acceptodds
Under review as a conference paper at ICLR 2027

A Stereochemistry-Aware Vector–Scalar Framework for Enzyme–Substrate Specificity Prediction

Abstract

Enzyme specificity determines which substrates are selectively recognized and catalyzed, but existing computational methods often provide limited modeling of pair-specific three-dimensional residue–substrate interactions. Here, we develop StereoVSE, a stereochemistry-aware Vector–Scalar framework for enzyme–substrate specificity prediction and candidate residue prioritization. StereoVSE integrates protein sequence representations, molecular features, and enzyme–substrate structural configurations through a Vector–Scalar graph neural network and Bidirectional Residue–Atom Interaction Alignment, enabling substrate-conditioned residue–atom interaction learning. Experiments on halogenase and glycosyltransferase benchmarks demonstrate improved specificity prediction performance over existing methods. Furthermore, StereoVSE generates substrate-conditioned residue–atom attribution signals that complement sequence conservation analysis and enable candidate residue prioritization.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Keywords: Enzyme–substrate specificity prediction, Protein–ligand interaction modeling, Geometric deep learning, Vector–scalar representation learning, Molecular representation learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_YUg2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces StereoVSE, a multimodal framework for predicting enzyme–substrate specificity and prioritizing candidate specificity-determining residues. It combines pretrained protein representations, molecular features, and docking-derived pocket–substrate structures. A coupled vector–scalar graph encoder captures directional geometric information, while bidirectional residue–atom cross-attention learns substrate-conditioned interaction representations. Experiments on halogenase and glycosyltransferase benchmarks report improvements over six baselines, supported by component ablations and additional halogenase cold-start evaluations. Residue-level analyses compare attention-derived rankings with evolutionary conservation and examine prediction changes under computational mutations. Overall, the framework demonstrates promising predictive performance, although broader generalization, architectural novelty, and biological interpretation require stronger independent validation and controlled comparisons.

Critical strengths
  • This paper reports consistent improvements over six baselines on both benchmarks, including halogenase AUPR increasing from 0.8664 to 0.9112 and glycosyltransferase MCC increasing from 0.6198 to 0.6813.
  • This paper combines complementary sequence, chemical, and structural representations within a coherent architecture, with ablations on both datasets supporting the contributions of the structural encoder and interaction module.
  • This paper extends prediction to substrate-conditioned residue prioritization and appropriately acknowledges that attention scores represent attribution signals and that its mutation analysis holds the structural context fixed.
Critical weaknesses
  • This paper provides insufficient evidence of robust generalization: cold-start experiments cover only halogenases, identity-based partitions do not exclude close homologs or similar molecular scaffolds, and “pair-cold” evaluation remains a seen-enzyme, seen-substrate setting rather than a jointly unseen setting.
  • This paper does not sufficiently isolate its architectural novelty from established vector–scalar message passing and bidirectional cross-attention. Matched-encoder, matched-capacity comparisons against alternative geometric encoders and simpler interaction modules would clarify whether the gains arise from the proposed mechanisms.
  • This paper provides limited independent validation of residue attribution. Spatial proximity, disagreement with conservation, and prediction changes from the same model do not establish biological specificity determinants; quantitative evaluation against experimentally characterized mutations or prospective biochemical validation is needed.
Questions for authors

NA

11 comments
AuthorsOpenReview ↗

Q1: Insufficient generalization evidence from limited cold-start splits

R1: Expanded cold-start evaluations and adoption of double-cold protocol for generalization assessment

We sincerely thank the reviewer for this insightful and constructive comment regarding the generalization capability of StereoVSE.

Regarding the evaluation scope, we have substantially extended our cold-start experiments beyond the halogenase benchmark. Specifically, we conducted three complementary cold-start evaluations—enzyme-cold, substrate-cold, and double-cold—on the glycosyltransferase benchmark as well. These settings assess the ability of StereoVSE to generalize to unseen enzyme identities, unseen substrate identities, and simultaneously unseen enzyme–substrate identities, respectively. The additional glycosyltransferase results are provided in Table R8 of the revised manuscript.

Table R8: Cold-start evaluation on the glycosyltransferase benchmark under enzyme-cold, substrate-cold, and double-cold settings. Best results are bolded.

ModelEnzyme-cold AUCEnzyme-cold AUPREnzyme-cold MCCSubstrate-cold AUCSubstrate-cold AUPRSubstrate-cold MCCDouble-cold AUCDouble-cold AUPRDouble-cold MCC
CPI0.8720.5690.5130.8280.4700.4100.7640.4490.422
ESP0.8640.6180.4540.8560.6410.5110.7760.4510.235
CATNIP0.8460.5660.2610.8390.5630.3510.7570.4850.329
CLIPZyme0.8670.6010.4840.8630.5940.4980.8030.4390.206
RAMMESI0.8780.5950.5040.8760.5890.4940.7960.4580.407
EZSpecificity0.8860.5910.5310.9090.6890.5780.8250.5060.429
StereoVSE0.8960.6210.5700.9130.7070.5830.8630.5290.485

Regarding the limitations of the original pair-cold setting, we replaced the original pair-cold protocol with a more stringent enzyme–substrate double-cold setting, in which both the enzyme and substrate identities appearing in the test set are strictly disjoint from those in the training set.

