Paper deep dive
Elucidating the Space of Enzymatic Reaction: A Unified Benchmark and Pretrained Model
Yutong Hu, Tianming Huang, Yanbo Zhao, Qiongyu Zhang, Shixiang Tang, Lei Bai, Ziyi Zhou, Liang Hong, Pan Tan
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 10/10/2026, 2:15:54 AM
Summary
This paper introduces VenusRX-Bench, a unified benchmark for enzymatic reaction prediction, and VenusRX, a pretrained T5-style sequence-to-sequence model. The benchmark integrates 126,361 reactions from multiple databases (BRENDA, PathBank, Rhea, KEGG) with EC annotations and atom mappings. VenusRX addresses the domain gap between chemical and enzymatic reactions through two-stage pretraining on template-expanded synthetic reactions and real data, optional EC conditioning, and Molecule Library-Constrained Decoding (MLCD). The model achieves state-of-the-art performance in forward prediction, retrosynthesis, and EC-number prediction.
Entities (12)
Relation Signals (12)
VenusRX â achievesbestperformanceon â VenusRX-Bench
confidence 95% · VenusRX achieves the best or competitive performance on most evaluated settings over representative chemical and enzymatic baselines.
VenusRX-Bench â containsdatafrom â PathBank
confidence 95% · We integrate EC-annotated reactions from BRENDA, PathBank, Rhea, and KEGG into a common representation
VenusRX-Bench â containsdatafrom â Rhea
confidence 95% · We integrate EC-annotated reactions from BRENDA, PathBank, Rhea, and KEGG into a common representation
VenusRX-Bench â containsdatafrom â KEGG
confidence 95% · We integrate EC-annotated reactions from BRENDA, PathBank, Rhea, and KEGG into a common representation
VenusRX-Bench â containsdatafrom â BRENDA
confidence 95% · We integrate EC-annotated reactions from BRENDA, PathBank, Rhea, and KEGG into a common representation
VenusRX-Bench â hassize â 126,361
confidence 95% · VenusRX-Bench comprises 126,361 unique enzymatic reactions
VenusRX-Bench â supportstask â EC-number prediction
confidence 95% · VenusRX-Bench, a unified benchmark for forward reaction prediction, single-step retrosynthesis, and EC-number prediction.
VenusRX-Bench â supportstask â
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Existing reaction models primarily learn molecular transformations, whereas enzy- matic 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, ret- rosynthesis, 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 on most evaluated settings over representative chem- ical 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
Tags
Links
- Source: https://arxiv.org/abs/2610.11694v1
- Canonical: https://arxiv.org/abs/2610.11694v1
Trouble viewing inline? Open PDF directly â
Full Text
98,651 characters extracted from source content.
Expand or collapse full text
Elucidating the Space of Enzymatic Reaction: A Unified Benchmark and Pretrained Model Yutong Hu Affiliation: Shanghai Artificial Intelligence Laboratory Email: sjtu1669165@sjtu.edu.cn Tianming Huang Affiliation: Shanghai Artificial Intelligence Laboratory Email: qzha0437@uni.sydney.edu.au Yanbo Zhao Email: tangshixiang@pjlab.org.cn Qiongyu Zhang Affiliation: The University of Sydney huyutong2020@sjtu.edu.cn, 1780693545@sjtu.edu.cn Email: bailei@pjlab.org.cn Shixiang Tang Affiliation: Shanghai Artificial Intelligence Laboratory Email: zy-zhou@sjtu.edu.cn Lei Bai Affiliation: Shanghai Artificial Intelligence Laboratory Email: tpan1039@gmail.com Ziyi Zhou â thanks: Corresponding authors: Ziyi Zhou, Pan Tan, and Liang Hong. Email: hongl3liang@sjtu.edu.cn Liang Hong11footnotemark: 1 Affiliation: Shanghai Artificial Intelligence Laboratory Pan Tan11footnotemark: 1 Affiliation: Shanghai Artificial Intelligence Laboratory Shanghai Jiao Tong University 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 on 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. $ $$ $footnotetext: These authors contributed equally to this work. 1 Introduction Enzymes are remarkable biocatalysts capable of accelerating complex chemical transformations with extraordinary regio- and stereoselectivity under mild physiological conditions. These superior catalytic properties position enzymatic reactions as a sustainable alternative to traditional chemical synthesis, driving transformative applications in synthetic biology, pharmaceutical manufacturing, and green chemistry (Turner & OâReilly, 2013; Sheldon & Woodley, 2018; Nielsen & Keasling, 2016). With the burgeoning demand for enzymatic synthesis, deciphering and modeling the underlying principles of enzymatic reactivity has emerged as a fundamental endeavor. In the pursuit of automated biocatalytic pathway design, data-driven approaches capable of predicting reaction outcomes, proposing biosynthetic routes, and inferring catalytic functions directly from molecular representations hold great potential (DelĂ©pine et al., 2018; Finnigan et al., 2021). Currently, machine learning and deep learning methods have achieved substantial progress in conventional organic reaction modeling, including reaction outcome prediction, retrosynthesis, and reaction classification. This progress has been largely driven by the availability of large-scale reaction datasets extracted from chemical patents (Lowe, 2012; Kearnes et al., 2021), which have enabled efficient benchmarking (Wan et al., 2022) and pretraining (Schwaller et al., 2021b; Irwin et al., 2022) of AI models. However, for enzymatic reaction, the scarcity of high-quality, large-scale datasets and benchmarks has become a major bottleneck. First, reported enzymatic reactions are limited in quantity and scattered across heterogeneous databases, requiring extensive standardization, correction, and curation (Heid et al., 2023). Second, existing studies for modeling enzymatic reactions vary in data sources, preprocessing pipelines, task definitions, and data splitting strategies (Probst et al., 2022a; Zheng et al., 2022; Heid et al., 2023), making their reported performances difficult to compare directly. Moreover, enzymatic reactions differ from organic reactions in a fundamental way: the enzyme catalyst is not merely a rate accelerator but an active determinant of reaction outcome. Organic reaction outcomes are primarily governed by substrate functional group chemistry, yielding relatively consistent pathways across different conditions. In contrast, the high functional diversity of enzymes introduces additional dimensions into enzymatic reactionsâincluding substrate specificity, stereoselectivity, and cofactor dependenceâsuch that reaction outcomes depend jointly on both substrate and enzyme. Since existing organic reaction models are not designed to incorporate enzymatic features, they are insufficient to capture the complexity of enzymatic transformations. In this work, we formulate the enzymatic reaction space as the joint relationships among reactants, products, and catalytic functions (Figure 1). Specifically, we employ Enzyme Commission (EC) numbers as a proxy for catalytic function, as their hierarchical taxonomy systematically encodes enzyme specificity while reaction-EC pairs are more abundant in existing databases than reaction-enzyme pairs. To chart the enzymatic reaction space, we first construct a comprehensive benchmark named VenusRX-Bench. It comprises 126,361 unique enzymatic reactions curated from multiple biochemical databases, all annotated with EC numbers and atom-atom mapping. To the best of our knowledge, it represents the largest curated dataset of enzymatic reactions to date. VenusRX-Bench covers three core tasks: forward reaction prediction, single-step retrosynthesis, and EC-number prediction. It employs leakage-controlled data splits to evaluate out-of-distribution generalization across distinct biochemical dimensions. Our evaluation reveals that directly transferring organic reaction models to enzymatic reactions yields suboptimal performance, primarily due to the organic-to-enzymatic domain gap: (1) the presence of enzymatic catalysis, and (2) the larger molecular size of biomolecules. Given the limited scale of enzymatic reaction data, this gap cannot be easily bridged by simple supervised learning. Accordingly, we propose VenusRX, a unified pretrained language model for enzymatic reactions. VenusRX adopts an encoder-decoder architecture and is pretrained under a T5-style (Raffel et al., 2020a) multi-task objective on enzymatic reaction SMILES and EC numbers, encompassing span corruption, forward reaction prediction, and retrosynthesis. To address the scarcity of enzymatic reaction data, we generate million-scale synthetic enzymatic reactions based on extracted reaction templates and conduct two-stage pretraining on both synthetic and real enzymatic reactions. To incorporate catalytic function context, we randomly append EC prompts to the input during pretraining, enabling the model to optionally condition its predictions on EC numbers at inference time. To improve the generation of large biomolecules, we further design a trie-based constrained sampling strategy that effectively reduces invalid outputs. Experimental results demonstrate that VenusRX achieves the best or competitive performance on most evaluated settings compared to both organic-chemistry and biochemical baselines, and that each of the above designs contributes critically to model performance. 2 VenusRX-Bench Figure 1: Overview of the enzymatic reaction space and our framework. VenusRX-Bench evaluates the relationships among reactants, products, and catalytic functions, while VenusRX learns a shared model of this space through domain-specific pretraining. Biochemical reaction databases differ in their formats, identifiers, and structural coverage. Study-specific tasks and processing procedures further limit direct comparison of reported results. VenusRX-Bench addresses these differences through a common reaction collection, standardized prediction tasks, and shared evaluation rules. 2.1 Data Construction and Splits Data Construction We integrate EC-annotated reactions from BRENDA, PathBank, Rhea, and KEGG into a common representation while retaining source identifiers for traceability. Records are deduplicated by normalized reaction text and EC annotation, preserving associations between the same reaction and different catalytic annotations. We then resolve molecular structures, use EnzymeMap (Heid et al., 2023) for reaction balancing where possible, and generate atom mappings with RXNMapper (Grandjean et al., 2026). Following ECReact (Probst et al., 2022a), product-side curation retains a single principal product and excludes records without reactants. Table 1(a) compares the integrated collection with existing biochemical reaction datasets. VenusRX-Bench has the largest reported reaction count among the listed datasets and provides both EC annotations and atom mappings. Splits We employ three split strategies for training, validation, and test sets. In each split, the grouping unit is disjoint across the training, validation, and test sets: âą Reactant split. Records are grouped by canonical reactant combinations to evaluate generalization to unseen combinations. âą Product split. Records are grouped by canonical principal products to evaluate generalization to unseen target products. âą EC split. Records are grouped by their original complete four-level EC annotations to evaluate generalization to unseen catalytic activities. Product disjointness applies to the product side, while parent EC classes may overlap under the EC split. Detailed curation procedures and additional grouping constraints are provided in Appendix A. 2.2 Tasks and Evaluation Each benchmark example is represented as (R,P,e)(R,P,e), where R denotes the reactants, P the retained principal product, and e the associated EC annotation. We define three tasks: Forward prediction: prediction: Râeâ(optional)P, R \,e\ (optional)\,P, Single-step retrosynthesis: -step retrosynthesis: Pâeâ(optional)R, P \,e\ (optional)\,R, EC-number prediction: -number prediction: (R,P)âŒe. (R,P) e. For forward prediction and retrosynthesis, we report multi-reference Top-k exact-match accuracy for kâ1,3,5,10kâ\1,3,5,10\. When multiple valid targets are associated with the same input in the test set, a prediction is counted as correct if any of the top-k candidates matches any recorded reference. Results with and without EC conditioning are reported separately. For EC-number prediction, we evaluate all four hierarchy levels (EC1âEC4) using top-1 accuracy, weighted F1, and macro F1. At level â , the target consists of the first â fields of the EC annotation. The EC label is used only as the prediction target and is never provided as input. 