acceptodds
Under review as a conference paper at ICLR 2027

Elucidating the Space of Enzymatic Reaction: A Unified Benchmark and Pretrained Model

Abstract

Existing reaction models primarily learn molecular transformations, whereas enzymatic reactions depend jointly on molecular structure and catalytic function. We formulate this problem as learning an enzymatic reaction space linking reactants, products, and Enzyme Commission (EC) annotations. To characterize this space, we introduce **VenusRX-Bench**, a unified benchmark for forward reaction prediction, single-step retrosynthesis, and EC-number prediction. **VenusRX-Bench** integrates reactions from multiple biochemical databases with standardized curation, leakage-controlled splits, and consistent evaluation. Benchmarking representative chemical and enzymatic models reveals a clear chemical-to-enzymatic domain gap, driven by limited domain data, catalytic-context dependency, and the difficulty of modeling large biomolecular structures. To bridge this gap, we develop **VenusRX**, a unified T5-style sequence-to-sequence model for enzymatic reactions. **VenusRX** jointly learns forward prediction, retrosynthesis, and reaction reconstruction, with two-stage training on millions of template-expanded reactions followed by real biochemical reactions. In addition, optional EC conditioning incorporates catalytic context, while Molecule Library-Constrained Decoding improves the generation of complex biomolecules. Across benchmark tasks and challenging generalization splits, **VenusRX** achieves the best or competitive performance in most evaluated settings over representative chemical and enzymatic baselines. Moreover, EC information consistently improves reaction prediction, while learned reaction representations support accurate EC prediction, revealing a bidirectional relationship between reaction structure and catalytic function. Together, **VenusRX-Bench** and **VenusRX** provide a unified framework for elucidating and modeling enzymatic reaction space.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Enzymatic Reaction Modeling, Reaction Benchmark, Multi-task Pretraining, Retrosynthesis, EC Conditioning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_1LxfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a unified benchmark and pretrained model for enzymatic reaction prediction. Through a two-stage pretraining process, transferable knowledge regarding enzymatic transformations is acquired and subsequently adapted to real-world benchmark reactions. Experiments show state-of-the-art performance in both forward reaction and retrosynthesis prediction, validating the effectiveness of this method.

Critical strengths

None

Critical weaknesses

Some of the details and experimental analysis should be clarified, which is described in "Questions For Authors".

Questions for authors
  • Why does VenusRX use a T5-style encoder-decoder architecture? What is the necessity for this?

  • The authors claimed "For each reaction, we additionally use R-SMILES augmentation (Zhong et al., 2022)". Did the other baseline methods also use this type of augmentation?

  • I think more details about the construction of Stage 1 corpus are needed. In addition, the differences from existing corpora should be outlined.

  • It's imporant to explain how data leakage during the testing phase will be prevented.

  • In Table 5, for the first two tasks, applying only Stage-1 shows poorer results than only Stage-2. However, the situation becomes opposite for EC prediction task. What is the reason for this? In addition, given the core argument that "Stage-1 provides transferable enzymatic transformation knowledge, while Stage-2 adapts this knowledge to real benchmark reactions", one would expect EC prediction results without Stage-2 to be pretty worse, but the performance gap compared to full model is pretty small.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 1Lxf (Part 1/2)

We thank Reviewer 1Lxf for the thoughtful review and helpful questions. Below, we address each of your points in detail.

Response to Q1:

A T5-style encoder-decoder architecture is well suited for unifying multiple input-to-output tasks such as forward prediction and retrosynthesis. At the same time, EC-number prediction requires a strong reaction representation model. An encoder-decoder architecture naturally supports both sequence-to-sequence generation and representation learning within a single model, which is difficult to achieve with encoder-only (e.g., BERT), decoder-only (e.g., GPT), or diffusion-based architectures. We have updated the main text to make this motivation clearer.

Response to Q2:

Whether a baseline uses R-SMILES augmentation depends on the design of the original method. Specifically, RSGPT and R-SMILES both use R-SMILES augmentation, following their official implementations.

