefits are assessed through empirical comparisons. **Proposition 1 and the learned gate.** Proposition 1 gives a sensitivity bound for perturbations to lower-scoring clusters, with composition scores and gate parameters held fixed. In $g_{ik}=\sigma(a\phi_{ik}+b)$, the learned slope is $a\approx0.05555$.
Reviewer_3M64OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose HATIS for predicting TCR–pMHC binding from peptide, MHC, and TCR CDR3β sequences. The model represents each component as a chemical atom graph, learns monomer representations through self-supervision, and trains pairwise peptide–MHC and peptide–TCR models before transferring their representations to a triplet predictor. These are sequence-derived chemical graphs; the method does not require predicted three-dimensional complexes. The second contribution, interface-aware stable learning (IASL), uses the molecular composition of learned clusters to guide sample reweighting and interface-gated pooling.

Beyond monomer pretraining, the model uses 341,610 positive pMHC binding examples, 170,250 pMHC affinity records, and 417,800 positive peptide–TCR examples. The triplet training pool contains 90,169 positives collected from PISTE, EPACT, human VDJdb, and the TCRCube training collection. Internal binding datasets use five sampled negatives per positive, and the labeled triplet training pool is divided 80:20 into training and validation sets.

Evaluation includes author-constructed peptide, TCR, and joint cluster holdouts, plus the published TCRCube external test set. HATIS reports substantial improvements on peptide and joint holdouts and achieves 26.39% AUPR on TCRCube, compared with 21.74% for the strongest listed baseline. These results support the usefulness of the overall pipeline. However, the negative construction and incomplete accounting of supervision across stages leave the scope of TCR-specific generalization unclear, while the ablations do not sufficiently isolate the contribution of the proposed IASL mechanism.

Critical strengths
  1. The evaluation makes a meaningful effort to test generalization beyond randomly held-out interactions. Peptides and CDR3β sequences are clustered separately with MMseqs2 using a minimum identity setting of 0.4 and coverage of 0.7, followed by cluster-level holdouts. The external TCRCube test retains its supplied 205 positive and 1,002 negative labels. HATIS improves internal peptide and joint AUPR by 12.86 and 13.98 percentage points over the strongest listed baselines, respectively. The external improvement provides additional evidence that the gains extend beyond the authors’ internal test construction, although performance is more modest there.

  2. The additional EPACT comparison addresses an important concern about unequal training data. Appendix F.1 adds HATIS’s pMHC binding data to EPACT and compares it with HATIS without pTCR transfer. These variants share the auxiliary datasets except for HATIS’s standalone MHC monomer pretraining. HATIS still improves peptide, joint, and external AUPR by 10.82, 13.67, and 5.61 percentage points. This is useful evidence that the gains cannot simply be attributed to access to the additional pMHC binding dataset.

Critical weaknesses

The evaluation does not sufficiently separate TCR specificity from peptide–MHC compatibility, and the meaning of “unseen” needs to be established across training stages.

Appendix C states that internal negatives independently sample peptide, MHC, and TCR, without preserving the peptide–MHC pair. Consequently, some negatives may be distinguishable because the peptide is incompatible with the MHC, regardless of which receptor is supplied. Presentation compatibility is relevant to the overall task, but this construction makes it difficult to determine how much of the reported improvement reflects recognition by a particular TCR. The concern is especially relevant because removing pMHC transfer causes the largest performance drop. This does not demonstrate that HATIS ignores the receptor; it identifies an alternative explanation that the current experiments do not resolve.

The paper also explicitly excludes held-out entities from triplet training, but does not provide an equivalent overlap audit for the supervised pairwise set. Exposure to a held-out peptide through pMHC training, or to a held-out recognition pair through pTCR training, changes the interpretation of generalization. These exposures should be reported separately. Unlabeled monomer exposure is also different from prior access to interaction labels. Without this accounting, it is difficult to tell whether the model generalizes to genuinely new recognition relationships or transfers previously supervised relationships into the triplet task.

