acceptodds
Under review as a conference paper at ICLR 2027

EnzyFuse: Cross-View Interaction over Spatial Neighborhoods for Catalytic Residue Prediction

Abstract

Predicting catalytic residues supports mechanistic annotation and enzyme engineering. When annotated homologs are scarce, prediction can draw on evidence from the protein's sequence, evolutionary and structural views. The challenge is to learn from both view-specific evidence and its organization across spatially neighboring residues. We introduce EnzyFuse, a task-specific multi-view model built on pretrained sequence and alignment encoders and equivariant geometric message passing. Directly supervised view branches provide residue predictions, while an interaction branch combines their representations over spatial neighborhoods. Catalytic partner retrieval reuses the residue annotations as structured auxiliary supervision, identifying catalytic partners near an annotated anchor during training. A learned gate combines branch logits during training; inference uses a validation-selected average of the available logits. We evaluate EnzyFuse under joint sequence–structure similarity control, with post-split audits and controlled ablations to assess generalization and component contributions. Ablations and inference-time view masking indicate that evolutionary and structural information both contribute, and that the gain over a three-branch late-fusion model arises when spatial interaction is combined with catalytic partner retrieval. Under this protocol, a single EnzyFuse model outperforms all evaluated baselines, including a multi-model ensemble, in protein-averaged area under the precision–recall curve. This advantage persists on the subset with no detected neighbor among catalytic-labeled training proteins.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Catalytic residue prediction, Enzyme function prediction, Protein language models, Geometric deep learning, Adaptive multimodal fusion

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 5 reviews, above 85% of the others.
Reviewer_Yk2pOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EnzyFuse predicts catalytic residues from three views: frozen ESM-2 and MSA Transformer embeddings and an SO(3)\mathrm{SO}(3)-equivariant structure encoder. An interaction branch combines the three representations on the structure graph. Each branch is directly supervised, and a training-only auxiliary task (catalytic partner retrieval) scores other catalytic residues within 15 Å of an annotated one. On a split controlling both sequence and structural similarity, EnzyFuse reaches an AUPRC of 0.757, compared with 0.701 for the five-model Squidly ensemble and 0.672 for retrained SCREEN; it also outperforms both on the subset with no detected neighbor among catalytic-labeled training proteins. Squidly retains higher F1 and MCC.

Critical strengths
  1. Careful evaluation: a joint sequence–structure split, a nearest-neighbour copy control showing that sequence-only splits reward annotation transfer (AUPRC 0.845 vs 0.299, App. A), and baselines retrained on the same data with paired statistics.
  2. Consistent gains, including over SCREEN, which uses comparable inputs (+0.085, Table 3).
Critical weaknesses
  1. Weak evidence in the low-homology setting. The paper is motivated by prediction when annotated homologs are scarce (L044–046). On the 3,322 enzymes with no detected neighbour, however, removing partner retrieval or the interaction branch lowers AUPRC by only 0.004 and 0.009, without confidence intervals, and the authors note that these comparisons do not establish significance (Table 7, L957–962). The "w/o structure" ablation also removes both components (L418–419), so it does not isolate them.
  2. The Squidly baseline configuration is underspecified. Appendix C states that Squidly was retrained using its released code but does not report which ESM-2 variant was used. Since the Squidly implementation supports 3B and 15B variants while EnzyFuse uses ESM-2 650M, reporting the variant would clarify the comparison and aid reproducibility.
Questions for authors
  1. Which ESM-2 size was used when retraining Squidly? How do Squidly and EnzyFuse compare when both are given the same BLAST fallback?
  2. Table 2 gives paired confidence intervals for these ablations on the full test set. Since the h=0h=0 proteins are a subset of it, why are paired intervals unavailable in Table 7, and do the 0.004 and 0.009 differences exclude zero?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Yk2p

We thank the reviewer for the specific suggestions on the low-homology analysis and the Squidly configuration. Changes in the revised PDF are marked in blue.

W1 / Q2. Component contributions on the low-homology subset

The h=0h=0 comparisons in the submission were computed from aggregate AUPRC only. We have now computed paired protein-level statistics on the same 3,322 enzymes, as for Table 2, and added the parameter-matched non-spatial control, which isolates spatial message passing while keeping the structure input and partner retrieval:

Variant (h=0h=0)AUPRCΔ\Delta95% CI
EnzyFuse0.6688±0.00910.6688_{\pm 0.0091}——
w/o partner retrieval0.6648±0.01070.6648_{\pm 0.0107}−0.0040-0.0040[−0.0081,+0.0000][-0.0081, +0.0000]
w/o spatial edges0.6547±0.00730.6547_{\pm 0.0073}−0.0141-0.0141[−0.0183,−0.0099][-0.0183, -0.0099]
w/o interaction branch†0.6601±0.00340.6601_{\pm 0.0034}−0.0088-0.0088[−0.0131,−0.0043][-0.0131, -0.0043]

Removing the interaction branch or its spatial message passing lowers AUPRC significantly on this subset, whereas the effect of removing partner retrieval does not exclude zero. For each variant, however, at least one of the three training runs favors the ablated model, so we describe component contributions on this subset as less certain than on the full test set. Unlike "w/o structure", the first two variants keep the structure input and isolate partner retrieval and spatial message passing, respectively.

The smaller effects are consistent with the smaller contribution of structural input on low-homology enzymes: on the 2,909 h=0h=0 enzymes with all three views, masking the structure at inference lowers AUPRC from 0.680 to 0.669, whereas masking the MSA lowers it to 0.551 (Figure 7). The low-homology claim of the paper rests on the full model, which exceeds the Squidly ensemble by +0.064+0.064 ([0.053,0.075][0.053, 0.075]) and SCREEN by +0.116+0.116 ([0.106,0.127][0.106, 0.127]) on these enzymes (Table 6). (Table 7 and Appendix E revised.)

W2 / Q1. Squidly configuration and the same BLAST fallback

Squidly was retrained with its default encoder, layer 36 of ESM-2 3B, whereas EnzyFuse uses the frozen 650M model; the advantage is therefore not due to a larger sequence encoder. Table 1 reports Squidly's model predictions without its BLAST-based ensemble.