Response to Q3:

  1. Thank you for the suggestion. We have expanded Section 3.2 in the revised manuscript to provide more details on the data sources, filtering criteria, and reaction generation pipeline for the Stage-1 corpus. Specifically, for each split of VenusRX-Bench, we extract local reaction templates from the atom-mapped enzymatic reactions in the corresponding training set, and decompose molecules from AtlasX BioDB, KEGG, and COCONUT using BRICS to build a candidate substrate fragment pool. We then apply the templates to compatible fragment combinations to generate candidate reactions, followed by standardization, deduplication, structural validity filtering (via RDKit), and thermodynamic filtering (via eQuilibrator). This pipeline produces 4.04M, 4.05M, and 3.85M full training reactions for the reactant split, product split, and EC split, respectively.

  2. Compared with existing corpora, our Stage-1 synthetic corpus differs in two main ways:

    (a) Compared with USPTO-FULL, which is widely used for chemical reaction pretraining:

     (1) USPTO-FULL consists of chemical synthesis reactions from patents, whereas our Stage-1 corpus is built from recorded enzymatic reactions and uses template-based generation to expand their substrate coverage.
    
     (2) Our Stage-1 corpus retains template provenance, EC categories, and thermodynamic estimates, where each inherited EC label indicates the catalytic category of the source reaction.
    
     (3) Our synthetic corpus contains around 4 million reactions, which is roughly 4 times larger than USPTO-FULL (about 1.04 million reactions).
    

    (b) Compared with the existing enzymatic reaction datasets listed in Table 1, our synthetic corpus is substantially larger in scale than all of them.

AuthorsOpenReview ↗

Response to Reviewer 1Lxf (Part 2/2)

Response to Q4:

Thank you for raising this important point. We will add a dedicated description in the paper to clearly summarize our test-set isolation procedure:

  1. For data splitting, the three splits group samples using canonical reactant combinations, principal products, or complete level-4 EC numbers as keys. Samples sharing overlapping keys are grouped into connected components and assigned as a whole to the training, validation, or test set. This guarantees that there are no overlapping keys across the training, validation, and test partitions.

  2. For Stage-1 pretraining, the corpus for each split is filtered independently: (a) synthetic reactions are generated using templates extracted from the corresponding training set, and (b) any generated reactions whose structures match validation or test reactions are removed from the Stage-1 corpus.

  3. For Stage-2 training, each split only continues training on its own training partition. For EC nearest-neighbor retrieval, the reference database consists solely of training reactions, and the number of neighbors k is selected on the validation set.

We will also clearly state the information used in each evaluation setting alongside the result tables: MLCD is evaluated using a benchmark-wide molecule vocabulary that includes test molecules, and round-trip evaluation is an auxiliary diagnostic computed with a forward oracle trained on the complete benchmark.

Response to Q5:

We believe this pattern occurs because Stage-1 and Stage-2 pretraining emphasize different capabilities. Stage-1 uses a very large set of synthetic reactions; while the exact substrate-product structures and combinations may not always match real benchmark distributions, the underlying reaction rules are broadly accurate. This makes Stage-1 particularly effective for learning global representations of enzymatic reactions. In contrast, Stage-2 trains directly on real benchmark reactions, which are smaller in scale but have exact molecular structures and combinations, making Stage-2 much more important for improving token-level generation accuracy.

Therefore, for EC prediction, which relies heavily on global reaction representations, Stage-1 pretraining plays the primary role and the gap without Stage-2 is small. For forward prediction and retrosynthesis, which require accurate token-by-token generation, Stage-2 adaptation is more critical. We will include this explanation in the revised manuscript.

We thank the reviewer again for the constructive comments, which helped us improve the clarity of our paper.

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

The paper introduces VenusRX-Bench for enzymatic reaction modeling and VenusRX, a T5-style model trained on template-expanded and curated reactions. It combines multitask learning, optional EC conditioning, and molecule-library constraints to support forward prediction, retrosynthesis, and EC inference.

