acceptodds
Under review as a conference paper at ICLR 2027

Chemistry-Informed Dual Encoders for Enzyme-Reaction Discovery

Abstract

Enzyme discovery often begins by uncovering promiscuous activities in known enzymes or characterizing the reactions catalyzed by newly identified sequences. Understanding an enzyme’s substrate and reaction scope requires connecting active-site features to reaction chemistry, a relationship that existing prediction methods often overlook. We propose Chemistry-Informed Reaction–Catalyst Embeddings (CIRCE), a contrastive dual encoder for bidirectional enzyme–reaction retrieval. CIRCE integrates function-aware protein representations with reaction representations that encode reactants, products, stoichiometry, cofactors, and additional chemical annotations. Incorporating chemical knowledge into both its representations and training objective enables CIRCE to learn features relevant to enzyme reactivity and promiscuity, with the aim of improving generalization to new reactions and accelerating enzyme discovery. Across enzyme screening and bidirectional retrieval benchmarks, CIRCE achieves state-of-the-art performance and robust generalization. These results establish its utility for enzyme discovery and provide a foundation for investigating the active-site determinants of catalytic promiscuity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI4Bio, enzyme discovery

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_GyZ3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Chemistry-Informed Reaction–Catalyst Embeddings (CIRCE), a dual encoder framework for bidirectional enzyme–reaction retrieval. The enzyme encoder combines global protein representations, functional residue priors, and learned residue views, while the reaction encoder integrates pretrained molecular representations and explicit chemical descriptors. Training combines an anchor-balanced, multi-positive contrastive objective with residual refinement over known enzyme–reaction associations. Experiments on EnzymeMap and ReactZyme demonstrate improved screening and retrieval performance against reported baselines. Additional analyses examine functional organization, inference sensitivity, and catalytic site conservation. However, experimental validation, stronger component attribution, and clearer evaluation protocols are needed to substantiate claims about enzyme discovery and generalization.

Critical strengths
  • This paper integrates function-aware protein pooling with chemically enriched reaction representations, providing a biologically motivated approach to connecting enzyme sequences with reaction compatibility.
  • This paper accommodates many-to-many enzyme–reaction associations through an anchor-balanced multi-positive objective, avoiding competition between recorded positive partners during initial alignment.
  • This paper reports substantial benchmark improvements, including increasing EnzymeMap BEDROC85 from 44.69 to 53.05, and supplements retrieval results with functional-neighborhood analyses and controlled residue-view ablations.
Critical weaknesses
  • This paper leaves evaluation reliability insufficiently resolved: Appendix D acknowledges test-set inspection during development, inference coefficients are inconsistently specified as two or three, and Appendix G.4 reports a differently budgeted Horizyn refit outperforming CIRCE on all six ReactZyme MRR settings. These issues require reconciliation before accepting broad state-of-the-art claims.
  • This paper does not sufficiently isolate its central contributions. The matched retraining ablation primarily removes learned residue views, without systematically separating functional-site priors, chemical feature sources, and the proposed alignment objective; the additional parameter capacity also confounds attribution.
  • This paper provides limited evidence for discovering new catalytic activities or promiscuity. EC3 agreement and active-site conservation indicate functional similarity but cannot establish substrate specificity or catalytic activity, while the case study lacks prospective biochemical validation.
Questions for authors

NA

5 comments
AuthorsOpenReview ↗

General comment.

We thank the reviewer for recognizing CIRCE's biological motivation, many-to-many formulation, and screening results. We have audited the configurations underlying the submitted results, completed a Horizyn comparison using CIRCE's first-stage schedule, and conducted additional component ablations.

To enhance clarity while reading the reply, we define two different configurations of the proposed model: CIRCE denotes the submitted model, i.e., the one in the original paper, while CIRCE* denotes the gated fingerprint extension evaluated during the rebuttal.

This paper leaves evaluation reliability insufficiently resolved: Appendix D acknowledges test-set inspection during development, inference coefficients are inconsistently specified as two or three, and Appendix G.4 reports a differently budgeted Horizyn refit outperforming CIRCE on all six ReactZyme MRR settings. These issues require reconciliation before accepting broad state-of-the-art claims.