We applied Squidly's own fallback to both models, with the same search results against the 75,847 catalytic training enzymes. In the released code, the fallback uses DIAMOND blastp (--ultra-sensitive) and transfers sites from the top hit with Clustal Omega; any hit triggers the transfer. We also tested a 30% identity cutoff and MMseqs2 retrieval:

Fallback applied to both modelsEnzymes using fallbackEnzyFuseSquidly ×5Δ\Delta95% CI
None (Table 1)—0.7570.7570.7010.701+0.056+0.056[0.049,0.062][0.049, 0.062]
Squidly's fallback, as released2,4140.7120.7120.6670.667+0.045+0.045[0.039,0.051][0.039, 0.051]
Squidly's fallback, identity > 30%2,0460.7240.7240.6750.675+0.049+0.049[0.043,0.055][0.043, 0.055]
MMseqs2, identity > 0.31,5750.7580.7580.7030.703+0.055+0.055[0.049,0.062][0.049, 0.062]

The fallback lowers both models: on the 2,414 enzymes where it applies, the transferred sites reach 0.797 AUPRC, compared with 0.924 for EnzyFuse and 0.893 for the Squidly ensemble. On these enzymes both models receive the same transferred sites, so the remaining difference comes from the 4,390 enzymes without a hit, where each model keeps its own predictions (0.665 vs. 0.596). Against SCREEN, the difference stays between +0.073+0.073 and +0.085+0.085 under all settings. (Appendix C revised.)

Thank you for the detailed rebuttal and the additional statistical analyses. These resolve my concerns; I will retain my rating.

Reviewer_14SwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies catalytic residue prediction using sequence, evolutionary, and structural information. EnzyFuse combines frozen ESM-2 and MSA Transformer representations with an equivariant structural encoder, directly supervised view branches, and a spatial cross-view interaction branch. A training-only catalytic partner retrieval objective reuses catalytic annotations to supervise local candidate ranking. Evaluation uses a joint sequence–structure similarity graph split, followed by residual similarity audits. The model improves protein-averaged AUPRC and top-k hit rates over the evaluated baselines, while Squidly retains higher F1 and MCC.

Critical strengths
  1. The task is practically relevant, and the combination of evolutionary conservation with spatial context is well motivated. The partner retrieval objective provides a task-specific use of existing labels without requiring additional annotations at inference.

  2. The main ranking improvement is substantial: AUPRC increases from 0.7012 for the Squidly ensemble to 0.7568, with a paired confidence interval excluding zero. Top-3 hit rate also improves from 78.5% to 84.8%.

  3. The experimental analysis includes three training seeds, retrained baselines, view removal, a parameter-matched non-spatial interaction control, a three-branch late-fusion comparison, and similarity-stratified evaluation. These are useful controls.

  4. The manuscript explicitly acknowledges residual similarity, incomplete labels, and weaker threshold-dependent metrics. The strict-subset comparison provides additional evidence that the overall advantage is not confined to the detected higher-similarity examples.

Critical weaknesses
  1. The role of partner retrieval could be clearer. Since all anchors and positive partners are catalytic, the loss may mainly improve local residue ranking. A candidate-only scorer would help show whether the anchor features add useful information.

  2. Component gains are smaller on the subset with no detected neighbor. In Table 7, removing partner retrieval lowers AUPRC by 0.0040, while removing the interaction branch and retrieval lowers it by 0.0087. Paired confidence intervals would help assess these gains.

  3. The similarity audit covers only training proteins with catalytic labels. Other training proteins also contribute through binding/site supervision. Extending the audit to all training inputs would strengthen the generalization analysis.

Questions for authors
  1. Does partner retrieval retain its benefit when the scorer omits the anchor representation and anchor–candidate feature product? This would help distinguish anchor-dependent information from localized candidate ranking.

  2. Can you provide paired confidence intervals and per-seed results for Table 7, and report the core component ablations on the strict subset used in Table 5? Including the existing parameter-matched non-spatial control in these analyses would be particularly informative.

  3. Can you extend the residual similarity audit to all supervised training proteins and explain which search or graph-construction settings account for the remaining above-threshold cross-split neighbors? Please also distinguish queries with no structural model from queries with a usable model but no detected structural hit.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 14Sw

We thank the reviewer for the precise questions, in particular the suggestion to test a candidate-only partner scorer. Changes in the revised PDF are marked in blue, and new analyses are in Appendix I.

W1 / Q1. Does partner retrieval need the anchor?

We trained the suggested candidate-only scorer, which omits the anchor representation and the anchor–candidate product and keeps the candidate representation and the distance encoding (three seeds, full test set):

VariantAUPRCΔ\Delta vs. EnzyFuse95% CI
EnzyFuse0.7568±0.00350.7568_{\pm 0.0035}——
Candidate-only partner scorer0.7541±0.00100.7541_{\pm 0.0010}−0.0027-0.0027[−0.0056,+0.0000][-0.0056, +0.0000]
w/o partner retrieval0.7360±0.00550.7360_{\pm 0.0055}−0.0208-0.0208[−0.0232,−0.0187][-0.0232, -0.0187]

The candidate-only scorer retains most of the gain: it exceeds the model without partner retrieval by +0.018+0.018 ([0.015,0.021][0.015, 0.021]), whereas its difference from EnzyFuse is not significant. The reviewer's reading is correct: the benefit arises mainly from ranking candidates within the neighborhood of each catalytic residue, with little added by anchor features. This agrees with Appendix G, where partner retrieval mainly lowers the scores of non-catalytic residues near catalytic sites. (Table 8, Section 5.4 and Appendix G revised.)

W2 / Q2. Paired CIs, per-seed results and the strict subset

We computed paired protein-level statistics on the h=0h=0 subset (Table 7) and on the strict subset of Table 5, including the parameter-matched non-spatial control (AUPRC per seed 42/43/44; † also removes partner retrieval):

h=0h=0 (n=3,322n=3{,}322)s42s43s44Δ\Delta95% CI
EnzyFuse0.65840.65840.66740.66740.68060.6806——
w/o partner retrieval0.66920.66920.67520.67520.65010.6501−0.0040-0.0040[−0.0081,+0.0000][-0.0081, +0.0000]
w/o spatial edges0.66370.66370.65450.65450.64590.6459−0.0141-0.0141[−0.0183,−0.0099][-0.0183, -0.0099]
w/o interaction branch†0.65980.65980.66430.66430.65610.6561−0.0088-0.0088[−0.0131,−0.0043][-0.0131, -0.0043]
Strict (n=5,746n=5{,}746)s42s43s44Δ\Delta95% CI
EnzyFuse0.72290.72290.71900.71900.72990.7299——
w/o partner retrieval0.71520.71520.71810.71810.70420.7042−0.0114-0.0114[−0.0143,−0.0085][-0.0143, -0.0085]
w/o spatial edges0.70400.70400.71930.71930.70720.7072−0.0137-0.0137[−0.0166,−0.0109][-0.0166, -0.0109]
w/o interaction branch†0.70200.70200.70990.70990.71680.7168−0.0143-0.0143[−0.0177,−0.0111][-0.0177, -0.0111]

On the strict subset, all three components contribute significantly. On h=0h=0, removing the interaction branch or its spatial message passing still lowers AUPRC significantly at the protein level, whereas removing partner retrieval does not, and for each variant at least one seed favors the ablated model; we therefore describe component contributions on this subset as less certain. Because the non-spatial control matches the number of parameters, its drop on both subsets (−0.014-0.014) is attributable to spatial message passing rather than to additional parameters. (Table 7 and Appendix E revised.)