Critical strengths
  1. The benchmark brings three related tasks and several generalization splits into a common evaluation framework.
  2. The compact model achieves strong results in many evaluated settings.
  3. Comparisons with and without EC conditioning help assess the value of catalytic context.
Critical weaknesses
  1. Access to held-out molecules and the test-trained round-trip oracle complicate interpretation of the evaluation results.
  2. Product overlap during pretraining is not fully established, leaving uncertainty about the generalization claims.
  3. Inconsistent dataset counts and underspecified synthetic-data filtering reduce clarity and reproducibility.
Questions for authors
  1. Appendix C.6 states that MLCD uses molecules from the training, validation, and test sets. Please make this access explicit in Tables 2–3 and add results using a molecule library constructed independently of the test set.
  2. The Product Split excludes test products from benchmark training, but not necessarily from Stage-1 pretraining (Appendix C.4). Please audit product overlap and repeat the evaluation with held-out products excluded from both training stages.
  3. The round-trip oracle is trained on the complete benchmark, including test reactions (Section 4.1). This evaluation would be more convincing with an independent forward model trained without those test reactions.
  4. The Introduction and Table 1 report 126,361 reactions, while Appendix A and Table 6 use 83,439. If these counts refer to different curation stages, please explain the distinction and ensure the dataset-size comparison uses comparable counts.
  5. Appendix D.5 mentions biochemical and thermodynamic filtering but does not specify the energy source, filtering thresholds, or retention counts. Please provide these details so readers can reproduce and assess the synthetic-data pipeline.
4 comments
AuthorsOpenReview ↗

Response to Reviewer SqDR (Part 1/4)

We thank Reviewer SqDR for the thoughtful review and constructive feedback. We appreciate your recognition of our unified evaluation framework, compact model performance, and EC-conditioned comparisons. Below, we address each of your questions in detail.

Response to Q1:

  1. We will clearly state the source of the MLCD molecule library wherever MLCD results are reported, including in Tables 2 and 3 of the manuscript.

  2. Following your suggestion, we evaluated MLCD using an independent external biochemical molecule library, ChEBI. The results are shown in the tables below.

Table R1. Forward Reaction Prediction (%)

SplitEC InputMethodMolecular LibraryTop-1Top-3Top-5Top-10
ReactantWithout ECVenusRXVenusRX-Bench73.06684.61188.04890.739
ReactantWithout ECVenusRX (MLCD)VenusRX-Bench77.07488.35091.41494.210
ReactantWithout ECVenusRX (MLCD)ChEBI42.74255.77859.56464.049
ReactantWith ECVenusRXVenusRX-Bench75.22186.77889.72592.474
ReactantWith ECVenusRX (MLCD)VenusRX-Bench80.05691.13593.97795.515
ReactantWith ECVenusRX (MLCD)ChEBI44.40858.42362.71066.484
ProductWithout ECVenusRXVenusRX-Bench50.95966.27571.39377.085
ProductWithout ECVenusRX (MLCD)VenusRX-Bench58.78574.41379.12384.552
ProductWithout ECVenusRX (MLCD)ChEBI20.63831.37636.13440.508
ProductWith ECVenusRXVenusRX-Bench55.82569.46374.04178.811
ProductWith ECVenusRX (MLCD)VenusRX-Bench62.59677.19381.55685.918
ProductWith ECVenusRX (MLCD)ChEBI21.46532.53837.15241.335
ECWithout ECVenusRXVenusRX-Bench31.10045.17053.23659.947
ECWithout ECVenusRX (MLCD)VenusRX-Bench39.27456.32864.75374.161
ECWithout ECVenusRX (MLCD)ChEBI17.62928.54733.49740.340
ECWith ECVenusRXVenusRX-Bench34.56450.24056.34064.058
ECWith ECVenusRX (MLCD)VenusRX-Bench42.91761.18269.12877.661
ECWith ECVenusRX (MLCD)ChEBI18.56430.09336.15842.378

Table R2. Retrosynthesis (%)