Appendix D acknowledges test-set inspection during development; inference coefficients are inconsistently specified as two or three.

Data preparation and test-set inspection. We acknowledge that in Appendix D we refer to a test set inspection, but we also refer to Appendix B, which clarifies that “test set inspection” refers to the analysis of the datasets. Appendix B documents the construction of the released benchmark splits and the overlap audit.

Exact enzyme–reaction association pairs do not overlap between training, validation, and test sets, although individual enzyme sequences or reactions can recur in different pairs; these entity overlaps are explicitly reported and are characteristic of the predetermined dataset splits. The chemistry preprocessing statistics and vocabulary are trained on training data and then frozen for validation and test, whose split is defined according to the released benchmarks.

We clarify that, during development, there was no hyperparameter tuning on the test set and no test-set exploitation: we constructed internal validation from the released training associations, while the test partition was not used for development. By the word “inspection”, we meant “inspecting possible overlaps" for extra fairness. Appendix D refers to this screening procedure when it mentions the test set inspection. We will make this clearer by rephrasing it.

Inference coefficient. We thank the reviewer for identifying the inconsistent description. The correct protein fusion multiplier is αinf=2\alpha_{\mathrm{inf}}=2, as specified at the start of the main scoring subsection and in the saved inference manifests. The references to αinf=3\alpha_{\mathrm{inf}}=3 in Appendix D are typographical errors; we will consistently report αinf=2\alpha_{\mathrm{inf}}=2.

Appendix G.4 reports a differently budgeted Horizyn refit outperforming CIRCE on all six ReactZyme MRR setting

Horizyn comparison.
We retrained Horizyn on the identical training associations using CIRCE's first-stage schedule: batch size 1,024, 20 epochs, AdamW learning rate 10−410^{-4}, and weight decay 10−210^{-2}. Validation selected epoch 20 on all three splits. This control also exceeds submitted CIRCE in all six settings. The following are test MRR values.

ConfigurationTime E→R MRRTime R→E MRREnzyme-Sim E→R MRREnzyme-Sim R→E MRRReaction-Sim E→R MRRReaction-Sim R→E MRR
CIRCE0.7650.4560.9590.6360.5280.377
Horizyn0.8350.5890.9770.6980.5310.449

EnzymeMap. CIRCE exceeds all three benchmark-trained Horizyn controls on all eight screening metrics. The table below shows the fixed epoch-15 Horizyn checkpoint, matching CIRCE's encoder endpoint. “Full” contains 261,907 candidate identifiers; “Excluded” contains 252,113 after removing training enzyme IDs. This is identifier exclusion, not homology exclusion. B85/B20 denote BEDROC85/BEDROC20 multiplied by 100; EF5/EF10 are fold enrichment.

ConfigurationFull B85Full B20Full EF5Full EF10Excluded B85Excluded B20Excluded EF5Excluded EF10
Horizyn44.4660.7013.387.6839.4256.3812.487.38
CIRCE53.0571.3216.038.6249.6468.7015.488.45

Case study and observed gains.

Under the revised recipe, we decided to inspect the method and to apply changes to the CIRCE framework.
We additionally evaluated an auxiliary reaction fingerprint under a revised training recipe. It concatenates two 512-bit, radius-3, chirality-aware Morgan fingerprints, which are computed separately for reactants and products and then aggregated with a Differential Reaction Fingerprint (DRFP). This is then applied with a gated, g>0g>0, attention to the feature fusion of the reaction modality. We denote this extension as CIRCE*.