W3 / Q3. Audit scope, residual neighbors and structural coverage

All supervised training proteins. With all 224,427 training proteins as the reference set, the fraction of test enzymes with a detected neighbor increases from 15.5% to 20.8%. On the remaining 5,392 enzymes, EnzyFuse exceeds the Squidly ensemble by +0.045+0.045 ([0.037,0.053][0.037, 0.053]) and SCREEN by +0.080+0.080 ([0.073,0.087][0.073, 0.087]). (Appendix I.1, Table 9.)

Residual neighbors. The split links proteins through cluster membership: MMseqs2 clustering at 30% identity with coverage ≥0.8\ge 0.8 of both sequences, and five Foldseek clusterings with TM-score thresholds from 0.5 to 0.9 and coverage ≥0.7\ge 0.7. The audit instead searches every test protein against all training proteins and accepts a coverage of 0.5. Similar proteins assigned to different clusters, and matches covering only part of a protein, are therefore not linked by the split but are detected by the audit. (Appendix A revised.)

Structural coverage. Test proteins by structural search outcome, where a model is an AlphaFold DB structure in the search set:

Reference setNo modelModel, no hitHit, qTM < 0.5Hit, qTM ≥ 0.5
Catalytic-labeled6132,9972,597597
All training proteins6132,9232,434834

Among the 3,322 h=0h=0 enzymes, 2,863 (86%) have a model but no structural hit, and 459 (14%) have no model, so the h=0h=0 group mainly reflects the absence of detected neighbors rather than missing structures. (Appendix A revised.)

Thank you for the additional experiments. W1 and W3 are resolved, and the evidence and revisions for W2 are appropriate. Overall, the model’s advantage over the baselines is clear, but the attribution of that advantage to the proposed components remains partial: the candidate-only scorer shows the anchor representation contributes little, and on the h = 0 subset the contribution of partner retrieval is not significant. I will retain my rating.

Reviewer_Gzn3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes EnzyFuse, a multi-view framework for catalytic residue prediction that combines protein sequence, evolutionary, and structural information. The method uses pretrained protein encoders, spatial message passing, and a catalytic partner retrieval objective to model local relationships between residues. The authors also introduce a joint sequence-structure similarity split to evaluate generalization under reduced homology. Experiments on 6,804 test proteins show improvements in AUPRC over several existing methods, supported by ablation studies and statistical analyses.

Critical strengths

The joint sequence-structure similarity split is a useful aspect of this work. The authors also retrain several baselines, report results across three seeds, and provide confidence intervals, making the evaluation more convincing than a standard benchmark comparison. EnzyFuse achieves an AUPRC of 0.7568 compared with 0.7012 for the Squidly ensemble. The ablations also suggest that spatial interaction and catalytic partner retrieval contribute to the improvements.

Critical weaknesses
  1. The framework mainly combines established components, including ESM-2, MSA Transformer, equivariant message passing, auxiliary supervision, and logit fusion. The proposed cross-view interaction is essentially feature concatenation followed by spatial message passing. While effective, the overall design feels more like a task-specific integration of existing techniques than a substantial methodological advance.

  2. Table 2 shows that removing the interaction branch reduces AUPRC from 0.7568 to 0.7254, but this ablation also removes catalytic partner retrieval. Although the authors provide separate ablations, a direct comparison with simpler fusion strategies, such as early fusion followed by a standard GNN or attention-based fusion, would better establish the value of the proposed design.

  3. The improvements are not consistent across evaluation metrics. EnzyFuse achieves the best AUPRC and AUROC, but Squidly performs better on F1 (0.5779 vs. 0.5521) and MCC (0.5950 vs. 0.5726). This matters because catalytic residue prediction is ultimately a classification problem, not just a ranking task. The paper should better explain why stronger ranking performance does not translate into better classification results.

  4. The similarity analysis only considers catalytic-labeled training proteins, rather than all proteins used during training. Therefore, the reported no-neighbor subset does not necessarily represent genuinely unseen protein families. In addition, the method still relies on pretrained sequence and MSA representations, whose training data may contain related proteins.

  5. The experiments mainly focus on benchmark metrics. There is little analysis of whether the predicted catalytic residues form chemically meaningful active sites or whether the method can identify functionally important residues in challenging enzyme families. Such evidence would make the biological relevance of the proposed partner retrieval objective more convincing.

  6. The model learns a gating mechanism during training but discards it in favor of simple logit averaging at inference. Although the authors report that averaging performs better, this raises questions about the actual role of the learned gate and whether the additional complexity is necessary.

Questions for authors
  1. Can the authors compare the proposed cross-view interaction against simpler fusion architectures, such as concatenation with a standard GNN or cross-attention, under matched parameter counts?
  2. Why does EnzyFuse outperform Squidly in AUPRC but underperform in F1 and MCC? Have the authors investigated score calibration or alternative threshold-selection strategies?
  3. Can the authors provide a stronger generalization analysis against all training proteins, rather than only those with catalytic annotations?
  4. How much of the performance improvement comes from the catalytic partner retrieval objective versus the additional listwise ranking loss? Can the authors further analyze their interaction?
  5. Why is the learned gate discarded at inference? Could the authors clarify what it learns during training and whether a simpler training strategy could achieve similar performance?
  6. Can the authors provide biological case studies showing how the proposed spatial interaction helps identify catalytic residues that existing methods miss?
3 comments
AuthorsOpenReview ↗

Response to Reviewer Gzn3 (1/3)

We thank the reviewer for the constructive suggestions, which led to several new comparisons and analyses in the revision. Changes in the revised PDF are marked in blue, and new analyses are in Appendix I.