SplitEC InputMethodMolecular LibraryTop-1Top-3Top-5Top-10
ReactantWithout ECVenusRXVenusRX-Bench8.44617.18323.77730.603
ReactantWithout ECVenusRX (MLCD)VenusRX-Bench9.44819.07026.58434.355
ReactantWithout ECVenusRX (MLCD)ChEBI2.5867.1538.92411.556
ReactantWith ECVenusRXVenusRX-Bench17.23026.39829.34534.425
ReactantWith ECVenusRX (MLCD)VenusRX-Bench17.30027.40032.25838.828
ReactantWith ECVenusRX (MLCD)ChEBI4.83511.51013.09415.506
ProductWithout ECVenusRXVenusRX-Bench20.55430.05834.92339.945
ProductWithout ECVenusRX (MLCD)VenusRX-Bench22.90334.63640.24446.093
ProductWithout ECVenusRX (MLCD)ChEBI6.48410.65412.62015.328
ProductWith ECVenusRXVenusRX-Bench28.84734.69638.26740.568
ProductWith ECVenusRX (MLCD)VenusRX-Bench29.30239.52543.67248.466
ProductWith ECVenusRX (MLCD)ChEBI7.75411.87713.92616.311
ECWithout ECVenusRXVenusRX-Bench4.8788.73710.48714.274
ECWithout ECVenusRX (MLCD)VenusRX-Bench7.40714.65718.38423.849
ECWithout ECVenusRX (MLCD)ChEBI3.2364.8905.7057.239
ECWith ECVenusRXVenusRX-Bench18.36024.60527.48130.872
ECWith ECVenusRX (MLCD)VenusRX-Bench20.37427.50530.60934.552
ECWith ECVenusRX (MLCD)ChEBI5.7418.48510.96613.255

The effectiveness of MLCD depends heavily on how well the molecule library covers the ground-truth molecules. ChEBI covers only 35.749% of the molecules in VenusRX-Bench, and in dry-lab evaluation, only molecules in VenusRX-Bench are counted as ground truth.

However, in practical bio-retrosynthesis applications, MLCD remains very useful because the goal is often not to generate novel molecules, but to identify or rank the correct candidate molecule within a known candidate set. For example, in natural product biosynthetic pathway elucidation and metabolic network completion, the target precursors are restricted to the metabolome of the corresponding host.

AuthorsOpenReview ↗

Response to Reviewer SqDR (Part 2/4)

Response to Q2:

Thank you for pointing this out. We audited the Product Split and found that 34.48% (895/2,596) of the test-set products appeared in the Stage-1 pretraining corpus, as template-based synthetic generation can produce some products that only occur in the test set.

Following your suggestion, we additionally removed all Stage-1 pretraining records that had product overlap with the test set (referred to as Strict Pretraining, Product-Disjoint) and re-ran the full two-stage training. The dataset sizes and updated evaluation results on the Product Split are shown below.

Table R3. Stage-1 Pretraining Dataset Sizes

Stage-1 Pretraining SettingNumber of Samples
Original Pretraining4,047,718
Strict Pretraining (Product-Disjoint)3,684,869

Table R4. Forward Reaction Prediction on Product Split (%)

EC InputMethodTop-1Top-3Top-5Top-10
Without ECR-SMILES30.84944.63150.00055.849
Without ECGraph2SMILES34.40844.42748.28651.702
Without ECVenusRX50.95966.27571.39377.085
Without ECVenusRX (Strict Pretrain)51.87067.06672.59177.589
Without ECVenusRX (MLCD)58.78574.41379.12384.552
Without ECVenusRX (MLCD, Strict Pretrain)59.42075.30079.81884.720
With ECECReact37.81251.55855.90859.456
With ECVenusRX55.82569.46374.04178.811
With ECVenusRX (Strict Pretrain)55.16569.37974.42578.919
With ECVenusRX (MLCD)62.59677.19381.55685.918
With ECVenusRX (MLCD, Strict Pretrain)62.94376.64281.58086.805

Table R5. Retrosynthesis on Product Split (%)

EC InputMethodTop-1Top-3Top-5Top-10
Without ECGraph2SMILES27.39734.19236.92539.022
Without ECRSGPT8.49715.58018.81623.730
Without ECVenusRX20.55430.05834.92339.945
Without ECVenusRX (Strict Pretrain)19.12828.03231.49637.224
Without ECVenusRX (MLCD)22.90334.63640.24446.093
Without ECVenusRX (MLCD, Strict Pretrain)21.34532.27537.47642.761
With ECECReact17.28224.52127.04927.457
With ECVenusRX28.84734.69638.26740.568
With ECVenusRX (Strict Pretrain)27.20533.67736.60140.089
With ECVenusRX (MLCD)29.30239.52543.67248.466
With ECVenusRX (MLCD, Strict Pretrain)27.39738.42343.58850.312
AuthorsOpenReview ↗

