PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2610.11694v1q-bio.QMcs.AI

TL;DR

Enzymatic reaction modeling needs to account for catalytic function alongside molecular structure, but available data and evaluation procedures are limited and heterogeneous. The paper introduces VenusRX-Bench and a unified pretrained model with optional EC conditioning and constrained decoding. VenusRX performs best or competitively on most evaluated settings, and the paper reports that reaction representations also encode catalytic information.

  • Problem

    Enzymatic reaction modeling lacks large, high-quality datasets and comparable benchmarks, while reaction outcomes depend jointly on molecular structure and catalytic function.

    Source: §1 · p. 2

  • Method

    VenusRX-Bench standardizes tasks and splits, while VenusRX uses shared text-to-text objectives, two-stage training, optional EC conditioning, and molecule-library constrained decoding.

    Source: §1 · p. 2, §3 · p. 5, §3.1 · p. 5

  • Results

    VenusRX achieves the best or competitive performance on most evaluated settings, and experiments indicate that reaction representations encode useful catalytic information.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    Across the benchmark’s tasks and controlled splits, domain-specific pretraining and catalytic context are important for modeling enzymatic reactions, while reaction representations also carry catalytic information.

    Source: §6 · p. 9

    EC numbers provide only a coarse description of catalytic function, and the benchmark focuses on single-step reactions while constrained decoding relies on a predefined molecular library.

    Source: §6 · p. 10

Abstract

from arXiv · show

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

1 INTRODUCTION

Enzymatic reaction modeling lacks large, standardized datasets and comparable benchmarks, and it must account for catalytic function as well as molecular structure. The paper introduces a benchmark and a unified pretrained model, reporting competitive performance across evaluated settings.

1 INTRODUCTION

  • Enzymatic modeling is hindered by scarce, dispersed data and inconsistent preprocessing, task definitions, and splits that make reported results difficult to compare directly.The paper notes that enzymatic reactions differ from conventional organic reactions because outcomes depend on catalytic function as well as substrate chemistry.Source: §1 · p. 2
  • Enzymatic reactions depend jointly on substrate and enzyme, whose specificity, stereoselectivity, and cofactor dependence can shape reaction outcomes.The paper uses hierarchical EC numbers as a proxy for catalytic function because reaction–EC pairs are more abundant than reaction–enzyme pairs.Source: §1 · p. 2
  • 126,361 unique enzymatic reactions form VenusRX-Bench, which covers forward prediction, single-step retrosynthesis, and EC-number prediction with leakage-controlled splits.The benchmark is curated from multiple biochemical databases and includes EC annotations and atom-atom mapping.Source: §1 · p. 2
  • VenusRX achieves the best or competitive performance on most evaluated settings against organic-chemistry and biochemical baselines.The model uses a two-stage pretraining approach and can optionally condition on EC numbers; trie-based constrained sampling is designed to reduce invalid outputs for large biomolecules.Source: §1 · p. 2

2 VENUSRX-BENCH

VenusRX-Bench standardizes enzymatic reaction data and evaluation to study links among reactants, products, and catalytic functions. Its tasks and splits probe generalization across reaction components, while dataset comparisons highlight the structural complexity of its reactants.

2 VENUSRX-BENCH

  • VenusRX-Bench addresses differences across biochemical databases and study-specific procedures with a common reaction collection, standardized tasks, and shared evaluation rules.It evaluates relationships among reactants, products, and catalytic functions.Source: §2 · pp. 2–3, Figure 1 · p. 3

2.1 DATA CONSTRUCTION AND SPLITS

  • VenusRX-Bench has the largest reported reaction count among listed datasets and provides both EC annotations and atom mappings.Table 1 compares the integrated collection with existing biochemical datasets.Source: §2.1 · p. 3, Table 1 · p. 4

2.2 TASKS AND EVALUATION

  • The benchmark evaluates forward prediction, single-step retrosynthesis, and EC-number prediction using multi-reference Top-k accuracy or EC1–EC4 accuracy and F1 metrics.Its reactant, product, and EC splits group examples by canonical reactant combinations, principal products, and complete four-level EC annotations, respectively.Source: §2.1 · p. 3, §2.2 · p. 4

2.3 DATASET PROPERTY

  • VenusRX-Bench has more reactant-side heavy atoms on average than USPTO50K, including a larger average size for the largest individual reactant.The structural increase reflects larger molecules, not simply more reactant components, and motivates enzymatic training and constrained generation.Source: §2.3 · p. 4

3 VENUSRX

VenusRX models enzymatic reactions through shared text-to-text objectives, with training that progresses from template-expanded reactions to real benchmark data. Optional EC conditioning and molecule-library constraints address catalytic context and complex molecular generation.

