SynRXN: An Open Benchmark and Curated Dataset for Computational Reaction Modeling.

SynRXN: An Open Benchmark and Curated Dataset for Computational Reaction Modeling - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Data . 2026 Apr 20;13:625. doi: 10.1038/s41597-026-07260-w Search in PMC Search in PubMed View in NLM Catalog Add to search SynRXN: An Open Benchmark and Curated Dataset for Computational Reaction Modeling Tieu-Long Phan Tieu-Long Phan 1 Bioinformatics Group, Department of Computer Science & Interdisciplinary Center for Bioinformatics & School for Embedded and Composite Artificial Intelligence (SECAI), Leipzig University, Härtelstraße 16-18, D-04107 Leipzig, Germany 2 Department of Mathematics and Computer Science, University of Southern Denmark, DK-5230 Odense M, Denmark Find articles by Tieu-Long Phan 1, 2, ✉ , Nhu-Ngoc Nguyen Song Nhu-Ngoc Nguyen Song 3 School of Pharmacy, University of Medicine and Pharmacy at Ho Chi Minh City, Dinh Tien Hoang, Ho Chi Minh City, Vietnam Find articles by Nhu-Ngoc Nguyen Song 3 , Peter F Stadler Peter F Stadler 1 Bioinformatics Group, Department of Computer Science & Interdisciplinary Center for Bioinformatics & School for Embedded and Composite Artificial Intelligence (SECAI), Leipzig University, Härtelstraße 16-18, D-04107 Leipzig, Germany 4 Max Planck Institute for Mathematics in the Sciences, Inselstraße 22, D-04103 Leipzig, Germany 5 Department of Theoretical Chemistry, University of Vienna, Währingerstraße 17, A-1090 Vienna, Austria 6 Facultad de Ciencias, Universidad National de Colombia, Bogotá, Colombia 7 Center for non-coding RNA in Technology and Health, University of Copenhagen, Ridebanevej 9, DK-1870 Frederiksberg, Denmark 8 Santa Fe Institute, 1399 Hyde Park Rd., Santa Fe, NM 87501 USA Find articles by Peter F Stadler 1, 4, 5, 6, 7, 8 Author information Article notes Copyright and License information 1 Bioinformatics Group, Department of Computer Science & Interdisciplinary Center for Bioinformatics & School for Embedded and Composite Artificial Intelligence (SECAI), Leipzig University, Härtelstraße 16-18, D-04107 Leipzig, Germany 2 Department of Mathematics and Computer Science, University of Southern Denmark, DK-5230 Odense M, Denmark 3 School of Pharmacy, University of Medicine and Pharmacy at Ho Chi Minh City, Dinh Tien Hoang, Ho Chi Minh City, Vietnam 4 Max Planck Institute for Mathematics in the Sciences, Inselstraße 22, D-04103 Leipzig, Germany 5 Department of Theoretical Chemistry, University of Vienna, Währingerstraße 17, A-1090 Vienna, Austria 6 Facultad de Ciencias, Universidad National de Colombia, Bogotá, Colombia 7 Center for non-coding RNA in Technology and Health, University of Copenhagen, Ridebanevej 9, DK-1870 Frederiksberg, Denmark 8 Santa Fe Institute, 1399 Hyde Park Rd., Santa Fe, NM 87501 USA ✉ Corresponding author. Received 2025 Dec 1; Accepted 2026 Apr 14; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ . PMC Copyright notice PMCID: PMC13096289  PMID: 42009670 Abstract We present SynRXN, a unified benchmark dataset resource for computer-aided synthesis planning (CASP). SynRXN decomposes end-to-end synthesis planning into five task families, covering reaction rebalancing, atom-to-atom mapping, reaction classification, reaction property prediction, and synthesis prediction. Curated, provenance-tracked reaction corpora are assembled from heterogeneous public sources into a harmonized representation and packaged as versioned datasets for each task family, with explicit source metadata, licence tags, and machine-readable manifests that record checksums and row counts. For every task, SynRXN provides predefined, leakage-aware partitions, and standardized evaluation metrics tailored to classification, regression, and structured prediction settings. For sensitive benchmarking, we combine public training and validation data with held-out gold-standard test sets, and contamination-prone tasks such as reaction rebalancing and atom-to-atom mapping are distributed only as evaluation sets and are explicitly not intended for model training. Scripted build recipes enable bitwise-reproducible regeneration of all corpora across machines and over time, and the entire resource is released under permissive open licences to support reuse and extension. By removing dataset heterogeneity and packaging transparent, reusable benchmark specifications, SynRXN enables fair longitudinal comparison of CASP methods, supports rigorous ablations and stress tests along the full reaction-informatics pipeline, and lowers the barrier for practitioners who seek robust and comparable performance estimates for real-world synthesis planning workloads. Background & Summary Computer-aided synthesis planning (CASP) assists chemists in designing feasible synthetic routes by combining mechanistic insight, experimental data, and algorithmic search. In parallel with model innovations for reaction prediction and retrosynthesis 1 – 4 , the field has been accelerated by the emergence of standardized open repositories and newly curated reaction databases 5 that enable large-scale, higher-quality training and evaluation. These resources have supported improved reaction prediction and retrosynthesis as well as more effective multi-step route design 6 – 9 . We decouple CASP into three branches according to the chemical object being modeled and the corresponding input to output transformation, namely reaction curation 10 , reaction characterization 11 , and synthesis prediction 1 . In reaction curation , heterogeneous reaction extractions are converted into chemically executable and template ready records by repairing incomplete equations such as missing counter ions, reagents, or coproducts, restoring mass balance, and producing atom-to-atom maps (AAM) that localize bond changes for reaction center identification and template extraction. Given curated records, reaction characterization assigns chemically meaningful identities and attributes to transformations such as reaction class or role labels and physicochemical properties. Finally, synthesis prediction models single-step chemistry via forward reaction prediction and single-step retrosynthesis. We omit multi-step route planning because route level