2.3 Dataset Property Table 1(b) compares VenusRX-Bench with USPTO50K. VenusRX-Bench contains more reactant-side heavy atoms on average. The largest individual reactant is also larger on average. Thus, the increase reflects larger molecular structures, not simply more reactant components. (a) Dataset comparison. Dataset # Rxns EC Mapping VenusRX-Bench (ours) 126,361 â â EnzymeMap (Heid et al., 2023) 47,974 â â ECReact (Probst et al., 2022a) 62,222 â Ă CARE (Yang et al., 2024) 48,392 â Ă EnzymeCage (Liu et al., 2026b) 29,608 â â BioNavi-NP (Zheng et al., 2022) 33,710 Ă Ă (b) Molecular statistics. Metric Dataset Mean Med. Max Sum of heavy atoms VenusRX-Bench 48.78 44 266 USPTO50K 28.29 28 97 SMILES length VenusRX-Bench 234.56 208 844 USPTO50K 189.83 186 707 molecule count VenusRX-Bench 2.11 2 5 USPTO50K 1.71 2 3 Max reactant heavy atoms VenusRX-Bench 36.23 33 265 USPTO50K 21.87 21 97 Table 1: Comparison of VenusRX-Bench with existing reaction datasets. (a) VenusRX-Bench provides the largest collection of unique enzymatic reactions together with EC annotations and atomâatom mappings. (b) Compared with USPTO50K, VenusRX-Bench contains larger and more complex reactant structures.All metrics are reactant-side. These structural differences affect reactant encoding in forward prediction and target generation in retrosynthesis. In particular, longer, multi-component retrosynthetic targets require models to generate larger molecular structures while maintaining valid representations. This motivates training directly on enzymatic reactions and constraining molecular generation, rather than assuming reliable transfer from conventional chemical reaction data. The paired structure-only and EC-conditioned settings separately examine how catalytic information guides prediction. These considerations motivate VenusRX in Section 3; their effects on transfer, prediction accuracy, and generation validity are assessed in the experiments. 3 VenusRX VenusRX is a unified model for learning enzymatic reaction space. It formulates reaction modeling as shared text-to-text learning over forward prediction, retrosynthesis, and reaction reconstruction. To address the limited amount of experimentally observed enzymatic reactions, we first train the model on a large template-expanded reaction corpus and then continue training on real VenusRX-Bench reactions. VenusRX also supports optional Enzyme Commission (EC) conditioning, allowing catalytic information to guide reaction prediction when it is available. For retrosynthesis and forward prediction, we further study constrained generation to reduce the difficulty of generating large and complex biochemical molecules. 3.1 Unified Multitask Modeling VenusRX uses a T5-style encoderâdecoder architecture (Raffel et al., 2020b), with all parameters shared across three text-to-text objectives: Forward: : <forward>âRâP, <forward>Râ P, Retrosynthesis: : <retro>âPâR, <retro>Pâ R, Span corruption: corruption: <mlm>â(Râ«P)~âmasked spans. <mlm> (R\! \!P) spans. The first two objectives learn both directions of an enzymatic transformation, while span corruption encourages shared reaction representations from the complete reaction context. EC conditioning. Because molecular structures alone may not uniquely determine the intended enzymatic transformation, we additionally treat the EC number as an optional condition. For EC-annotated reactions, <ec_number>e is appended to the input with probability 0.50.5 during training. This allows the same model to operate with or without catalytic information at inference time. Detailed task construction and optimization settings are provided in Appendix C. 3.2 Training Procedure VenusRX is trained in two stages. Stage-1 learns general enzymatic reaction patterns from a large template-expanded reaction corpus. Stage-2 then continues training on the real reactions in the corresponding VenusRX-Bench training split, adapting the pretrained model to the benchmark distribution. Both stages use the same model architecture and multitask objectives. For each reaction, we additionally use R-SMILES augmentation (Zhong et al., 2022) and randomly sample one representation during training. This exposes the model to multiple sequence views of the same chemical reaction without changing the number of unique reactions. 3.2.1 Large-Scale Enzymatic Reaction Pretraining Experimentally observed enzymatic reactions are limited in both scale and substrate diversity. To expand the reaction space available for pretraining, we construct a Stage-1 corpus containing about four million reaction records. The corpus is built in four steps: 1. Template extraction. We extract reaction templates from atom-mapped enzymatic reactions. Each template captures the local structural transformation around the reaction center and can be reused with new substrates. 2. Candidate substrate collection. We assemble a large molecule library from biochemical and natural product databases, standardize molecular structures, and remove duplicated or unsuitable candidates. 3. Template-based reaction generation. For each template, we identify structurally compatible substrates and apply the reaction rule to generate new reactantâproduct pairs. This expands known enzymatic transformations to a broader range of substrate contexts. 4. Reaction filtering. Generated reactions are checked for structural validity, chemical feasibility, and duplication. To prevent data leakage, we further remove reactions overlapping with the validation or test set of each benchmark split before pretraining. This process expands the training data along the chemical dimension, while R-SMILES augmentation provides multiple sequence representations of each retained reaction. The two augmentations therefore play complementary roles: template-based generation increases reaction diversity, whereas R-SMILES augmentation increases representation diversity. Full details of molecule sources, template generation, filtering, and split-specific processing are provided in Appendix D. 3.3 Molecule Library-Constrained Decoding Enzymatic reactions in VenusRX-Bench often involve larger and more complex molecules than conventional organic reaction datasets (Table 1(b)), making autoregressive generation more prone to invalid outputs. To address this issue, we introduce Molecule Library-Constrained Decoding (MLCD), which restricts molecular generation to a predefined library of valid molecules. MLCD builds a prefix tree (trie) over tokenized molecules in the library. During beam search, only next tokens that remain consistent with at least one library molecule are allowed. For reactions containing multiple molecules, the constraint is applied independently to each generated molecule. The trie constrains the molecular search space, while VenusRX still scores and selects candidates according to the input reaction context. Detailed decoding procedures and pseudocode are provided in Appendix G. 4 Experiments 4.1 Experimental Setup We evaluate forward prediction, single-step retrosynthesis, and EC-number prediction using the tasks, splits, and metrics defined by VenusRX-Bench. All VenusRX experiments use the same 25.1M-parameter configuration. For reaction generation, we report multi-reference Top-k accuracy, while weighted F1 is the primary metric for EC prediction. Results with and without EC conditioning are reported separately. The best result is shown in bold and the second-best result is underlined. Baselines. We compare VenusRX with representative methods for conventional chemical reaction modeling, enzymatic reaction modeling, and EC prediction. Trainable baselines are adapted to the corresponding VenusRX-Bench training splits, while zero-shot methods use their released pretrained checkpoints. Baseline-specific training and inference details are provided in the appendix. EC prediction. We evaluate VenusRX representations in both frozen and task-specific settings. The frozen setting directly uses pretrained reaction representations for EC retrieval, while the task-specific setting replaces CLAIREâs reaction features with the same frozen VenusRX representations and retrains its downstream predictor. Leakage control and round-trip evaluation. For each benchmark split, we remove reactions overlapping with its validation or test sets from the Stage-1 corpus before pretraining. For round-trip evaluation, we use a fixed VenusRX forward oracle trained on the complete benchmark data and measure whether retrosynthetic candidates can recover the target product. The oracle is used only for evaluation and does not affect retrosynthetic generation. Further implementation details are provided in the appendix. 4.2 Results 4.2.1 Generalization across Enzymatic Reaction Space Table 2: Forward reaction prediction results on three data splits. We report Top-k exact-match accuracy (%). Methods marked with â are evaluated using open-source checkpoints without fine-tuning. Method Reactant Split Product Split EC Split T1 T3 T5 T10 T1 T3 T5 T10 T1 T3 T5 T10 Zero-shot R-SMILESâ (Zhong et al., 2022) 0.000 0.711 0.979 1.841 0.000 0.683 1.354 2.313 0.000 0.324 0.923 1.594 Without EC Input R-SMILES 70.562 82.048 85.543 88.618 30.849 44.631 50.000 55.849 15.568 28.104 34.899 44.643 Graph2SMILES (Tu & Coley, 2022) 68.453 80.219 82.863 84.576 34.408 44.427 48.286 51.702 19.823 29.794 33.401 37.548 MolTransformer (Schwaller et al., 2019) 52.796 66.344 71.307 75.221 12.776 22.471 26.330 29.051 9.396 15.808 18.456 21.560 MEGAN (Sacha et al., 2021) 28.285 39.247 44.676 46.202 19.128 24.676 26.366 27.421 7.646 12.308 14.166 14.801 VenusRX 73.066 84.611 88.048 90.739 50.959 66.275 71.393 77.085 31.100 45.170 53.236 59.947 VenusRX (MLCD) 77.074 88.350 91.414 94.210 58.785 74.413 79.123 84.552 39.274 56.328 64.753 74.161 With EC Input ECReact (Probst et al., 2022a) 61.440 74.080 78.285 81.885 37.812 51.558 55.908 59.456 26.246 40.400 45.602 51.989 VenusRX 75.221 86.778 89.725 92.474 55.825 69.463 74.041 78.811 34.564 50.240 56.340 64.058 VenusRX (MLCD) 80.056 91.135 93.977 95.515 62.596 77.193 81.556 85.918 42.917 61.182 69.128 77.661 Table 3: Retrosynthesis results on three data splits. We report Top-k exact-match accuracy (%). Methods marked with â are evaluated using open-source checkpoints without fine-tuning. Method Reactant Split Product Split EC Split T1 T3 T5 T10 T1 T3 T5 T10 T1 T3 T5 T10 Zero-shot SciReasonerâ (LLM) (Wang et al., 2025) 1.445 1.957 2.528 3.355 0.947 1.570 2.121 2.876 0.204 0.611 0.839 1.282 R-SMILESâ 0.000 0.687 2.004 2.656 0.000 1.007 2.073 2.601 0.000 0.228 0.527 0.899 RSGPTâ (Deng et al., 2025a) 0.058 0.082 0.128 0.373 0.000 0.048 0.132 0.204 0.024 0.048 0.192 0.527 Without EC Input Graph2SMILES 7.025 11.708 15.040 18.802 27.397 34.192 36.925 39.022 3.320 4.626 5.693 6.388 Retroformer (Wan et al., 2022) 6.710 12.488 16.170 22.938 9.863 12.800 15.688 18.253 1.558 2.601 3.248 4.219 MEGAN 5.510 9.914 12.022 15.645 10.127 14.190 17.581 19.343 0.683 1.534 2.061 2.469 R-SMILES 5.359 9.425 12.127 16.146 4.147 7.299 9.084 11.314 1.174 3.667 4.986 6.364 BioNavi (Zheng et al., 2022) 2.901 5.440 8.889 13.292 9.767 15.484 18.552 21.632 1.007 1.642 1.930 2.900 G2Retro (Guo et al., 2023) 1.351 2.365 2.971 5.720 2.385 5.177 7.227 9.012 0.503 0.923 1.594 1.822 RSGPT 2.854 7.712 10.869 16.181 8.497 15.580 18.816 23.730 1.151 2.601 3.679 6.628 VenusRX 8.446 17.183 23.777 30.603 20.554 30.058 34.923 39.945 4.878 8.737 10.487 14.274 VenusRX (MLCD) 9.448 19.070 26.584 34.355 22.903 34.636 40.244 46.093 7.407 14.657 18.384 23.849 With EC Input ECReact 11.149 17.568 21.226 25.757 17.282 24.521 27.049 27.457 5.837 8.689 10.606 11.086 VenusRX 17.230 26.398 29.345 34.425 28.847 34.696 38.267 40.568 18.360 24.605 27.481 30.872 VenusRX (MLCD) 17.300 27.400 32.258 38.828 29.302 39.525 43.672 48.466 20.374 27.505 30.609 34.552 Tables 2 and 3 evaluate reaction prediction across the Reactant, Product, and EC Splits. For forward prediction, VenusRX achieves consistently strong performance across all three settings, with particularly clear gains on the Product and EC Splits. Similar improvements are observed in retrosynthesis, where VenusRX generalizes effectively across changes in molecular targets and catalytic functions. Molecule Library-Constrained Decoding (MLCD) provides further gains in both prediction directions. These results show that domain-specific pretraining transfers across multiple dimensions of enzymatic reaction space rather than being tied to a particular reaction distribution. They also demonstrate the value of explicitly modeling the structural characteristics of enzymatic reactions during generation. 