acceptodds
Under review as a conference paper at ICLR 2027

PIT-GCL: Protein Interaction using Topological Graph Contrastive Learning

Abstract

Protein binding prediction is central to target identification, therapeutic binder design, and large scale screening, yet remains challenging because binding depends on sequence, three dimensional geometry, and global structural organization. Recent folding models such as AlphaFold3 and Boltz-2 have substantially improved structure prediction, but their confidence outputs (pLDDT, pTM, ipTM) are not specifically designed for binary binding prediction, and dedicated structure aware predictors often require bound complex structures that are unavailable at screening scale. We introduce PIT-GCL, a dual tower structure aware framework that encodes each protein independently from its amino acid sequence, Cα point cloud, and a global persistent homology descriptor. Each tower combines residue ESM-2 embeddings with a topological summary computed from the H0 and H1 persistence landscapes of a Vietoris-Rips filtration, and processes the resulting tokens with a structure aware Transformer in which pairwise Cα distances enter as a learned attention bias. A bidirectional cross attention module then performs latent space soft docking between the two per-protein representations, and the model is trained with a combined binary cross entropy and NT-Xent contrastive objective. On three binary interaction prediction benchmarks, general PPI on PPIRef, TCRpMHC binding on STAG, and whole chain pairs on PPB-Affinity, PIT-GCL outperforms representative sequence based, structure aware, and task specific baselines on general PPI under our evaluation, and is the only method above chance on PPB-Affinity; on TCR-pMHC it leads at a fixed decision threshold but is outranked by a task specific sequence model. Because each protein is encoded independently in the first phase, its representation can be precomputed and reused across candidate pairs, which is convenient for large scale screening.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein Binding, Protein Protein Interaction, Graph Learning, Topological Data Analysis

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

The authors propose PIT-GCL which predicts whether two proteins interact using a shared per-protein encoder followed by pairwise fusion. For each protein, frozen ESM-2 residue embeddings are combined with a global persistent-homology descriptor computed from C-alpha coordinates. Four transformer layers optionally use intra-chain distance biases. Bidirectional cross-attention then fuses the independently encoded proteins, and training combines binary cross-entropy with NT-Xent. Evaluation is done on PPIRef, a constructed PPB-Affinity classification task, and STAG TCR–pMHC.

Critical strengths
  1. The separation between cacheable single-protein encoding and lightweight pairwise fusion is easy to understand and has a plausible high-throughput benefit when done at scale.

  2. The no-topology, zeroed-topology, shuffled-topology, no-distance-bias, and no-contrastive-loss controls make the empirical contribution of each proposed component clear. The shuffled control seems useful, although the zero-input arm is not a true functional-capacity match. The paper acknowledges that the distance-attention bias contributes no measurable gain.

  3. The manuscript reports component-level timings and distinguishes first-time structural preprocessing from reuse of cached protein encodings.

  4. Results span three interaction settings, and the appendix explicitly acknowledges bound-conformation dependence, synthetic negative artifacts, limited support for the attention machinery, and the lack of conformational-change stratification.

  5. The full model's mean PPIRef performance is above the paper's ESM mean-pooling control, while topology removal causes a mean drop over the three reported seeds.

Critical weaknesses
  1. All coordinates come from resolved complexes. PPIRef inputs are interface fragments extracted with a 6 Å contact rule from the known bound complex, with a median length of only 24 residues. At screening time the binding interface is unknown, so this preprocessing is an oracle that might be unavailable for novel candidate pairs. The paper itself notes this limitation in Appendix A, but still motivates the method as a practical alternative to pairwise cofolding. The architectural fact that each protein is encoded independently does not resolve this mismatch: the input fragments and conformations were selected using the holo complex.

  2. PPIRef negatives are created by randomly re-pairing chains from different complexes. These are not verified non-interactors: absence from the same deposited complex does not imply non-binding. A logistic regression on five chain-length features reaches AUC 0.8138, and a simple pair persistence descriptor reaches 0.7905. Then much of the benchmark is predictable from construction cues rather than demonstrated physical non-interaction. The length-matched negative analysis is helpful but incomplete. The paper does not retrain and compare all baselines, topology ablations, or simple geometry controls under this harder distribution. Comparing 0.8863 on hard negatives against baseline numbers measured on the easier test distribution does not establish relative superiority.

  3. The manuscript describes seeded 80/10/10 splits of PPIRef cluster representatives but does not establish protein, chain, PDB, complex, or sequence-disjointness. The same or closely related proteins may therefore occur across train and test in different complexes. It is also unclear whether negative re-pairing is performed strictly within each split and whether a chain used in a training negative can appear in a test pair.

  4. Binary protein interaction is invariant to swapping the two inputs, but the model uses side-specific cross-attention projections and concatenates the pooled A and B representations in an ordered way. No architectural invariance, bidirectional averaging, augmentation with both orders, or swap-sensitivity test is reported.