We evaluated $g\in{0,0.25,0.5,0.75,1,1.5,2$ for value selection, using as a criterion the mean of the six validation MRR values for ReactZyme and the mean across the four full-bank validation screening metrics for EnzymeMap. The recorded choices are g=1g=1 for ReactZyme and g=0g=0 for EnzymeMap. While for EnzymeMap the performance of CIRCE* reduces to the CIRCE ones, the ReactZyme test results for the extension are:

ConfigurationTime E→R MRRTime R→E MRREnzyme-Sim E→R MRREnzyme-Sim R→E MRRReaction-Sim E→R MRRReaction-Sim R→E MRR
CIRCE*, g=1g=10.8390.5910.9820.6990.4630.423
Horizyn, 20-epoch budget0.8350.5890.9770.6980.5310.449

CIRCE* improves five of six settings over submitted CIRCE and four of six over the Horizyn model. Both configurations remain below Horizyn on the two Reaction-Sim directions. We will revise the unrestricted state-of-the-art claim to reflect these benchmark-, split-, and direction-specific comparisons

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

The paper proposes CIRCE, a dual encoder for bidirectional enzyme–reaction retrieval. It combines functional-residue-guided protein representations, chemical reaction features, multi-positive alignment, and residual refinement. CIRCE reports improved screening on EnzymeMap and the highest MRR across all ReactZyme settings.

Critical strengths
  1. Combining global protein context with functional-residue features and reaction chemistry reflects the biological requirements of enzyme–reaction discovery.
  2. The alignment objective accommodates multiple positive pairs. Known positive pairs do not compete against one another, making the objective appropriate for the many-to-many structure of enzyme–reaction associations.
Critical weaknesses
  1. The ablations do not isolate the contributions of SLEEC, chemical descriptors, or the multi-positive alignment loss. Their individual benefits remain unclear.
  2. CIRCE combines features from several pretrained encoders, whereas the baselines use fewer features. Without comparisons using the same encoders, it is difficult to attribute the gains to the proposed architecture.

typo: Line 71, UniMol2 was written as UniMol2D

Questions for authors
  1. How does performance change when SLEEC and chemical descriptors are individually removed, and when the multi-positive alignment loss is replaced with a standard contrastive loss?
  2. How do CIRCE and the baselines compare when using the same pretrained encoders?
4 comments
AuthorsOpenReview ↗

General comment.

We thank the reviewer for recognizing the biological motivation for combining functional protein representations with reaction chemistry and for accommodating many-to-many enzyme–reaction associations.

As suggested by the reviewer, we conducted controlled retraining experiments to isolate the contributions of the functional-site prior, chemical inputs, alignment objective, and aggregation architecture.

The strongest evidence concerns explicit chemistry and learned aggregation, while the effects of SLEEC and the alignment objective depend on the benchmark.

We want to point out that in the reply, the reviewer will find, in the experiments, two versions of the model: CIRCE is the submitted model, while the fingerprint-enabled extension CIRCE* is introduced in the context of this rebuttal. Implementation and gate-selection details that describe the CIRCE* implementation are provided in our response to Reviewer GyZ3.

Typographical correction.

Thank you for identifying the typo on Line 71. The intended name is UniMol2, and we will correct “UniMol2D” accordingly.

Q1/W1. How does performance change when SLEEC and chemical descriptors are individually removed, and when the multi-positive alignment loss is replaced with a standard contrastive loss?

We report the ablations using the same metric groups as the main paper: EnzymeMap BEDROC85, BEDROC20, EF5 and EF10 for both candidate pools, and ReactZyme all-positive MRR in both directions on Time, Enzyme-Sim and Reaction-Sim. Within each model recipe and benchmark, the controls retain the training associations, splits, candidate banks, training schedule, refinement procedure and inference settings, apart from the specified intervention. Encoder endpoints are 20 epochs for ReactZyme and 15 for EnzymeMap, followed by 100 full-graph refinement updates.

ReactZyme: fingerprint-enabled CIRCE* at g=1, all-positive MRR. These controls use CIRCE*'s own training and inference recipe on each split. The fingerprint remains active when explicit descriptors or a pretrained feature source are removed. The full table includes standard paired InfoNCE and the identical-input fixed-pooling control, so their effects are visible across all three splits rather than only Reaction-Sim.