Regarding the experimental results (Table R9), StereoVSE consistently maintains its performance advantages under the expanded cold-start evaluations. On the halogenase benchmark, StereoVSE achieves an AUPR of 0.762 under the double-cold setting, compared with 0.620 for EZSpecificity. Similarly, on the glycosyltransferase benchmark (Table R8), StereoVSE achieves an AUPR of 0.529, compared with 0.506 for EZSpecificity. Furthermore, StereoVSE outperforms the evaluated baselines under the enzyme-cold and substrate-cold settings on both benchmarks. These additional results provide stronger evidence that the predictive improvements are not restricted to previously observed enzyme or substrate identities.

Table R9: Cold-start evaluation on the halogenase benchmark under the enzyme–substrate double-cold setting. Best results are bolded.

ModelAUC$\uparrow$AUPR$\uparrow$MCC$\uparrow$
CPI0.4950.2670.156
ESP0.6750.2830.148
CATNIP0.7860.4240.349
CLIPZyme0.8400.5370.366
RAMMESI0.5920.2130.066
EZSpecificity0.8700.6200.335
StereoVSE0.8820.7620.394

We sincerely appreciate the reviewer's valuable suggestions. The revised cold-start protocols and additional results are presented in Section 4.3 and Appendix E.1 of the revised manuscript. The concerns regarding homologs and similar molecular scaffolds will be addressed in Response R4 (Rigorous cold-start evaluations controlling for homologous enzymes and substrate scaffold similarity are ongoing, and we will report these results once they are available) .

AuthorsOpenReview ↗

Q2: Disentangling architectural novelty from existing geometric and attention modules

A2: Clarification of architectural novelty via controlled substitution ablations

We sincerely thank the reviewer for this insightful and constructive comment regarding the architectural novelty of StereoVSE and the need to distinguish its contributions from established geometric message-passing and cross-attention mechanisms.

First, we would like to clarify that our contribution is not the introduction of vector–scalar message passing or bidirectional cross-attention as entirely new operations. Rather, StereoVSE integrates these mechanisms into a unified framework for enzyme–substrate specificity prediction. Specifically, the framework constructs a joint three-dimensional pocket–substrate graph that explicitly represents intermolecular spatial relationships, employs coupled scalar and directional vector representations to encode the local binding geometry, and combines these structural representations with bidirectional residue–atom interaction features and global chemical information for specificity prediction. This design aims to capture complementary geometric and sequence–chemical information within a single predictive framework.

To better isolate the architectural contributions, we have conducted additional controlled ablation experiments in which individual components are replaced with alternative implementations while the remaining components and pretrained representations of StereoVSE are retained.

In particular, we considered three architectural alternatives (Table R10):

1.Intra-molecular 3D: Removing cross-molecular pocket–substrate edges while retaining the intramolecular three-dimensional information.

2.BRIA → dual cross-attention: Replacing BRIA with EZSpecificity's dual cross-attention interaction module.

3.V–S → EGNN: Replacing the coupled Vector–Scalar geometric encoder with EZSpecificity's distance-based EGNN.

Table R10: Ablation results on the halogenase benchmark. The best results are shown in bold.

VariantAUCAUPRF1ACCMCC
intra-only 3D0.96050.87230.78700.90980.7298
BRIA$\rightarrow$dual cross-atten0.96570.88170.79360.91710.7431
V-S$\rightarrow$EGNN0.96190.86940.77750.91250.7260
StereoVSE0.97280.91120.81380.92380.7666

The results consistently favor the complete StereoVSE architecture on both benchmarks. On the halogenase dataset, the full model achieves an AUPR of 0.9112, whereas replacing BRIA with dual cross-attention reduces AUPR to 0.8817, replacing the Vector–Scalar encoder with EGNN reduces it to 0.8694, and removing cross-molecular structural edges reduces it to 0.8723. Similarly, on the glycosyltransferase dataset (Table R11), the full model achieves an AUPR of 0.7338, compared with 0.7036 for the dual cross-attention variant, 0.7010 for the EGNN variant, and 0.6973 for the intra-molecular 3D variant.

Table R11: Architectural ablation results on the glycosyltransferase benchmark. The best results are shown in bold.

Ablation ModuleAUC$\uparrow$AUPR$\uparrow$F1$\uparrow$ACC$\uparrow$MCC$\uparrow$
intra-only 3D0.92590.69730.72120.87120.6410
BRIA$\rightarrow$dual cross-atten0.92670.70360.72750.86530.6506
V-S$\rightarrow$EGNN0.92900.70100.73750.87190.6632
StereoVSE0.93640.73380.75000.87620.6813

These comparisons are conducted within the same overall framework, with the remaining pretrained representations and modeling components preserved. Therefore, the observed performance differences provide evidence that the proposed architectural choices contribute to predictive performance beyond the selection of pretrained protein and molecular representations.