outcomes depend strongly on stock definitions and search heuristics instead of isolated model performance 12 , 13 . Anchoring the reaction curation phase, the pipeline begins with raw reaction data, which are often noisy or incomplete. Extractions from patents, electronic lab notebooks (ELNs), and the literature can omit solvents, counterions, stoichiometric reagents, or byproducts and can contain inconsistent stoichiometry or charge assignments 10 . Such errors corrupt model inputs and bias learned representations 14 . Automated reaction rebalancing methods 15 restore elemental and charge balance using rule-based and graph-based corrections, thereby producing cleaner inputs for subsequent stages. Because rebalancing is a corrective preprocessing step, standardized rebalancing test sets and diagnostics are necessary to quantify correction accuracy and to assess how residual inconsistencies may influence downstream tasks. To complete the reaction curation process once stoichiometry is addressed, AAM establishes the structural lineage that reveals the microscopic changes defining each transformation. Accurate AAM is essential for identifying reaction centers, extracting mechanistic templates, and supervising models that reason about bond changes. Foundational studies have established robust automated mapping methods for complex reactions 16 , and contemporary toolchains include transformer-based mappers such as RXNMapper 17 , graph-based mappers such as GraphormerMapper 18 and LocalMapper 19 , and heuristic mappers such as Indigo 20 and RDTool 21 . Ensemble strategies that arbitrate among multiple mappers 22 , 23 increase coverage and flag low-confidence correspondences. Because mapping errors propagate into template extraction and mechanistic features, AAM benchmarking should ideally rely on curated, held-out gold standards and report exact match accuracy. Where such standards are unavailable, consensus among multiple mapping tools can serve as a proxy for identifying high-confidence subsets 23 . It should be kept in mind, however, that this approach does not replace the need for human-verified validation. For reaction characterization, models assign taxonomic identity to these standardized records resulting from structural curation and template extraction. When accurate atom-to-atom mappings are available, template extraction and mechanistic clustering are straightforward. Because AAM are not universally available or reliable 16 for many corpora, reaction classification performed without atom mapping is still of practical importance and widely used. Benchmarks, therefore, should evaluate both regimes and their sensitivity to mapping quality. Reaction classification groups transformations by mechanism, functional-group change, or named-reaction taxonomy, and supports search, curation, and downstream prediction. Methods range from engineered fingerprints and compact differential descriptors to learned embeddings. Engineered fingerprints enabled the first large-scale categorization efforts 24 . The compact, alignment-free descriptor DRFP encodes bond-change information via hashed circular substructure differences and remains competitive in small-data, interpretable settings 11 . The learned representation RXNFP derives attention-based reaction embeddings from Molecular Transformer sequence models 25 . Molecule-level cross-attention GNNs such as SynCat explicitly model intermolecular context and reagent roles 26 . Consequently, classification benchmarks require representative label taxonomies and transparent splitting procedures, together with repeated resampling to reveal robustness, calibration, and per-class behavior across method families. Beyond categorical labels, reaction property prediction targets continuous and probabilistic quantities that guide experimental decisions. These include yields and physicochemical properties such as activation barriers and transition-state features. Graph- and sequence-based deep models have shown promise for yield and condition recommendations 27 , 28 , and representation-learning approaches (e.g., Condensed Graph of Reactions 29 or Imaginary Transition State 30 embeddings) extend to thermochemical and kinetic targets 31 . Public barrier corpora such as QMrxn/QMrxn20 32 and curated SN2/E2 collections, as well as community benchmarking suites (e.g., Chemprop datasets 33 ) provide training and test splits for systematic barrier modeling and evaluation 34 . These studies show that machine learning predictors can approximate higher level quantum calculations when trained on high quality labels. However, they also reveal limited out-of-distribution transferability for transition-state properties, which motivates the use of QM-augmented descriptors and transition-state-based architectures 31 , 34 . The final stage is synthesis prediction , which composes single-step predictions into multi-step routes under feasibility, cost, and experimental constraints. Single-step models based on templates, sequence-to-sequence architectures, and graph neural networks have improved both forward prediction and retrosynthesis 1 , 4 , 35 . Hybrid planners that combine learned policies 36 with symbolic search produce efficient multi-step routes and expose explicit trade-offs between search budget and route quality 3 . Evaluating planning algorithms is complex because route quality depends on budget, stopping criteria, route-cost models, and practical feasibility checks. Community efforts such as PaRoutes 12 and Syntheseus 13 demonstrate the value of shared route corpora and standardized metrics, though broader harmonization across single- and multi-step settings remains necessary. These observations highlight a critical disparity: while the molecular machine learning community has flourished through standardized ecosystems like MoleculeNet 37 and the Therapeutics Data Commons 38 , reaction informatics remains fragmented. Current studies often rely on bespoke subsets, inconsistent preprocessing, and opaque splitting strategies, rendering cross-paper comparisons nearly impossible. Although broad aggregation efforts like the Open Reaction Database (ORD) 5 solve the data access problem, they do not provide the unified benchmarking