Response to Reviewer SqDR (Part 3/4)

Response to Q3:

  1. The reason we originally trained the round-trip oracle on the complete benchmark is as follows. In retrosynthesis, a target product can have several plausible sets of precursors, only some of which are recorded in the benchmark. A forward reaction model provides a way to credit alternative candidates by checking whether they lead back to the target product. Because this check is approximate, if the forward model fails to recover the target from a chemically plausible candidate due to limited training coverage, that error is attributed to the retrosynthesis method. We therefore trained a single, fixed VenusRX forward model on the full benchmark to give it the broadest available coverage of known enzymatic reactions and applied the same model to every method and split after candidate generation.

  2. However, we agree that exposure to test reactions can make recorded precursor-product mappings easier for the oracle to recover and may affect methods differently. Following your suggestion, we report the round-trip results for each split using an oracle model trained without access to the corresponding test set.

Table R6. Round-trip Accuracy (%)

Round-trip accuracy (%) across the Reaction, Product, and EC splits. For each split, we employ a split-specific forward prediction oracle trained without access to the corresponding test set. Accuracy@k measures the proportion of round-trip-consistent predictions among all Top-k candidate positions, with invalid and missing candidates counted as failures. Models marked with * are evaluated in the zero-shot setting.

MethodEC InputReaction T1T3T5T10Product T1T3T5T10EC T1T3T5T10
R-SMILES*Without EC2.7143.6934.0563.8404.1232.8482.6172.3294.3383.3322.5892.502
RSGPT*Without EC12.38411.2039.5977.1687.2755.9164.8183.2727.5146.3645.8654.394
SciReasoner*Without EC27.44621.69218.51416.24817.09012.62811.1708.80316.75512.51611.1729.034
R-SMILESWithout EC29.12425.38423.48820.62614.37011.91311.0599.48213.19511.64510.6099.202
Graph2SMILESWithout EC38.43233.47730.10325.61425.09619.57116.67512.98225.70722.15219.73915.581
RetroformerWithout EC24.10322.53422.00820.9559.6008.5017.8866.63613.65112.44011.41710.735
MEGANWithout EC34.15732.46330.39127.08325.20422.00020.94218.51422.85520.65418.97917.367
BioNaviWithout EC36.47535.71433.91930.5849.8518.4457.4526.06922.54319.23917.32714.400
G2RetroWithout EC26.02522.84921.35119.64921.60816.97415.40713.39917.96515.68413.87612.090
RSGPTWithout EC22.16921.34620.31917.5109.1087.5426.8225.61811.84111.36110.5158.663
VenusRXWithout EC37.61631.68332.32830.77030.17726.19824.33422.45631.56827.94826.10524.228
ECReactWith EC44.58342.35039.89335.37227.63726.04325.65723.76833.72532.35931.01929.178
VenusRXWith EC48.08940.24936.39331.25838.36329.88626.29221.12328.04423.33021.54118.227
AuthorsOpenReview ↗

Response to Reviewer SqDR (Part 4/4)

Table R7. Round-trip Coverage (%)

Round-trip coverage (%) across the Reaction, Product, and EC splits. Coverage@k measures the percentage of test instances for which at least one of the Top-k retrosynthesis candidates is validated by the corresponding split-specific forward prediction oracle. All experiments use canonical single-view representations, beam size 10, Top-1 forward predictions, and FP32 precision. Models marked with * are evaluated in the zero-shot setting.