  5. The strongest general PPI sequence baselines are relatively old, while PIT-GCL benefits from ESM-2. The paper's ESM mean-pooling MLP is valuable, but it is described as untuned, is not run on PPB-Affinity or STAG, and stronger modern PLM pair classifiers and matched pair-fusion baselines are needed.

  6. The paired bootstrap uses predictions from one seed and resamples test rows. It therefore ignores split/training variability and likely treats examples sharing proteins as independent. The very small reported p-values should not be interpreted as uncertainty for the full training-and-data procedure.

Questions for authors
  1. How much does the score change under input reversal? Please report f(A,B) versus f(B,A) and the result of symmetrized inference or training.

  2. What exact entities are disjoint across PPIRef train, validation, and test? Please report chain/protein/PDB/complex/sequence overlap and explain whether all negative sampling happens strictly within each split.

  3. Can all principal baselines and component ablations be trained and evaluated on the length-matched negatives, ideally alongside curated non-interactors or physically plausible decoys?

  4. What performance remains when inputs are full independently predicted or experimentally determined unbound monomers, without using an oracle interface from the bound complex?

  5. How does persistent homology compare with parameter-matched simple global geometry descriptors under both easy and hard negative construction?

2 comments

W1&Q4) “all coordinates come from resolved complexes; the 6 Å interface crop is an oracle.” We have measured it. Table in our response to Reviewer J543 separates the two oracles: knowing which residues form the interface is worth −0.0952 AUROC with crystal coordinates and −0.0923 with predicted unbound ones, and −0.2833 to −0.3080 in accuracy. Replacing bound coordinates with coordinates predicted for each chain alone changes nothing measurable (+0.0024 to +0.0038 across all four ). Of the two objections the crop invites, one is confirmed at that size and one is not supported here.

W2-1)“the negatives are not verified non-interactors.” Conceded for PPIRef; we will cite the convention and mark it a limitation rather than defend itWhat does address it is a wet-lab measured set: 1,320 AlphaFold-3 designed candidates against 15 targets, 354 measured binders and 966 verified non-binders, where our model reaches 0.5013 ± 0.0253 (macro-average of per-target AUC). We will report that negative outcome. 1,320 AlphaFold-3 designed candidates against 15 targets, 354 measured binders and 966 verified non-binders, where our model reaches 0.5013 ± 0.0253 (macro-average of per-target AUC). We will add report that negative outcome.

W2-2) “0.8138 and 0.7905 are construction cues.”

ControlAUC
Chain length alone, PPIRef with uniform negatives0.8138
Chain length alone, PPIRef with length-matched negatives0.5456
Chain length alone, wet-lab measured set0.5111
Pair persistence descriptor, as reported in Appendix B0.7905
Pair persistence descriptor, concatenated form0.5000

Appendix B reports rows one and four. We add two and three: the length cue collapses both when negatives are length-matched and when they are measured non-binders rather than re-paired chains. The last row is why we retain the Appendix B note, 0.7905 arises from the interaction terms that expose the length cue, not the descriptor, and the concatenated form scores exactly chance. The underlying point stands: much of the submitted benchmark’s difficulty was construction, which is why every ablation and baseline comparison here is reported under both constructions.

W2-3&Q3) “the length-matched analysis is incomplete.” All 17 ablation arms under both protocols (2CYy W1-2); DeepDTA, PIPR and PLM-interact under both constructions (J543 Q3). Geometry controls under Q5.

W3&Q2) “no protein / chain / PDB / sequence disjointness.” Answered by the protein-disjoint split under 2CYy W2-1: complex, PDB-ID, chain-instance and exact fragment-sequence overlap all exactly zero across every partition pair for all three seeds, largest component 1.06% of the dataset. Retrained on it the full model scores 0.9285 ± .0027 against 0.9305 ± .0048, topology margin +0.0114 ± .0016, positive on every seed. Re-pairing is strictly within partition. What the submission bounded, we now measure, and we state the limit: this is exact-sequence disjointness, not homology or fold disjointness.

W4 & Q1) swap test. Reported in full under J543 Q4: 7.9–9.0% of pairs change their 0.5 decision, the non-canonical ordering costs 0.0075–0.0095 AUROC with the same sign on every seed, and symmetrised inference removes the dependence at no cost (within 0.0002 of the forward pass on every seed). We adopt symmetrised inference as the recommended protocol.

W5) “the baselines are old; the ESM MLP is untuned.” We added PLM-interact (2025), an end-to-end fine-tuned ESM-2 pair model, and retrained it on each benchmark. PIT-GCL achieves higher mean AUROC across all five evaluation settings, including length-matched negatives and protein-disjoint evaluation.