W1. Methodological contribution

We agree that EnzyFuse builds on established encoders and techniques, and the paper presents it as a task-specific design rather than a new general architecture. The revision provides two pieces of evidence that the components alone do not explain the gains. First, concatenating the view features and applying a standard GNN on the same residue graph, the design this description suggests, is significantly worse at a matched number of parameters, as is attention-based fusion (W2 below). Second, the training objectives of EnzyFuse do not improve Squidly when transferred to it: −0.0009-0.0009 [−0.0028,+0.0009][-0.0028, +0.0009], with EnzyFuse still ahead by +0.053+0.053 [0.047,0.060][0.047, 0.060] (Appendix I.5).

W2 / Q1. Comparison with simpler fusion at matched parameter counts

We replaced the interaction branch with two generic designs. All other parts, including the training objectives, partner retrieval and data, were kept the same, with 4.20M parameters against 4.16M for EnzyFuse (three seeds, full test set):

  • Early fusion + standard GNN: per-residue concatenation of the ESM-2, MSA and structural features, followed by a GINE network on the same C$\alpha$ kk-NN graph.
  • Attention fusion + standard GNN: the three views of each residue as tokens in a Transformer layer, followed by the same GNN. Learning rate and dropout were selected on the validation set, because the EnzyFuse settings overfit.
VariantAUPRCΔ\Delta vs. EnzyFuse95% CI
EnzyFuse0.7568±0.00350.7568_{\pm 0.0035}——
Early fusion + standard GNN0.7310±0.00470.7310_{\pm 0.0047}−0.0258-0.0258[−0.0306,−0.0213][-0.0306, -0.0213]
Attention fusion + standard GNN0.7025±0.01370.7025_{\pm 0.0137}−0.0542-0.0542[−0.0588,−0.0496][-0.0588, -0.0496]

Both are significantly below EnzyFuse, so a generic GNN on the combined views does not reach the interaction branch at a matched number of parameters. Because the per-view branches and objectives are unchanged, the variants differ only within the interaction branch, and the comparison measures this design as a whole, including its SO(3)-equivariant message passing. (Table 8 and Appendix E revised.)

W3 / Q2. AUPRC vs F1 and MCC

A single global threshold requires scores to be comparable across proteins, whereas in practice predictions are used to select a few candidate residues within each enzyme for experimental follow-up. Neither a test-optimal nor a length-binned global threshold closes the gap. Selecting residues within each protein, with the rule chosen on the validation set, gives EnzyFuse higher F1:

Selection ruleEnzyFuseSquidly ×5Δ\Delta F1 [95% CI]
Global, validation-selected (Table 1)0.5520.5520.5780.578−0.026-0.026 [−0.033,−0.019][-0.033, -0.019]
Global, test-optimal0.5550.5550.5780.578−0.024-0.024 [−0.031,−0.017][-0.031, -0.017]
Global, per length bin0.5600.5600.5740.574−0.014-0.014 [−0.020,−0.006][-0.020, -0.006]
Within protein: relative to maximum0.6300.6300.5980.598+0.032+0.032 [0.025,0.039][0.025, 0.039]
Within protein: top-10.5720.5720.5240.524+0.048+0.048 [0.041,0.056][0.041, 0.056]

MCC shows the same pattern (Table 11): +0.032+0.032 and +0.051+0.051 under the two within-protein rules. Two observations place the gap in the scale of scores across proteins. Under the validation-selected global threshold, EnzyFuse predicts no catalytic residue for 14.9% of test enzymes, against 7.5% for the Squidly ensemble, and a separate threshold per length bin halves the F1 gap. A likely contributor is the listwise ranking term, which is unchanged when all logits of a protein shift together and therefore improves the order within each protein without fixing the scale across proteins. The lower global-threshold F1 and MCC thus reflect how scores compare across proteins, not how residues are ranked within each protein. (Appendix I.3, Table 11; Section 5.2.)

AuthorsOpenReview ↗

Response to Reviewer Gzn3 (2/3)

W4 / Q3. Similarity to all training proteins and pretrained encoders

With all 224,427 training proteins as the reference set, the fraction of test enzymes with a detected neighbor increases from 15.5% to 20.8%. On the remaining 5,392 enzymes, EnzyFuse exceeds the Squidly ensemble by +0.045+0.045 ([0.037,0.053][0.037, 0.053]) and SCREEN by +0.080+0.080 ([0.073,0.087][0.073, 0.087]). (Appendix I.1, Table 9.)

ESM-2 and the MSA Transformer were pretrained without labels on sequence databases that may contain test proteins. This exposure provides no catalytic labels, which the split controls. Squidly uses ESM-2 3B, pretrained on the same database, so the comparison with Squidly is not explained by pretraining data. We agree that no detected neighbor does not mean an unseen family. The paper therefore describes this subset as having no detected neighbor, and the Limitations now also note the pretraining data. (Section 6 revised.)

Q4. Listwise ranking term and partner retrieval

Listwise termPartner retrievalAUPRCΔ\Delta vs. EnzyFuse95% CI
✓✓0.7568±0.00350.7568_{\pm 0.0035}——
✓✗0.7360±0.00550.7360_{\pm 0.0055}−0.0208-0.0208[−0.0232,−0.0187][-0.0232, -0.0187]
✗✓0.7068±0.00920.7068_{\pm 0.0092}−0.0500-0.0500[−0.0564,−0.0438][-0.0564, -0.0438]
✗✗0.6856±0.01150.6856_{\pm 0.0115}−0.0712-0.0712[−0.0802,−0.0622][-0.0802, -0.0622]

Removing both lowers AUPRC by 0.07120.0712, close to the sum of the individual effects (0.07080.0708). Partner retrieval adds 0.0210.021 with or without the listwise term, so the two contributions are approximately additive and neither substitutes for the other. (Table 8 revised.)

W6 / Q5. Role of the gate

The gate has a training role. Replacing it with equal weights during training, with the same averaging at inference, lowers test AUPRC from 0.75680.7568 to 0.74180.7418 (−0.0150-0.0150 [−0.0196,−0.0106][-0.0196, -0.0106]; Table 8), so this simpler strategy does not reach the same performance. At inference, equal averaging gives higher validation AUPRC than the gate (0.72270.7227 vs. 0.71390.7139; −0.0088-0.0088 [−0.0112,−0.0065][-0.0112, -0.0065]; Appendix B), so the gate is not used there and adds no inference cost.