ConfigurationTime E→R MRRTime R→E MRREnzyme-Sim E→R MRREnzyme-Sim R→E MRRReaction-Sim E→R MRRReaction-Sim R→E MRR
CIRCE*0.8390.5910.9820.6990.4630.423
Uniform prior instead of SLEEC0.8370.5880.9770.6990.4920.420
Without explicit descriptors0.8410.5910.9670.6950.4460.403
Multi-positive softmax CE0.8250.5940.9770.6980.4540.416
K=0, parameter-matched0.8340.5880.9780.6980.4790.423
Without ReactionT5v20.8270.5880.9860.7010.4800.419
Without UniMol20.8420.5930.9780.6990.4900.418
Without ChIRo0.8380.5970.9680.6970.4580.417
Standard paired InfoNCE0.8410.5930.9730.6980.4730.416
Fixed pooling, same inputs/capacity0.8470.5850.9820.6990.4050.415

EnzymeMap: complete screening metrics. “Full” denotes the 261,907-identifier library; “Excluded” denotes the 252,113-identifier library after removing training enzyme IDs. This is identifier exclusion, not homology exclusion. B85 and B20 are BEDROC85 and BEDROC20 multiplied by 100; EF5 and EF10 are fold enrichment. Higher is better for every column. The shared CIRCE / CIRCE* reference row reuses the original model: CIRCE* with g=0 retains the original recipe and scores. All controls in this panel therefore apply to the shared EnzymeMap configuration; they are not independent CIRCE* replications.

ConfigurationFull B85Full B20Full EF5Full EF10Excluded B85Excluded B20Excluded EF5Excluded EF10
CIRCE*53.0571.3216.038.6249.6468.7015.488.45
Uniform prior instead of SLEEC55.2473.4816.388.8351.5670.6415.938.67
Without explicit descriptors44.9263.7814.438.0740.7360.3213.797.79
Multi-positive softmax CE53.0371.3816.008.6649.5768.7515.508.47
K=0, parameter-matched51.6469.0715.128.3546.9065.2914.448.09
Without ReactionT5v250.5169.8215.748.6046.0266.5315.238.41
Without UniMol248.4168.4115.488.5144.0165.2215.008.30
Without ChIRo49.9069.6615.778.5145.5666.3715.168.30
Uniform prior + multi-positive CE55.7573.7416.448.8352.1070.9515.968.67
Standard paired InfoNCE53.3371.9916.228.7150.1869.4815.738.53
Fixed pooling, same inputs/capacity46.1363.5114.008.0340.7359.1813.227.76

a. Contribution of the SLEEC functional-site prior.

We replace the SLEEC-derived residue scores with a uniform prior while retaining the protein features and trainable pooling architecture. This isolates the information supplied by the functional-site prior without reducing trainable capacity. For fingerprint-enabled CIRCE*, SLEEC improves MRR in four of six ReactZyme settings: both Time directions, Enzyme-Sim E→R, and Reaction-Sim R→E.

On EnzymeMap, the uniform prior instead improves all eight reported screening metrics, including full-bank BEDROC85 from 53.05 to 55.24. These results indicate that SLEEC provides a task-dependent functional prior rather than a uniform improvement across benchmarks.

b. Contribution of chemical descriptors and pretrained reaction features.

We remove the explicit chemical descriptors while retaining the pretrained reaction representations and, for fingerprint-enabled CIRCE*, the auxiliary fingerprint.

On EnzymeMap, this reduces all eight screening metrics, including full-bank BEDROC85 from 53.05 to 44.92 and training-ID-excluded BEDROC85 from 49.64 to 40.73.

On ReactZyme, removing descriptors lowers both retrieval directions on Enzyme-Sim and Reaction-Sim. In particular, Reaction-Sim MRR decreases from 0.463 to 0.446 for E→R and from 0.423 to 0.403 for R→E. Time E→R improves slightly, while Time R→E is unchanged at the displayed precision. These results support a contribution from explicit chemistry that is not fully replaced by the pretrained representations or auxiliary fingerprint under the tested recipes.