BenchmarkPIT-GCLPLM-interactPaired difference
PPIRef, uniform0.9305 ± .00480.9207 ± .0040+0.0099 ± .0028, CI [+0.0074, +0.0124], p = 0.025
PPIRef, length-matched0.8956 ± .00360.8902 ± .0061+0.0053 ± .0029, CI [+0.0023, +0.0085], p = 0.088
PPIRef, protein-disjoint (pooled)0.9249 ± .00550.9150 ± .0091+0.0099 ± .0036, CI [+0.0071, +0.0126], p = 0.042
TCR–pMHC, full fold0.8844 ± .00320.8607 ± .0070+0.0237 (difference of means) †
PPB-Affinity0.6612 ± .01960.6463 ± .0070+0.0149 (difference of means)

AUROC. PLM-interact was retrained eighteen times on PPIRef (three blocks × three seeds × two negative draws); margins are per-seed paired differences. Each PPIRef row carries two statistics: a paired t over three seeds (df=2), asking whether a fresh run reproduces the ordering, and a row-level paired bootstrap over 2,000 resamples with models fixed, asking whether the margin holds on these rows. The protein-disjoint row is pooled over all components to stay row-matched to PLM-interact; excluding the 542-chain component gives 0.9285 ± 0.0027 and a margin of +0.0114 ± 0.0016. TCR–pMHC and PPB-Affinity use five seeds each, with verified test-row alignment and matched inputs.

AuthorsOpenReview ↗

Re:Review2

The TCR-pMHC fold re-uses training structures and (TCR β, peptide) pairs: on the 2,402 rows where neither appears, the margin reverses to −0.0221 [−0.0399, −0.0043], and rank-averaged ensemble gains (+0.0348 [+0.0261, +0.0433] and +0.0210 [−0.0160, +0.0579]) share the caveat. A 2025 PLM narrows our margin over the best baseline from +0.0478 to +0.0099 (uniform) and from +0.0684 to +0.0053 (length-matched), without closing it: PIT-GCL leads all nine paired PPIRef comparisons and the pooled interval excludes zero throughout. Under length matching we take the conservative reading: 0.005 against a single-fold SE near 0.006 is consistent in sign, significant when pooled, not decisive. On PPB we claim no statistical superiority.

W6) single seed row bootstrap. Replaced by a cluster bootstrap resampling complexes. Applying that standard uniformly, we withdraw the A8a claim (per-seed −0.0054, −0.0048, −0.0013; three-seed interval crosses zero) and report the structural attention bias as unattributed on the same grounds (+0.0014, −0.0010, −0.0006; p = 0.27). Details under 2CYy W2-3.

Q5) persistent homology against parameter-matched geometry descriptors. We do not report this comparison, and we state that as a limitation rather than a result. We have removed the two descriptors from the released source so the archive matches what we report. We agree it is a reasonable control to ask for.

Reviewer_2CYyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PIT-GCL for binary protein–protein interaction prediction. It encodes proteins independently using frozen ESM-2 features, a C-alpha distance attention bias, and persistent-homology descriptors, followed by bidirectional cross-attention and a joint classification/contrastive objective. Experiments cover PPIRef, STAG, and PPB-Affinity. The main finding is that topology provides predictive signal on PPIRef, while the structural bias adds little. Whether that signal reflects generalizable binding features rather than dataset-construction cues, and whether the full architecture is needed, remains unresolved.

Critical strengths
  1. The two-stage design separates monomer encoding from pairwise prediction, allowing each protein’s representation to be cached and reused. The stage-wise runtime breakdown in Appendix G supports this practical motivation.

  2. The parameter-matched zeroed and shuffled topology ablations are useful controls (Table 3). They support a contribution from descriptor content rather than parameter count alone. The authors also report adverse evidence directly, including PPIRef chain-length shortcuts, ineffective structural biases, limited architectural gain over ESM mean-pooling, weaker ranking metrics than epiTCR on STAG, and modest PPB-Affinity performance. This transparency helps identify what the current results do and do not establish.

Critical weaknesses
  1. The screening motivation, evaluation inputs, and attribution of the topology gain are not yet aligned. PPIRef uses 6 Å interface fragments cropped from known complexes, with a median of 24 residues per side (Appendix A, p. 14, lines 732–743). This preprocessing supplies partner-specific binding locations that would be unavailable when screening unknown partners. It is an input limitation, not established cross-split leakage. PPB-Affinity uses full chains but still relies on complex-derived bound conformations and achieves 0.6612 AUROC (Table 2). The current experiments therefore do not establish screening performance from unbound or predicted monomer structures.

On PPIRef, removing topology lowers AUROC from 0.9305 to 0.9149; the zeroed and shuffled controls score 0.9165 and 0.9157 (Table 3). Yet chain length alone reaches 0.8138, and logistic regression on descriptor interactions reaches 0.7905 (Appendix B). These controls leave open how much of the topology gain reflects cropping or size cues. The length-matched negative experiment reports only the full model (0.8863), without the no-topology model, ESM baselines, or other controls on those same pairs. It therefore does not isolate the topology contribution after mitigating that shortcut.