To show what the gate learns, we recorded its weights on the 6,259 test enzymes with all three views (mean over three runs):

ResiduesSequenceMSAStructureInteraction
All0.3890.3890.1390.1390.0020.0020.4690.469
Catalytic0.3530.3530.1030.1030.0140.0140.5300.530

The gate relies mainly on the interaction and sequence branches and gives almost no weight to the structure branch. This is likely because the interaction branch builds on the structure view: it combines the structural representation of each residue with the sequence and evolutionary ones over the structure graph, and it is available wherever the structure branch is (Section 4.2). Structural evidence therefore reaches the combined prediction mainly through the interaction branch. At annotated catalytic residues, the gate shifts weight further toward the interaction branch (+0.062+0.062 [0.058,0.065][0.058, 0.065], paired over proteins), consistent with the role of spatial context at catalytic sites. The dependence on the protein's sequence identity to the training set is weak: the interaction branch receives 0.0310.031 [0.026,0.036][0.026, 0.036] more weight on proteins with a detected hit, and no branch weight correlates with sequence identity beyond ∣ρ∣=0.17|\rho| = 0.17. The gate thus mainly learns which branches to rely on at which residues, with a small adjustment for similarity.

AuthorsOpenReview ↗

Response to Reviewer Gzn3 (3/3)

W5 / Q6. Biological relevance

Two analyses address whether predictions form active sites and recover functionally important residues. All annotated catalytic residues of an enzyme rank within its top 2n2n residues (nn catalytic residues) for 73.6% of test enzymes, against 66.5% for the Squidly ensemble (+0.072+0.072 [0.063,0.081][0.063, 0.081]). On 69 test enzymes with M-CSA reference residues, EnzyFuse ranks 44.2% of the 169 experimentally characterized catalytic residues that lack a Swiss-Prot annotation among the top 5 of their enzyme (Squidly ×5: 37.7%; Appendix I.2).

We also added two case studies on enzymes with h=0h=0 (Appendix I.6, Figure 10). In Lam81A (GH81), D457 is about 80 residues from the other two catalytic residues in sequence, yet the three line the same pocket, and mutating each reduces or abolishes activity. EnzyFuse ranks all three as its top three, while the Squidly ensemble ranks D457 sixth. In neprilysin-1 (M13), the proton donor D653 is 65 residues from E588 of the zinc-binding motif HExxH but 8.9 Å away in space. EnzyFuse ranks the pair first and second, while the Squidly ensemble ranks D653 fifth. Removing spatial edges at a matched number of parameters drops D653 to third in all three runs, and removing the interaction branch or partner retrieval also drops it out of the top two.

Reviewer_cDqeOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper introduces EnzyFuse, a multimodal model for predicting catalytic residues in enzymes by combining protein sequence, evolutionary information and predicted 3D structures. Each modality is processed separately, and their representations are integrated through spatial interactions between neighboring residues (based on a k-NN graph). The training objectives include residue classification, listwise ranking and catalytic-partner retrieval.

The authors expand the evaluation dataset, split it based on both sequence and structural similarity, and retrain the baselines for comparison.

EnzyFuse consistently outperforms the baselines in protein-averaged AUPRC, including on proteins with no detected sequence or structural homologs among catalytic-annotated training proteins, yet performs worse than Squidly on F1 and MCC. The experiments are supported by extensive ablation studies looking into the contributions of individual modalities, architectural components and training objectives.

Critical strengths

S1. Generalization when there is no reference

In my opinion the most important result in the paper is Appendix D, Table 6.

It shows the performance on a split with no detected neighbors, meaning that neither the sequence nor the structure similarity search returned a qualified match. This is a deployment scenario that is particularly relevant for annotating uncharacterised enzymes.

hnEnzyFuse AUPRCBaselineBaseline AUPRCΔ95% CIp
none (h = 0)3,3220.6688Squidly ×50.6050+0.0639[+0.0526, +0.0750]3.3 × 10⁻³⁶
none (h = 0)3,3220.6688SCREEN0.5527+0.1161[+0.1062, +0.1265]9.2 × 10⁻¹¹¹

EnzyFuse shows significant improvements compared to both baselines in this setting.

I'd encourage the authors to highlight this result more in the main part.


S2. Ablations

The ablation studies are extensive and informative. They examine the contributions of individual modalities, architectural choices, training objectives etc. For most comparisons, the ablated models are also retrained multiple times, giving uncertainty estimates.

Critical weaknesses

W1. Generalization across reactions

The dataset is split using both sequence and structural similarity, rather than relying only on one of them. This is largely inherited from Squidly and applied to build a larger jointly controlled evaluation set.

The authors also look at first-class EC (enzyme commission) for stratification - which aims to preserve the distribution of reaction classes across training, validation and test sets. It does not prevent proteins catalyzing the same reaction from appearing in both training and test sets.

Within a specific reaction, active sites are usually completely conserved in terms of sequence, with small variation in structure - meaning if a model has seen a reaction during its training, it can probably memorize it.

Since the motivation of this work is to generalize to unfamiliar enzymes, it is important to show whether the model can also generalize beyond familiar reactions.


W2. Complete catalytic site recognition

The main reported metric is protein-averaged AUPRC which evaluates how well the model predicts individual catalytic residues. However, it does not directly measure whether the whole catalytic site has been correctly identified. This matters because the activity depends on the coordinated action of multiple residues. A useful additional metric would be the fraction of proteins for which all annotated catalytic residues are recovered within the top-k predictions, stratified by the number of annotated catalytic residues per protein.


W3. Multiple active sites

An enzyme may have one active site with several catalytic residues, multiple distinct active sites (performing the same or different reactions), or an active site spanning multiple protein chains.

The proposed solution represents the problem as a binary classification task, where each residue is either catalytic or not. It allows the model to identify residues belonging to multiple sites, but it does not distinguish between separate sites. Also, the representation operates on individual chains, which suggests that catalytic sites formed at interfaces between chains would be beyond the scope of the method.


W4. Training objectives

The secondary ablation Appendix E, Table 8, shows that one of the biggest contributors to the performance is listwise ranking loss, contributing to AUPRC roughly the entire performance gap between EnzyFuse and Squidly.

Since the authors retrained Squidly on their dataset, it would be interesting to evaluate whether using the same training objectives (i.e. residue classification, listwise ranking and catalytic-partner retrieval ) on Squidly narrows this gap. This would help distinguish improvements from the architecture from the training objectives.


W5. Presentation

The presentation is mixed. Some sections are well written and the figures are generally very well made, but the abstract and the introduction are hard to follow.

Although the idea introduced in the paper is relatively straightforward, the abstract reads like a list of implementation details. It introduces concepts such as views, branches, anchors and gates before explaining their meaning or motivation.