4.2.2 Catalytic Function and Reaction Structure Are Coupled Table 4: EC-number prediction on the Reactant and Product splits. Weighted F1 (%) is reported at EC levels 1â4. Method Reactant Split Product Split EC1 EC2 EC3 EC4 EC1 EC2 EC3 EC4 Zero-Shot crxnfp (Zhang et al., 2025) 92.3892.38 85.7385.73 80.9080.90 46.5446.54 89.1589.15 80.8380.83 74.3574.35 41.9041.90 DRFP (Probst et al., 2022b) 92.57ÂŻ 92.57 85.75ÂŻ 85.75 81.21ÂŻ 81.21 49.40ÂŻ 49.40 91.48ÂŻ 91.48 84.1084.10 77.8477.84 44.61ÂŻ 44.61 rxnfp (Schwaller et al., 2021b) 92.1192.11 85.5985.59 80.5280.52 47.1147.11 88.8888.88 79.8079.80 73.0373.03 41.0941.09 VenusRXâ 94.7694.76 87.7687.76 82.8182.81 51.1751.17 92.1192.11 83.80ÂŻ 83.80 77.68ÂŻ 77.68 47.2747.27 Task Specific Model Theia (Probst, 2023) 93.09±0.2893.09± 0.28 87.72±0.4687.72± 0.46 83.68±0.2483.68± 0.24 52.49±0.2652.49± 0.26 92.48±0.4792.48± 0.47 84.86±0.3884.86± 0.38 79.33¯±0.45 79.33± 0.45 47.44±0.3647.44± 0.36 CLAIRE (Zeng et al., 2025) 94.88±0.1494.88± 0.14 89.36±0.2789.36± 0.27 84.14¯±0.40 84.14± 0.40 49.49±0.9649.49± 0.96 92.41±0.6292.41± 0.62 86.18±0.3086.18± 0.30 78.86±0.5178.86± 0.51 43.50±0.6043.50± 0.60 BEC-Pred (Qian et al., 2024) 96.16¯±0.33 96.16± 0.33 89.78¯±0.14 89.78± 0.14 81.45±0.5081.45± 0.50 46.79±0.4346.79± 0.43 94.67¯±0.36 94.67± 0.36 87.02¯±0.99 87.02± 0.99 76.13±0.6076.13± 0.60 42.14±0.7642.14± 0.76 VenusRX 96.45±0.1196.45± 0.11 90.99±0.1690.99± 0.16 86.03±0.2486.03± 0.24 51.33¯±0.59 51.33± 0.59 94.85±0.3394.85± 0.33 87.40±0.2287.40± 0.22 80.23±0.6480.23± 0.64 47.02¯±0.28 47.02± 0.28 VenusRX denotes frozen VenusRX representations followed by a task-specific predictor. Task-specific results are mean ± std over three seeds. Comparing the EC-conditioned and structure-only settings in Tables 2 and 3 shows that catalytic information consistently improves reaction prediction. The benefit is especially clear in retrosynthesis, where EC annotations provide useful context for identifying plausible transformations from a target product. The reverse relationship is also observed. As shown in Table 4, frozen VenusRX reaction representations encode sufficient catalytic information to support effective EC prediction, while using the same representations in a task-specific predictor further improves performance across the EC hierarchy. Together, these results reveal a bidirectional relationship between reaction structure and catalytic function: catalytic information guides reaction prediction, while reaction representations recover catalytic information. This empirical coupling supports our formulation of enzymatic reaction space as a joint relationship among reactants, products, and catalytic functions. 4.2.3 Consistency and Constrained Generation MLCD consistently improves exact-match performance across forward prediction and retrosynthesis, showing that constraining the molecular output space helps the model recover benchmark targets more reliably. The validity analysis in Appendix G.2 further shows that MLCD substantially reduces invalid molecular generation, supporting its effectiveness for decoding large and complex biochemical structures. Round-trip evaluation in Appendix F.1 provides a complementary view of the generated candidate space. While unconstrained decoding retains a broader set of alternative hypotheses, MLCD focuses generation on molecules supported by the predefined library. This more targeted search improves recovery of recorded reaction targets and reduces structurally invalid outputs. Overall, these results show that MLCD provides an effective inference-time strategy for navigating the large molecular output space of enzymatic reactions, improving both generation validity and target recovery. 4.2.4 Ablation Study Table 5 examines the main components of VenusRX. The full two-stage training configuration provides the strongest overall performance, showing that large-scale reaction pretraining and benchmark-domain continued training play complementary roles. Stage-1 provides transferable enzymatic transformation knowledge, while Stage-2 adapts this knowledge to real benchmark reactions. The span-corruption objective provides additional benefits, particularly for retrosynthesis and EC prediction. This suggests that reconstructive learning strengthens the shared reaction representation beyond the two directional generation tasks. Overall, the ablation results support the combination of large-scale pretraining, domain adaptation, and multitask learning used by VenusRX. Table 5: Ablation study on the Reactant Split. Reaction generation is evaluated with EC conditioning and without MLCD; EC prediction reports zero-shot weighted F1 (%). Variant Training Objectives Stage-1 Stage-2 Forward (with EC) Retrosynthesis (with EC) EC Prediction Fwd. Retro. MLM Top-1 Top-5 Top-10 Top-1 Top-5 Top-10 EC1 EC2 EC3 EC4 Full â â â â â 75.221 89.725 92.474 17.230 29.345 34.425 94.76 87.76 82.81 51.17 w/o MLM â â â â â 75.175 89.119 92.020 17.160 27.691 32.013 92.53 85.14 80.31 49.54 w/o Stage-1 â â â â â 47.192 71.086 77.178 7.130 17.882 22.682 89.86 80.78 75.49 45.17 w/o Stage-2 â â â â â 35.450 54.578 60.881 6.454 14.003 17.999 94.41 87.01 82.03 49.38 5 Related Work 5.1 Organic Reaction Modeling Forward reaction prediction and retrosynthesis. A dominant line of work casts reaction modeling as SMILES-to-SMILES translation with Transformers. Molecular Transformer pioneered this paradigm for forward prediction (Schwaller et al., 2019), later extended to single-step retrosynthesis (Karpov et al., 2019); Retroformer further adds local attention over reactive topology (Wan et al., 2022), while R-SMILES aligns input-output roots via atom mapping to ease sequence learning (Zhong et al., 2022). In parallel, graph-based models encode molecular topology more explicitly: Graph2SMILES uses a permutation-invariant graph encoder with a SMILES decoder (Tu & Coley, 2022), whereas MEGAN and Graph2Edits model reactions as sequences of graph edits (Sacha et al., 2021; Zhong et al., 2023). These approaches provide useful architectures and baselines for our study. Chemical reaction pretraining. To learn transferable reaction knowledge, recent work pretrains models on large chemical corpora before adapting them to downstream tasks. On the representation side, RXNFP learns reaction fingerprints via BERT-style masked language modeling for classification and reaction-space mapping (Schwaller et al., 2021b), while RxnRep uses chemically consistent contrastive pretraining for classification and regression (Wen et al., 2022). On the generative side, Chemformer applies BART-style pretraining for forward prediction and retrosynthesis (Irwin et al., 2022); ReactSeq introduces a reaction description language for retrosynthesis, retrieval, and yield prediction (Xiong et al., 2025); and RSGPT scales generative pretraining to ten-billion synthetic reactions for retrosynthesis planning (Deng et al., 2025a). Their strong performance motivates us to develop a pretrained model tailored specifically to enzymatic reactions. 5.2 Enzymatic Reaction Data and Modeling Data resources and benchmarks. Existing efforts have constructed enzymatic reaction datasets and associated evaluation protocols. ECREACT aggregates EC-annotated reactions from Rhea, BRENDA, PathBank, and MetaNetX for forward prediction and retrosynthesis (Probst et al., 2022a). EnzymeMap provides a pipeline for correcting, validating, and atom-mapping enzymatic reactions, and evaluates the curated data on forward prediction, retrosynthesis, and regioselectivity (Heid et al., 2023). ReactZyme instead focuses on bidirectional enzyme-reaction retrieval, assessing generalization under enzyme and reaction similarity splits (Hua et al., 2024). CARE further introduces a benchmark for enzyme classification and retrieval, where Task 2 retrieves EC numbers from query reactions (Yang et al., 2024). While complementary, these resources differ substantially in task definitions, data processing, and curation quality, making them difficult to merge into a unified benchmark. Bio-retrosynthesis and catalytic function prediction. Beyond dataset curation, several methods model enzymatic transformations by incorporating catalytic context into neural networks. Kreutter et al. adapt a Molecular Transformer via transfer learning, conditioning product prediction on textual enzyme descriptions (Kreutter et al., 2021). ECREACT instead encodes EC numbers as class tokens for forward prediction and jointly predicts reactants and EC classes in retrosynthesis (Probst et al., 2022a). BioNavi-NP trains Transformers on organic and biosynthetic reactions and couples single-step predictions with multi-step search for natural-product pathway planning (Zheng et al., 2022). For catalytic function annotation, CLAIRE employs contrastive learning to predict EC numbers for enzymatic reactions (Zeng et al., 2025). More recently, Enzyformer applies two-stage molecular and organic-reaction pretraining before fine-tuning for enzymatic retrosynthesis, together with a separate contrastive model for EC prediction (Liu et al., 2026a). Despite this progress, existing approaches are largely confined to task-specific supervised learning on individual datasets, and a unified pretraining framework with a comprehensive benchmark dedicated to enzymatic reactions remains lacking. 