3 VENUSRX

  • VenusRX jointly models forward prediction, retrosynthesis, and reaction reconstruction, training first on a template-expanded corpus and then on real VenusRX-Bench reactions.Optional EC conditioning can provide catalytic information, and constrained generation is studied for large, complex biochemical molecules.Source: §3 · pp. 4–5

3.1 UNIFIED MULTITASK MODELING

  • The T5-style encoder–decoder shares parameters across three text-to-text objectives, while EC numbers can be appended as optional input conditions.The objectives learn both directions of an enzymatic transformation and use span corruption to encourage shared reaction representations; EC prompts are appended with probability 0.5 during training.Source: §3.1 · p. 5

3.2 TRAINING PROCEDURE

  • Two-stage training first learns general enzymatic reaction patterns and then adapts the model to the real reactions in the corresponding benchmark training split.Both stages use the same architecture and multitask objectives; R-SMILES augmentation supplies multiple sequence views without changing the number of unique reactions.Source: §3.2 · p. 5

3.2.1 LARGE-SCALE ENZYMATIC REACTION PRETRAINING

  • The Stage-1 corpus contains about four million reaction records generated by applying extracted templates to compatible substrates and filtering for validity, feasibility, duplication, and split leakage.Template-based generation expands reaction diversity, whereas R-SMILES augmentation increases representation diversity.Source: §3.2.1 · pp. 5–6

3.3 MOLECULE LIBRARY-CONSTRAINED DECODING

  • Molecule Library-Constrained Decoding restricts generation to a predefined library because larger, more complex enzymatic molecules make autoregressive outputs more prone to invalidity.A trie permits only continuations consistent with a library molecule, while VenusRX scores candidates using the input reaction context.Source: §3.3 · p. 6

4 EXPERIMENTS

Experiments assess VenusRX across reaction-generation and EC-prediction tasks, including generalization splits, catalytic-function conditioning, constrained decoding, and training ablations. Results indicate strong performance across splits and a bidirectional link between reaction structure and catalytic function.

4.1 EXPERIMENTAL SETUP

  • Evaluation covers forward prediction, single-step retrosynthesis, and EC prediction, with multi-reference Top-k accuracy for generation and weighted F1 as the primary EC metric.Experiments report results with and without EC conditioning; leakage control removes validation- or test-overlapping reactions from Stage-1 pretraining.Source: §4.1 · p. 6

4.2.1 GENERALIZATION ACROSS ENZYMATIC REACTION SPACE

  • VenusRX performs strongly across Reactant, Product, and EC splits, with particularly clear forward-prediction gains on Product and EC and effective retrosynthesis generalization.The reported pattern supports transfer across reaction distributions; MLCD adds further gains in both prediction directions.Source: §4.2.1 · pp. 6–7, Table 2 · p. 7, Table 3 · p. 7

4.2.2 CATALYTIC FUNCTION AND REACTION STRUCTURE ARE COUPLED

  • EC conditioning consistently improves reaction prediction, while frozen VenusRX representations support EC prediction and task-specific prediction improves performance across the EC hierarchy.The two directions together support a bidirectional relationship between reaction structure and catalytic function.Source: §4.2.2 · p. 7, Table 4 · p. 8

4.2.3 CONSISTENCY AND CONSTRAINED GENERATION

  • MLCD improves exact-match performance and target recovery while reducing invalid molecular generation by restricting outputs to molecules supported by a predefined library.Unconstrained decoding retains a broader set of alternative hypotheses, whereas MLCD focuses generation on library-supported molecules.Source: §4.2.3 · p. 8

4.2.4 ABLATION STUDY

  • The full two-stage training configuration performs strongest overall, with Stage-1 transferable transformation knowledge and Stage-2 adaptation to real benchmark reactions.Span corruption adds benefits particularly for retrosynthesis and EC prediction, supporting the combined pretraining and multitask approach.Source: §4.2.4 · p. 8, Table 5 · p. 8

5 RELATED WORK

Prior work provides reaction-modeling architectures, pretraining methods, enzymatic datasets, and catalytic-function models. The paper situates VenusRX against these efforts and identifies the lack of a unified enzymatic benchmark and pretraining framework.