MethodEC InputReaction T1T3T5T10Product T1T3T5T10EC T1T3T5T10
R-SMILES*Without EC2.7149.36613.29219.0824.1237.61010.10315.4004.3387.8149.46815.280
RSGPT*Without EC12.38422.84527.26032.4217.27512.50015.36419.0207.51414.72919.43923.873
SciReasoner*Without EC27.44640.54146.86656.66417.09027.18134.25242.65316.75528.36834.79143.025
R-SMILESWithout EC29.12441.12345.59650.96714.37025.65932.17939.96913.19523.80230.60937.548
Graph2SMILESWithout EC38.43252.00457.43263.38525.09631.05234.14438.17125.70735.10337.94341.719
RetroformerWithout EC24.10332.56140.92551.0379.60013.23116.83821.47713.65118.28921.94428.476
MEGANWithout EC34.15749.63955.89562.91925.20438.49547.59153.53522.85536.07441.37149.868
BioNaviWithout EC36.47551.92257.56163.3159.85115.54418.86422.95122.54329.72233.61737.452
G2RetroWithout EC26.02546.52856.59466.03021.60834.81542.62952.55317.96532.82638.95048.921
RSGPTWithout EC22.16942.49849.66258.8199.10816.95821.51228.58311.84121.76428.09235.283
VenusRXWithout EC37.61656.71065.87875.05830.17747.16053.55963.26731.56849.32956.68765.628
ECReactWith EC44.58358.69163.75869.39727.63734.60039.51345.12233.72541.41944.48751.043
VenusRXWith EC48.08959.02862.82668.25538.36348.89753.18859.67228.04439.22644.34351.989

Response to Q4:

Thank you for pointing this out. The two counts refer to different curation stages: Table 1 of the manuscript reports the number of full reactions in VenusRX-Bench (126,361), whereas Table 6 of the manuscript reports the number of single-product reactions (83,439).

In our benchmark, full reactions are used for EC-number prediction, span corruption, and template extraction, while single-product reactions are curated for forward prediction and retrosynthesis. Specifically, starting from the full enzymatic reaction dataset, we further process and filter the reactions so that each sample retains only a single non-trivial principal product, decomposing reactions with multiple principal products into separate single-product entries. The number of single-product samples is smaller than that of full reactions due to this product filtering strategy. We have clarified this distinction in the revised manuscript and included both counts in Table 6 of the manuscript.

Response to Q5:

We apply structural filtering via RDKit and thermodynamic filtering via eQuilibrator (Beber et al., 2022) to the generated reactions, requiring all reactant and product molecules of a reaction to be structurally valid and the reaction Gibbs free energy to be negative (\Delta_r G < 0). This filtering pipeline yields 4.04M, 4.05M, and 3.85M full training reactions for the reactant split, product split, and EC split, respectively. We will include more detailed filtering statistics in the appendix.

We thank the reviewer again for the valuable comments, which helped us strengthen the evaluation and improve the clarity of the paper.

Reviewer_6D9fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper frames enzymatic reaction modeling as learning a joint space over reactants, products and EC numbers. It introduces VenusRX-Bench, a curated, atom-mapped and EC-annotated collection of enzymatic reactions with reactant, product and EC grouped splits for forward prediction, single-step retrosynthesis and EC prediction, and VenusRX, a small T5-style model trained first on millions of template-expanded reactions and then on real reactions, with optional EC prompts and molecule-library constrained decoding. VenusRX reports the best Top-k accuracy in most generation settings and competitive EC prediction.

Critical strengths

The benchmark integrates several biochemical sources and provides EC labels and atom mappings under shared splits, addressing real fragmentation in enzymatic-reaction evaluation.

Baseline coverage is broad, several gains still good when no library constraints given. EC-conditioned and structure-only results are reported separately, and the appendix documents processing in detail

Critical weaknesses

W1.Comparison is not fair. VenusRX is trained on millions of template-expanded reactions and uses multiple R-SMILES views at training and test time, while baselines are trained only on the real split. And even worse, removing Stage-1 drops forward Top-1 below the R-SMILES baseline, so the gain is mainly attributable to the extra corpus, which could equally be given to the baselines.

W2. the modeling component offers limited methodological novelty. It is a standard T5-style model with template-based data expansion and constrained decoding. The main value lies in the curated resource, so the paper's contribution to the ICLR audience is mainly a domain benchmark, and it may be better suited to a chemistry or bioinformatics venue. ICLR usually prefers chemical knowledge naturally turned into some network structure, so that following work can reuse it.