6 Conclusions and Future Work We introduced VenusRX-Bench, a unified benchmark for enzymatic reaction modeling across forward prediction, retrosynthesis, and EC-number prediction, with controlled splits that evaluate generalization across reactants, products, and catalytic functions. Based on this benchmark, we developed VenusRX, a unified pretrained encoderâdecoder model that combines large-scale template-expanded pretraining, continued training on real enzymatic reactions, optional EC conditioning, and molecule-library constrained decoding. Experiments across multiple tasks and splits show that domain-specific pretraining and catalytic context are important for modeling enzymatic reactions, while reaction representations themselves also encode useful catalytic information. Several directions remain for future work. First, EC numbers provide only a coarse description of catalytic function; incorporating enzyme sequences, structures, cofactors, and reaction conditions may enable more precise modeling of enzyme-dependent reaction outcomes. Second, the current benchmark focuses on single-step reactions, and extending the framework to multi-step biosynthetic pathway planning would provide a more complete evaluation of biocatalytic synthesis. Finally, constrained decoding currently relies on a predefined molecular library. Developing generation methods that preserve chemical validity while allowing exploration beyond a fixed candidate space may further improve generalization to previously unseen biochemical molecules. AI use statement AI was used to provide feedback on experimental design, assist with code implementation and checking, conduct related-work research, and support manuscript drafting. The correctness of the implementation and experimental results was independently verified by the authors. The authors take full responsibility for the final content of this work. Reproducibility Statement. We provide detailed descriptions of the benchmark construction, data splits, model architecture, training procedure, inference settings, and evaluation protocols in the main text and appendix. Hyperparameters and implementation details necessary to reproduce the reported experiments are documented where applicable. Upon acceptance, we will publicly release the source code, benchmark processing pipeline, evaluation scripts, and model checkpoints to facilitate reproducibility and future research. References Degen et al. (2008) Jörg Degen, Christof Wegscheid-Gerlach, Andrea Zaliani, and Matthias Rarey. On the art of compiling and using âdrug-likeâ chemical fragment spaces. ChemMedChem, 3(10):1503â1507, 2008. doi: 10.1002/cmdc.200800178. DelĂ©pine et al. (2018) Baudoin DelĂ©pine, Thomas Duigou, Pablo Carbonell, and Jean-Loup Faulon. Retropath2.0: A retrosynthesis workflow for metabolic engineers. Metabolic Engineering, 45:158â170, 2018. doi: 10.1016/j.ymben.2017.12.002. Deng et al. (2025b) Yafeng Deng, Xinda Zhao, Hanyu Sun, Yu Chen, Xiaorui Wang, Xi Xue, Liangning Li, Jianfei Song, Chang-Yu Hsieh, Tingjun Hou, Xiandao Pan, Taghrid Saad Alomar, Xiangyang Ji, and Xiaojian Wang. RSGPT: a generative transformer model for retrosynthesis planning pre-trained on ten billion datapoints. Nature Communications, 16:7012, 2025b. doi: 10.1038/s41467-025-62308-6. Deng et al. (2025a) Yafeng Deng, Xinda Zhao, Hanyu Sun, Yu Chen, Xiaorui Wang, Xi Xue, Liangning Li, Jianfei Song, Chang-Yu Hsieh, Tingjun Hou, Xiandao Pan, Taghrid Saad Alomar, Xiangyang Ji, and Xiaojian Wang. Rsgpt: a generative transformer model for retrosynthesis planning pre-trained on ten billion datapoints. Nature Communications, 16(1):7012, 2025a. doi: 10.1038/s41467-025-62308-6. URL https://doi.org/10.1038/s41467-025-62308-6. Finnigan et al. (2021) William Finnigan, Lorna J. Hepworth, Sabine L. Flitsch, and Nicholas J. Turner. Retrobiocat as a computer-aided synthesis planning tool for biocatalytic reactions and cascades. Nature Catalysis, 4(2):98â104, 2021. doi: 10.1038/s41929-020-00556-z. Grandjean et al. (2026) Yves Grandjean, Sacha Raffaud, Annie M. Westerlund, Philippe Schwaller, Samuel Genheden, and Jean-Louis Reymond. Rxnmapperv2: updated validation and evaluation on expanded chemical space. ChemRxiv, 2026(0626), 2026. doi: 10.26434/chemrxiv.15005247/v1. URL https://chemrxiv.org/doi/abs/10.26434/chemrxiv.15005247/v1. Guo et al. (2023) Zhongliang Guo, Yan Wu, Ke Liu, Yuxin Sun, Yujing Yang, Peng Yang, Jun Wang, Yidan Zhang, Jian Wang, and Wen Du. G2Retro as a two-step graph generative models for retrosynthesis prediction. Communications Chemistry, 6:102, 2023. doi: 10.1038/s42004-023-00897-3. Heid et al. (2023) Esther Heid, Daniel Probst, William H. Green, and Georg K. H. Madsen. Enzymemap: Curation, validation and data-driven prediction of enzymatic reactions. Chemical Science, 14(48):14229â14242, 2023. doi: 10.1039/D3SC02048G. Hua et al. (2024) Chenqing Hua, Bozitao Zhong, Sitao Luan, Liang Hong, Guy Wolf, Doina Precup, and Shuangjia Zheng. ReactZyme: A benchmark for enzyme-reaction prediction. In Advances in Neural Information Processing Systems, volume 37, p. 26415â26442, 2024. URL https://proceedings.neurips.c/paper_files/paper/2024/hash/2e68b2367d2e0bc8d6f0f86e07c2eb-Abstract-Datasets_and_Benchmarks_Track.html. Irwin et al. (2022) Ross Irwin, Spyridon Dimitriadis, Jiazhen He, and Esben Jannik Bjerrum. Chemformer: A pre-trained transformer for computational chemistry. Machine Learning: Science and Technology, 3(1):015022, 2022. doi: 10.1088/2632-2153/ac3ffb. Karpov et al. (2019) Pavel Karpov, Guillaume Godin, and Igor V. Tetko. A transformer model for retrosynthesis. In Artificial Neural Networks and Machine Learning â ICANN 2019: Workshop and Special Sessions, volume 11731 of Lecture Notes in Computer Science, p. 817â830, Cham, 2019. Springer. ISBN 978-3-030-30492-8. doi: 10.1007/978-3-030-30493-5_78. Kearnes et al. (2021) Steven M. Kearnes, Michael R. Maser, Michael Wleklinski, Anton Kast, Abigail G. Doyle, Spencer D. Dreher, Joel M. Hawkins, Klavs F. Jensen, and Connor W. Coley. The open reaction database. Journal of the American Chemical Society, 143(45):18820â18826, 2021. doi: 10.1021/jacs.1c09820. URL https://doi.org/10.1021/jacs.1c09820. Kreutter et al. (2021) David Kreutter, Philippe Schwaller, and Jean-Louis Reymond. Predicting enzymatic reactions with a molecular transformer. Chemical Science, 12(25):8648â8659, 2021. doi: 10.1039/D1SC02362D. Liu et al. (2026a) Tiantao Liu, Jiangcheng Xu, Xinke Zhan, Shaolong Lin, and Shirley W. I. Siu. Enzyformer: a two-stage pretrained model for enzymatic retrosynthesis. Journal of Cheminformatics, 18:35, 2026a. doi: 10.1186/s13321-026-01164-y. Liu et al. (2026b) Yong Liu, Chenqing Hua, Minghao Xu, Tao Zeng, Jiahua Rao, Zhaoyang Zhang, others, and Shuangjia Zheng. A geometric foundation model for enzyme retrieval with evolutionary insights. Nature Catalysis, 9(2):148â160, 2026b. doi: 10.1038/s41929-026-01478-y. Lowe (2012) Daniel Mark Lowe. Extraction of Chemical Structures and Reactions from the Literature. PhD thesis, University of Cambridge, 2012. Nielsen & Keasling (2016) Jens Nielsen and Jay D. Keasling. Engineering cellular metabolism. Cell, 164(6):1185â1197, 2016. doi: 10.1016/j.cell.2016.02.004. Probst (2023) Daniel Probst. Explainable prediction of catalysing enzymes from reactions using multilayer perceptrons. bioRxiv, 2023. doi: 10.1101/2023.01.28.526009. Probst et al. (2022a) Daniel Probst, Matteo Manica, Yves Gaetan Nana Teukam, Alessandro Castrogiovanni, Federico Paratore, and Teodoro Laino. Biocatalysed synthesis planning using data-driven learning. Nature Communications, 13(1):964, 2022a. doi: 10.1038/s41467-022-28536-w. Probst et al. (2022b) Daniel Probst, Philippe Schwaller, and Jean-Louis Reymond. Reaction classification and yield prediction using the differential reaction fingerprint DRFP. Digital Discovery, 1(2):91â97, 2022b. doi: 10.1039/D1D00006C. Qian et al. (2024) Wenjia Qian, Xiaorui Wang, Yu Kang, et al. A general model for predicting enzyme functions based on enzymatic reactions. Journal of Cheminformatics, 16(1):38, 2024. doi: 10.1186/s13321-024-00829-x. Raffel et al. (2020a) Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1â67, 2020a. Raffel et al. (2020b) Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1â67, 2020b. URL https://jmlr.org/papers/v21/20-074.html. Sacha et al. (2021) MikoĆaj Sacha, MikoĆaj BĆaĆŒ, Piotr Byrski, PaweĆ DÄ browski-TumaĆski, MikoĆaj ChromiĆski, RafaĆ Loska, PaweĆ WĆodarczyk-PruszyĆski, and StanisĆaw JastrzÄbski. Molecule edit graph attention network: Modeling chemical reactions as sequences of graph edits. Journal of Chemical Information and Modeling, 61(7):3273â3284, 2021. doi: 10.1021/acs.jcim.1c00537. Schwaller et al. (2019) Philippe Schwaller, Teodoro Laino, ThĂ©ophile Gaudin, Peter Bolgar, Christopher A. Hunter, Costas Bekas, and Alpha A. Lee. Molecular transformer: A model for uncertainty-calibrated chemical reaction prediction. ACS Central Science, 5(9):1572â1583, 2019. doi: 10.1021/acscentsci.9b00576. Schwaller et al. (2021a) Philippe Schwaller, Benjamin Hoover, Jean-Louis Reymond, Hendrik Strobelt, and Teodoro Laino. Extraction of organic chemistry grammar from unsupervised learning of chemical reactions. Science Advances, 7(15):eabe4166, 2021a. doi: 10.1126/sciadv.abe4166. Schwaller et al. (2021b) Philippe Schwaller, Daniel Probst, Alain C. Vaucher, Vishnu H. Nair, David Kreutter, Teodoro Laino, and Jean-Louis Reymond. Mapping the space of chemical reactions using attention-based neural networks. Nature Machine Intelligence, 3(2):144â152, 2021b. doi: 10.1038/s42256-020-00284-w. Sheldon & Woodley (2018) Roger A. Sheldon and John M. Woodley. Role of biocatalysis in sustainable chemistry. Chemical Reviews, 118(2):801â838, 2018. doi: 10.1021/acs.chemrev.7b00203. Tu & Coley (2022) Zhengkai Tu and Connor W. Coley. Permutation invariant graph-to-sequence model for template-free retrosynthesis and reaction prediction. Journal of Chemical Information and Modeling, 62(15):3503â3513, 2022. doi: 10.1021/acs.jcim.2c00321. Turner & OâReilly (2013) Nicholas J. Turner and Elaine OâReilly. Biocatalytic retrosynthesis. Nature Chemical Biology, 9(5):285â288, 2013. doi: 10.1038/nchembio.1235. Wan et al. (2022) Yue Wan, Chang-Yu Hsieh, Benben Liao, and Shengyu Zhang. Retroformer: Pushing the limits of end-to-end retrosynthesis transformer. In Proceedings of the 39th International Conference on Machine Learning, volume 162, p. 22475â22490. PMLR, 2022. URL https://proceedings.mlr.press/v162/wan22a.html. Wang et al. (2025) Yizhou Wang, Chen Tang, Han Deng, Jiabei Xiao, Jiaqi Liu, Jianyu Wu, Jun Yao, Pengze Li, Encheng Su, Lintao Wang, Guohang Zhuang, Yuchen Ren, Ben Fei, Mingmao Hu, Xin Chen, Dongzhan Zhou, Junjun He, Xiangyu Yue, Zhenfei Yin, Jiamin Wu, Qihao Zheng, Yuhao Zhou, Huihui Xu, Chenglong Ma, Yan Lu, Wenlong Zhang, Chunfeng Song, Philip H. S. Torr, Shixiang Tang, Xinzhu Ma, Wanli Ouyang, and Lei Bai. SciReasoner: Laying the scientific reasoning ground across disciplines. arXiv preprint arXiv:2509.21320, 2025. doi: 10.48550/arXiv.2509.21320. Wen et al. (2022) Y. Wen, Z. Li, X. Xiang, D. Zhang, and S. Kim. Improving machine learning performance on small chemical reaction data with unsupervised contrastive pretraining. Chemical Science, 13(5):1448â1459, 2022. doi: 10.1039/D1SC05469D. URL https://doi.org/10.1039/D1SC05469D. Xiong et al. (2025) Jiacheng Xiong, Wei Zhang, Yinquan Wang, Jiatao Huang, Yuqi Shi, Mingyan Xu, Manjia Li, Zunyun Fu, Xiangtai Kong, Yitian Wang, Zhaoping Xiong, and Mingyue Zheng. Bridging chemistry and artificial intelligence by a reaction description language. Nature Machine Intelligence, 7:782â793, 2025. doi: 10.1038/s42256-025-01032-8. URL https://doi.org/10.1038/s42256-025-01032-8. Yang et al. (2024) Jason Yang, Ariane Mora, Shengchao Liu, Bruce J. Wittmann, Anima Anandkumar, Frances H. Arnold, and Yisong Yue. CARE: A benchmark suite for the classification and retrieval of enzymes. In Advances in Neural Information Processing Systems, volume 37, p. 3094â3121, 2024. doi: 10.52202/079017-0101. URL https://proceedings.neurips.c/paper_files/paper/2024/hash/05a7ad45d75a3082d7a3a70de8743140-Abstract-Datasets_and_Benchmarks_Track.html. Zeng et al. (2025) Zishuo Zeng, Jin Guo, Jiao Jin, and Xiaozhou Luo. CLAIRE: a contrastive learning-based predictor for EC number of chemical reactions. Journal of Cheminformatics, 17:2, 2025. doi: 10.1186/s13321-024-00944-8. Zhang et al. (2025) Kunyang Zhang, Thierry D. Marti, Silke I. Probst, Serina L. Robinson, and Kathrin Fenner. Enzyme association for environmental biotransformation reactions through contrastive learning of reaction center-specific fingerprints. Bioinformatics, 2025. doi: 10.1093/bioinformatics/btag142. Zheng et al. (2022) Shuangjia Zheng, Tao Zeng, Chengtao Li, Binghong Chen, Connor W. Coley, Yuedong Yang, Ruibo Wu, et al. Deep learning driven biosynthetic pathways navigation for natural products with BioNavi-NP. Nature Communications, 13(1):3342, 2022. doi: 10.1038/s41467-022-30970-9. Zhong et al. (2023) Weihe Zhong, Ziduo Yang, and Calvin Yu-Chian Chen. Retrosynthesis prediction using an end-to-end graph generative architecture for molecular graph editing. Nature Communications, 14:3009, 2023. doi: 10.1038/s41467-023-38851-5. Zhong et al. (2022) Zipeng Zhong, Jie Song, Zunlei Feng, Tiantao Liu, Lingxiang Jia, Shaolun