The introduction and related work would benefit from a clearer biological motivation, for example, a discussion of when catalytic residue prediction is genuinely difficult (e.g., when annotated homologs are unavailable or catalytic mechanisms are poorly understood). Although the authors briefly discuss these challenges, they provide few concrete examples illustrating the limitations of existing methods.

Similarly, while previous multimodal approaches are reviewed, it is not clear which of their limitations EnzyFuse specifically addresses, and why the proposed architectural components are necessary to overcome them.

Questions for authors

Q1. How well does your method generalize across different catalytic reactions? EC numbers form a four-level hierarchy, going from broad reaction classes to specific reactions. How does the model perform when training and test sets are made disjoint at different levels of this hierarchy? (W1)

Q2. Have you measured the success rate of complete active site recognition? If so, what are they for different methods? (W2)

Q3. Does the dataset contain enzymes with multiple distinct catalytic sites? If so, how frequent are they and how well EnzyFuse performs on these proteins? (W3)

Q4. How would the model distinguish separate catalytic sites? How would it handle active sites that span multiple chains? (W3)

Q5. How would Squidly perform if retrained with the same training objectives as EnzyFuse, particularly listwise ranking loss? (W4)

6 comments
AuthorsOpenReview ↗

Response to Reviewer cDqe (1/2)

We thank the reviewer for the constructive suggestions on presentation and for highlighting the result on enzymes without detected neighbors. Changes in the revised PDF are marked in blue, and new analyses are in Appendix I.

W1 / Q1. Generalization across reactions

We grouped the test enzymes by the deepest EC level that also occurs among catalytic-labeled training proteins. Every first- and second-level class in the test set occurs in training. In the table, A.B.C.D means the full EC number occurs in training, A.B.C.X that only the fourth level is new, and A.B.X.X that the third level is new. AUPRC:

EC in trainingnnEnzyFuseSquidly ×5Δ\Delta [95% CI]
A.B.C.D2,1650.7710.7710.8170.817−0.046-0.046 [−0.057,−0.037][-0.057, -0.037]
A.B.C.X2,8510.7120.7120.5720.572+0.141+0.141 [0.130,0.152][0.130, 0.152]
A.B.X.X3390.6870.6870.6050.605+0.082+0.082 [0.059,0.105][0.059, 0.105]

The pattern matches the reviewer's hypothesis. When the full EC number occurs in training, the Squidly ensemble is higher. EnzyFuse leads most on enzymes whose reaction is absent from training. The same holds within the h=0h=0 subset (A.B.C.X: +0.153+0.153 [0.138,0.168][0.138, 0.168]; A.B.C.D: −0.087-0.087 [−0.113,−0.063][-0.113, -0.063]), so detected sequence or structural similarity does not explain it. (Appendix I.4, Table 12; Section 5.5.)

W2 / Q2. Recovery of the complete catalytic site

We counted a protein as recovered when all of its annotated catalytic residues rank within the top kk residues of that protein:

kkEnzyFuseSquidly ×5SCREENΔ\Delta vs Squidly ×5 [95% CI]
nn0.5830.5830.5530.5530.5380.538+0.031+0.031 [0.022,0.040][0.022, 0.040]
2n2n0.7360.7360.6650.6650.6160.616+0.072+0.072 [0.063,0.081][0.063, 0.081]
50.7980.7980.7280.7280.6720.672+0.070+0.070 [0.062,0.079][0.062, 0.079]
100.8810.8810.8160.8160.7240.724+0.066+0.066 [0.059,0.073][0.059, 0.073]

Here nn is the number of annotated catalytic residues of the protein. Stratified by nn at k=2nk=2n, EnzyFuse is significantly higher than the Squidly ensemble for n=1n=1 (4,067 proteins; +0.021+0.021 [0.012,0.032][0.012, 0.032]), n=2n=2 (1,428; +0.249+0.249 [0.224,0.272][0.224, 0.272]) and n=3n=3 (1,184; +0.066+0.066 [0.048,0.084][0.048, 0.084]). It is significantly lower for n=4n=4 (−0.355-0.355 [−0.458,−0.256][-0.458, -0.256]; 91 proteins, 64 of them HMG-CoA reductases, viral VP3 proteins or OTU1 deubiquitinases), and the difference is not significant for n≥5n\ge 5 (−0.010-0.010; 34 proteins).

W3 / Q3, Q4. Multiple active sites and multi-chain sites

We grouped catalytic residues into sites by clustering their C$\alpha$ positions with a 15 Å radius, the partner-retrieval radius. Of the 2,578 test enzymes with at least two catalytic residues and a usable structure, 543 (8.0% of all test enzymes) have two or more sites. With a 10 Å or 20 Å radius, the count is 1,123 or 400. Two families account for 59% of the 543 (MurQ, 194; cyanate hydratase, 124). Both form homo-oligomers, and where an active site is assembled from several chains, clustering the residues of a single chain may split it into two; we have not verified this for these families. On the remaining 225 enzymes from 96 families, EnzyFuse reaches 0.6390.639 AUPRC against 0.5330.533 for the Squidly ensemble (+0.106+0.106 [0.063,0.149][0.063, 0.149]). All methods score lower on these enzymes than on single-site enzymes (EnzyFuse 0.8310.831, Squidly ×5 0.7660.766).

As the reviewer notes, EnzyFuse scores residues without assigning them to distinct sites, and it operates on single chains. Sites could be separated afterwards by clustering high-scoring residues in the structure, as in the analysis above, but the model does not learn this assignment. Sites formed at chain interfaces are outside its scope. We state both points in the Limitations. Learning to group predicted residues into sites, and extending the structural view to predicted complex structures so that interface sites can be modeled, are directions we plan to pursue. (Section 6 revised.)

AuthorsOpenReview ↗

Response to Reviewer cDqe (2/2)

W4 / Q5. Squidly with the training objectives of EnzyFuse

We retrained Squidly (three runs, five members each) with its classifier loss replaced by our masked classification loss plus the listwise ranking term and partner retrieval. The partner scorer is the same as in EnzyFuse, reads Squidly's per-residue states and uses the same structures. Squidly's encoder and inference are unchanged. The per-branch auxiliary losses of EnzyFuse do not apply to a single-branch model.