Questions for authors
  1. What information about each test set is available at each training stage? Please provide an overlap table for the monomer, pMHC binding, pMHC affinity, pTCR binding, and triplet corpora, reporting exact entity overlap, cluster overlap, and relevant interaction-pair overlap with each test partition. In particular, can test peptide–TCR relationships occur in supervised pTCR training? If recognition-label overlap exists, please report results after excluding it. A stricter supervised-stage holdout, or results stratified by auxiliary-data exposure, would help distinguish different transfer settings. I am not treating unlabeled monomer exposure as equivalent to label leakage.

  2. Does the improvement persist when peptide–MHC compatibility cannot distinguish positives from negatives? Please preserve observed peptide–MHC pairs and generate negatives by changing only the TCR, excluding known positive combinations. Use this construction consistently for training and evaluation, with source-matched receptor backgrounds where possible. Compare HATIS with EPACT under shared auxiliary supervision and add a pMHC-only ablation: remove the TCR input and train a prediction head on the same triplet labels using the pretrained pMHC representation. Report both pooled AUPR and performance within individual pMHC groups. This would directly measure how much predictive value comes from receptor-specific information. Such sampled negatives would still be presumed non-binders, rather than experimentally confirmed negatives.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 3M64 (1/3)

We thank the reviewer for the constructive comments.

W2 & Q1: Training-stage overlap and generalization

To address this concern, we conducted an additional overlap audit, reporting exact entity and pair overlap in Tables R1–R2 and cluster overlap in Tables R3–R4. We include both training and validation data. Since triplet training excludes the peptides or TCRs held out in each test set, we focus on monomer and pairwise pretraining data. We also conducted an additional retraining experiment after removing exact overlap along the held-out peptide and TCR axes from the pairwise data; Table R5 reports the results.

Exact overlap. Table R1 covers test sets with unseen peptides. Overlap occurs only in pMHC affinity data. Table R2 covers test sets with unseen TCRs. TCR overlap occurs in monomer and pTCR data for TCR-only shift and Joint shift, but not External test. Peptide–TCR pair overlap occurs only for TCR-only shift (5.44%).

Table R1. Sample-level overlap, peptide-unseen partitions.

Test setMonomer EntitypMHC binding EntitypMHC binding PairpMHC affinity EntitypMHC affinity PairpTCR binding EntitypTCR binding Pair
Peptide-only shift0/41,226 (0%)0/41,226 (0%)0/41,226 (0%)7,615/41,226 (18.47%)4,783/41,226 (11.60%)0/41,226 (0%)0/41,226 (0%)
Joint shift0/3,066 (0%)0/3,066 (0%)0/3,066 (0%)890/3,066 (29.03%)590/3,066 (19.24%)0/3,066 (0%)0/3,066 (0%)
External test0/1,207 (0%)0/1,207 (0%)0/1,207 (0%)721/1,207 (59.73%)356/1,207 (29.49%)0/1,207 (0%)0/1,207 (0%)

Table R2. Sample-level overlap, TCR-unseen partitions.

Test setMonomer EntitypTCR binding EntitypTCR binding Pair
TCR-only shift24,869/41,112 (60.49%)8,225/41,112 (20.01%)2,238/41,112 (5.44%)
Joint shift1,778/3,066 (57.99%)9/3,066 (0.29%)0/3,066 (0%)
External test0/1,207 (0%)0/1,207 (0%)0/1,207 (0%)

Cluster overlap. Table R3 shows limited peptide-cluster overlap with monomer, pMHC binding, and pTCR binding data (0–6.45%), while overlap with pMHC affinity data is higher (13.11–28.50%). Table R4 shows higher TCR-cluster overlap with both monomer and pTCR data. In particular, External test has no exact TCR overlap but shares 67.44% with pTCR data.

Table R3. Cluster-level overlap, peptide axis.

Test setMonomerpMHC bindingpMHC affinitypTCR binding
Peptide-only shift54/1,785 (3.03%)2/1,785 (0.11%)234/1,785 (13.11%)6/1,785 (0.34%)
Joint shift20/379 (5.28%)0/379 (0%)108/379 (28.50%)2/379 (0.53%)
External test4/62 (6.45%)0/62 (0%)17/62 (27.42%)0/62 (0%)

Table R4. Cluster-level overlap, TCR axis.