Yao, Min Wu, Tingjun Hou, and Mingli Song. Root-aligned SMILES: a tight representation for chemical reaction prediction. Chemical Science, 13(31):9023â9034, 2022. doi: 10.1039/D2SC02763A. Appendix A Benchmark Construction Details This section provides the construction details of VenusRX-Bench. The pipeline consists of four stages: multi-source harmonization, structural processing and deduplication, single-product curation, and leakage-controlled data partitioning. A.1 Multi-source Harmonization We collect enzymatic reactions with EC annotations from BRENDA, PathBank, Rhea, and KEGG. Because these databases differ in record format and reaction representation, all entries are first converted into a common schema containing the source identifier, EC annotation, reaction text, substrates, products, and molecular structures when available. Source-specific representations are normalized before merging. Positive integer stoichiometric coefficients in PathBank are expanded into repeated molecular components, Rhea component separators are converted into the common reaction format, and the balanced reaction-level structures provided by KEGG are retained directly. Reaction text is normalized by standardizing arrow notation and whitespace. EC annotations are cleaned only at the representation level; missing EC numbers are removed rather than inferred or completed. We first deduplicate records using the joint key of normalized reaction text and EC annotation. Thus, records describing the same reaction with different EC annotations remain distinct. Starting from 310,209 source records with valid EC annotations, this procedure yields 89,605 reactionâEC records, corresponding to 80,558 unique reaction texts. A.2 Structural Processing and Deduplication For records represented by compound names, molecular structures are resolved through a shared name-to-SMILES mapping and converted into structured reactions. Reaction balancing is then applied where possible. Records that already provide a balanced reaction-level representation, such as the retained KEGG reactions, are used directly. Atom mapping is generated with RXNMapper. When a balanced reaction candidate is available, mapping is performed on that candidate; otherwise, a structurally valid reaction may be mapped as a fallback. Records for which no usable mapped reaction can be obtained are removed from subsequent construction. A reaction can admit multiple mapped candidates. For candidate mapping m, we define the edit score as SâĄ(m)=nremoveâ(m)+naddâ(m)+natomâ(m),S(m)=n_remove(m)+n_add(m)+n_atom(m), (1) where nremoven_remove and naddn_add denote the numbers of removed and added bonds, respectively, and natomn_atom denotes the number of atoms whose local environments change. Candidates with the minimum score are retained. If several candidates attain the same minimum score, all tied candidates are preserved rather than selecting one arbitrarily. We subsequently canonicalize every molecule while retaining atom-map information, sort molecular components independently on the two sides of the reaction, and reconstruct the normalized mapped reaction. This removes differences caused only by component ordering. Structure-level deduplication is then performed using the EC annotation together with the normalized mapped reaction representation. A.3 Single-product Curation Many biochemical reactions contain multiple products, cofactors, or small by-products and therefore do not directly define a unique target for reaction prediction. Following the single-product curation protocol used by ECReact (Probst et al., 2022a), we construct one principal product for each retained reaction. Specifically, we apply the following rules: 1. remove products that also occur on the reactant side of the same reaction; 2. for reactions containing multiple products, remove known coenzymes and common by-products; 3. remove product molecules containing fewer than four heavy atoms; and 4. discard reactions that retain zero or multiple products, or contain no reactants after curation. The coenzyme and by-product filtering follows the ECReact curation lists, including common phosphate- and nucleotide-related species, NAD-related cofactors, coenzyme A, SAM, flavins, heme, ironâsulfur clusters, acetate, and pyruvate. This filtering is applied only when the original reaction contains multiple products, so that a reaction whose sole product belongs to one of these categories is not removed solely by this rule. We additionally require consistency between the name-level, unmapped-structure, and mapped-structure representations. The mapped and unmapped versions therefore contain exactly the same examples in the same order and differ only in the presence of atom-map indices. After structural processing, product curation, and final single-product deduplication, the benchmark contains 83,439 reaction examples. A.4 Leakage-controlled Data Splits All three benchmark splits are constructed from the same 83,439 single-product reactions. Rather than independently assigning individual rows, we first define canonical structural keys. Atom maps are removed, each molecule is canonicalized, molecular components are sorted, and the two reaction sides are represented as kR k_R =canonâĄ(R), =canon(R), (2) kP k_P =canonâĄ(P), =canon(P), (3) krxn k_rxn =kRâ„>>â„kP. =k_R \| >> \|k_P. (4) We additionally use normalized reaction text and the hierarchical EC annotation as grouping keys. Samples sharing a constrained key are grouped into indivisible connected components. Each component is assigned wholly to the training, validation, or test partition, targeting an approximately 80/10/1080/10/10 ratio. This prevents records connected by the corresponding leakage criterion from appearing in different partitions. We construct three complementary splits: Reactant Split. Samples sharing the same canonical reactant set or normalized reactant text are kept in the same partition. This prevents identical substrate combinations and identical reactant descriptions from crossing partitions. Product Split. In addition to the Reactant Split constraints, samples sharing the same canonical product are kept in the same partition. The test set therefore evaluates prediction on target products that are not present in the training partition. EC Split. Samples are grouped according to their complete four-level EC annotation together with the canonical full reaction. Consequently, an EC class assigned to the validation or test partition does not occur in the training partition, while identical complete reactions are also prevented from crossing partitions. Because complete connected components are assigned jointly, the final partition sizes can deviate slightly from the nominal 80/10/1080/10/10 ratio. The resulting split sizes are shown in Table 6. Table 6: Statistics of the three leakage-controlled benchmark splits. All splits are constructed from the same 83,439 curated reactions. Split Train Validation Test Reactant Split 66,672 8,183 8,584 Product Split 66,751 8,344 8,344 EC Split 66,751 8,344 8,344 A.5 Mapped and Unmapped Representations For each partition, we provide both mapped and unmapped reaction representations. The two versions share the same examples, filtering decisions, and row ordering. The unmapped representation is used by models that operate directly on molecular strings, whereas the mapped representation supports methods that require atom correspondence for reaction-center identification or graph-edit supervision. This paired construction ensures that models requiring different reaction representations are evaluated on exactly the same underlying examples. ââ Appendix B EC Prediction B.1 EC Prediction: Implementation Details Data and label spaces. Reaction SMILES are standardized with RDKit, while preserving one complete reaction per record and retaining all products. The supplied validation partitions are used without resplitting. After label normalization, the training vocabularies contain 7, 67, 230, and 4,074 classes at EC1âEC4 for the Reactant Split, and 7, 67, 231, and 3,943 classes for the product split. Labels appearing only in validation or test data are not added to the prediction candidates. Test examples with unseen labels are retained in the reported evaluation. Pretrained checkpoints. The VenusRX model has 25.1M parameters, hidden size 384, and six layers in each of the encoder and decoder. EC prediction uses the encoder only. We use the final stage-two checkpoints after 80,100 and 72,100 steps for the reaction and product splits, respectively, rather than validation-best checkpoints. Stage-two pretraining uses the corresponding training partition with 20-fold R-SMILES augmentation. The sampling weights for forward prediction, retrosynthesis, and span corruption are 0.35, 0.35, and 0.30, respectively. EC prompting is enabled with probability 0.5, and the EC contrastive objective is disabled. Thus, the downstream zero-shot setting does not imply that EC annotations were absent during pretraining. Reaction representations and zero-shot retrieval. For a reaction r, let xr(v)x_r^(v) be its v-th complete R-SMILES view with the <mlm> prefix. The frozen encoder EΞE_Ξ produces the reaction representation âĄ(r)=110ââv=110MeanPoolâĄ(EΞâ(xr(v))).z(r)= 110 _v=1^10MeanPool\! (E_Ξ(x_r^(v)) ). (5) Pooling first averages token embeddings within a view and then averages the 10 view-level vectors. The same precomputed vectors are used by VenusRX â and VenusRX. The 10-view representation extraction is distinct from the 20-fold augmentation used during pretraining. VenusRX â uses cosine k-nearest-neighbor retrieval with labeled training reactions as its reference database. The value of k is selected on validation data, without using test scores. DRFP and rxnfp also use cosine k-nearest-neighbor retrieval without downstream parameter training, with k selected on validation data. crxnfp uses its released encoder weights without downstream training and performs 1-nearest-neighbor retrieval against the training set. CLAIRE and VenusRX. CLAIRE uses its original concatenated DRFP and rxnfp features, with DRFP folded to 256 dimensions. VenusRX replaces these inputs with âĄ(r)z(r) while keeping the VenusRX encoder frozen. The original CLAIRE implementation is reused for projection training; only feature construction is replaced. Both methods use the same threefold permutation-augmented training set and train a separate projection network for each EC level. The five-layer projection network has hidden width 1,280, output dimension 128, LayerNorm, and dropout 0.1. Training uses a learning rate of 5Ă10â45Ă 10^-4, batch size 6,000, and triplet margin 1.0. Both methods train for a fixed 2,000 epochs and select the checkpoint with the lowest training loss within the final 20% of epochs. This checkpoint-selection rule does not use validation scores or validation-based early stopping. Retrieval uses Euclidean distances in the learned projection space. Other trained baselines. Theiaâs classifier and CLAIREâs projection network are randomly initialized and trained on the corresponding training partition, rather than initialized from their released task-trained weights. Theia uses 10,240-bit DRFP features with radius 2 and a single hidden layer of 1,664 units with Tanh activation. It is trained with Adam, learning rate 10â310^-3, batch size 64, and an exponential learning-rate decay factor of 0.9. Multi-label training records are expanded for this baseline. Its official stopping rule requires the current validation loss to improve by at least 10â310^-3 relative to the mean of the preceding five epochs, with a maximum of 50 epochs. BEC-Pred is initialized from the official rxnfp bert_pretrained weights and trained for 48 epochs, using batch size 8, learning rate 10â510^-5, maximum sequence length 512, and no warmup. Where paper descriptions and released code differ, we follow the official code configurations. CLAIRE is additionally evaluated at