We sincerely appreciate the reviewer's valuable suggestion, which has helped us clarify the architectural contribution and the interpretation of our experimental evidence. The additional architectural ablation results are presented in Section 4.4 and Appendix D of the revised manuscript.

AuthorsOpenReview ↗

R3: Independent experimental validation for residue attribution is required

R3: Retrospective Mutagenesis Validation for Residue Prioritization

We sincerely thank the reviewer for this insightful and constructive suggestion regarding the independent experimental validation of residue rankings.

To address this concern, we conducted an additional retrospective evaluation using an independently published mutagenesis screen of the RebH variant 3-LSR [1]. In this study, 13 positions near the substrate-binding pocket were experimentally screened for improved bromination activity toward indole-6-carboxylic acid. Among these positions, four (Y455, F465, W466, and N470) were identified as activity-enhancing positions, each having at least one mutation that increased bromination activity by at least twofold relative to the parental RebH 3-LSR enzyme.

Table R11. Experimentally Identified Activity-Enhancing Mutations in RebH 3-LSR [1]

Experimentally identified positionActivity-enhancing single mutantsReported activity relative to 3-LSR
Y455Y455W, Y455C≥2-fold
F465F465K, F465N≥2-fold
W466W466I, W466Y≥2-fold
N470N470S, N470K≥2-fold

We evaluated BRIA-derived residue rankings against these experimentally characterized outcomes. Specifically, using wild-type RebH, we calculated a substrate-pair differential attention score based on the absolute difference in residue-level BRIA attention between L-tryptophan and indole-6-carboxylic acid. All ranking methods were evaluated on the same 13 experimentally screened positions, with the four activity-enhancing positions treated as positives, ensuring a consistent comparison.

Notably, BRIA substrate-pair differential attention ranked all four experimentally identified activity-enhancing positions above the remaining nine screened positions, achieving an AUROC of 1.000. In comparison, global BRIA ranking achieved an AUROC of 0.722, minimum distance to bound L-tryptophan achieved 0.694, distance to the L-tryptophan carboxyl group achieved 0.528, and alignment-based sequence conservation achieved 0.222.

Table R12. Quantitative Comparison of Residue-Ranking Methods Against Experimental RebH Mutagenesis Data

Ranking methodAUROCExact permutation p-value
BRIA substrate-pair differential attention1.0000.0028
Global BRIA ranking0.7220.26
Minimum distance to bound L-tryptophan0.6940.33
Distance to the L-tryptophan carboxyl group0.5280.94
Alignment-based sequence conservation0.2220.15

The two-sided exact permutation test yielded a p-value of 0.0028 for the BRIA substrate-pair ranking, supporting the statistical significance of the observed separation within the 13 screened positions. Furthermore, when ranked across the complete 530-residue sequence, W466, N470, Y455, and F465 occupied positions 37, 39, 62, and 64, respectively.

To assess potential training-data overlap, we repeated the analysis after excluding N470, as an N470S variant was included in the training data. The BRIA substrate-pair ranking maintained an AUROC of 1.000 across the remaining 12 positions (two-sided exact permutation p = 0.0091), indicating that the observed result was not solely driven by N470.

Ref.

[1] Sana B, Ho T, Kannan S, Ke D, Li EHY, Seayad J, Verma CS, Duong HA, Ghadessy FJ. Engineered RebH Halogenase Variants Demonstrating a Specificity Switch from Tryptophan towards Novel Indole Compounds. ChemBioChem. 2021;22(18):2791–2798. doi: 10.1002/cbic.202100210. PMID: 34240527.

Reviewer_wcHbOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

StereoVSE predicts enzyme–substrate specificity by combining protein and molecular representations with docking-derived pocket–substrate graphs. It uses vector–scalar geometric learning and bidirectional residue–atom attention, which also provides candidate residue rankings.

Critical strengths
  1. The framework combines complementary sequence, molecular, and structural information.
  2. Experiments cover two enzyme families and include cold-start evaluation on halogenases.
  3. Ablations provide useful evidence that structural features contribute substantially to prediction.
Critical weaknesses
  1. Architectural improvements are not fully separated from pretrained-encoder improvements.
  2. Generalization and stereochemical sensitivity need more targeted evaluation.
  3. Residue prioritization lacks independent validation, and some metric definitions need clarification.
Questions for authors
  1. The ESM-2 ablation falls below EZSpecificity on glycosyltransferase AUPR: 0.6719 versus 0.6805. Please compare with EZSpecificity using the same protein and chemical encoders to separate encoder gains from architectural gains.
  2. The cold splits exclude exact identities, but close enzyme homologs or substrate analogues may still cross partitions. Please report cross-split similarities and add homology- and scaffold-separated evaluations.
  3. The stereochemistry claim is supported by general 3D ablations rather than a direct stereochemical test. Please evaluate matched stereoisomers and clarify what geometric information remains in the “w/o V–S” baseline.
  4. The T467A experiment measures sensitivity of the same model that ranks the residues, while keeping its docking graph fixed. Please add random-residue and conservation-based controls, and validate the rankings against independent mutagenesis evidence where available.
  5. Please reconcile Table A2’s F1 of 0.7500 and accuracy of 0.8762 with the 7.79% positive ratio reported in Table A1. Specify the test-set class balance, metric averaging, and any negative sampling, and provide confusion matrices.
5 comments
AuthorsOpenReview ↗