Test setMonomerpTCR binding
TCR-only shift4,888/6,817 (71.70%)2,190/6,817 (32.13%)
Joint shift361/508 (71.06%)90/508 (17.72%)
External test132/172 (76.74%)116/172 (67.44%)

Retraining after removing overlap. We remove held-out peptides from pMHC data and held-out peptides or TCRs from pTCR data, as required by each test split. We then retrain the affected pairwise models and the triplet model using the same triplet training and test data. Table R5 reports the main-table Full model alongside HATIS retrained after removing this exact overlap.

Table R5. Main-table HATIS and HATIS retrained after overlap removal. Metrics are AUROC / AUPR (%).

Test setHATISHATIS after overlap removal
Peptide-only shift79.53 / 46.0778.56 / 43.85
TCR-only shift92.34 / 66.2492.39 / 67.15
Joint shift79.51 / 44.5479.42 / 45.19
External test62.51 / 26.3957.91 / 23.61

HATIS still outperforms the strongest baseline in the main table on Peptide-only shift, Joint shift, and External test after removing exact overlap. The larger drop on External test may partly reflect its greater overlap with pMHC affinity data.

AuthorsOpenReview ↗

Response to Reviewer 3M64 (2/3)

W1 & Q2: TCR specificity and pMHC compatibility

Fixed-pMHC retraining. Following the reviewer’s suggestion, we conducted an additional experiment, retraining HATIS and EPACT+binding with negatives that preserve observed peptide–MHC pairs and replace only the TCR, excluding known positive combinations. The same construction is applied to internal training and evaluation; External test retains its published labels. We have added this experiment to Appendix F.3 of the revised manuscript (Table 7 and Figure 5); Table R6 below summarizes pooled performance. HATIS performs better on Joint shift and External test, whereas EPACT performs better on TCR-only and Peptide-only shifts.

Table R6. Fixed-pMHC retraining. Pooled metrics are AUROC / AUPR (%).

Test setHATISEPACT+binding
TCR-only shift60.60 / 30.1365.73 / 35.19
Peptide-only shift62.37 / 21.6368.16 / 27.89
Joint shift51.38 / 17.2249.19 / 16.51
External test53.73 / 18.8253.15 / 18.02

Within-pMHC performance. Figure 5 in Appendix F.3 of the revised manuscript reports AUPR for individual peptide–MHC groups containing both classes, with one point per group. This evaluates TCR discrimination within a fixed pMHC; the relative performance varies across groups.

pMHC-only control on the original benchmark. We added a control experiment that removes TCR input and trains a prediction head on the original triplet labels using frozen pretrained pMHC representations. Full HATIS outperforms this control on all four test sets, with AUPR gains of 11.98, 9.22, 6.89, and 7.88 percentage points, respectively.

Table R7. Original-benchmark comparison. Metrics are AUROC / AUPR (%).

Test setHATIS pMHC-onlyFull HATIS
TCR-only shift89.70 / 54.2692.34 / 66.24
Peptide-only shift76.99 / 36.8579.53 / 46.07
Joint shift77.93 / 37.6579.51 / 44.54
External test52.55 / 18.5162.51 / 26.39

These results support predictive value beyond the frozen pMHC representation, while confirming that pMHC-level information contributes to the original benchmark.

AuthorsOpenReview ↗

Response to Reviewer 3M64 (3/3)

Summary comment: IASL ablation

The summary notes that the ablations do not sufficiently isolate the proposed IASL mechanism.

We added ablation between Full IASL and Plain HSIC (uniform cluster weighting), reported in Table R8. IASL gives modest improvements under peptide-only and joint shifts, little change under TCR-only shift, and slightly lower external AUPR.

Table R8. Full IASL versus Plain HSIC. Metrics are AUROC / AUPR (%).

Test setFull IASLPlain HSIC
TCR-only shift92.335 / 66.24492.334 / 66.106
Peptide-only shift79.535 / 46.07579.392 / 45.899
Joint shift79.514 / 44.54479.159 / 43.771
External test62.507 / 26.39461.809 / 26.638

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.