EC4, and BEC-Pred at EC1, EC2, and EC4, by adapting the label space while keeping their training configurations unchanged. BEC-Pred retains its raw-string label indexing; for reaction-split EC4, this produces 4,081 output classes before normalization, compared with 4,074 normalized benchmark labels. Evaluation and strict-subset analysis. All methods are scored with the same score_ec.py implementation and scoring parameters. Let i,âY_i, be the set of reference labels for reaction i at EC level â , and y^i,â y_i, the top-ranked prediction. The hit-any top-1 accuracy is Accâ=1Nââi=1Nâ[y^i,ââi,â],Acc_ = 1N_ _i=1^N_ 1\! [ y_i, _i, ], (6) where NâN_ is the number of records evaluated at level â . Weighted F1 is the primary metric, with macro F1 and top-1 accuracy also reported. The supplementary strict-subset scores exclude 278 reaction-split test records and 18 product-split test records flagged as input-structure-equivalent to training reactions. These records remain in the main test evaluation. Strict-subset filtering uses the test_leaked_ids_input_only flags in leakage.json. Theia, CLAIRE, BEC-Pred, and VenusRX are run with seeds 13, 21, and 42, and their results are summarized by the mean and standard deviation. crxnfp, DRFP, rxnfp, and VenusRX â are evaluated as deterministic methods and have no across-seed standard deviation. Appendix C VenusRX Implementation Details C.1 Architecture and Reaction Representation Architecture and initialization. The T5-style model shares input embeddings across the encoder and decoder and ties the output embedding to them. Table 7 lists the configurations used for reaction-split scaling; Tiny is used for cross-split comparisons. All configurations use ReLU feed-forward layers, dropout 0.10.1, 32 relative-attention buckets, a maximum relative-distance parameter of 128, and layer-normalization epsilon 10â610^-6. The relative-distance parameter is not a maximum sequence length. The decoder starts with padding token ID 0; EOS has ID 1. Table 7: VenusRX configurations. Parameter counts are recorded wrapper totals, including the inactive projection head. Model Enc./Dec. dmodeld_model dffd_f Heads dkvd_kv Parameters Micro 4/4 256 1,024 4 64 7,554,816 Compact 6/6 320 1,280 5 64 17,495,296 Tiny 6/6 384 1,536 6 64 25,147,648 Medium 6/6 768 3,072 12 64 100,135,168 The wrapper also instantiates a two-layer projection head, dmodelâdmodelâ256d_modelâ d_modelâ 256, with a tanh activation. Its contrastive-loss coefficient is zero, so it does not contribute to the sequence-training objective. For Tiny, this head contains 246,400 parameters and the T5 backbone contains 24,901,248 parameters. No enzyme-sequence encoder or additional EC-prediction loss is used in this training protocol. Tokenizer and length limits. Both stages and standard inference load the same fixed 302-token vocabulary. A custom regular expression identifies SMILES tokens, including bracketed atoms, multi-character elements, bonds, branches, ring indices, stereochemical symbols, and >>. If the expression does not cover an entire string, tokenization falls back to individual characters; unseen tokens map to <unk>. Input whitespace is removed. Maximum source and target lengths are 1,200 and 768 tokens, including special tokens and EOS. Truncation forces the last retained token to EOS. Batches are dynamically right-padded; source padding is attention-masked and target padding is assigned label â100-100 and excluded from the loss. Molecules, cofactors, and stereochemistry. Molecules on the same side are separated by ., and reaction sides by >>. An original R>agents>P record contributes its reactant and product sides; the middle agents field is not modeled. The reactant side retains all components present in the prepared view, including recorded cofactors. Forward targets and retrosynthesis inputs use one principal product, whereas span corruption retains the full product side. Principal-product selection first prefers components not recognized as common cofactors and then maximizes a SMILES-based size heuristic. The heuristic counts uppercase non-hydrogen letters rather than computing an exact heavy-atom count. If every component is identified as a cofactor, the highest-scoring component is retained; ties follow input order. Cofactor matching uses an explicit list and canonical or uncharged forms. Atom-map identifiers are removed from model inputs and targets. Existing chirality and bond-direction symbols are retained; this does not imply complete stereochemical annotation of every record. Forward and retrosynthesis use their respective aligned views and need not be literal string inverses. Task serialization. Table 8 gives the with-EC formats before EOS is appended. Removing the complete <ec_number>e suffix gives the without-EC format. Here, C replaces masked spans with sentinel tokens while protecting special tokens and the EC suffix. Table 8: Task-specific source and target formats. EOS is appended to both sequences. Task Source Target Forward <forward>R>><mask_product><ec_number>e pâp Retro <retro>pâp <ec_number>e R Span âĄ(CLOSE C(<mlm>R>>P)<ec_number>e Masked spans C.2 Task Sampling and Span Corruption Sampling and loss reduction. Training samples records randomly, with distributed sharding across devices. For each presentation, the collator samples one available view uniformly by list position and one task with probabilities (0.35,0.35,0.30)(0.35,0.35,0.30). The task, view, EC inclusion, and corruption mask can change between presentations. Duplicate stored views are retained, so repeated strings can have greater sampling probability. The training loss is computed for each device-local batch; distributed training averages the resulting gradients across devices. This is not necessarily a globally token-weighted average when devices have different target-token counts. EOS and sentinel tokens contribute to the loss, and longer targets contribute more supervised positions. No additional task-loss weights, label smoothing, or EC contrastive loss are applied. Corruption procedure. The span task uses a forward R-SMILES view of the full reaction, with an original-reaction fallback when no valid view is available. After tokenization, the source is initially restricted to 1,199 tokens. Special tokens and all tokens from <ec_number> onward are protected; molecular tokens, ., and >> remain maskable. For M>0M>0 maskable positions, the target number of masked tokens is K=maxâĄ(1,â0.15âMâ).K= \! (1, 0.15M ). (7) Span lengths are sampled as â=maxâĄ(1,roundâĄ(z)),zâŒExponentialâĄ(rate=1/3). = \! (1,round(z) ), z (rate=1/3). (8) Thus, 3 is the scale of the underlying exponential distribution, not a measured mean length of the final merged spans. Sampling extends rightward from a selected start and may revisit previously selected positions. It stops when the token budget is reached, no eligible positions remain, or the number of attempts reaches ten times the sequence length. Adjacent masked positions are merged, ordered by their original positions, and restricted to at most 100 spans. Each retained span is replaced with <extra_id_i> in the source. The target concatenates each sentinel and its missing span in source order, followed by EOS. No additional terminal sentinel is appended after the final span, and the target does not reconstruct the full reaction. An empty maskable set yields an EOS-only target. The final source and target obey the 1,200/768-token limits. Reaction-center-dependent start weights require atom-map identifiers in the selected view. The bundled R-SMILES generator removes these identifiers, so views produced through that path use base weights. The configured center/neighbor/base weights of 4/2/1 do not multiply the decoder loss. C.3 EC Serialization and Label Handling When available, a label is sampled uniformly from the cleaned, deduplicated, sorted ec4 list; otherwise, selection falls back to the recordâs ec_number field. The chosen string is appended as <ec_number>e before source EOS with probability 0.50.5. The tokenizer represents EC digits and periods using its existing tokens, rather than assigning one token to each EC class. Missing annotations always yield unconditioned inputs, so the corpus-wide conditioned fraction need not equal 50%50\%. Because the suffix is appended before truncation, it may be partially or entirely removed from an overlength source. The implementation does not impose a complete numerical-validity check on all four-part labels; strings containing an unspecified component such as 1.1.1.- are not categorically excluded. Stage-two preparation deduplicates canonical substrate/product pairs without EC in the key and retains the first record, rather than necessarily preserving every EC annotation for that pair. In grouped retrosynthesis inference, the first sorted available EC annotation is selected. Inference removes an optional EC prefix and uses the first four components; labels with fewer than four components do not produce a suffix. The conditions come from benchmark annotations, not from a separately predicted EC label. C.4 Pretraining Corpus and Split-Specific Filtering Common preprocessing. The available pipeline starts from merged upstream reaction records and cluster assignments. It selects the designated cluster-center records, applies RXNMapper (Schwaller et al., 2021a), and constructs up to 20 task-specific R-SMILES views with seed 13 and mixed forward mode. Mapping checks examine the presence of atom maps, uniqueness of map identifiers within each side, and agreement between the two sidesâ map-ID sets. These checks are not a complete elemental or charge-balance test. Records with mapping or R-SMILES failures are retained and can fall back to their original reaction representation. One stored view contains four task-specific strings; these are not four distinct reactions. The common pre-filtering corpus contains 4,168,335 records. Structural overlap detection. For each benchmark split, comparisons remove atom maps and sort components within each reaction side. Five matching levels are used: canonical isomeric SMILES; canonical SMILES without stereochemistry; uncharged SMILES without stereochemistry; the preceding form after removal of small inorganic components; and that form with repeated components collapsed to a set. The small-component rule removes molecules with at most one heavy atom, or carbon-free molecules with at most five heavy atoms. Matching considers both reaction directions and supports matching a complete reactant side with an individual product. The last two levels additionally require matching EC annotations when the pretraining record has an EC annotation; otherwise, matching is structural only. Uncharging failures retain the original molecule, and there is no explicit tautomer-canonicalization level. These are overlap-detection rules, not the upstream definition of a unique reaction. Template-associated exclusion. Let TtrainT_train, TvalT_val, and TtestT_test be the template-ID sets collected from upstream records that match the corresponding benchmark partitions. Define E=(TvalâȘTtest)âTtrain.E=(T_valâȘ T_test) T_train. (9) We remove records that either directly match a validation/test reaction or carry any template ID in E. A scalar template field is used when the template-ID list is unavailable. Templates shared with training matches are not removed merely because they also occur in held-out matches. This is post-generation filtering based on matched-record provenance, not training-only template extraction. Table 9: Split-specific stage-one corpora. Stored views include duplicates and exclude dynamically constructed fallback views. Records are not audited counts of unique synthetic reactions. Split Retained records 20-view records Zero-view records Stored views Reaction 4,044,455 3,660,831 383,624 73,216,620 EC 3,853,604 3,471,243 382,361 69,424,860 Product 4,047,718 3,665,111 382,607 73,302,220 Audit scope and counting. The final audits report