layer necessary to rigorously evaluate the full CASP pipeline. To bridge this gap, we introduce SynRXN, a unified, FAIR (Findable, Accessible, Interoperable, and Reusable) 39 benchmarking data resource (see Figure 1 ). SynRXN goes beyond simple data hosting by enforcing strict reproducibility: it supplies deterministic split functions that, when paired with our versioned manifest files and RNG seeds, recreate identical train-test partitions without the overhead of massive index files. SynRXN is complementary to raw corpora such as USPTO, and its goal is not to rehost reactions but to make model comparisons meaningful by standardizing task definitions, splits, and metrics across multiple upstream and downstream components, including rebalancing, atom-to-atom mapping (AAM), classification, property prediction, and single-step synthesis prediction. By providing dedicated benchmarks for both upstream and downstream tasks, SynRXN lets users evaluate each component under a consistent benchmark specification and report upstream quality alongside downstream predictive performance. SynRXN targets component-level, single-step benchmarking rather than multi step route planning, which depends on inventory and search protocol choices beyond the benchmark specification. We further provide held-out gold standards for sensitive upstream tasks like rebalancing and AAM. To avoid hidden assumptions, preprocessing choices are explicit and versioned, so users can adopt the reference specification for head to head comparison or substitute their own upstream methods while keeping the benchmark interface fixed. Available via PyPI ( https://pypi.org/project/synrxn/ ) and Zenodo, SynRXN bundles standardized metrics, reference baselines, and CI-enabled regression checks. By codifying data hygiene and split transparency, SynRXN enables the first truly fair, head-to-head comparison of methods across the task landscape of reaction modeling. Fig. 1. Open in a new tab The SynRXN benchmark. ( A ) Benchmark suite overview. Individual tasks include ( B ) reaction rebalancing , ( C ) atom-to-atom mapping , ( D ) reaction classification , ( E ) reaction property prediction , and ( F ) synthesis prediction . All tasks provide curated datasets, predefined splits, and evaluation metrics. Methods Dataset construction We assembled the SynRXN corpus from publicly available reaction repositories and community benchmarks, including USPTO patent-extracted reaction records 40 and widely used derived sets such as USPTO_50K 41 , USPTO_MIT 42 , and USPTO_500 43 . A complete inventory of upstream sources and their assignment to SynRXN tasks is provided in the following sections and in the archived build manifest. We retrieved raw inputs from their original public release locations on GitHub and Zenodo and from supporting information associated with the source publications. For each component, we recorded the dataset identifier, resolved retrieval location, version or release tag when available, access date, file checksums, and redistribution license in a single manifest distributed alongside the curated release. Using this manifest, we executed deterministic build scripts to fetch inputs, validate checksums, and convert heterogeneous upstream formats into a unified reaction table schema. The curation pipeline applied molecular standardization, record-level validity checks, canonicalization to stable reaction identifiers, and deduplication, with stoichiometric rebalancing applied where required by the task definition. The entire pipeline can be reproduced from the version-pinned manifest by running the accompanying scripts in the script folder deposited with the Zenodo release and mirrored in the project repository. Redistribution follows the upstream licensing terms recorded per component in the manifest. Most components are provided under CC BY 4.0, while the SNAr subset retains CC BY 3.0 attribution requirements. Task datasets Reaction rebalancing Chemical reaction records mined from patent literature frequently lack stoichiometric fidelity, often omitting necessary inorganic reagents, solvents, or byproducts. Restoring mass balance in these records is critical for downstream modeling (Figure 1B ), because missing reactants/products can yield under-specified (or atom-nonconserving) transformations and reduce the chemical executability of extracted retrosynthesis templates. In addition, mechanistic reaction modeling requires stoichiometrically faithful equations 44 . To construct a chemically robust reaction rebalancing benchmark, we derived targeted perturbation sets from the USPTO_50K corpus 35 , partitioning records into three specific modes of stoichiometric violation: MNC (Missing Non-Carbon), MOS (Missing One Side), and MBS (Missing Both Sides). Additionally, we curated a Complex set (1892 examples) from manually validated Golden and Jaworski collections 16 , 22 to capture transformations involving significant skeletal rearrangements. The final compendium included: MNC (33147), MOS (12781), MBS (491), and Complex (1748) (Table 1 a). Each subset was processed using SynRBL 15 , a hybrid rule- and graph-based algorithm for reaction completion. We retained only stoichiometric corrections resolved with a confidence of ≥90%. Table 1. Rebalancing and atom-to-atom mapping benchmark datasets. (a) Rebalancing task Dataset Size Reference MNC 33147 35 MOS 12781 35 MBS 491 35 Complex 1748 16 , 22 (b) Atom-mapping task Dataset Type Size Reference Golden Chem 1785 19 , 22 NatComm Chem 491 16 USPTO_3K Chem 3000 19 Recon3D Bio 382 45 EColi Bio 273 46 Open in a new tab Atom-to-atom mapping To benchmark atom-to-atom mapping (Figure 1C ), we stratified the reaction corpus into two distinct domains: synthetic chemical reactions and biochemical transformations . The collection integrated diverse reference sets, comprising the Golden dataset (1785) 19 , 22 , the manually curated Jaworski subset (491) 16 , and a USPTO_3K partition (3000) sampled from USPTO_50K 19 , totaling 5,276 reactions. The biochemical partition aggregated metabolic data from Recon3D (382) 45 and a validated EColi dataset (273) 46 , yielding 655 reactions. Prior to mapping, all records were subjected to a deterministic standardization and