Architectural necessity is also unclear. Removing the structural bias barely changes performance, and A10 exceeds the ESM mean-pooling baseline C1 by only 0.0026. The frozen ESM + topology + lightweight prediction-head baseline is missing, as the Limitations acknowledge. A matched comparison on the same harder test set would be more informative than attributing the full improvement to the attention architecture.

  1. Split integrity, reproducibility, and baseline comparability need further clarification. The manuscript describes STAG’s original train/test split and experimentally resolved structural inputs. Please document the exact structural sources and protocol used. The submitted reconstruction script requires an absent test_predictions.csv, describes 31,753 rows over 28,625 unique keys, and recovers train/test membership through row-occurrence counts. train_stag.py then uses a row-level random_split for validation without explicit pair grouping. This leaves cross-partition duplicate overlap unverified; it does not establish that leakage occurred. For PPIRef, iDist interface deduplication likewise does not ensure disjoint protein identities or sequence families. Constant-input control A5 cannot rule out identity-based leakage when the control removes both genuine signals and memorizable identifiers.

The archive excludes the core ablation switches, baseline code, paired-bootstrap script, and threshold analysis, as the reproducibility statement confirms. On PPB-Affinity, the absence of a topology ablation or a simple baseline sharing PIT-GCL’s ESM features makes representation quality difficult to separate from architectural gains. Several baselines are near chance, so convergence, validation, and tuning details matter. Supervised PIT-GCL versus the released, unadapted PPIFormer checkpoint is also not a matched architectural comparison. On STAG, PIT-GCL trails epiTCR in AUROC (0.8844 vs. 0.9031) and PR-AUC (0.7637 vs. 0.8187); its accuracy/F1 advantage uses a fixed 0.5 threshold. Appendix D applies validation-selected threshold comparisons to PPIRef, leaving that comparison open for STAG.

Two technical details also need clarification. The single-seed row-level bootstrap (Section 4.3) does not capture shared-protein dependencies or training/split variability, which matters for effects around 0.003–0.005. Also, topology.py fits Landscape separately per protein/dimension without an explicit sample_range. Please explain whether a common filtration grid is intended and how it is enforced.

I lean toward weak rejection on the current evidence. Universal SOTA or a gain from every module is not necessary. Consistent topology gains under stricter splits, mitigated shortcuts, and realistic uncropped inputs, even with a simpler architecture, would materially improve my assessment.

Questions for authors
  1. Can you provide complete split manifests, cross-split duplicate and sequence-family overlap statistics, the STAG reconstruction input, and the exact STAG structural provenance and evaluation protocol? These are needed to audit the reported setup without inferring leakage from missing documentation.

  2. On the exact same length-matched negative test set, does the topology gain over the no-topology variant persist, and how does frozen ESM + topology + a lightweight head perform under comparable tuning? Please report multi-seed paired differences and uncertainty that accounts for shared proteins or complexes.

  3. On full-length monomer inputs without ground-truth interface cropping, how much advantage does PIT-GCL retain over baselines using the same pretrained ESM representations? Unbound or predicted monomer inputs would directly test the screening motivation.

4 comments
AuthorsOpenReview ↗

Re: Review

W1-1) Answered by the 2×2 reported under J543 W2. We commit to replacing contents based on measurement.

W1-2) We now have every arm under both protocols, and we draw a distinction the submission did not: a zero-shot row and a retrained row are not comparable, and pairing them incorrectly reverses conclusions. Paired the wrong way, our model zero-shot at 0.8775 against C2 retrained at 0.8892, the frozen-feature control appears to win by 0.0117; the correct retrained to retrained pairing gives +0.0064 the other way.

ArmPublished splitLength-matched (ZS)
Full model0.9305 ± .00480.8775 ± .0083
A1 topology removed0.9149 ± .00390.8590 ± .0034
A1z branch retained, fed zeros0.9165 ± .00460.8607 ± .0043
A1s another protein's descriptor0.9157 ± .00430.8574 ± .0025
A9r fixed random fingerprint0.8882 ± .00720.8186 ± .0067
A10 ESM-only dual tower0.9152 ± .00550.8579 ± .0055
A5 constant-input floor0.5401 ± .00400.5040 ± .0067
Full − A1+0.0156+0.0185

AUROC, 3 seeds per arm. Each arm was trained once, on uniform negatives. Both columns score those same checkpoints on the same test keys; the only difference between the columns is the negative construction used at test time. Switching the construction collapses the chain-length cue from AUC 0.8138 to 0.5456.

ArmLength-matched (retrained)
Full model0.8956 ± .0036
A1 topology removed0.8809 ± .0068
Full − A1+0.0147 ± .0032
AUROC, 3 seeds, trained and tested on length matched negatives. We retrained two arms only (the full model and A1) which is why Table 2 has two columns rather than others; the remaining arms exist as zero-shot evaluations alone, and we do not present retrained figures we do not have.