We also remove each pretrained reaction-feature source individually. On EnzymeMap, full-bank BEDROC85 decreases from 53.05 to 50.51 without ReactionT5v2, 48.41 without UniMol2, and 49.90 without ChIRo; each removal also reduces the other screening metrics in both banks. On ReactZyme, all three removals lower Reaction-Sim R→E MRR, but their effects elsewhere are mixed. For example, removing ReactionT5v2 improves both Enzyme-Sim directions, while removing UniMol2 improves both Time directions.

c. Contribution of the alignment objective. Replacing the first-stage loss with multi-positive softmax CE reduces MRR in five of six ReactZyme settings, while EnzymeMap performance remains similar. Standard paired InfoNCE produces mixed ReactZyme results and slightly improves EnzymeMap screening. Thus, the proposed objective benefits several retrieval settings and accommodates multiple recorded partners without competition during initial alignment.

Q2/W2. How do CIRCE and the baselines compare when using the same pretrained encoders? Does additional parameter capacity explain the gains?

We evaluate two architectural controls using identical pretrained features and approximately matched trainable capacity:

  • Fixed pooling replaces learned aggregation with fixed averages while retaining trainable projections.

  • K=0 removes the four learned auxiliary residue views while retaining learned fusion.

ReactZyme: fingerprint-enabled CIRCE* at g=1, all-positive MRR. Controls are compared within this recipe. Fixed pooling was evaluated only for CIRCE*.

ConfigurationTime E→RTime R→EEnzyme-Sim E→REnzyme-Sim R→EReaction-Sim E→RReaction-Sim R→E
CIRCE*0.8390.5910.9820.6990.4630.423
K=00.8340.5880.9780.6980.4790.423
Fixed pooling0.8470.5850.9820.6990.4050.415

Removed capacity is reallocated to an MLP on the global protein summary. Fixed pooling matches total trainable parameters within 0.005%, while K=0 matches protein-side parameters within 0.006%. Both controls retain the same pretrained inputs as their corresponding reference.

EnzymeMap. CIRCE and CIRCE* at g=0 share the original checkpoint. B85/B20 denote BEDROC85/20 multiplied by 100; EF denotes enrichment factor. “Excluded” removes training enzyme IDs.

ConfigurationFull B85Full B20Full EF5Full EF10Excluded B85Excluded B20Excluded EF5Excluded EF10
CIRCE*53.0571.3216.038.6249.6468.7015.488.45
K=051.6469.0715.128.3546.9065.2914.448.09
Fixed pooling46.1363.5114.008.0340.7359.1813.227.76

On EnzymeMap, CIRCE* outperforms both controls across all eight metrics. Relative to fixed pooling, BEDROC85 improves by 6.92 points on the full bank and 8.91 points after training-ID exclusion.

These comparisons support a contribution from learned aggregation when pretrained inputs and parameter counts are controlled. They complement the published-method comparisons, which retain each method’s feature collection; we have not retrained CREEP or CLIPZyme with CIRCE’s complete inputs. We will clarify this distinction in Section 5 and report the controls and parameter counts in Appendix F.2.

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

The authors developed CIRCE, which is a tool for enzyme discovery. They recognize the gap that there is a small fraction of sequences in large protein databases that have characterized reactions and claim that this can be improved by incorporating chemical knowledge on the reaction and functionally-relevant residues on the enzyme side. They seek to discover promiscuous enzyme activity and novel enzymatic functions.

To this end, the authors developed a contrastive dual encoder for bidierctional enzyme-reaction retrieval.They encode using published tools and embed the enzyme and the reaction separately, score their match and provide a ranking for both reactions per enzymes and enzymes per reaction.

CIRCE is evaluated on EnzymeMap and ReactZyme and show improvement.

Critical strengths

The main strength to my understanding is conceptual, changing the relatioinship that is being modeled from one-to-one unidirectional to many-to-many bidirectional. This makes biological sense and is thus justifiable. The authors show some convincing benchmarking results.

Critical weaknesses

Having little experience in this field, I elected to leave this empty and provide only questions.