canonicalization procedure 47 . This pipeline enforced structural integrity by filtering anomalies (e.g., malformed SMILES strings or valence/charge violations) and ensuring uniform canonicalization. Topological fidelity is evaluated using two criteria: graph isomorphism of Imaginary Transition State (ITS) graphs 23 , 48 , 49 , and exact-match accuracy of canonicalized reaction SMILES verified via SynKit. Dataset statistics are detailed in Table 1 b. Reaction classification Reaction classification requires mapping raw reaction inputs to predefined classes based on their structural or functional signatures (Figure 1D ). We assembled a benchmark suite spanning multiple levels of granularity. The USPTO_TPL collection served as a fine-grained standard (1000 classes), annotated by deriving SMARTS templates from RXNMapper atom-maps 17 , 50 . For high-level categorization, we utilized the Schneider corpus (50 classes) 24 , which follows the hierarchical RSC reaction ontology. The USPTO_50K dataset 40 provides two label sets: the legacy manual curation (10 classes) 35 and a modern structural relabeling via SynTemp 23 . The latter enforces center-specific isomorphism and extends the reaction core to controlled radii (R0–R2) to capture subtle mechanistic variations. Finally, biochemical diversity is addressed via ECREACT 51 , which provides Enzyme Commission (EC) number hierarchies. To ensure rigorous evaluation, stratified splitting strategies were employed across all corpora to maintain label density across all folds (see Table 2 ). Table 2. Reaction-classification benchmarks. Splits are stratified by reaction class and generated deterministically. The Complete column indicates whether reactions are fully specified or may be missing components. Dataset Size Split ratio Classes Complete Reference Schneider_U 50000 9:1:40 50 No 24 Schneider_B 50000 9:1:40 50 Yes 24 , 26 USPTO_TPL_U 445115 8:1:1 1000 No 50 USPTO_TPL_B 445115 8:1:1 1000 Yes 26 , 50 USPTO_50K_U 50016 8:1:1 10 No 35 USPTO_50K_B 50016 8:1:1 10 Yes 26 , 35 SynTemp_R0 43441 8:1:1 143 Yes 23 , 35 SynTemp_R1 43441 8:1:1 356 Yes 23 , 35 SynTemp_R2 43441 8:1:1 680 Yes 23 , 35 ECREACT_1st 185734 8:1:1 7 No 51 ECREACT_2nd 185734 8:1:1 63 No 51 ECREACT_3rd 185734 8:1:1 175 No 51 Open in a new tab Reaction property prediction The reaction property prediction task (Figure 1E ) targets the quantification of continuous chemical attributes. We assembled a comprehensive benchmark suite by aggregating data from public repositories (including Zenodo) and the literature. The suite encompasses ab initio kinetics datasets (e.g., B97XD3, LogRate), specific mechanistic classes (SNAr, SN2, E2), and high-throughput experimental results (e.g., RGD1). A significant portion of the data was sourced from the Heid collection and related works 31 , 52 , with additional datasets curated from the publications listed in Table 3 . To ensure the integrity of the benchmark, all cleaning, standardization, and filtering operations were fully automated via SynRXN scripts. Datasets are provided in standardized formats containing either rxn (raw SMILES) or aam (atom-mapped SMILES) keys, mapped to specific property labels (e.g., “ea” for barrier height, “dh” for enthalpy). Table 3. Reaction property datasets included in the SynRXN benchmark. The H column indicates whether the dataset includes explicit hydrogen atoms. The Complete column indicates whether reactions are fully specified (Yes) or may be missing components (No). Dataset Size Split AAM H Complete Reference B97XD3 16365 8:1:1 Yes Yes No 58 , 59 SNAr 503 8:1:1 No No Yes 60 E2SN2 3625 8:1:1 Yes Yes Yes 31 , 32 Rad6Re 31923 8:1:1 Yes Yes Yes 31 , 61 LogRate 778 8:1:1 Yes Yes Yes 31 , 62 Phosphatase 33354 8:1:1 Yes No Yes 31 , 63 E2 1264 8:1:1 Yes Yes Yes 52 SN2 2361 8:1:1 Yes Yes Yes 52 RDB7 23852 8:1:1 Yes Yes Yes 52 CycloAdd 5269 8:1:1 Yes Yes * Yes 52 RGD1 353984 8:1:1 Yes Yes Yes 52 Open in a new tab * Can be expanded to include explicit hydrogens (conversion available in our preprocessing scripts). Synthesis prediction The synthesis prediction task consolidates essential benchmarks for algorithmic single-step reaction prediction. We relied on three established subsets of the USPTO patent literature: USPTO_50K, the primary benchmark for template-based and template-free retrosynthesis 53 ; USPTO_MIT, a high-volume corpus optimized for molecular transformer training 54 ; and USPTO_500, a specialized dataset targeting reagent and catalyst inference 43 . These subsets are summarized in Table 4 . Crucially, we recommend standardized, deterministic splits to resolve prevalent issues with benchmark comparability. Our accompanying evaluation suite standardizes reporting protocols, focusing on conventional top- k accuracy alongside structural similarity metrics. Table 4. Reaction prediction corpora used in SynRXN. Dataset Size Split AAM Task Reference USPTO_50K 50016 8:1:1 Yes forward / backward 19 , 35 USPTO_MIT 479035 41:3:4 Yes forward / backward 54 USPTO_500 143535 9:1:1.1 No reagent prediction 43 Open in a new tab Data Records The SynRXN dataset is published as a machine-readable archive and source repository. Canonical releases are available from Zenodo 55 and mirrored on GitHub at https://github.com/TieuLongPhan/SynRXN . Each release includes an authoritative manifest.json file that enumerates all data files under the project root, recording for each file its relative path, cryptographic checksum, row and column counts, column names, a short human-readable description, and file-level license information. This manifest serves as the primary source of provenance and is used by the build and verification scripts to check the internal consistency of each release. The top-level layout, file formats, and evaluation metrics are summarized in Figure 2 . All data files in each release have an explicit license tag recorded in the top-level manifest.json under the license field. Fig. 2. Open in a new tab Overview of the SynRXN benchmark. (A) Data organization under the Data/ root, showing task-specific subdirectories and main tabular records. (B) Evaluation metrics employed for the different SynRXN tasks. All data records reside in the top-level Data/ directory, partitioned into