Each protocol, the topology margin is +0.0156 on the published split, +0.0185 zero-shot and +0.0147 ± .0032 retrained. The band is +0.015 to +0.019 across all three settings and does not collapse under any of them. We added A9r to test whether the branch contributes only capacity or protein identity. A fixed random fingerprint scores below removing the branch altogether (0.8186 against 0.8590), so the branch requires real content rather than an identifier. A1, A1z and A1s fall within 0.0033 of one another, which indicates the effect is attributable to the descriptor’s content rather than to parameter count.

W1-3) Answered by under J543 W1

C2 = 0.9202 ± .0036 over 5 seeds is frozen ESM-2 + persistence landscape.

W2-1) We built a strictly protein-disjoint split and retrained on it.

SplitFull modelA1Topology margin
Published random (complex-keyed)0.9305 ± .00480.9149 ± .0039+0.0157 ± .0009
Protein-disjoint, pooled0.9249 ± .00550.9130 ± .0068+0.0119 ± .0013
Protein-disjoint, excluding the largest component0.9285 ± .00270.9170 ± .0043+0.0114 ± .0016

AUROC, both models retrained on the strict split, 3 seeds. The strict split costs 0.0020 AUROC, and the topology margin shrinks by roughly a quarter while remaining positive on every seed. Two limits we state rather than leave implicit: this enforces exact-sequence disjointness, not homology or fold disjointness, and structural redundancy is untouched.

W2-2) topology ablation We ran arm A1 on PPB-Affinity at all five published seeds, with identical flags to the main arm (100 epochs, batch 16, lr 1e-4, d_model 256, 4 layers, 8 heads, α=β=0.5, H0+H1, 3 landscapes, ESM-2 t6_8M, patience 20). The main arm is the already published run, not a re-run.

seedmainA1 (− topology)margin
420.67350.6466+0.0269
430.63610.6560−0.0198
440.68120.6347+0.0465
450.64480.5988+0.0460
460.67030.6598+0.0106
mean0.6612 ± .01960.6392 ± .0246+0.0220 ± .0278

AUROC on the shipped 1,034pair test set (517 positives, 517 negatives). All five A1 runs carry the completion marker and early-stopped between epochs 29 and 56 of 100.

The main model outperforms the topology-removed arm in four of five seeds, with a mean AUROC gain of +0.0220, although seed 43 shows a reversal. A row-level paired bootstrap on the five-seed ensemble likewise estimates a positive gain of +0.0200 (95% CI: [−0.0059, +0.0464]; one-sided p = 0.065). The paired t-test (t(4) = 1.77, p = 0.15) and Wilcoxon test (p = 0.19) do not reach statistical significance.

These results provide additional directional support for the contribution of topology, consistent with PPIRef. The positive mean gain is observed across both the per-seed comparison and the ensemble analysis, strengthening the evidence beyond the original benchmark. Nevertheless, uncertainty remains substantial on the 1,034-pair PPB-Affinity test set, and we therefore present this result as supporting evidence rather than a statistically significant improvement.

AuthorsOpenReview ↗

Re:Review2

W2-2) Code ablation switch This is already done. The released archive has been rebuilt so that every arm in Table 3 and in this response is a single flag on one code path.

W2-2) threshold analysis on TCR-pMHC Applying the Appendix D procedure over five published seeds, beside the PPIRef figures from the submitted paper:

PPIRef (Appendix D)TCR–pMHC, full foldTCR–pMHC, file-unseen subset
F1 at the fixed 0.50.8413 ± .00520.7152 ± .00850.6790 ± .0087
F1 at a threshold not fitted on the scored rows0.8439 ± .0051 (validation)0.7207 (sample split)0.6838 (sample split)
F1 at the oracle threshold0.8458 ± .00510.7253 ± .00830.6894 ± .0091
oracle headroom over 0.5+0.0045+0.0101+0.0104
share of it a blind threshold recovers58%48%37%
test-oracle threshold0.4026 ± .04190.2022 ± .08180.1536 ± .1116

Full fold n = 6,350; the file-unseen subset is the n = 5,343 rows whose structure file does not appear in training. AUROC and PR-AUC are threshold-free and do not move, so they are omitted.

The fixed cut-off costs 0.0101 F1 on this benchmark against 0.0045 on PPIRef, roughly 2.2×. We state one limitation of the middle row explicitly: the PPIRef figure is a genuine validation-to-test transfer, whereas this benchmark provides no per-row validation predictions, so our surrogate is a sample split within the test fold and is not comparable in kind. The oracle row is therefore an upper bound rather than an achievable operating point. The duplication described below inflates the metrics but does not move the threshold.

W2-2) Table 9 decomposed for every method We went beyond what the question asked. A collaborator supplied per-row test predictions for four of Table 9’s eight baselines, epiTCR, epiTCR-MHC, NetTCR-2.0 and ConTCR, five seeds each, so we decomposed the fold for every method rather than only for ours. The six methods share 6,347 of the 6,350 test rows, and we work on that common set throughout this subsection. It splits into three cells: the structure file appears in training (1,007 rows, 15.87%), the (TCR β, peptide) pair appears in training but the file does not (2,938, 46.29%), and neither appears (2,402, 37.84%). The file-unseen subset used in the threshold table above is the union of the second and third.