zero validation and test reaction overlaps under the specified matching procedure for all three splits. Validation and test structures are used for exclusion and auditing, not for standard model gradient updates. This does not establish that the upstream substrate pool excludes held-out molecules, or that all pretraining products and scaffolds are disjoint from the benchmark held-out sets. The reaction-split filter removes 123,880 records: 9,398 direct reaction matches and 114,482 additional template-associated records. Table 9 distinguishes retained records from stored views. Unique real/synthetic reaction counts and their intersection remain unaudited; no separate real/synthetic sampling ratio is configured. C.5 Optimization and Checkpoint Selection Table 10 reports the standard Tiny training settings. Stage two loads the selected stage-one weights but resets optimizer, scheduler, step counter, and early-stopping state. The training seed is 13; internal-validation and fixed benchmark-validation sampling seeds are 10013 and 20013, respectively. Stage-two preprocessing deduplicates canonical substrate/product pairs before an internal training/validation subdivision. These prepared-record counts are not substituted for the benchmark scorerâs original test-row denominators. Table 10: Standard Tiny optimization settings. Step limits are maximum budgets, not actual training durations. Setting Stage one Stage two Initialization Random Stage-one best weights Optimizer AdamW AdamW Adam betas (0.9,0.999)(0.9,0.999) (0.9,0.999)(0.9,0.999) Weight decay 0 0 Peak learning rate 2Ă10â42Ă 10^-4 5Ă10â55Ă 10^-5 Warmup steps 1,000 1,000 Schedule Cosine decay to zero Cosine decay to zero Gradient clipping Norm 1.0 Norm 1.0 Per-device batch size 32 32 Devices / accumulation 8 / 1 8 / 1 Nominal global batch 256 records 256 records Precision BF16 autocast BF16 autocast Maximum steps 700,000 100,000 Validation interval 16,000 steps 1,000 steps Early-stopping patience 5 validation checks 5 validation checks Internal validation. The historical training configuration groups records connected by shared template IDs and assigns groups to training or validation using a seeded stable hash. Records without template IDs use a reaction-key fallback. Hash thresholds are 0.010.01 in stage one and 0.050.05 in stage two; these are not guarantees of exact validation-record fractions. Validation uses batch size 32 and a cap of 200 batches per validation loader. Reported loss averages batch-level scalar losses rather than computing one corpus-wide token-weighted mean. Distributed sampler padding can increase the reported sample count above the number of distinct records. Checkpoint selection. For the standard Tiny runs, checkpoint selection uses internal val_loss; the fixed benchmark validation set is monitored separately as stage2_val_loss in stage two. For example, the Tiny reaction-split checkpoint selected at step 75,000 records internal loss 0.0986220.098622 and fixed benchmark validation loss 0.1169060.116906. These are different metrics and must not be interchanged. Table 11 distinguishes selected and final training steps. The no-span ablation samples forward/retro tasks with probabilities 0.5/0.50.5/0.5; the stage-two-only ablation starts from random initialization without stage-one training. These are single-seed experiments with different realized training budgets, not strictly compute-matched ablations. Table 11: Selected checkpoint step / final training step for each run. Run Stage one Stage two Micro, reaction 544,000 / 630,080 86,000 / 91,136 Compact, reaction 400,000 / 488,312 72,000 / 77,074 Tiny, reaction 288,000 / 378,048 75,000 / 80,100 Medium, reaction 240,000 / 330,792 38,000 / 43,076 Tiny, EC 336,000 / 420,196 72,000 / 77,000 Tiny, product 176,000 / 268,022 67,000 / 72,100 Tiny, no span corruption 176,000 / 267,784 47,000 / 52,154 Tiny, stage two only Not applicable 93,000 / 98,078 C.6 Inference and Separately Reported Diagnostics Direct inference. Forward and retrosynthesis evaluation use the selected stage-two checkpoint with and without supplied EC annotations. The configuration requests 20 R-SMILES views, mixed mode, seed 13, beam size 10, and 10 returned candidates per view, retaining at most 10 final candidates. Source/target limits remain 1,200/768 tokens and inference batch size is 2. The candidate count is not a top-k sampling parameter. There is no additional task-specific fine-tuning after stage two; this is not zero-shot evaluation with respect to benchmark training data. Retrosynthesis fusion. Candidates are canonicalized and deduplicated within each view, so a structure contributes at most once per view. For a candidate a, let âĄ(a)V(a) contain views that generate it, and let rjâ(a)r_j(a) be its best rank in view j. Define bâĄ(a)=minjââĄ(a)âĄrjâ(a),qâĄ(a)=âjââĄ(a)1rjâ(a).b(a)= _j (a)r_j(a), q(a)= _j (a) 1r_j(a). (10) The implementation sorts by decreasing sfusionâ(a)=qâĄ(a)â108â(bâĄ(a)â1).s_fusion(a)=q(a)-10^8 (b(a)-1 ). (11) This prioritizes the best rank obtained in any view, then the reciprocal-rank sum; remaining ties retain first-occurrence order. It combines representations of one model, not independent models or averaged sequence log probabilities. Molecule-library constraints. The constrained-decoding diagnostic applies token-level constraints from a split-specific molecule library containing train, validation, and test R-SMILES molecules, preserving their SMILES strings. It measures performance with access to a benchmark-wide molecular vocabulary and is kept separate from unrestricted held-out evaluation. Library membership does not establish reaction feasibility or enzyme activity, and this setting is not a guarantee of an accuracy upper bound. Forward rescoring. Forward rescoring uses the forward task of the same Tiny stage-two checkpoint, with length-normalized teacher-forced scoring of the target product given a candidate reactant mixture. It does not use a separately trained forward model. Two diagnostics are distinguished: the earlier raw-score setting with λ=0.5λ=0.5, and a rank-blend setting with sblendâ(a)=â(rfusionâ(a)â1)âλâĄ(rforwardâ(a)â1),s_blend(a)=- (r_fusion(a)-1 )-λ (r_forward(a)-1 ), (12) where larger scores rank first and the fixed main rank-blend setting is λ=0.25λ=0.25. The test-set sweep over λ is a sensitivity diagnostic, not validation-based parameter selection. Neither rescoring variant is assumed to improve every condition. Appendix D Synthetic Enzymatic Reaction Generation Stage-1 pretraining uses a large synthetic reaction corpus constructed with a template-based generation framework following RSGPT (Deng et al., 2025b). The overall pipeline is enzymatic reactionsâreaction templates,molecular collectionâBRICS fragments,templates+fragmentsâcandidate reactionsâfiltered synthetic corpus. splitenzymatic reactions& templates,\\ molecular collection& fragments,\\ templates+fragments& reactions synthetic corpus. split (13) The following sections describe each stage in detail. D.1 Reaction Template Construction A reaction template represents a transferable local transformation rather than a complete reaction instance. Starting from an atom-mapped reaction, we compare the reactant and product sides to identify atoms and bonds whose local environments change. The transformation together with the required neighboring structural context defines a reaction template. Conceptually, a template T with K reactant patterns can be written as T:(Q1,Q2,âŠ,QK)âQP,T: (Q_1,Q_2,âŠ,Q_K )â Q_P, (14) where QiQ_i denotes a reactant-side substructure query and QPQ_P encodes the corresponding product-side transformation. The query patterns preserve the local chemical constraints required for template application, including relevant atom and bond properties. Templates are normalized and duplicate transformation rules are merged while retaining their provenance to the source enzymatic reactions. Each template additionally retains its association with the catalytic annotations of its source reactions when available. Such an inherited EC annotation describes the reaction category from which the template was obtained. It should not be interpreted as experimental evidence that a particular enzyme catalyzes every synthetic reaction generated from that template. D.2 Candidate Substrate Pool To expose each transformation to a larger molecular context, we construct a candidate substrate pool following the fragment-based strategy used in RSGPT (Deng et al., 2025b). Molecules from the upstream molecular collection are decomposed using the BRICS fragmentation algorithm (Degen et al., 2008). BRICS identifies chemically meaningful bonds that can be disconnected to produce smaller molecular fragments. After fragmentation, attachment points are converted into chemically valid terminated structures, yielding independently parseable fragments. The resulting structures are standardized and deduplicated before template matching. Let â±=f1,f2,âŠ,fNF=\f_1,f_2,âŠ,f_N\ (15) denote the resulting fragment pool. These fragments provide candidate substrate contexts; BRICS connection points themselves do not define the enzymatic reaction center. D.3 TemplateâFragment Matching For every reactant pattern QiQ_i of template T, we identify fragments whose molecular structures satisfy the corresponding substructure query. We define the candidate set â±âĄ(T,i)=fââ±âŁfâ§Qi,F(T,i)= \f f Q_i \, (16) where fâ§Qif Q_i indicates that fragment f contains a valid match to the templateâs i-th reactant pattern. For a template containing K reactant components, the candidate combination space is therefore (T)=â±(T,1)Ăâ±(T,2)ĂâŻĂâ±(T,K).C(T)=F(T,1)ĂF(T,2)Ă·sĂF(T,K). (17) This combinatorial construction is the main source of data scaling: a single experimentally observed transformation can be instantiated with many compatible substrate combinations. Importantly, matching is based on the structural constraints encoded by the reaction template rather than molecular similarity alone. D.4 Template-Based Reaction Generation For each compatible reactant combination =(f1,âŠ,fK)ââĄ(T),f=(f_1,âŠ,f_K) (T), (18) we apply template T in the forward direction to obtain one or more candidate products: Tâ()=TâĄ(f1,âŠ,fK).P_T(f)=T(f_1,âŠ,f_K). (19) Each successful template application produces a candidate reaction f1.f2.âŻ.fKâ¶p,pâTâ().f_1.f_2.