task-specific subdirectories for the five benchmark tasks (Figure 2A ): rebalancing (rbl), atom-to-atom mapping (aam), reaction classification (class), reaction property prediction (prop), and synthesis prediction (synthesis). Each dataset is provided as a gzip-compressed CSV file (.csv.gz) in UTF-8 with a single header row and one reaction per decompressed line. Missing values are encoded as empty fields. Datasets include metadata and predefined train/validation/test splits supplied either as an in-table split column or as companion split files, and may contain task-specific columns such as label, property, mapping-completeness or hydrogen-explicitness flags, and external-source identifiers. Exact file-level metadata (paths, checksums, row/column counts, column names and license tags) are recorded in the top-level manifest.json. Across all tasks, a small set of core columns is used consistently. The r_id column is a string that serves as a stable record identifier (e.g. uspto_00001), unique within each dataset. The rxn column contains the canonical, unmapped reaction SMILES, and aam holds the corresponding atom-mapped reaction SMILES when available. Downstream task labels are stored in either a label column (integer class codes for classification tasks) or a property column (floating-point reaction properties such as activation energies ea, enthalpies dh, or logarithmic rate constants lograte). Some datasets include additional, task-specific columns, for example flags indicating mapping completeness or hydrogen-explicitness, or identifiers linking back to external source corpora. Technical Validation Our goal is to provide dataset-level sanity checks and reproducible reference baselines that contextualize the released benchmarks, rather than to establish optimized or state-of-the-art task performance. To achieve this, raw reaction records were retrieved from their original sources and ingested without manual curation or augmentation. Each entry passed through an automated, deterministic canonicalization and chemical-sanity pipeline implemented in SynKit 47 , which is illustrated in Figure 3 . The pipeline normalizes charges and valence states, standardizes aromaticity, and enforces a consistent SMILES canonicalization for all reactants, reagents, and products. We also perform automated duplicate detection during ingestion via (i) exact SMILES string matches and (ii) structure-level equivalence via isomorphism checks. Duplicate records identified by these checks are removed during ingestion; the dataset manifest records only the final counts so that the provenance and filtering outcome are reproducible. Records failing canonicalization or basic sanity checks (e.g. invalid SMILES, unparsable fields, impossible element/atom counts, or inconsistent valence) were excluded; no manual corrections were applied to excluded items. Finally, we provide reproducible baselines with fixed random seeds (default: 42) and fully specified preprocessing and training configurations. Unless stated otherwise, all models consume deterministic, canonicalized inputs, and all evaluations apply the same standardization and canonicalization to both predictions and curated references. Fig. 3. Open in a new tab Technical validation workflow for the SynRXN benchmark. For the reaction rebalancing task, the test split is curated to provide ground truth: all candidate corrections proposed by SynRBL on the initial test pool were manually inspected, and only reactions that passed verification were retained. Because the final test set consists exclusively of these verified predictions, SynRBL serves as a reference baseline with 100% accuracy by design, reflecting a manually constrained subset of reliable labels rather than unconstrained automatic performance. For atom-mapped subsets, we retain atom maps from the original sources, and intentionally do not rebalance reactions because current atom-mapping models are typically trained or pre-trained on raw, often incomplete data. Mapped reactions are therefore only filtered by basic parsing and valence checks, such that invalid mapped SMILES or impossible valences are removed, while chemically plausible but stoichiometrically imbalanced reactions are kept, preserving the distributional characteristics of contemporary mapping corpora (baseline mapping accuracies for four tools across five held-out datasets are summarized in Table 5 ). Table 5. Atom-mapping accuracy (%) across five datasets. EColi Recon3D USPTO_3K Golden NatComm RXNMapper 0.4.1 72.53 48.69 93.53 87.43 87.58 Graphormer * 42.12 34.82 95.10 89.59 92.87 LocalMapper 0.1.5 69.96 50.79 97.77 89.08 92.67 RDTool 2.4.1 78.02 54.97 90.87 82.54 84.11 Open in a new tab *Graphormer built with Cython 1.7.8. For classification tasks, we report reference baselines using RXNFP and DRFP embeddings as fixed input features to a RandomForest classifier implemented in scikit-learn 56 . Performance is evaluated using repeated stratified cross-validation (5 repeats of 5-fold CV, i.e. 5 × 5 k-fold), following the statistical testing procedure of Ash et al . 57 . Stratified folds preserve class proportions and mitigate variance due to class imbalance. Primary evaluation metrics are the weighted F 1 score 51 and the multiclass Matthews correlation coefficient (MCC) 11 . Results for the stratified split are reported in Table 6 . Table 6. Reaction classification on stratified splits using DRFP and RXNFP embeddings with a RandomForest reference baseline. Values are mean  ± std; higher mean indicates better performance. Significance: NS ( p > 0.05), * ( p < 0.05), ** ( p < 0.01), *** ( p < 0.001), **** ( p < 0.0001). Dataset Level F1 weighted ↑ MCC ↑ DRFP RXNFP p DRFP RXNFP p Schneider_U — 0.968 ± 0.002 0.962  ± 0.002 **** 0.968 ± 0.002 0.961  ± 0.002 **** Schneider_B — 0.953 ± 0.002 0.936  ± 0.002 **** 0.952 ± 0.002 0.935  ± 0.002 **** USPTO_TPL_U — 0.968 ± 0.002 0.962  ± 0.002 **** 0.968 ± 0.002 0.961  ± 0.002 **** USPTO_TPL_B — 0.953 ± 0.002 0.936  ± 0.002 **** 0.952 ± 0.002 0.935  ± 0.002 **** USPTO_50K_U — 0.953  ± 0.002 0.958 ± 0.002 **** 0.943  ± 0.003 0.949 ± 0.002 **** USPTO_50K_B — 0.966 ± 0.002 0.952  ± 0.002 **** 0.958 ± 0.002 0.941  ± 0.002 **** SynTemp 0 0.952 ± 0.001 0.920  ± 0.002 **** 