The third cell is the one clean comparison, and it is clean for every method by construction: if the (TCR β, peptide) pair is unseen then every coarser sequence unit is unseen too. It is also where the exposure asymmetry sitsm the sequence baselines meet a training input on 61.9–62.2% of rows against our 15.87%.

Two findings follow, and the second is against us. First, Table 9 inflates every method, by 0.12 to 0.17 AUROC. That is a property of the benchmark and it is the part of this analysis worth keeping. Second, our own inflation is the largest of the six at 0.1692, larger than epiTCR’s 0.1396 despite one quarter of the exposure, and on the unseen cell PIT-GCL ranks fourth of six:

ours − theirs, paired row bootstrapfull foldunseen cell
vs epiTCR−0.0187 [−0.0256, −0.0119]−0.0477 [−0.0692, −0.0267]
vs epiTCR-MHC−0.0101 [−0.0168, −0.0037]−0.0544 [−0.0734, −0.0356]
vs PLM-interact+0.0237 [+0.0160, +0.0316]−0.0221 [−0.0399, −0.0043]
vs NetTCR-2.0+0.0842 [+0.0750, +0.0931]+0.0664 [+0.0448, +0.0880]
vs ConTCR+0.1812 [+0.1673, +0.1946]+0.0013 [−0.0219, +0.0235]

AUROC, 2,000 resamples. We lose to both epiTCR variants by more than twice our published margin, we lose to PLM-interact on this cell although we lead it on the full fold, and we only tie ConTCR, the weakest method in Table 9. The published F1 advantage of +0.0386 over epiTCR becomes −0.0746 at the fixed cut-off, and it stays reversed when every method is given its own best threshold per cell (ours 0.4111 against epiTCR’s 0.4486), so it is not an artifact of calibration. This is not leakage inside our training rows. Retraining on a sequence-disjoint split raises our unseen-cell AUROC from 0.7152 to 0.7359 ± 0.0085, so the figure is real; it is simply not a win. We will restate the Table 9 paragraph in the appendix accordingly, and we note the limit of the audit: four of eight baselines, since per-row scores for STAG, TEPCAM, TRAP and TITAN were not available to us.

AuthorsOpenReview ↗

Re:Review3

W2-2) comparability of the baseline table DeepDTA, PIPR, and GVP were trained from scratch on PPIRef, with DeepDTA and PIPR evaluated across the same five seeds as PIT-GCL; GVP used one seed. PPIformer instead measures zero-shot transfer from its released checkpoint. We clarify two remaining protocol differences: method specific 80/10/10 splits yield non row matched test sets, and sequence baselines generate negatives before splitting, whereas PIT-GCL splits first.

BaselineTrained on PPIRef?HowTest rows scored on
DeepDTAyes, from scratch1.9M params, batch 256, early-stopped at epoch 22, seeds 42–4610,219
PIPRyes, from scratch58,433 params, batch 256, 46 epochs, seeds 42–4610,219
GVPyes, from scratch50 epochs, batch 32, seed 42 onlya different partition again
PPIformerno zero-shotreleased masked_modeling.ckpt, 12.3M params, scored directly10,206
PIT-GCL (ours)yes50 epochs, patience 10, seeds 42–4610,242

W2-3) the single-seed row bootstrap, and sample_range We have replaced row-level bootstrapping with a complex-level cluster bootstrap throughout. The resulting interval widths are 0.97–1.00× those of the row-level analysis, indicating that the uncertainty estimates remain similar under complex-level resampling. We do not interpret this agreement as evidence that protein sharing is negligible.

ArmPer-seed difference vs the full modelIntervalp
A8a, contrastive term removed−0.0054, −0.0048, −0.0013three-seed interval crosses zero (seed 42 alone: [−0.0079, −0.0028])—
A6z, structural bias zeroed+0.0014, −0.0010, −0.0006[−0.0030, +0.0029]0.27
A11, common [0,20] landscape gridpaired −0.0004 ± .0014, t(2) = −0.53[−0.0039, +0.0031]0.65

Removing the contrastive term (A8a) reduces AUROC in all three seeds, providing consistent directional support for its contribution. However, although the seed-42 interval excludes zero, the three-seed interval does not; we withdraw the A8a claim to distinguish this positive trend from a statistically established improvement. For the structural attention bias (A6z), the differences are small and mixed in direction, so its independent contribution remains unresolved. To address the reviewer’s concern about the implicit sampling grid, we retrained A11 using a common landscape grid of [0, 20] for every protein and homology dimension, matching the Rips max_edge_length. Performance is essentially unchanged: 0.9301 ± 0.0038 versus 0.9305 ± 0.0048 for the published model, with a paired difference of −0.0004 ± 0.0014. This directly supports the robustness of the reported performance to the grid choice and argues against adaptive per-protein sampling as the explanation for the observed gains. We will clarify the grid configuration and add A11 to Table 3.