Q1: Disentangling encoder and architectural gains via matched-encoder comparison with EZSpecificity

R1: Disentangling architectural contributions from pretrained encoder effects

We thank the reviewer for this valuable comment regarding the distinction between architectural improvements and gains from pretrained encoders. We agree that the contribution of the proposed architecture should be evaluated separately from the influence of the protein encoder.

Initially, the ESM-2 variant reused the hyperparameters optimized for ESMC (learning rate = 3 × 10⁻⁴, cutoff = 10 Å, 48 neighbors, and 6 GNN layers), without encoder-specific tuning. Nevertheless, its AUC (0.9157) already exceeded that of EZSpecificity (0.9147), while its AUPR (0.6719) was close to EZSpecificity (0.6805) on the glycosyltransferase benchmark (Table R5).

To ensure a fairer comparison, we independently optimized the ESM-2 variant, selecting a learning rate of 3 × 10⁻⁴, a cutoff of 12 Å, 32 neighbors, and 4 GNN layers. This improved its AUC to 0.9223 and AUPR to 0.6811, both numerically exceeding EZSpecificity.

Table R5: Architectural ablation results on the glycosyltransferase benchmark. The best results are shown in bold.

Ablation ModuleAUC$\uparrow$AUPR$\uparrow$F1$\uparrow$ACC$\uparrow$MCC$\uparrow$
ESMC →\rightarrow ESM-20.92230.68110.71210.86230.6289
intra-only 3D0.92590.69730.72120.87120.6410
BRIA$\rightarrow$dual cross-atten0.92670.70360.72750.86530.6506
V-S$\rightarrow$EGNN0.92900.70100.73750.87190.6632
StereoVSE0.93640.73380.75000.87620.6813

The updated experimental results and corresponding analysis are provided in Section 4.4 and Appendix D of the revised manuscript.

AuthorsOpenReview ↗

Q2: Independent validation and controls for residue prioritization

R2: Retrospective mutagenesis validation for BRIA residue prioritization

We sincerely thank the reviewer for this insightful and constructive suggestion regarding the experimental validation of BRIA-derived residue rankings. The model-based perturbation sensitivity alone is insufficient to establish the biological relevance or superiority of the identified residues.

To address this concern, we have added a retrospective evaluation using an independently published mutagenesis screen of the RebH variant 3-LSR [1]. In the published study, 13 positions near the substrate-binding pocket were experimentally screened for improved bromination activity toward indole-6-carboxylic acid. Among these positions, four (Y455, F465, W466, and N470) yielded variants exhibiting at least a twofold improvement in activity over the parental enzyme.

Table R6. Experimentally Validated Activity-Enhancing Mutations in RebH 3-LSR [1]

Experimentally identified positionActivity-enhancing single mutantsReported activity relative to 3-LSR
Y455Y455W, Y455C≥2-fold
F465F465K, F465N≥2-fold
W466W466I, W466Y≥2-fold
N470N470S, N470K≥2-fold

We evaluated BRIA-derived residue rankings against these experimental outcomes. Specifically, we calculated a substrate-pair differential attention score using the absolute difference in residue-level BRIA attention between L-tryptophan and indole-6-carboxylic acid. All ranking methods were evaluated on the same 13 experimentally screened positions to ensure a consistent comparison.

The BRIA substrate-pair differential attention successfully ranked all four experimentally identified activity-enhancing positions above the remaining nine positions, achieving an AUROC of 1.000. In comparison, global BRIA ranking achieved an AUROC of 0.722, minimum distance to bound L-tryptophan achieved 0.694, distance to the L-tryptophan carboxyl group achieved 0.528, and alignment-based sequence conservation achieved 0.222.

Table R7. Quantitative Comparison of Residue-Ranking Methods Against Experimental RebH Mutagenesis Data

Ranking methodAUROCExact permutation p-value
BRIA substrate-pair differential attention1.0000.0028
Global BRIA ranking0.7220.26
Minimum distance to bound L-tryptophan0.6940.33
Distance to the L-tryptophan carboxyl group0.5280.94
Alignment-based sequence conservation0.2220.15

The two-sided exact permutation test yielded a p-value of 0.0028 for the BRIA substrate-pair ranking. Importantly, when ranked across the complete 530-residue sequence, W466, N470, Y455, and F465 occupied positions 37, 39, 62, and 64, respectively.