5.1 ORGANIC REACTION MODELING

  • Organic reaction modeling includes Transformer SMILES translation for forward prediction and retrosynthesis, as well as graph encoders and graph-edit approaches.Molecular Transformer and retrosynthesis extensions and graph-based methods provide architectures and baselines.Source: §5.1 · pp. 8–9
  • Chemical reaction pretraining spans representation learning and generative models, including Chemformer and RSGPT for generative pretraining.RXNFP and RxnRep learn reaction representations; Chemformer and RSGPT apply generative pretraining.Source: §5.1 · p. 9

5.2 ENZYMATIC REACTION DATA AND MODELING

  • Existing enzymatic resources cover reaction prediction, retrosynthesis, retrieval, regioselectivity, and EC annotation, but differ in tasks, processing, and curation.These differences make the resources difficult to merge into a unified benchmark.Source: §5.2 · p. 9
  • Enzymatic models incorporate catalytic context or predict EC function, yet existing approaches largely use task-specific supervised learning on individual datasets.The reviewed work includes enzyme-conditioned prediction, EC-token conditioning, pathway planning, and contrastive EC prediction; a unified enzymatic pretraining framework and comprehensive benchmark remain lacking.Source: §5.2 · p. 9

6 CONCLUSIONS AND FUTURE WORK

VenusRX-Bench and VenusRX provide a unified setup for enzymatic reaction modeling across prediction, retrosynthesis, and EC tasks. Future work would extend catalytic context, pathway scope, and constrained generation beyond the current boundaries.

6 CONCLUSIONS AND FUTURE WORK

  • VenusRX-Bench evaluates forward prediction, retrosynthesis, and EC prediction across controlled generalization splits, while VenusRX combines template-expanded pretraining, real-reaction training, optional EC conditioning, and constrained decoding.Experiments show that domain-specific pretraining and catalytic context matter, and that reaction representations encode catalytic information.Source: §6 · p. 9
  • Current scope is limited to coarse EC descriptions, single-step reactions, and constrained decoding over a predefined molecular library.Future directions include enzyme sequences, structures, cofactors, and conditions; multi-step pathway planning; and generation beyond a fixed candidate space.Source: §6 · p. 10

AI USE STATEMENT

The benchmark construction pipeline harmonizes biochemical records, processes structures and mappings, and curates each retained reaction to one principal product. The resulting collection preserves reaction–EC associations while standardizing representations.

AI USE STATEMENT

A BENCHMARK CONSTRUCTION DETAILS

  • The construction pipeline has four stages: multi-source harmonization, structural processing and deduplication, single-product curation, and leakage-controlled partitioning.Source: A BENCHMARK CONSTRUCTION DETAILS · p. 13

A.1 MULTI-SOURCE HARMONIZATION

  • BRENDA, PathBank, Rhea, and KEGG records are converted to a common schema, with source-specific formats normalized and missing EC annotations removed rather than inferred.Source: §A.1 · p. 13
  • 89,605 reaction–EC records from 310,209 source records correspond to 80,558 unique reaction texts after deduplication by normalized reaction text and EC annotation.Records for the same reaction text remain distinct when their EC annotations differ.Source: §A.1 · p. 13

A.2 STRUCTURAL PROCESSING AND DEDUPLICATION

  • Structures are resolved and balanced where possible, mapped reactions are generated, and minimum-edit candidate mappings are retained, preserving ties before canonicalization and structure-level deduplication.The edit score counts removed and added bonds and atoms with changed local environments.Source: §A.2 · pp. 13–14

A.3 SINGLE-PRODUCT CURATION

  • Product curation removes reactant-side products, filters coenzymes and common by-products from multi-product reactions, excludes products with fewer than four heavy atoms, and retains only reactions with one product and at least one reactant.The coenzyme and by-product filter applies only when the original reaction has multiple products.Source: §A.3 · p. 14
  • 83,439 reaction examples remain after structural processing, product curation, and final single-product deduplication.Mapped and unmapped versions contain the same examples in the same order.Source: §A.3 · p. 14

A.4 LEAKAGE-CONTROLLED DATA SPLITS

The benchmark uses complementary grouping constraints to control leakage across reactant, product, and EC dimensions. Mapped and unmapped representations preserve identical examples and row ordering.