Q1) manifests, overlap statistics, split provenance All three are included in the rebuilt archive: per-partition manifests, the overlap statistics reported under W2-1, and the provenance of the TCR–pMHC split. We confirm that all five seeds of our TCR–pMHC runs survive at row level and reproduce the submitted table to four decimal places.

Q2) topology gain on length-matched negatives, and the frozen-feature control Both are reported above. In the full architecture the topology margin is +0.0185 zero-shot and +0.0147 ± .0032 retrained; in the frozen-feature head it is +0.0069 and +0.0065 respectively. The frozen ESM + topology control is C2 = 0.9202 ± .0036.

Q3) full-length monomers against matched ESM baselines on PPIRef We have carried out the first half of this experiment and report it below. The baseline comparison it asks for we were not able to complete, and we state that explicitly rather than approximate it. To evaluate the model on full-length monomers we re-folded 549 chains from 263 test complexes individually with AlphaFold 3, with no partner present in any input, and scored the published checkpoint on four arms that cross the two forms of prior knowledge a PPIRef input carries: the coordinates are taken from a bound complex, and the residue selection is the true interface.

AUROCResidues = true 6 Å cropResidues = whole chainΔ interface oracle
Coordinates = crystal0.94020.8450−0.0952
Coordinates = AF3 unbound fold0.94360.8513−0.0923
Δ coordinate oracle+0.0034+0.0063

All arms are row-matched (526 pairs, seed-42). Unbound AF3 coordinates match crystal (+0.0034/+0.0063), so the bound prior is not load-bearing; knowing the interface is worth 0.0952/0.0923. We will quote 0.8513 for screening; ACC 0.5456 there is calibration, not ranking (median p=0.9841). Limits: one checkpoint, 263 complexes; 9x longer inputs make 0.09 an upper bound; coordinates compare only when cropped (89.9% vs 99.99%). No baseline was scored unbound; cropped comparisons (PLM-interact, C1, C2) ship with the code.

Thank you for the thorough experiments and candid reporting across all three comments. The new controls on cropped PPIRef (length-matched retraining, exact protein-disjoint splits, and the fixed [0, 20] grid A11) do confirm a consistent +0.011 to +0.019 AUROC signal from the topological descriptors in that cropped setting, and the TCR-pMHC overlap audit is very informative.

However, looking at the complete picture, I will maintain my current score (2, Weak rejection) for three reasons established by the new results:

  1. On realistic uncropped AF3 monomers (Q3), removing the 6 A interface oracle drops AUROC from 0.9436 to 0.8513, and neither the no-topology arm (A1) nor matched ESM baselines (C1/C2) were evaluated on whole chains, leaving the core monomer-screening advantage unverified.
  2. Even on cropped length-matched PPIRef, the full architecture (0.8956) beats the simple retrained frozen ESM-2 + landscape baseline C2 (0.8892) by only +0.0064 AUROC while both the structural bias (A6z) and contrastive term (A8a) are non-significant over 3 seeds, and external PPIRef baselines still use non-matched splits.
  3. Beyond cropped PPIRef, the topology gain on full-chain PPB-Affinity is not statistically significant (p = 0.15), and on the clean unseen (TCR beta, peptide) subset of TCR-pMHC, PIT-GCL trails epiTCR, epiTCR-MHC, and PLM-interact while only tying ConTCR.
Reviewer_J543OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PIT-GCL, a model that predicts whether two proteins interact. Each protein is encoded separately using ESM-2 embeddings, persistent homology features, and attention with a structure-based bias. A cross-attention module then combines the two proteins. Training uses a classification loss and a contrastive loss. The model is tested on PPIRef, PPB-Affinity, and TCR-pMHC. It does well on PPIRef, but results on the other two are weaker or mixed. The paper also includes ablation studies.

Critical strengths
  1. The problem is important. Encoding each protein separately is practical: features can be computed once and reused for many pairs, without predicting complex structures.
  2. The experiments cover three settings. The ablations (shuffled or zeroed topology features, negative sampling, dataset artifacts) are useful and help show what works and what does not.
Critical weaknesses
  1. The structure part does not seem to help. Most components already exist (ESM-2, persistent homology, Transformer, contrastive learning). In Table 3, removing the structure-aware attention bias barely changes AUC (0.9305 vs. 0.9302). Removing the topology feature lowers AUC by about 0.016. So the gain seems to come from a global feature, not from the residue-level structure modeling that the paper presents as its main idea.
  2. The PPIRef setup does not match real screening. PPIRef uses interface regions taken from known complexes. When screening new pairs, the interface is usually unknown. On PPB-Affinity, which uses full chains, AUC is only 0.6612. It is unclear whether the PPIRef results would hold in practice.
Questions for authors
  1. Removing the structure-aware bias changes almost nothing. What shows that residue-level 3D structure actually helps? Can you separate its effect from ESM-2 and the global topology feature?
  2. How would the model work on new pairs where the interface is unknown? Can you report results using full chains on PPIRef or another dataset?
  3. Can you add stronger pretrained sequence baselines and test all methods on the same length-matched negatives?
  4. The two towers use different attention projections, and the pair features are concatenated. Does the prediction change if the two proteins are swapped? If so, how do you handle it?
1 comment
AuthorsOpenReview ↗