ModelAUPRCΔ\Delta [95% CI]
Squidly ×5, own loss$^*$0.70470.7047—
Squidly ×5, objectives of EnzyFuse0.70380.7038−0.0009-0.0009 [−0.0028,+0.0009][-0.0028, +0.0009] vs. above
EnzyFuse0.75680.7568+0.053+0.053 [0.047,0.060][0.047, 0.060] vs. above

∗^*A repeat of the retraining in Table 1, which reports 0.70120.7012.

The objectives do not improve Squidly, and the difference is not significant on the h=0h=0 or strict subsets either. The large effect of the listwise term in EnzyFuse (Table 8) thus appears to depend on the multi-view architecture it trains, and the training objectives alone do not explain the advantage over Squidly. (Appendix I.5, Table 13.)

W5 and S1. Presentation and the low-homology result

We agree and revised the three parts. The abstract is rewritten to give the problem and the idea before the components: catalytic residues are evolutionarily conserved and cluster in the folded structure, prediction remains difficult for enzymes with few annotated homologs, and EnzyFuse captures these two properties from the sequence, evolutionary and structural views. Each component is introduced with its purpose, views are defined where first used, and branches, anchors and gates no longer appear. The revised abstract reports the same main results as the submitted version and makes no new claim, so it is not marked in blue. The introduction now motivates the low-homology setting: many enzymes encoded in sequenced genomes have few annotated homologs, locating their catalytic residues can suggest the reactions they catalyze and prioritize key residues for experiments, and annotation transfer requires a matching reference that such enzymes often lack (Section 1). The related work now states the limitation EnzyFuse addresses: existing multi-view predictors merge the view features in one network trained with per-residue labels, so that no view is supervised on its own and the relations among neighboring catalytic residues are not directly supervised. EnzyFuse instead gives each view its own residue prediction and combines the views over spatial neighborhoods in an interaction branch trained with partner retrieval (Section 2).

We also thank the reviewer for pointing to the no-detected-neighbor result, which corresponds to annotating uncharacterized enzymes. It is now stated in the abstract and in the introduction, which gives the subset size (3,322 enzymes, 48.8% of the test set) and notes that the advantage is largest on enzymes whose reaction is absent from training (W1 above).

Thank you for your detailed response and adjustments. I consider W3, W4 and W5 resolved.

Two follow up questions:

  • [W1 / Q1] Your catalytic test subset (6,804 proteins) is smaller in this experiment (2,165 + 2,851 + 339). What happened to the remaining proteins?
  • [W2 / Q2] Why do you think your model performs poorly on proteins with exactly four annotated catalytic residues? You mention that this subset consists largely of HMG-CoA reductases, viral VP3 proteins or OTU1 deubiquitinases. Why are these enzymes more challenging for your model?
Reviewer_pemxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces EnzyFuse, a multi-view model for catalytic residue prediction. It combines three sources of protein information: sequence representations from ESM-2, evolutionary information from multiple sequence alignments, and structural information from predicted protein structures. Its main contribution is a structure-guided cross-view interaction branch that combines information from spatially neighboring residues. The authors also introduce catalytic partner retrieval as an auxiliary training objective, using existing catalytic residue annotations to supervise local relationships. The model is evaluated on a homology-controlled dataset containing 6,804 test proteins. EnzyFuse achieves an AUPRC of approximately 0.757, outperforming the evaluated baselines, including the Squidly ensemble. Ablation experiments further examine the contributions of individual views and spatial interaction.

Critical strengths
  1. Strong attention to data leakage and generalization One of the most convincing aspects of this work is its evaluation protocol. The authors control both sequence similarity and structural similarity when constructing training, validation, and test splits. This is important because sequence-based separation alone may leave structurally similar proteins across different partitions. The additional similarity audits make the generalization claims more carefully supported than those based on conventional random splits.
  2. Meaningful model design and component analysis The combination of directly supervised view branches, spatial interaction, and catalytic partner retrieval is well motivated by the biological characteristics of catalytic sites. The ablation results are also informative: removing spatial edges decreases AUPRC by approximately 0.034, while removing partner retrieval decreases it by approximately 0.021. These experiments provide evidence that the improvement is not simply due to combining several pretrained protein representations.
Critical weaknesses
  1. Incomplete catalytic annotations may affect evaluation The paper acknowledges that unannotated residues are treated as negative examples. However, the absence of a catalytic annotation does not necessarily mean that a residue has no catalytic function. This introduces potential false negatives and may affect the measured AUPRC. The model is therefore evaluated against existing annotations rather than a fully verified experimental ground truth. This limitation should be considered when interpreting its absolute performance.
  2. The generalization claim needs further qualification The similarity audit focuses on catalytic-labeled training proteins rather than every protein used during training. The authors explicitly acknowledge that having no detected neighbor in this subset does not establish evolutionary independence from all training inputs. Additionally, although EnzyFuse achieves the best AUPRC, the Squidly ensemble performs better on global-threshold F1 and MCC. SCREEN also performs better in the highest-similarity group. Therefore, the advantage is not uniform across evaluation settings.
Questions for authors
  1. Can the authors evaluate EnzyFuse on an independent experimentally validated catalytic-residue dataset to estimate how incomplete annotations affect the reported performance?
  2. Can the authors provide an additional similarity audit against all training proteins, including those without catalytic annotations, to better support the claim of generalization to proteins without close annotated homologs?
2 comments
AuthorsOpenReview ↗

Response to Reviewer pemx

We thank the reviewer for the constructive questions on label completeness and generalization, which led to the evaluation with M-CSA labels and the extended audit. Changes in the revised PDF are marked in blue, and new analyses are in Appendix I.

W1 / Q1. Incomplete annotations and an independent, experimentally validated dataset

We agree that unannotated residues may include true catalytic residues. We evaluated on M-CSA, an independent, manually curated database of experimentally characterized catalytic residues (reference entries only, no homology-transferred annotations). Of the 488 enzymes in our data with M-CSA reference residues, 69 are test proteins; the others are in the training or validation splits and cannot be used without breaking the similarity control. On these 69 enzymes, M-CSA lists 169 catalytic residues that Swiss-Prot does not annotate and that our main evaluation counts as negatives. Keeping all predictions fixed and changing only the labels (AUPRC):

Labels (69 test enzymes)EnzyFuseSquidly ×5SCREEN
Swiss-Prot0.7100.7100.6750.6750.6810.681
M-CSA0.6360.6360.5700.5700.5110.511

Under M-CSA labels, EnzyFuse exceeds the Squidly ensemble by +0.066+0.066 (95% CI [0.020,0.116][0.020, 0.116]) and SCREEN by +0.126+0.126 ([0.079,0.175][0.079, 0.175]), whereas neither difference is significant under Swiss-Prot labels on this subset. EnzyFuse also ranks 44.2% of the 169 unannotated residues among the top 5 of their enzyme (Squidly ×5: 37.7%; SCREEN: 30.8%; random: 1.4%), where the main evaluation counts them as false positives. On this subset, incomplete Swiss-Prot labels appear to understate rather than inflate the advantage of EnzyFuse. (Appendix I.2, Table 10; Limitations.)