We additionally considered potential overlap with the training data. Because the training variants included N470S, we repeated the analysis after excluding N470. The BRIA substrate-pair ranking maintained an AUROC of 1.000 across the remaining 12 positions (two-sided exact permutation p = 0.0091), providing further support for the robustness of this retrospective observation.

Overall, the additional RebH analysis provides retrospective experimental support for the potential utility of substrate-conditioned BRIA attention in prioritizing positions associated with improved catalytic activity toward a specified substrate.

Ref.

[1]. Sana B, Ho T, Kannan S, Ke D, Li EHY, Seayad J, Verma CS, Duong HA, Ghadessy FJ. Engineered RebH Halogenase Variants Demonstrating a Specificity Switch from Tryptophan towards Novel Indole Compounds. Chembiochem. 2021 Sep 14;22(18):2791-2798. doi: 10.1002/cbic.202100210. Epub 2021 Jul 29. PMID: 34240527; PMCID: PMC8518859.

AuthorsOpenReview ↗

Q3: Some metric definitions need clarification

R3: Clarification of dataset sampling, class distribution and evaluation metrics

We sincerely thank the reviewer for this careful observation and for highlighting the need to clarify the relationship between the dataset class distribution and the reported evaluation metrics. We appreciate the opportunity to clarify the sampling procedure, test-set composition, metric definitions, and aggregation protocol.

1. Original dataset versus evaluation benchmark. The positive ratio of 7.79% reported in Table A1 corresponds to the original candidate pool, which contains 30,865 enzyme–substrate pairs, including 2,405 positive and 28,460 negative pairs. Importantly, this is not the class distribution of the test sets used to calculate the reported metrics.

For evaluation, we constructed a negative-subsampled benchmark by retaining all 2,405 positive pairs and sampling approximately four reaction-matched negative candidates per positive pair from the EC 2.4.1 enzyme pool. Negative candidates were sampled independently for each positive entry. Because multiple positive entries could share the same reaction, some negative candidates were selected more than once. After deduplication by the (enzyme, reaction) identifier, 8,873 unique negative pairs remained, resulting in 11,278 benchmark pairs with a positive ratio of 21.32%. This configuration keeps the evaluation protocol consistent with that of the other benchmark.

2. Actual test-set class balance and structural filtering. For example, in Fold-1, the initially held-out test set contained 1,410 pairs, comprising 300 positives and 1,110 negatives (21.28% positives). After excluding pairs without valid structural inputs, the final test set contained 1,351 pairs, including 289 positives and 1,062 negatives (21.39% positives). All reported test metrics were computed on these structurally valid held-out samples.

Metrics were calculated over the complete filtered test set for each evaluation run. The values reported in Table A2 are the arithmetic means of the corresponding metrics across four experimental.

The confusion matrix for the structurally valid test set in Fold-1 is:

Actual classPredicted positivePredicted negative
PositiveTP = 239FN = 50
NegativeFP = 111TN = 951

These counts yield an F1 score of 0.7480, an accuracy of 0.8808, and an MCC of 0.6762. Its AUROC and AUPR are 0.9368 and 0.7477, respectively. The corresponding values in Table A2: F1 = 0.7500, accuracy = 0.8762, and MCC = 0.6813, represent the averages across the four experimental runs.

3. Metric definitions and aggregation. We used standard binary classification metrics with label 1 denoting a compatible enzyme–substrate pair. Binary predictions were obtained using a fixed probability threshold of 0.5, without threshold optimization on the test set. Accuracy was calculated as the proportion of correctly classified pairs; precision and recall were calculated for the positive class; F1 was their harmonic mean; and MCC was calculated from all four confusion-matrix entries. Please refer to Response R5 to Reviewer 1 for the definitions of these equations.

We will clarify the distinction between the original candidate-pool distribution and the actual evaluation distribution, document the negative-sampling and structural-filtering procedures, and provide the run-level confusion matrices and metric-aggregation details in the revised manuscript.

We sincerely appreciate the reviewer's valuable comment, which has helped us improve the transparency and reproducibility of our evaluation protocol.

Reviewer_bU1YOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces StereoVSE for enzyme–substrate specificity prediction and candidate residue prioritization. It combines ESMC protein embeddings, molecular representations, and docking-derived pocket–substrate graphs. Its main components are a coupled vector–scalar geometric encoder and Bidirectional Residue–Atom Interaction Alignment (BRIA). Evaluation includes four-fold pair-level validation on halogenase and glycosyltransferase benchmarks, additional halogenase holdout settings, and component ablations. StereoVSE shows improvements over six baselines. Residue-level analyses compare BRIA rankings with conservation across 19 enzymes and examine selected structural contacts and one sequence-perturbation example. These analyses provide exploratory attribution evidence, although the usefulness of the residue rankings remains insufficiently established.