0.954 ± 0.001 0.927  ± 0.002 **** SynTemp 1 0.940 ± 0.002 0.897  ± 0.002 **** 0.943 ± 0.002 0.903  ± 0.002 **** SynTemp 2 0.913 ± 0.003 0.737  ± 0.004 **** 0.907 ± 0.003 0.714  ± 0.005 **** ECREACT 1 0.977 ± 0.001 0.905  ± 0.001 **** 0.966 ± 0.001 0.862  ± 0.002 **** ECREACT 2 0.964 ± 0.001 0.857  ± 0.002 **** 0.961 ± 0.001 0.846  ± 0.002 **** ECREACT 3 0.949 ± 0.001 0.840  ± 0.001 **** 0.947 ± 0.001 0.835  ± 0.001 **** Open in a new tab For datasets with numerical targets (e.g., yields, rates, or energies), we propagate the target values and their units exactly as provided by the original sources, each of which uses a single documented unit for the corresponding endpoint; any residual unit inconsistencies are thus inherited from the original data. To establish reference baselines for property prediction, we apply the same evaluation framework 57 described for classification. The fixed embeddings are instead passed to a RandomForestRegressor 56 , with overall performance reported as mean  ± std. As a necessary exception for the exceptionally large RGD1 dataset, we use a single train-validation-test split evaluated over five random seeds. Empirical baseline performance for these regression tasks, reported as MAE and MSE 31 in Table 7 , indicates that the target distributions are learnable and do not exhibit obvious pathologies such as degenerate ranges or pervasive outliers. Table 7. Reaction property prediction on random splits using DRFP and RXNFP embeddings with a RandomForest reference baseline. Values are mean  ± std; lower values indicate better performance. Significance: NS ( p > 0.05), * ( p < 0.05), ** ( p < 0.01), *** ( p < 0.001), **** ( p < 0.0001). Dataset Prop MAE ↓ MSE ↓ DRFP RXNFP p DRFP RXNFP p B97XD3 dh 19.838  ± 0.262 19.323 ± 0.214 **** 649.814  ± 21.305 599.521 ± 15.667 **** B97XD3 ea 14.617 ± 0.268 15.324  ± 0.239 **** 376.803 ± 13.839 396.723  ± 12.360 **** CycloAdd act 5.853 ± 0.157 6.115  ± 0.157 **** 57.696 ± 4.071 63.569  ± 4.261 **** CycloAdd r 11.790 ± 0.306 12.081  ± 0.312 **** 227.691 ± 12.297 236.306  ± 12.085 ** E2 ea 3.247 ± 0.206 7.377  ± 0.354 **** 20.067 ± 3.174 91.161  ± 8.376 **** E2SN2 ea 4.150 ± 0.126 7.116  ± 0.133 **** 30.667 ± 2.074 81.454  ± 2.857 **** LogRate lograte 1.054 ± 0.068 1.077  ± 0.059 NS 1.970 ± 0.284 2.149  ± 0.355 ** Phosphatase Conversion 0.098 ± 0.001 0.099  ± 0.001 **** 0.019  ± 0.000 0.019  ± 0.000 **** Rad6Re dh 1.126  ± 0.019 0.908 ± 0.013 **** 2.585  ± 0.083 1.612 ± 0.052 **** RDB7 ea 30.136  ± 0.210 18.812 ± 0.240 **** 1362.068  ± 16.817 579.031 ± 15.282 **** RGD1 ea 16.704  ± 0.074 15.953 ± 0.032 NS 495.386  ± 3.867 453.876 ± 2.628 NS SN2 ea 4.433 ± 0.161 6.940  ± 0.234 **** 34.664 ± 2.426 75.566  ± 4.393 **** SNAr ea 1.402 ± 0.158 1.447  ± 0.139 NS 4.348 ± 1.496 4.355  ± 1.032 NS Open in a new tab Usage Notes The SynRXN loader supports three sources for programmatic access: Zenodo (stable, citable archive and recommended for publications), GitHub release tag (release artifacts) and GitHub commit (exact snapshot; may be unstable unless archived). Note that Zenodo queries can occasionally be delayed; enable GitHub fallback when immediate access is required. We recommend using Zenodo for publication workflows (set source="zenodo" and provide version). Use source="github" for workflows driven by releases. For exact reproducibility, use source="commit" and provide the full commit_id; if those results are published, archive the snapshot (create a GitHub release or deposit on Zenodo) so it is citable. For tutorials, splitting strategies, and dataset construction reproducibility, see our documentation. ( https://synrxn.readthedocs.io/en/latest/tutorials_and_examples.html ). Acknowledgements This project has received funding from the European Unions Horizon Europe Doctoral Network programme under the Marie-Skłodowska-Curie grant agreement No 101072930 (TACsy – Training Alliance for Computational systems chemistry). Open Access funding enabled and organized by Projekt DEAL. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union. Neither the European Union nor the granting authority can be held responsible for them. Author contributions Contributions are reported according to the CRediT taxonomy. T.L.P. conceptualized the study, curated and analyzed the data, developed the methods and software, validated the results, and wrote and revised the manuscript. N.N.N.S. designed figures, conducted benchmarking, and reported baseline results. P.F.S. secured funding and resources, supervised the project, and drafted and reviewed the manuscript. Funding Open Access funding enabled and organized by Projekt DEAL. Data availability All data supporting this study are available in the SynRXN project repository ( https://github.com/TieuLongPhan/SynRXN/ ) and as a versioned archive on Zenodo (SynRXN v0.0.8): 10.5281/zenodo.17672847 Code availability The SynRXN source code is available on GitHub: https://github.com/TieuLongPhan/SynRXN . Archived releases are available on Zenodo (SynRXN v0.0.8, DOI: 10.5281/zenodo.17672847). Comprehensive documentation is hosted at https://synrxn.readthedocs.io/en/latest/ . Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Coley, C. W., Barzilay, R., Jaakkola, T. S., Green, W. H. & Jensen, K. F. Prediction of organic reaction outcomes using machine learning. ACS Central Science 3 , 434–443 (2017). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Schwaller, P., Gaudin, T., Lányi, D., Bekas, C. & Laino, T. “found in translation”: predicting outcomes of complex organic chemistry reactions using neural sequence-to-sequence models. Chemical Science 9 , 6091–6098 (2018). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Segler, M. H. S., Preuss, M. & Waller, M. P. Planning chemical syntheses with deep neural networks and symbolic ai. Nature 555 , 604–610 (2018). [ DOI ] [ PubMed ] [ Google Scholar ] 4. Schwaller, P. et al . Predicting retrosynthetic pathways using transformer-based models and a hyper-graph exploration strategy. Chemical Science 11 , 3316–3325 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Kearnes, S. M. et al . The open reaction database. Journal of the American Chemical Society 143 , 18820–18826 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 6. Heid, E., Probst, D., Green, W. H. & Madsen, G. K. H. Enzymemap: curation, validation and data-driven prediction of enzymatic reactions. Chemical Science 14 , 14229–14242 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Chen, S., Babazade, R., Kim, T., Han, S. & Jung, Y. A large-scale reaction dataset of mechanistic pathways of organic reactions. Scientific Data 11 , 863 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Kreutter, D. & Reymond, J. Chemoenzymatic multistep retrosynthesis with transformer loops. Chemical Science 15 , 18031–18047 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Deng, Y. et al . Rsgpt: a generative transformer model for retrosynthesis planning pre-trained on ten billion datapoints. Nature Communications 16 , 7012 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Gimadiev, T. R. et al . Reaction data curation i: chemical structures and transformations standardization. Molecular Informatics 40 , 2100119 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 11. Probst, D., Schwaller, P. & Reymond, J.-L. Reaction classification and yield prediction using the differential reaction fingerprint drfp. Digital Discovery 1 , 91–97 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Genheden, S. & Bjerrum, E. Paroutes: towards a framework for benchmarking retrosynthesis route predictions. Digital Discovery 1 , 527–539 (2022). [ Google Scholar ] 13. Maziarz, K. et al . Re-evaluating retrosynthesis algorithms with Syntheseus. Faraday Discuss. 256, 568–586 10.1039/D4FD00093E (2025). [ DOI ] [ PubMed ] 14. Patel, H., Bodkin, M. J., Chen, B. & Gillet, V. J. Knowledge-based approach to de novo design using reaction vectors. Journal of chemical information and modeling 49 , 1163–1184 (2009). [ DOI ] [ PubMed ] [ Google Scholar ] 15. Phan, T.-L. et al . Reaction rebalancing: a novel approach to curating reaction databases. Journal of Cheminformatics 16 , 82 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Jaworski, W. et al . Automatic mapping of atoms across both simple and complex chemical reactions. Nature Communications 10 , 1434 (2019). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Schwaller, P., Hoover, B., Reymond, J.-L., Strobelt, H. & Laino, T. Extraction of organic chemistry grammar from unsupervised learning of chemical reactions. Science Advances 7 , eabe4166 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Nugmanov, R., Dyubankova, N., Gedich, A. & Wegner, J. K. Bidirectional graphormer for reactivity understanding: neural network trained to reaction atom-to-atom mapping task. Journal of Chemical Information and Modeling 62 , 3307–3315 (2022). [ DOI ] [ PubMed ] [ Google Scholar ] 19. Chen, S., An, S., Babazade, R. & Jung, Y. Precise atom-to-atom mapping for organic reactions via human-in-the-loop machine learning. Nature Communications 15 , 2250 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Savelev, A., Puzanov, I., Samoilov, V. & Karnaukhov, V. Indigo toolkit https://lifescience.opensource.epam.com/indigo/index.html . Software (2019). 21. Rahman, S. A. et al . Reaction decoder tool (rdt): extracting features from chemical reactions. Bioinformatics 32 , 2065–2066 (2016). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Lin, A. et al . Atom-to-atom mapping: a benchmarking study of popular mapping algorithms and consensus strategies. Molecular Informatics 41 , 2100138 (2022). [ DOI ] [ PubMed ] [ Google Scholar ] 23. Phan, T.-L. et al . Syntemp: Efficient extraction of graph-based reaction rules from large-scale reaction databases. Journal of Chemical Information and Modeling 65 , 2882–2896 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Schneider, N., Lowe, D. M., Sayle, R. A. & Landrum, G. A. Development of a novel fingerprint for chemical reactions and its application to large-scale reaction classification and similarity. Journal of Chemical Information and Modeling 55 , 39–53 (2015). [ DOI ] [ PubMed ] [ Google Scholar ] 25. Schwaller, P. et al . Molecular transformer: a model for uncertainty-calibrated chemical reaction prediction. ACS central science 5 , 1572–1583 (2019). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Van Nguyen, P.-C. et al . Syncat: Molecule-level attention graph neural network for precise reaction classification. Digital Discovery 5 , 241–253 10.1039/D5DD00367A (2025). 27. Schwaller, P., Vaucher, A. C., Laino, T. & Reymond, J.-L. Prediction of chemical reaction yields using deep learning. Machine Learning: Science and Technology 2 , 015016 (2021). [ Google Scholar ] 28. Gao, H. et al . Using machine learning to predict suitable conditions for organic reactions. ACS Central Science 4 , 1465–1476 (2018). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Nugmanov, R. I. et al . Cgrtools: Python library for molecule, reaction, and condensed graph of reaction processing. Journal of chemical information and modeling 59 , 2516–2521 (2019). [ DOI ] [ PubMed ] [ Google Scholar ] 30. Fujita, S. Description of organic reactions based on imaginary transition structures. 1. introduction of new concepts. Journal of Chemical Information and Computer Sciences 26 , 205–212 (1986). [ Google Scholar ] 31. Heid, E. & Green, W. H. Machine learning of reaction properties via learned representations of the condensed graph of reaction. Journal of Chemical Information and Modeling 62 , 2101–2110 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. von Rudorff, G. F., Heinen, S. N., Bragato, M. & von Lilienfeld, O. A. Thousands of reactants and transition states for competing e2 and s mathrmn 2 reactions. Machine Learning: Science and Technology 1 , 045026, 10.1088/2632-2153/aba822 (2020). [ Google Scholar ] 33. Heid, E. et al . Benchmark data for chemprop (includes reaction barrier subsets: Sn2, e2, rdb7, ...). Zenodo dataset 10.5281/zenodo.8174268 (2023). 34. Spiekermann, K. A., Pattanaik, L. & Green, W. H. High accuracy barrier heights, enthalpies, and rate coefficients for chemical reactions. Scientific Data 9 , 417 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Liu, B. et al . Retrosynthetic reaction prediction using neural sequence-to-sequence models. ACS Central Science 3 , 1103–1113 (2017). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Akhmetshin, T. et al . Synplanner: an end-to-end tool for synthesis planning. Journal of Chemical Information and Modeling 65 , 15–21 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 37. Wu, Z. et al . Moleculenet: a benchmark for molecular machine learning. Chemical Science 9 , 513–530 (2018). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. Huang, K. et al . Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development. NeurIPS Datasets and Benchmarks (2021). ArXiv:2102.09548. 39. Wilkinson, M. D. et al . The fair guiding principles for scientific data management and stewardship. Scientific data 3 , 1–9 (2016). [ DOI ] [ PMC free article ] [ PubMed ] 40. Lowe, D. M. Extraction of chemical structures and reactions from the literature . Ph.D. thesis, University of Cambridge. http://www.repository.cam.ac.uk/handle/1810/244727 (2012). 41. Schneider, N., Lowe, D. M., Sayle, R. A., Tarselli, M. A. & Landrum, G. A. Big data from pharmaceutical patents: a computational analysis of medicinal chemists’ bread and butter. Journal of Medicinal Chemistry 59 , 4385–4402 (2016). [ DOI ] [ PubMed ] [ Google Scholar ] 42. Jin, W., Coley, C. W., Barzilay, R. & Jaakkola, T. Predicting organic reaction outcomes with weisfeiler–lehman networks. In Advances in Neural Information Processing Systems (NeurIPS) 2017 , 2604–2613, https://papers.nips.cc/paper/6854-predicting-organic-reaction-outcomes-with-weisfeiler-lehman-network (2017). 43. Lu, J. & Zhang, Y. Unified deep learning model for multitask reaction predictions with explanation. Journal of chemical information and modeling 62 , 1376–1387 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. Joung, J. F. et al . Electron flow matching for generative reaction mechanism prediction. Nature 645 , 115–123 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] 45. Litsa, E. E. et al . Machine learning guided atom mapping of metabolic reactions. Journal of Chemical Information and Modeling 59 , 1121–1135 (2018). [ DOI ] [ PubMed ] [ Google Scholar ] 46. Beier, N., Gatter, T., Andersen, J. L. & Stadler, P. F. Computing double-pushout graph transformation rules and atom-to-atom maps from KEGG RCLASS data. Algorithms Mol Biol 21 , 3 10.1186/s13015-025-00294-6 (2026). [ DOI ] [ PMC free article ] [ PubMed ] 47. Phan, T.-L. et al . SynKit: A graph-based python framework for rule-based reaction modeling and analysis. Journal of Chemical Information and Modeling 65 , 13012–13019 10.1021/acs.jcim.5c02123 (2025). [ DOI ] [ PMC free article ] [ PubMed ] 48. Laffitte, M. E. G., Beier, N., Domschke, N. & Stadler, P. F. Comparison of atom maps. MATCH: Comm. Math. Comp. Chem 90 , 75–102 (2023). [ Google Scholar ] 49. González Laffitte, M. E. et al . Partial imaginary transition state (its) graphs: A formal framework for research and analysis of atom-to-atom maps of unbalanced chemical reactions and their completions. Symmetry 16 , 1217 (2024). [ Google Scholar ] 50. Schwaller, P. et al . Mapping the space of chemical reactions using attention-based neural networks. Nature machine intelligence 3 , 144–152 (2021). [ Google Scholar ] 51. Zeng, Z., Guo, J., Jin, J. & Luo, X. Claire: a contrastive learning-based predictor for ec number of chemical reactions. Journal of Cheminformatics 17 , 2 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 52. Heid, E. et al . Benchmark data for chemprop. 10.5281/zenodo.10078142 (2023). 53. Chen, S. & Jung, Y. Deep retrosynthetic reaction prediction using local reactivity and global attention. JACS Au 1 , 1612–1620 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 54. Jin, W., Coley, C., Barzilay, R. & Jaakkola, T.Predicting organic reaction outcomes with weisfeiler-lehman network. Advances in neural information processing systems 30 (2017). 55. Phan, T. L. synrxn: A benchmarking framework and open data repository for computer-aided synthesis planning. 10.5281/zenodo.17672847 (2025). 56. Kramer, O. Scikit-learn. In Machine learning for evolution strategies , 45–53 (Springer, 2016). 57. Ash, J. R. et al . Practically significant method comparison protocols for machine learning in small molecule drug discovery. Journal of chemical information and modeling 65 , 9398–9411 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] 58. Grambow, C. A., Pattanaik, L. & Green, W. H. Reactants, products, and transition states of elementary chemical reactions based on quantum chemistry. Scientific data 7 , 137 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 59. Grambow, C. A., Pattanaik, L. & Green, W. H. Reactants, products, and transition states of elementary chemical reactions based on quantum chemistry. 10.5281/zenodo.3715478 (2020). [ DOI ] [ PMC free article ] [ PubMed ] 60. Jorner, K., Brinck, T., Norrby, P.-O. & Buttar, D. Machine learning meets mechanistic modelling for accurate prediction of experimental activation energies. Chemical Science 12 , 1163–1175 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 61. Stocker, S., Csányi, G., Reuter, K. & Margraf, J. T. Machine learning in chemical reaction space. Nature communications 11 , 5505 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 62. Bhoorasingh, P. L., Slakman, B. L., Seyedzadeh Khanshan, F., Cain, J. Y. & West, R. H. Automated transition state theory calculations for high-throughput kinetics. The Journal of Physical Chemistry A 121 , 6896–6904 (2017). [ DOI ] [ PubMed ] [ Google Scholar ] 63. Huang, H. et al . Panoramic view of a superfamily of phosphatases through substrate profiling. Proceedings of the National Academy of Sciences 112 , E1974–E1983 (2015). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement All data supporting this study are available in the SynRXN project repository ( https://github.com/TieuLongPhan/SynRXN/ ) and as a versioned archive on Zenodo (SynRXN v0.0.8): 10.5281/zenodo.17672847 The SynRXN source code is available on GitHub: https://github.com/TieuLongPhan/SynRXN . Archived releases are available on Zenodo (SynRXN v0.0.8, DOI: 10.5281/zenodo.17672847). Comprehensive documentation is hosted at https://synrxn.readthedocs.io/en/latest/ . Articles from Scientific Data are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (3.2 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top