A.4 LEAKAGE-CONTROLLED DATA SPLITS

  • Canonical structural keys, normalized reaction text, and hierarchical EC annotations define indivisible connected components assigned to train, validation, or test, targeting an approximately 80/10/10 ratio.Final partition sizes can deviate slightly from that target.Source: §A.4 · pp. 14–15
  • The Reactant Split groups shared reactant sets or text, the Product Split additionally groups shared canonical products, and the EC Split groups complete four-level EC annotations with canonical full reactions.The Product Split tests target products absent from training; the EC Split holds out EC classes from training.Source: §A.4 · pp. 14–15

A.5 MAPPED AND UNMAPPED REPRESENTATIONS

  • Mapped and unmapped representations use the same examples, filtering decisions, and row order, allowing models with different representation needs to be evaluated on matched data.The table caption specifies that all three splits derive from the same 83,439 curated reactions.Source: §A.5 · p. 15, Table 6 · p. 15

B.1 EC PREDICTION: IMPLEMENTATION DETAILS

The implementation details specify EC label spaces, encoder representations, retrieval and projection procedures, and evaluation protocols. The downstream zero-shot setting uses checkpoints that were exposed to EC annotations during pretraining.

B.1 EC PREDICTION: IMPLEMENTATION DETAILS

  • EC prediction uses final stage-two checkpoints rather than validation-best checkpoints, and EC prompting is enabled during pretraining despite the downstream zero-shot setting.The checkpoints follow 80,100 and 72,100 stage-two steps for the reaction and product splits, respectively.Source: §B.1 · p. 15
  • A frozen encoder represents each reaction from 10 R-SMILES views, averaging token embeddings within each view and then averaging the view vectors.This 10-view extraction is distinct from the 20-fold augmentation used during pretraining.Source: §B.1 · pp. 15–16
  • VenusRX* uses cosine k-nearest-neighbor retrieval with k selected on validation data, while DRFP and rxnfp also use cosine retrieval and crxnfp uses 1-nearest-neighbor retrieval.Source: §B.1 · p. 16
  • For EC prediction, VenusRX replaces CLAIRE’s DRFP and rxnfp features with frozen VenusRX encoder representations while retaining CLAIRE’s projection-training setup.Separate projection networks are trained for each EC level.Source: §B.1 · p. 16
  • Theia and BEC-Pred are trained on the corresponding training partition using the specified feature and model configurations, with released code configurations followed when descriptions differ.Source: §B.1 · p. 16
  • Weighted F1 is the primary EC-prediction metric, with macro F1 and top-1 accuracy also reported; test examples with unseen labels remain in evaluation but those labels are not prediction candidates.Training label vocabularies differ by EC level and benchmark split.Source: §B.1 · pp. 15–16

C.1 ARCHITECTURE AND REACTION REPRESENTATION

This section describes VenusRX’s reaction representations, training and inference procedures, split-specific filtering, and template-and-fragment pipeline for expanding enzymatic transformations into synthetic reactions. It distinguishes synthetic pretraining examples from experimentally verified reactions, outlines the synthetic-example interpretation and unconstrained-decoding protocol for VenusRX-Tiny, and notes that EC labels and molecule-library constraints do not establish catalytic activity or feasibility.

C.1 ARCHITECTURE AND REACTION REPRESENTATION

  • The shared-embedding T5-style model trains on task-specific reaction views, sampling tasks with probabilities (0.35, 0.35, 0.30); span corruption targets masked spans rather than reconstructing the full reaction.The model uses a fixed 302-token vocabulary, and task, view, EC inclusion, and corruption mask can change across presentations.Source: §C.1 · pp. 16–17, §C.2 · p. 18

C.3 EC SERIALIZATION AND LABEL HANDLING

  • EC conditioning appends a sampled annotation when available, but labels may be incomplete or partially truncated, and conditions come from benchmark annotations rather than a separately predicted EC label.Stage-two preparation deduplicates canonical substrate/product pairs without EC in the key and retains the first record.Source: §C.3 · p. 18

C.4 PRETRAINING CORPUS AND SPLIT-SPECIFIC FILTERING

  • Audits found zero validation and test reaction overlaps under the specified matching rules; the reaction-split filter removed 123,880 records, without establishing held-out molecule or scaffold disjointness.The removed records comprised 9,398 direct matches and 114,482 template-associated records.Source: §C.4 · p. 19

C.5 OPTIMIZATION AND CHECKPOINT SELECTION

  • Standard Tiny checkpoints are selected using internal validation loss, while benchmark validation is monitored separately; the reported ablations are single-seed experiments with unequal realized training budgets.The stage-two-only ablation starts from random initialization, and the no-span ablation samples forward and retrosynthesis tasks equally.Source: §C.5 · pp. 19–20

C.6 INFERENCE AND SEPARATELY REPORTED DIAGNOSTICS

  • Molecule-library decoding exposes the model to benchmark-wide molecules, while forward rescoring uses the same checkpoint; neither diagnostic guarantees improved accuracy in every condition.The constrained setting is reported separately from unrestricted held-out evaluation; the lambda sweep is a sensitivity diagnostic, not validation-based parameter selection.Source: §C.6 · p. 21