Critical strengths
  1. StereoVSE reports promising improvements over the compared baselines on the two evaluated enzyme-family benchmarks.
  2. The framework coherently integrates molecular features and directional geometry, with ablations suggesting benefits from its structural and interaction modules.
Critical weaknesses
  1. The stereochemistry-specific contribution is not demonstrated. The title and framing emphasize stereochemistry, but the manuscript does not establish what capability this adds beyond general three-dimensional geometry modeling. With reflection-invariant input attributes, the geometric operations in Eqs. 11–19 produce a representation invariant to reflection of the entire complex. This does not preclude distinguishing substrate stereoisomers in a fixed chiral enzyme environment, but it does not establish stereochemical discrimination either. The paper provides no controlled comparisons of stereoisomeric substrates or stereochemistry-stratified evaluation and does not specify how stereochemical information is preserved in the molecular features. Consequently, the claimed stereochemistry-specific benefit remains unsubstantiated.
  2. The substantive advance over the closest architecture is insufficiently isolated. BRIA uses bidirectional multihead cross-attention, a mechanism already present in EZSpecificity. Replacing ESMC with ESM-2 nearly eliminates the AUROC advantage over EZSpecificity on both benchmarks, while glycosyltransferase AUPR falls below baseline. A comparison using matched protein encoders, structural inputs, and training protocols is needed to establish the architectural contribution. The residue-prioritization results remain exploratory: disagreement with conservation and a selected in-model perturbation do not establish that BRIA rankings identify informative residues better than suitable controls.
  3. The evaluation scope and training protocols do not establish the breadth of the claimed improvement. Evaluation is restricted to two enzyme families without a clear rationale for omitting the broader ESIBank setting and remaining families evaluated by EZSpecificity. The manuscript does not clearly identify whether the EZSpecificity baseline uses the general model, family-specific fine-tuning, or training from scratch, making the comparison difficult to interpret. Cold-start evaluation is limited to halogenases. The "pair-cold" setting holds out associations while retaining both constituent identities in training. It does not test simultaneous generalization to unseen enzymes and substrates. These limitations leave both the comparability of the reported gains and their generality unresolved.
Questions for authors
  1. What is the substantive architectural distinction between BRIA and EZSpecificity‘s bidirectional cross-attention? Given that the ESMC to ESM-2 ablation largely removes the AUROC advantage, does the improvement persist when the protein encoder, structural inputs, and training protocol are matched? Please clarify what evidence attributes the gains to the proposed architecture rather than the encoder upgrade.
  2. What supervised training data and initialization were used for StereoVSE and the EZSpecificity baseline? Was StereoVSE trained separately on each family, jointly across families, or using additional enzyme-substrate data? Does the EZSpecificity result correspond to the general model, family-specific fine-tuning, or training from scratch? These distinctions are necessary to interpret the comparison.
  3. What motivates restricting evaluation to these two families rather than the broader ESIBank setting and family coverage of EZSpecificity? What evidence supports the generality of the improvements, given that cold-start evaluation is limited to halogenases and pair-cold retains both enzyme and substrate identities in training?
  4. Can you reconcile the glycosyltransferase class distribution in Table A1 with the F1, accuracy, and MCC values in Table A2? Please specify metric definitions, evaluation sampling, and aggregation across folds.
10 comments
AuthorsOpenReview ↗

Q1: The manuscript has not demonstrated a stereochemistry-specific gain of the model.

R1: Clarification on stereochemical sensitivity and results of supplementary architectural ablation experiments

We thank the reviewer for this insightful comment. To address this concern, we have incorporated the additional ablation experiments and corresponding analysis into the revised manuscript. Please refer to Section 4.4 and Appendix D for further details. These new contents are highlighted in blue.

To assess whether StereoVSE benefits from pair-specific three-dimensional pocket–substrate interactions rather than intramolecular structure alone, we conduct two complementary sets of ablation experiments: architectural ablation and component removal. The former evaluates the proposed architectural designs against controlled alternatives, while the latter quantifies the contribution of individual components.

For architectural ablation, we consider four variants: (1) ESMC→ESM-2 replaces the pretrained protein encoder; (2) Intra-only 3D retains intramolecular 3D geometry but removes all cross-molecular pocket–substrate edges, preventing explicit geometric message passing between the enzyme pocket and substrate; (3) BRIA→dual cross-attention replaces BRIA with EZSpecificity's interaction module; and (4) V–S→EGNN replaces the coupled Vector–Scalar encoder with an EGNN-based alternative.

Table R1: Architectural ablation results on the halogenase benchmark. The best results are shown in bold.

VariantAUCAUPRF1ACCMCC
ESMC → ESM20.95980.87430.78240.91340.7303
intra-only 3D0.96050.87230.78700.90980.7298
BRIA → dual cross-attention0.96570.88170.79360.91710.7431
V-S → EGNN0.96190.86940.77750.91250.7260
StereoVSE0.97280.91120.81380.92380.7666

Importantly, the Intra-only 3D variant preserves the three-dimensional information within each molecule, yet its AUPR decreases from 0.9112 to 0.8723 on the halogenase benchmark (Table R1) and from 0.7338 to 0.6973 on the glycosyltransferase benchmark (Appendix Table A2). The corresponding MCC values decrease from 0.7666 to 0.7298 and from 0.6813 to 0.6410, respectively. These results indicate that the predictive gains cannot be attributed solely to intramolecular geometry; explicitly modeling the relative spatial arrangement of enzyme-pocket and substrate atoms provides additional information for specificity prediction. Removing the entire structural branch causes larger performance losses (Tables R2 and Appendix Table A4), further supporting the contribution of three-dimensional structural information. The other architectural variants also underperform the complete model.