Minor

The domain-gap claim relies on zero-shot models reaching near-zero Top-1, which more likely reflects input-format mismatch

Questions for authors

See weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer 6D9f (Part 1/3)

Response to Reviewer 6D9f (Part 1/3)

We thank Reviewer 6D9f for the thoughtful review and helpful feedback. We appreciate your recognition of the value of VenusRX-Bench, our broad baseline coverage, and our detailed documentation. Below, we address each of your concerns in detail.

Response to W1:

  1. Regarding pretraining data, the three pretrained baselines used in our paper—RSGPT, R-SMILES, and SciReasoner (LLM)—were already pretrained on large-scale corpora before evaluation. Specifically, RSGPT and R-SMILES were pretrained on the much larger organic reaction dataset USPTO-FULL, and SciReasoner was pretrained on scientific literature, which may even contain data similar to our test set.

  2. Our implementation of the R-SMILES baseline strictly follows its official paper and codebase, and it also uses multiple R-SMILES views during both training and testing. Therefore, the comparison is fair in this regard.

  3. In the ablation study of VenusRX, the performance drop after removing Stage-1 only shows that pretraining on large-scale reaction data is important for VenusRX, rather than proving that the gain comes mainly from synthetic domain data. Similarly, R-SMILES itself also relies on pretraining on large-scale reaction data to perform well.

  4. In organic chemistry, large-scale databases such as USPTO-FULL already exist for pretraining. In contrast, enzymatic reaction modeling suffers from severe data scarcity. Constructing a large-scale synthetic enzymatic reaction corpus to address this bottleneck is itself one of our key contributions.

  5. Following your suggestion for a more controlled comparison, we pretrained and fine-tuned the strongest pretrained baseline, R-SMILES, on the exact same Stage-1 and Stage-2 datasets as VenusRX. The results are shown in the tables below. Even under matched training corpora, VenusRX still clearly outperforms R-SMILES (Same Corpus) across almost all settings. This confirms that the Stage-1 synthetic data is not the sole reason for our improvement, and our advantage does not come only from data differences.

Comparison of R-SMILES and VenusRX under matched training corpora. R-SMILES (Same Corpus) is pretrained and fine-tuned using the same Stage-1 and Stage-2 datasets as VenusRX. All models are evaluated without EC input. We report Top-k accuracy (%).

Table 1. Forward Reaction Prediction

MethodReactant Split T1T3T5T10Product Split T1T3T5T10EC Split T1T3T5T10
R-SMILES70.56282.04885.54388.61830.84944.63150.00055.84915.56828.10434.89944.643
R-SMILES (Same Corpus)70.60883.37686.68589.39933.29348.63455.09361.91325.95940.59246.29754.830
VenusRX73.06684.61188.04890.73950.95966.27571.39377.08531.10045.17053.23659.947

Table 2. Retrosynthesis

MethodReactant Split T1T3T5T10Product Split T1T3T5T10EC Split T1T3T5T10
R-SMILES5.3599.42512.12716.1464.1477.2999.08411.3141.1743.6674.9866.364
R-SMILES (Same Corpus)6.82312.49714.68218.2677.67014.28617.39022.4235.44110.60611.75712.979
VenusRX8.44617.18323.77730.60320.55430.05834.92339.9454.8788.73710.48714.274
AuthorsOpenReview ↗

Response to Reviewer 6D9f (Part 2/3)

Response to Reviewer 6D9f (Part 2/3)

Response to W2:

  1. Our starting point is that enzymatic reaction modeling still lacks sufficient domain training data and a shared evaluation benchmark for systematic comparison. Existing studies differ in data sources, processing pipelines, and task definitions, and they mostly rely on transferring models pretrained on general molecules or organic reactions. In this work, we aim to investigate what predictive and representational capabilities can be gained by large-scale pretraining directly on enzymatic reactions, and whether these capabilities generalize to unseen reactants, products, and catalytic functions.

  2. Our main contributions are as follows:

    (1) We build VenusRX-Bench to bring forward prediction, retrosynthesis, and EC prediction under a consistent data and evaluation protocol, using different splits to separate distinct generalization challenges.