D SYNTHETIC ENZYMATIC REACTION GENERATION

  • Templates encode local transformations and retain source EC annotations, while BRICS fragments supply substrate contexts; generated reactions are then mapped and represented with direction-specific R-SMILES views.Inherited EC labels describe source reaction categories, not experimental evidence that an enzyme catalyzes each generated reaction; BRICS connection points do not define the reaction center.Source: §D.1 · p. 21, §D.2 · p. 22, §D.4 · p. 22

D.3 TEMPLATE–FRAGMENT MATCHING

  • Template–fragment matching creates combinatorial substrate combinations, scaling one observed transformation into many candidates using template-encoded structural constraints rather than molecular similarity alone.The candidate combination space is the product of compatible-fragment sets for each reactant pattern.Source: §D.3 · p. 22

D.5 REACTION FILTERING

  • Synthetic candidates undergo structural-validity, reaction-consistency, biochemical, and thermodynamic filtering, followed by split-specific exclusion of held-out reaction overlaps and held-out-only templates.These stages address the fact that template application alone does not ensure a plausible biochemical reaction.Source: §D.5 · pp. 22–23

D.6 FROM SYNTHETIC REACTIONS TO MODEL INPUTS

  • Approximately four million synthetic reaction instances were used for Stage-1 pretraining before counting their alternative R-SMILES representations.R-SMILES increases the number of sequence realizations for each instance.Source: §D.6 · p. 23

D.7 INTERPRETATION OF SYNTHETIC REACTIONS

  • The generated corpus is large-scale pretraining data, not experimentally verified enzymatic reactions; template compatibility and filtering do not establish experimental feasibility.The corpus broadens the reaction space seen before continued Stage-2 training on experimentally observed VenusRX-Bench reactions.Source: §D.7 · p. 23
  • An EC label inherited from a template identifies the source transformation’s catalytic category but does not show that an enzyme with that annotation accepts the generated substrates.Source: §D.7 · p. 23

E VENUSRX-TINY RESULTS WITHOUT CONSTRAINED GENERATION

  • Table 12 reports VenusRX-Tiny Top-k exact-match accuracy (%) across three data splits using standard unconstrained decoding.Source: Table 12 · p. 24

F ADDITIONAL EXPERIMENTAL RESULTS

The additional results section reports round-trip retrosynthesis and EC-prediction evaluations, while the implementation section describes how MLCD constrains molecular decoding.

F.1 RETROSYNTHESIS

  • The retrosynthesis tables report round-trip coverage and accuracy across three data splits, grouped by whether EC information is provided.Coverage measures whether at least one of the top-k candidates maps back to the target; accuracy measures the fraction of the first k candidates that do.Source: Table 13 · p. 25, Table 14 · p. 25

F.2 EC PREDICTION

  • The EC-prediction table reports Macro F1 (%) at EC levels 1–4 on the Reactant and Product splits.Source: Table 15 · p. 26

G.1 IMPLEMENTATION

  • MLCD uses a token-prefix trie from a predefined molecule library to constrain molecular tokens, resetting the trie after each molecule separator.Reaction-level control tokens and separators are generated normally; model probabilities and beam ranking remain determined by VenusRX.Source: §G.1 · p. 26

G.2 ANALYSIS OF MOLECULE LIBRARY-CONSTRAINED DECODING

The analysis finds that MLCD eliminates invalid raw molecular candidates, while unconstrained decoding generally scores higher on round-trip evaluation.

G.2 ANALYSIS OF MOLECULE LIBRARY-CONSTRAINED DECODING

  • MLCD eliminates invalid raw molecular candidates in all evaluated retrosynthesis settings.The raw invalid rate counts candidates that cannot be parsed as valid SMILES before final filtering; missing Top-k slots are excluded.Source: §G.2 · p. 26, Table 16 · p. 27
  • Round-trip evaluation applies a fixed forward oracle to test whether retrosynthesis candidates map back to the target product.RT Accuracy@k is the fraction of the first k candidates that map back; RT Coverage@k is the fraction of test reactions with at least one such candidate.Source: §G.2 · p. 26
  • Unconstrained decoding generally achieves higher round-trip scores than MLCD, while MLCD has higher exact-match accuracy and reduces the diversity of retained alternatives.Source: §G.2 · p. 26

Related papers

Loading 2610.11694v1…