Table R2. Ablation by component removal results on the halogenase benchmark. The best results are shown in bold.

Ablation ModuleAUC$\uparrow$AUPR$\uparrow$F1$\uparrow$ACC$\uparrow$MCC$\uparrow$
w/o Pocket-Substrate 3D0.76150.49750.49510.80850.3819
w/o BRIA0.95500.85250.78190.90240.7216
w/o Vector-Scalar0.91640.77570.73170.88860.6615
StereoVSE0.97280.91120.81380.92380.7666

Additionally, regarding the issues related to the Eqs. (11)–(19) and chirality, we provide the following explanation. We acknowledge that the current Vector–Scalar geometric encoder, as formulated in Eqs. (11)–(19), produces reflection-invariant scalar representations and does not explicitly incorporate chirality-sensitive features. We also agree that reflection invariance does not necessarily preclude discrimination between substrate stereoisomers within a fixed chiral enzyme environment.

We would like to clarify that the primary objective of StereoVSE is to model the three-dimensional spatial interactions between enzyme-pocket and substrate atoms for specificity prediction, rather than to explicitly distinguish substrate enantiomers or other stereoisomers. We recognize this as a limitation of the current work. In future research, we plan to investigate chirality-sensitive geometric representations and evaluate their effectiveness in distinguishing substrate stereoisomers.

AuthorsOpenReview ↗

Q2: The core incremental contribution of the model has not been adequately isolated, and the substantive architectural difference between BRIA and the bidirectional cross-attention module of EZSpecificity needs to be clarified.

R2: Architectural contribution of BRIA and hyperparameter re-optimization for the ESM-2 ablation

We thank the reviewer for this constructive comment regarding the architectural contributions of StereoVSE and the influence of the pretrained protein encoder.

Regarding the distinction between BRIA and EZSpecificity, the contribution of StereoVSE does not lie in introducing a new attention operation, but rather in how the interaction representations are integrated with the other components of the framework. Specifically, StereoVSE employs a coupled Vector–Scalar encoder to capture directional geometric information from the enzyme pocket–substrate complex. The resulting structural representation is integrated with BRIA-derived interaction-aware representations and global chemical features in the prediction head. Additionally, BRIA retains head-averaged atom-to-residue attention weights for downstream candidate residue prioritization.

To further examine the architectural contribution, we included a controlled ablation in which BRIA is replaced with EZSpecificity's dual cross-attention module while retaining the remaining components of StereoVSE. As shown in Table R1 (Q1) and Appendix Table A2, this replacement reduces AUPR from 0.9112 to 0.8817 on the halogenase benchmark and from 0.7338 to 0.7036 on the glycosyltransferase benchmark. These results provide evidence that the proposed integration of interaction representations contributes to prediction performance beyond the choice of protein encoder.

Regarding the ESMC-to-ESM-2 ablation, we thank the reviewer for highlighting the importance of distinguishing encoder contributions from architectural improvements. Initially, the ESM-2 variant reused the hyperparameters optimized for ESMC (learning rate = 3 × 10⁻⁴, cutoff = 10 Å, 48 neighbors, and 6 GNN layers), without encoder-specific tuning. Nevertheless, its AUC (0.9157) already exceeded that of EZSpecificity (0.9147), while its AUPR (0.6719) was close to EZSpecificity (0.6805) on the glycosyltransferase benchmark.

To ensure a fairer comparison, we independently optimized the ESM-2 variant, selecting a learning rate of 3 × 10⁻⁴, a cutoff of 12 Å, 32 neighbors, and 4 GNN layers. This improved its AUC to 0.9223 and AUPR to 0.6811, both numerically exceeding EZSpecificity.

These results acknowledge the StereoVSE remains competitive without ESMC. The updated ESM-2 ablation results are reported in Appendix D of the revised manuscript.

AuthorsOpenReview ↗

Q3: The evaluation scope and training settings are insufficient to demonstrate the generality of performance improvements; the broader ESIBank dataset used in EZSpecificity is not adopted.

R3: Evaluation scope, dataset description and rationale for selected enzyme families

We thank the reviewer for this constructive comment regarding the evaluation scope and the clarity of the training protocols. We agree that these details are important for interpreting the comparison with EZSpecificity and understanding the scope of our conclusions.