    (2) Because real enzymatic reaction data is scarce, we use enzymatic reaction templates to construct a large-scale synthetic pretraining corpus and combine it with real reactions to produce a reusable pretrained model and reaction representations.

    (3) Adopting a mature T5-style architecture is a deliberate design choice. Our primary goal is to study the value of domain-specific pretraining for enzymatic reactions. Using a proven architecture avoids confounding effects from architectural modifications and makes it easier for future work to reproduce, use, and extend our model.

    (4) Our experiments further examine how catalytic context helps reaction prediction and whether the learned reaction representations can support catalytic function prediction. The results demonstrate both the predictive value of catalytic information and the reusability of our pretrained representations for downstream EC classification.

  3. Similar types of contributions have been well received at ICLR in recent years:

  • Mol-Instructions (ICLR 2024) focuses on building a biomolecular instruction resource and validating its value with existing models, which is similar to our enzymatic pretraining resource and model validation.
  • VenusX (ICLR 2026) contributes a fine-grained protein function benchmark, which is similar to our unified benchmark with disjoint reactant, product, and full-EC splits for evaluating generalization.
  • SaProt (ICLR 2024 Spotlight) keeps the standard ESM-2 backbone and improves representations through a structure-aware vocabulary and pretraining, showing that meaningful contributions can be built on mature architectures.
  1. Our work aligns with the ICLR 2027 scope, which explicitly includes "datasets and benchmarks" and "applications to physical sciences (physics, chemistry, biology, etc.)." Our unified enzymatic reaction benchmark, domain-specific pretraining, and reusable reaction representations contribute directly to these areas. We therefore believe the work is appropriate for ICLR and should be evaluated on these contributions alongside its modeling component. We hope this work provides a solid foundation for future research in enzymatic reaction modeling through our unified benchmark, our analysis of cross-domain transfer limits, and our empirical findings on mature architectures.
AuthorsOpenReview ↗

Response to Reviewer 6D9f (Part 3/3)

Response to Minor Comment:

In all zero-shot evaluations, we ensured that the input and output formats were strictly aligned with each model's original specification. We will add these formatting details to the appendix.

  1. For RSGPT, we used the exact format specified in the RSGPT paper:
TaskModel Input (Instruction)Model Output (Response)
Retrosynthesis<Isyn><O>products<F>reactants<RXN>templates
Forward<Syn><F>reactants<O>products<RXN>templates
  1. For R-SMILES, we followed the official Seq2Seq implementation for both tasks:
TaskModel Input (Source)Model Output (Target)
Retrosynthesis (P2R)Product SMILESReactants SMILES
Forward (R2P)Reactants SMILESProduct SMILES
  1. For SciReasoner, which is a language model, we constructed the following prompt for retrosynthesis:

SYSTEM: You are an expert in enzymatic retrosynthesis. Given an EC number and a target product SMILES, predict only the primary transformed substrate: the consumed reactant that shares the largest conserved atom scaffold with the target product and undergoes the EC-consistent reverse functional-group change. Exclude enzymes, products, auxiliary cofactors, co-substrates, salts, water, and inorganic species. Do not copy the unchanged target product. Return exactly one valid molecular SMILES enclosed in <SMILES>...</SMILES>.

USER: EC number: {EC_NUM} ({EC_CLASS_NAME}) With the provided product, recommend the single primary reactant that was most likely transformed into it.<SMILES> {PRODUCT_SMILES} </SMILES>

We thank the reviewer again for the constructive comments, which helped us strengthen the experimental comparison and clarify our presentation.

Thank you for the additional experiments and clarifications. However, my main concern about the methodological contribution remains. The matched-corpus comparison does not establish whether model size, training compute, and tuning budgets are comparable, or identify which modeling component explains the remaining advantage.

The response primarily reframes the contribution around the benchmark and domain-specific pretraining resource, rather than establishing a generalizable methodological insight beyond the value of domain data. AC may treat my score as 2.5~3 if AC buys the contribution, while my score is just based on my view on ICLR AI4S papers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.