Re: Review

W1&Q1) We added several more ablation study, frozen features into a lightweight MLP, no attention, no coordinates.

controlAUROC
C1 frozen ESM-2 mean-pool0.9126 ± 0.0032
C2 frozen ESM-2 + persistence landscape0.9202 ± 0.0036
C3 persistence landscape only0.8152 ± 0.0049
full model0.9305 ± 0.0048

C1–C3 are five seeds; the full-model row is the published three seed figure.

Neither modality alone suffices. 0.8152 topology alone, 0.9126 sequence alone and together, 0.9202. The descriptor contributes in both architectures: +0.0076 in a mean pooling MLP, +0.0156 in ours.

W2&Q2) We agree that a 6 Å interface fragment cut from a solved complex contains two separate forms of privileged information: the coordinates are a bound conformation, and the choice of residues is the true interface, which a screening user cannot know. §4.5 separates them with a 2×2 holding task, split and checkpoint fixed.

armcoordsresiduesAUROCAUPRCACCF1
B-crop (published setting)crystal6 Å crop0.94020.94630.84980.8501
U-cropAF3 unbound6 Å crop0.94360.94870.85360.8533
B-fullcrystalwhole chain0.84500.86760.56650.6842
U-full (screening-realistic)AF3 unboundwhole chain0.85130.87970.54560.6792

526 pairs from 263 complexes drawn from the seed-42 test fold (300 sampled; 37 dropped for length, unmappable crops or missing fragments), scored with the published seed-42 checkpoint, inference only, single seed. 549 unique chains were re-folded with AlphaFold 3 individually, with no partner present in any input.

  1. The bound conformation assumption is not load-bearing on this benchmark. Replacing crystal coordinates with coordinates predicted for each chain in isolation changes nothing measurable, and does so consistently across all four metrics (+0.0034, +0.0024, +0.0038, +0.0032).
  2. Knowing the interface is worth about 0.09 AUROC, and considerably more for decisions, accuracy falls by 0.28 to 0.31. We report the screening realistic corner, 0.8513 AUROC with ACC 0.5456 and F1 0.6792, as the honest figure, and note that the 0.5 threshold was calibrated on ~30residue fragments and does not transfer to ~270 residue chains. Any deployment claim would require recalibration.

Q3) We retrained three baselines under our split, our keys and both negative constructions, including a 2025 PLM pair classifier that finetunes ESM-2 end to end on the protein pair which is the direct answer to “untuned”. PLM-interact : https://www.nature.com/articles/s41467-025-64512-w

Methoduniformlength-matched (RT)loss to the shift
PIT-GCL (ours)0.9305 ± .00480.8956 ± .0036−0.0349
PLM-interact (2025)0.9207 ± .00400.8902 ± .0061−0.0305
PIPR0.8827 ± .00390.8272 ± .0109−0.0555
DeepDTA0.8780 ± .00520.7962 ± .0426−0.0818
Chain length alone (no model)0.81380.5456−0.2682

AUROC, three seeds (42-44) throughout. DeepDTA and PIPR were re-run on our splits and our keys, so their uniform figures differ slightly from the five-seed values in Table 1 of the paper. The final row is a single logistic regression on chain length, not a trained model.

Two findings here run against the submitted paper, and we report them rather than omit them. First, the margin over the strongest baseline narrows under length-matched negatives, from +0.0099 to +0.0053; the claim that our margin grows holds against PIPR and DeepDTA only, and we will qualify it accordingly. Second, the ranking does not change: PIT-GCL remains the highest-scoring method under both negative constructions, and the two pre-PLM baselines degrade two to three times as much as either PLM-based model (-0.0555 and -0.0818 against -0.0349 and -0.0305). The shift is a property of the benchmark, not of our method in particular

Q4) Yes, Here is swapped result. Although the architecture does not explicitly enforce invariance to input order, swapping the two proteins preserves strong overall discrimination across all three seeds. The mean AUC changes from 0.9305 for f(A,B) to 0.9221 for f(B,A), an absolute decrease of 0.0084. Symmetrizing the predictions yields a mean AUC of 0.9306, closely matching the reported performance, with differences of at most 0.0003 per seed. These results suggest that the model’s overall ranking performance is largely robust to input order, while symmetrization provides order invariant predictions without compromising AUC. However, 8-9% of pairs cross the 0.5 decision threshold when the inputs are swapped, indicating that input order can still affect individual binary predictions.

seedf(A,B) as reportedf(B,A)symmetrised
420.93580.92830.9359
430.92950.92130.9298
440.92630.91680.9261

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.