W2 / Q2. Generalization: audit against all training proteins

We repeated the audit with all 224,427 training proteins as the reference set and the same thresholds (sequence identity ≥0.3\ge 0.3 or query-normalized TM-score ≥0.5\ge 0.5). The fraction of test enzymes with a detected neighbor increases from 15.5% to 20.8%:

Reference setTest enzymes retainedΔ\Delta vs Squidly ×5Δ\Delta vs SCREEN
Catalytic-labeled (Table 5)5,746+0.066+0.066 [0.059,0.074][0.059, 0.074]+0.102+0.102 [0.095,0.110][0.095, 0.110]
All training proteins (Table 9)5,392+0.045+0.045 [0.037,0.053][0.037, 0.053]+0.080+0.080 [0.073,0.087][0.073, 0.087]

The advantage is smaller but remains significant. We agree that the absence of a detected neighbor does not establish evolutionary independence and keep this qualification in the Limitations. (Appendix I.1; Section 1; Limitations.)

W2. Generalization: global-threshold F1 and MCC

We agree that the Squidly ensemble achieves higher F1 and MCC at a single global threshold. A global threshold requires scores to be comparable across proteins, whereas in practice predictions are used to prioritize a few candidate residues within each enzyme. Neither a test-optimal nor a length-binned global threshold closes the gap, but selecting residues within each protein, with the rule chosen on the validation set, gives EnzyFuse higher F1:

Selection ruleEnzyFuseSquidly ×5Δ\Delta F1 [95% CI]
Global, validation-selected (Table 1)0.5520.5520.5780.578−0.026-0.026 [−0.033,−0.019][-0.033, -0.019]
Global, test-optimal0.5550.5550.5780.578−0.024-0.024 [−0.031,−0.017][-0.031, -0.017]
Global, per length bin0.5600.5600.5740.574−0.014-0.014 [−0.020,−0.006][-0.020, -0.006]
Within protein: relative to maximum0.6300.6300.5980.598+0.032+0.032 [0.025,0.039][0.025, 0.039]
Within protein: top-10.5720.5720.5240.524+0.048+0.048 [0.041,0.056][0.041, 0.056]

MCC shows the same pattern (Table 11): EnzyFuse is lower under both global thresholds (−0.022-0.022, −0.021-0.021) and higher under both within-protein rules (+0.032+0.032, +0.051+0.051), with all four differences significant. The gap thus reflects how scores compare across proteins rather than how residues are ranked within each protein, consistent with the higher top-kk hit rates of EnzyFuse (Table 4). (Appendix I.3, Table 11; Section 5.2.)

W2. Generalization: highest-similarity group

As stated in Section 5.5, the gains are not uniform across similarity ranges. SCREEN is significantly higher only for the 243 test enzymes (3.6%) with h≥0.6h \ge 0.6, where all three methods exceed 0.93 AUPRC (Table 6). In the largest group, enzymes with no detected neighbor (h=0h = 0; 3,322 enzymes, 48.8%), EnzyFuse exceeds the Squidly ensemble by +0.064+0.064 ([0.053,0.075][0.053, 0.075]) and SCREEN by +0.116+0.116 ([0.106,0.127][0.106, 0.127]). This is the low-homology setting that the method targets.

Thank you for the additional information. My concerns are solved. I will keep my rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Summary of Revisions

We thank all reviewers for their constructive feedback. Reviewers noted the joint sequence–structure similarity control and the additional audits (pemx, Gzn3, 14Sw, Yk2p), the extensive ablations with retrained models and confidence intervals (pemx, cDqe, Gzn3, 14Sw), and the result on enzymes with no detected neighbor, which corresponds to annotating uncharacterized enzymes (cDqe, 14Sw). In the revised PDF, changes are marked in blue and new analyses are collected in Appendix I. Below we summarize the main points raised and how the revision addresses them; details are in the individual replies.

1. Generalization beyond detected similarity (pemx, cDqe, Gzn3, 14Sw, Yk2p)

  • An audit against all 224,427 training proteins, rather than only catalytic-labeled ones: on the 5,392 test enzymes without a detected neighbor, EnzyFuse still exceeds the Squidly ensemble by 0.045 and SCREEN by 0.080 in AUPRC (Appendix I.1).
  • A grouping by EC number: the advantage is largest on enzymes whose reaction is absent from training, also within the h=0h=0 subset (Appendix I.4).
  • Paired protein-level tests of the components on the h=0h=0 and strict subsets (Table 7, Appendix E).

2. Label completeness (pemx)

  • With M-CSA labels of experimentally characterized residues, EnzyFuse remains the best method, and its lead over the Squidly ensemble becomes significant (0.066; Appendix I.2).

3. AUPRC versus F1 and MCC (pemx, Gzn3)

  • When residues are selected within each protein, as in experimental prioritization, EnzyFuse achieves higher F1 and MCC than the Squidly ensemble; the gap in Table 1 reflects how scores compare across proteins under one global threshold (Appendix I.3).

4. Source of the advantage (cDqe, Gzn3, 14Sw)

  • Generic fusion designs, early fusion or attention fusion followed by a standard GNN, are significantly worse than the interaction branch at a matched number of parameters (Table 8).
  • Retraining the Squidly ensemble with the training objectives of EnzyFuse does not change its AUPRC significantly, so the advantage depends on the multi-view architecture (Appendix I.5).
  • A partner scorer without the anchor representation retains most of the gain of partner retrieval (Table 8, Section 5.4), and Table 2 now identifies the variant without the interaction branch as the three-branch late-fusion model.

5. Biological relevance (Gzn3)

  • Two case studies on enzymes with h=0h=0 whose catalytic residues are far apart in sequence but close in space (Appendix I.6).

6. Presentation and experimental details (cDqe, 14Sw, Yk2p)

  • Following Reviewer cDqe's comments, the abstract, introduction and related work were revised for clarity and biological motivation. The abstract reports the same main results as the submitted version and is therefore not marked in blue.
  • The Squidly configuration and its BLAST fallback applied to both models (Appendix C), and the split construction and structural coverage (Appendix A).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.