Questions for authors
  1. Why did the authors show only the recovery of EC3? What happens with EC4, for example?
  2. The results show recovery@10 of ~30% in EC3, is that biologically meaningful?
2 comments

We thank the reviewer for the constructive points raised and for recognizing the biological motivation of the work in treating the specific enzyme reaction mapping problem as a many-to-many bidirectional enzyme–reaction cross-modal relationship.

We are going to further enrich the analysis of the enzyme encoder embedding space and its capability of aggregating data according to their biochemical functional relationships. The paper will incorporate new results shown below.

Q1/W1: Why did the authors show only the recovery of EC3? What happens with EC4, for example?

In biochemistry, enzymes are annotated using Enzyme Commission (EC) numbers. In particular, the EC hierarchy, made up of 4 levels, is the mechanism to specify the class of reactions the enzymes catalyze at different levels of granularity with respect to the EC level considered. The first three EC fields identify a reaction sub-subclass, specifying the general catalytic chemistry, reaction class, bond or group acted upon, and cofactor/acceptor type, whereas the complete four-field EC number specifies a finer enzymatic activity category.

In our evaluation, EC3 agreement requires matching the first three fields (which also preserves proteins annotated at the sub-subclass level, such as 1.1.1.-), whereas EC4 requires an exact match across all four fields.

Following your suggestion, we repeated the cross-modal RE recovery analysis at EC4 on EnzymeMap. Before ranking, we exclude identical sequences and enzymes with at least 50% identity and 80% coverage of both sequences. We then compute class-macro Recovery@10 across all queries with valid EC4 annotations.

On EnzymeMap, the comparison is:

ModelEC3 Recovery@10 (%)EC4 Recovery@10 (%)
CIRCE30.116.97
CLIPZyme25.465.93
CREEP21.494.64
Random1.620.0365

Both EC3 and EC4 experiments use the same 1,067 eligible reaction queries and the same 131,409 annotated candidate identifiers before query-specific exclusions. The class cardinality differs, though: EC3 has 68 represented classes and EC4 has 309.

CIRCE maintains the highest recovery among all evaluated methods when compatibility requires an exact four-digit EC match. The lower absolute percentages at EC4 reflect the much stricter matching criterion and the scarcity of compatible alternatives after homology filtering: whereas only 3 of 1,067 queries have no remaining compatible candidates at EC3, 240 of 1,067 queries (22.5%) have zero remaining EC4-compatible candidates in the bank and are scored as zero. CIRCE's 6.97% EC4 Recovery@10 represents a ~191% gain over random retrieval (0.0365%) and outperforms CLIPZyme by +1.04 percentage points (+17.5% relative improvement). We will report these outcomes as well in Appendix G.

Q2/W2: The results show recovery@10 of ~30% in EC3, is that biologically meaningful?

The EC3 agreement is an indicator of retrieved enzymes that share an annotated reaction sub-subclass with enzymes associated with the query reaction. It can then be interpreted as a proxy metric to measure recovery of related catalytic functions, while allowing differences in substrate specificity.

Our Recovery@10 measures the fraction of top-ten candidates sharing at least one EC3 annotation with the query’s known partners, averaged within each represented EC3 class and then across classes. We exclude recorded partners, identical-sequence matches, and their detected close-sequence components before ranking (at least 50% identity).

Under this protocol, CIRCE achieves 30.11% recovery, compared with 1.62% expected under random retrieval, 25.46% for CLIPZyme, and 21.49% for CREEP. This corresponds to approximately 18.6x enrichment over the random reference and a 4.65-point improvement over CLIPZyme.

These results support the biological relevance of the learned cross-modal relationships beyond recovery of known partners and close sequence relatives. We interpret this enrichment as evidence that CIRCE can prioritize candidates having functionally compatible annotated catalytic functions, which can be useful in case of experimental screening. It does not establish that 30.11% of retrieved candidates catalyze the exact query reaction, which requires biochemical validation. We will make this distinction explicit in the analysis and conclusions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.