Regarding the supervised training data and model initialization, StereoVSE is evaluated on two curated family-specific enzyme–substrate benchmarks: halogenases and glycosyltransferases. The halogenase dataset contains 6,206 labeled enzyme–substrate pairs, while the glycosyltransferase evaluation uses 11,278 pairs, including all 2,405 positive pairs and 8,873 sampled negative pairs. Four-fold pair-wise cross-validation is performed within each benchmark, with no identical enzyme–substrate pair shared across training, validation, and test subsets within a fold.

Regarding the choice of enzyme families, we selected halogenases and glycosyltransferases because they provide complementary settings for evaluating enzyme–substrate specificity. Halogenases exhibit pronounced catalyst-dependent substrate recognition, making them suitable for studying fine-grained specificity. Glycosyltransferases involve diverse substrate-recognition patterns and provide a larger benchmark. Together, these two families enable evaluation under distinct biochemical recognition scenarios and different levels of class imbalance. These two enzyme families are representative among those covered by ESIBank. We will conduct broader experiments on ESIBank in future work.

The corresponding clarifications are incorporated into Section 4.1 (Datasets), Sections 4.2–4.3 (Baselines and Evaluation), and Appendix A (Datasets and Baselines) of the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of changes: what is unchanged, what was added, and where

Strengths of the Paper

1. We propose StereoVSE, a stereochemistry-aware multimodal framework that integrates enzyme sequence representations, substrate chemical features, and three-dimensional binding geometries for accurate enzyme–substrate specificity prediction.

2. We develop a Vector–Scalar Graph Neural Network that jointly models invariant chemical features and equivariant directional geometric information to capture three-dimensional interactions between enzyme binding pockets and substrates.

3. We introduce Bidirectional Residue–Atom Interaction Alignment (BRIA) to explicitly model substrate-conditioned residue–atom interactions and provide interpretable attribution signals for prioritizing candidate specificity-determining residues.

4. Our method consistently outperforms all baselines on both enzyme–substrate benchmarks, demonstrates strong generalization across multiple challenging cold-start settings, and provides compelling ablation evidence highlighting the critical role of structural features in improving prediction performance.

New Additions Compared to the Original Manuscript

We have supplemented extensive experiments to address the reviewers’ concerns, which fall into three main parts:

  1. Module innovation and stereochemical sensitivity ablation: We performed architectural ablation and component removal experiments. The results verify that performance gains mainly come from the explicit modeling of cross-molecular 3D spatial interactions between the enzyme pocket and substrate, rather than intramolecular geometry alone. The integration of BRIA and the vector–scalar encoder outperforms alternative modules. We also re-optimized hyperparameters for the ESM-2 variant, confirming that our model’s superiority does not rely on the ESMC protein encoder.

  2. Double-cold-start generalization experiments: We expanded cold-start evaluations (enzyme-cold, substrate-cold, and double-cold) on both halogenase and glycosyltransferase benchmarks. The original pair-cold protocol was replaced with a stricter double-cold split where both test enzymes and substrates are unseen during training. StereoVSE consistently outperforms baselines across all cold-start settings, showing its advantages are not limited to previously observed enzymes or substrates.

  3. Strengthened cold-start evaluation controlled for enzyme homology and substrate scaffolds: We conducted split experiments with homology and scaffold controls for both datasets. On the enzyme side, MMseqs2 clustering at 30% and 50% sequence identity cutoffs eliminates leakage from close homologs. On the substrate side, Butina clustering and Bemis–Murcko scaffold decomposition ensure test substrates have unseen core scaffolds. We further built a rigorous double-cold setting: test enzymes share ≤30% sequence identity with training enzymes and are paired with substrates bearing novel Murcko scaffolds. Statistics confirm no train–test leakage, and StereoVSE achieves an even larger performance margin over baselines under this challenging setting.

Location of Revisions

  • Module innovation and stereochemical sensitivity ablation: Revisions are provided in the blue-highlighted text of Section 4.4 (main text) and Appendix D.
  • Double-cold-start generalization experiments: Revisions are provided in the blue-highlighted text of Section 4.3 (main text) and Appendix E.1.
  • Strengthened cold-start evaluation controlled for enzyme homology and substrate scaffolds: Revisions are provided in the blue-highlighted text of Appendix E.2.

Positive Feedback

We are pleased to receive the positive feedback from the reviewers:

  • Reviewer bU1Y: StereoVSE’s promising performance and valid module design.
  • Reviewer wcHb: Most of the reviewer’s concerns have been addressed. The reviewer will maintain the original score of 4 (Clear Accept).
  • Reviewer YUg2: The extended double-cold evaluations for glycosyltransferases and newly added architectural ablations are provided by the authors. These results largely resolve the reviewer’s concern that StereoVSE merely assembles existing modules, as the full StereoVSE consistently outperforms all ablated variants across benchmarks.

Closing Remarks

We sincerely thank you for your valuable time and careful guidance throughout the review and discussion phase. We have made every effort to respond thoroughly to all the concerns raised by the reviewers, including substantial additional ablation and cold-start generalization experiments. Given the positive feedback we have received, we respectfully ask you to favorably consider our manuscript for acceptance. We greatly appreciate your consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.