·s.f_K\; \;p, p _T(f). (20) A reactant combination may fail to produce a valid product, and a single combination may yield several products because of multiple matching sites. Such outcomes are handled before the reaction is admitted to the pretraining corpus. For traceability, generated reactions retain their template provenance and, where available, the catalytic annotation inherited from the source enzymatic transformation. D.5 Reaction Filtering Template application enforces a local reaction rule, but it does not guarantee that every generated instance represents a plausible biochemical reaction. We therefore apply several filtering stages before pretraining. Structural validity. Reactants and products must be parseable molecular structures and the template application must produce a non-empty product. Invalid molecular structures and failed template applications are discarded. Reaction consistency. We remove degenerate transformations and normalize component ordering and molecular representations before reaction-level comparison. Generated reactions are deduplicated after normalization while retaining template and EC provenance when multiple templates produce the same transformation. Biochemical and thermodynamic filtering. We further filter the generated candidates using the biochemical constraints and reaction-energy information available in our upstream processing pipeline. This step removes candidates that are inconsistent with the target enzymatic reaction space and prioritizes thermodynamically plausible transformations. Benchmark isolation. Finally, synthetic pretraining reactions are filtered independently for each VenusRX-Benchsplit. We remove direct overlaps with held-out reactions and additionally exclude records associated with templates that occur exclusively in held-out matches. Let TtrainT_train, TvalT_val, and TtestT_test denote template sets associated with reactions matched to each benchmark partition. The excluded held-out template set is Theld=(TvalâȘTtest)âTtrain.T_held= (T_valâȘ T_test ) T_train. (21) Synthetic records directly matching validation or test reactions, or associated with a template in TheldT_held, are removed from the corresponding pretraining corpus. D.6 From Synthetic Reactions to Model Inputs After generation and filtering, the retained synthetic reactions enter the same representation-processing pipeline used by VenusRX. We first obtain atom mappings with RXNMapper when a valid mapping is not already available (Schwaller et al., 2021a). We then construct up to 20 root-aligned SMILES views for each reaction using R-SMILES (Zhong et al., 2022). Forward and retrosynthetic views are generated according to their respective prediction directions and stored together for downstream multitask training. This stage should be distinguished from synthetic reaction generation. Template application increases the number of distinct reaction instances, |reaction|,|D_reaction|, (22) whereas R-SMILES increases the number of sequence realizations of each instance, râ¶sr(1),âŠ,sr(M),Mâ€20.r \s_r^(1),âŠ,s_r^(M)\, M†20. (23) Therefore, the approximately four million synthetic reactions used for Stage-1 pretraining refer to reaction instances before counting their alternative R-SMILES representations. D.7 Interpretation of Synthetic Reactions The generated corpus is intended as large-scale pretraining data rather than a database of experimentally verified enzymatic reactions. Each generated example satisfies an extracted local transformation rule and the implemented filtering criteria, but template compatibility alone does not establish experimental feasibility. Similarly, an EC label inherited from a template identifies the catalytic category associated with the source transformation. It does not establish that a specific enzyme carrying that EC annotation accepts the newly generated substrates. The synthetic corpus is therefore used to expose VenusRX to a broader enzymatic reaction space before the model is adapted to experimentally observed VenusRX-Benchreactions during Stage-2 continued training. Appendix E VenusRX-Tiny results without constrained generation Table 12: VenusRX-Tiny results without constrained generation on the three data splits. We report Top-k exact-match accuracy (%). These results use the standard unconstrained decoding protocol. Task EC Input Reactant Split Product Split EC Split T1 T3 T5 T10 T1 T3 T5 T10 T1 T3 T5 T10 Forward Without EC 73.066 84.611 88.048 90.739 50.959 66.275 71.393 77.085 31.100 45.170 53.236 59.947 With EC 75.221 86.778 89.725 92.474 55.825 69.463 74.041 78.811 34.564 50.240 56.340 64.058 Retrosynthesis Without EC 8.446 17.183 23.777 30.603 20.554 30.058 34.923 39.945 4.878 8.737 10.487 14.274 With EC 17.230 26.398 29.345 34.425 28.847 34.696 38.267 40.568 18.360 24.605 27.481 30.872 Appendix F Additional Experimental Results F.1 Retrosynthesis Table 13: Round-trip coverage on three data splits. We report RT Coverage@k (%), defined as the percentage of test reactions for which at least one of the top-k retrosynthesis candidates can be mapped back to the target product by the forward oracle. Methods marked with â are zero-shot methods and are evaluated directly using open-source checkpoints. Methods are grouped by whether EC information is provided to the retrosynthesis model. Within each group and column, the best result is bold and the second-best result is underlined. Method Reactant Split Product Split EC Split C@1 C@3 C@5 C@10 C@1 C@3 C@5 C@10 C@1 C@3 C@5 C@10 Zero-shot SciReasonerâ (LLM) 31.163 45.853 52.260 61.125 20.206 30.489 36.266 44.487 17.953 29.782 36.937 46.045 RSGPTâ 14.247 25.524 29.555 34.098 8.881 14.861 16.982 19.427 8.198 17.797 21.201 24.904 R-SMILESâ 3.856 9.879 13.898 19.385 2.721 6.807 10.139 14.849 4.830 8.401 12.548 19.691 Without EC Input Graph2SMILES 38.269 52.551 58.504 63.781 28.667 34.552 37.140 42.174 25.443 34.744 38.698 43.313 Retroformer 24.814 34.098 41.344 51.083 10.331 14.669 17.965 22.651 15.292 21.453 25.791 31.771 MEGAN 37.710 53.763 60.252 66.577 30.513 44.451 50.803 57.083 21.213 35.031 41.587 49.269 R-SMILES 29.928 42.661 47.029 52.575 15.832 28.164 35.031 42.581 10.930 22.016 28.092 37.452 BioNavi 39.108 52.365 58.388 64.189 11.397 17.414 20.686 24.724 20.973 29.279 33.030 38.219 G2Retro 31.139 51.549 58.854 67.754 27.205 40.448 49.029 58.293 22.699 39.070 46.069 56.843 RSGPT 22.623 42.428 49.219 60.217 12.021 21.908 26.282 33.869 13.627 24.065 30.717 39.250 VenusRX 43.383 60.648 66.787 75.128 41.107 54.602 60.954 69.607 28.883 48.394 55.081 64.753 VenusRX (MLCD) 43.057 58.586 63.269 70.282 38.255 49.940 56.520 62.105 28.655 47.711 53.152 60.990 With EC Input ECReact 44.851 60.135 65.284 70.701 28.152 36.601 41.407 47.291 32.862 43.349 48.370 55.633 VenusRX 53.308 66.880 72.064 76.887 47.915 56.544 63.015 69.307 42.006 53.428 58.221 65.964 VenusRX (MLCD) 50.233 63.723 67.719 74.348 43.996 53.248 57.251 62.668 38.231 50.096 53.955 60.822 Table 14: Round-trip accuracy on three data splits. We report RT Accuracy@k (%), defined as the fraction of the first k retrosynthesis candidates that can be mapped back to the target product by the forward oracle. Methods marked with â are zero-shot methods and are evaluated directly using open-source checkpoints. Methods are grouped by whether EC information is provided to the retrosynthesis model. Within each group and column, the best result is bold and the second-best result is underlined. Method Reactant Split Product Split EC Split A@1 A@3 A@5 A@10 A@1 A@3 A@5 A@10 A@1 A@3 A@5 A@10 Zero-shot SciReasonerâ (LLM) 31.163 24.282 20.503 17.672 20.206 14.533 12.081 9.340 17.953 13.774 12.157 9.939 RSGPTâ 14.247 12.950 10.848 7.780 8.881 6.935 5.297 3.496 8.198 7.674 6.764 4.715 R-SMILESâ 3.856 4.042 4.268 3.929 2.721 2.621 2.812 2.641 4.830 3.356 3.490 3.476 Without EC Input Graph2SMILES 38.269 33.652 30.594 25.495 28.667 22.064 18.502 14.403 25.443 22.108 19.578 15.511 Retroformer 24.814 22.930 22.260 21.553 10.331 9.416 8.758 7.296 15.292 14.446 13.320 11.956 MEGAN 37.710 35.089 32.968 29.065 30.513 25.559 23.418 20.150 21.213 19.451 18.068 15.690 R-SMILES 29.928 26.130 24.117 21.262 15.832 12.568 11.644 9.950 10.930 10.067 9.267 8.201 BioNavi 39.108 35.702 33.979 30.325 11.397 9.368 8.205 6.510 20.973 18.333 16.285 13.424 G2Retro 31.139 27.512 24.497 21.454 27.205 20.877 19.456 16.419 22.699 20.254 17.543 14.880 RSGPT 22.623 20.857 20.315 18.113 12.021 9.907 8.411 6.874 13.627 12.672 11.663 9.884 VenusRX 43.383 37.508 34.434 30.494 41.107 31.516 28.346 25.345 28.883 25.803 23.027 20.152 VenusRX (MLCD) 43.057 35.038 31.379 27.340 38.255 27.744 24.295 18.948 28.655 24.549 21.026 16.710 With EC Input ECReact 44.851 42.280 39.814 35.216 28.152 27.201 26.515 24.895 32.862 32.882 31.853 30.128 VenusRX 53.308 43.480 38.844 34.287 47.915 36.326 32.728 27.096 42.006 31.819 28.032 23.688 VenusRX (MLCD) 50.233 39.418 33.684 28.879 43.996 30.661 25.686 20.147 38.231 26.127 22.090 17.338 F.2 EC prediction Table 15: EC-number prediction on the Reactant and Product splits. Macro F1 (%) is reported at EC levels 1â4. Method Reactant Split Product Split EC1 EC2 EC3 EC4 EC1 EC2 EC3 EC4 Zero-Shot crxnfp (Zhang et al., 2025) 86.64ÂŻ 86.64 71.1771.17 57.2857.28 27.3527.35 68.4568.45 55.8555.85 46.1246.12 21.1221.12 DRFP (Probst et al., 2022b) 85.7585.75 72.80ÂŻ 72.80 57.53ÂŻ 57.53 31.94ÂŻ 31.94 71.11ÂŻ 71.11 60.03ÂŻ 60.03 51.23ÂŻ 51.23 26.15ÂŻ 26.15 rxnfp (Schwaller et al., 2021b) 83.7183.71 70.5670.56 56.3656.36 28.4628.46 66.9066.90 54.3254.32 45.6645.66 21.1121.11 VenusRXâ 87.3887.38 73.0573.05 61.0561.05 33.3133.31 73.4673.46 61.1461.14 51.9651.96 26.9026.90 Task Specific Model Theia (Probst, 2023) 83.72±1.7083.72± 1.70 71.34±0.9571.34± 0.95 62.15±1.7562.15± 1.75 36.29±0.5336.29± 0.53 73.72±2.0373.72± 2.03 63.79±1.2363.79± 1.23 52.92¯±1.07 52.92± 1.07 27.81±0.0227.81± 0.02 CLAIRE (Zeng et al., 2025) 89.80±0.6689.80± 0.66 72.32±2.9372.32± 2.93 59.51±1.5759.51± 1.57 31.74±0.7431.74± 0.74 74.50±0.2274.50± 0.22 57.82±0.6857.82± 0.68 49.99±0.8249.99± 0.82 23.96±0.5623.96± 0.56 BEC-Pred (Qian et al., 2024) 85.35±6.2585.35± 6.25 61.58±1.1061.58± 1.10 42.64±0.5942.64± 0.59 28.13±0.4328.13± 0.43 76.70±1.4976.70± 1.49 60.23±1.0360.23± 1.03 43.30±0.8943.30± 0.89 21.18±0.2721.18± 0.27 VenusRX 88.35¯±1.54 88.35± 1.54 71.65¯±0.64 71.65± 0.64 62.10¯±0.63 62.10± 0.63 33.93¯±0.93 33.93± 0.93 76.01¯±0.69 76.01± 0.69 62.12¯±1.88 62.12± 1.88 54.95±0.5754.95± 0.57 26.09¯±0.21 26.09± 0.21 VenusRX denotes frozen VenusRX representations followed by a task-specific predictor. Task-specific results are mean ± std over three seeds. Appendix G Details of Molecule Library-Constrained Decoding G.1 implementation MLCD constrains autoregressive decoding with a prefix tree constructed from the tokenized molecular representations in the predefined molecule library. Each trie node corresponds to a molecular token prefix, and its outgoing edges specify the valid next tokens. A molecule can terminate only when the current prefix reaches a complete library entry. For multi-molecule outputs, the trie is reset after each molecule separator so that the constraint is applied independently to every generated molecule. Reaction-level control tokens and separators are generated normally. MLCD therefore changes only the set of admissible molecular tokens; sequence probabilities and beam ranking remain determined by VenusRX. Algorithm 1 summarizes the decoding procedure. Algorithm 1 Molecule Library-Constrained Decoding (MLCD) 1: Input x, model pΞp_Ξ, molecule trie T, beam size B 2: Ranked output sequences 3: Initialize beam set âŹB with the start token 4: while not all beams are finished do 5: âŹnewââ B_newâ 6: for each partial sequence yââŹy do 7: if y is currently generating a molecule then 8: hâhâ current molecular prefix 9: âAâ valid next tokens from T given h 10: else 11: âAâ valid reaction-level tokens 12: end if 13: Expand y using tokens in A 14: Score expansions with pΞ(â âŁx,y)p_Ξ(· x,y) 15: end for 16: Keep the top-B scored expansions 17: Update âŹB 18: end while 19: return completed sequences ranked by model score G.2 Analysis of Molecule Library-Constrained Decoding We provide additional analysis of Molecule Library-Constrained Decoding (MLCD) for retrosynthesis. Generation validity. We report the raw invalid rate, defined as the percentage of generated molecular candidates that cannot be parsed as valid SMILES before final filtering. Missing Top-k slots are not counted as invalid molecules. Table 16: Raw invalid SMILES rate (%) for retrosynthesis with and without MLCD. Lower is better. Split EC Unconstrained MLCD Reactant â 2.277 0.000 â 1.599 0.000 Product â 3.641 0.000 â 2.756 0.000 EC â 3.086 0.000 â 2.210 0.000 As shown in Table 16, MLCD eliminates invalid raw molecular candidates in all evaluated retrosynthesis settings. This supports its role in constraining generation to structurally valid molecular outputs. Round-trip evaluation. For each retrosynthesis candidate, we apply a fixed forward oracle and test whether the predicted product matches the target product. We report RT Accuracy@k, the fraction of the first k candidates that map back to the target, and RT Coverage@k, the fraction of test reactions for which at least one of the first k candidates maps back to the target. Tables 14 and 13 show that unconstrained decoding generally achieves higher round-trip scores than MLCD. Combined with the higher exact-match accuracy of MLCD, this indicates that constrained decoding improves recovery of recorded benchmark targets while reducing the diversity of alternative candidates retained by unconstrained decoding.