Mechanistic Reaction Datasets

• Mechanistic reaction datasets are structured, machine-readable collections that detail reaction pathways through explicit annotations of atom mapping, electron flow, and stepwise transformations.
• They enable precise mechanism prediction, reactive-site identification, and impurity pathway inference by supporting diverse ML applications with benchmark metrics.
• Challenges include limited coverage of rare mechanisms, standardization of arrow codes and annotation schemes, and integration of 3D transition-state data for enhanced interpretability.

A mechanistic reaction dataset is a structured, machine-readable collection of chemical reactions where each entry is explicitly annotated with the underlying mechanistic steps, including atom and electron movements, atom mappings, reactive-site assignments, and associated metadata. Unlike conventional "reactant–product" reaction datasets, mechanistic datasets encode the individual elementary transformations—usually as arrow-pushing events or graph-edit operations—that constitute the reaction pathway from reactants to products. Such datasets provide the foundation for interpretable reaction prediction, mechanistic modeling, impurity pathway inference, and machine learning of chemical reactivity.

1. Mechanistic Dataset Types and Core Representation

Mechanistic reaction datasets can be grouped according to the type of reaction mechanisms they annotate—polar/ionic, radical, organometallic, or transition state–level physical data—and by the granularity of annotation (elementary-step vs. multi-step pathway). A typical mechanistic entry comprises:

• Reactant(s) and product(s) with explicit atom mapping
• A sequence of elementary steps, each described by electron flow (arrow-pushing), bond changes, or graph edits
• Mechanistic type/classification (e.g., nucleophilic substitution, oxidative addition, hydrogen abstraction)
• Optionally, byproducts, intermediates, and reaction conditions

Common data formats include structured JSON with SMILES, atom maps, and explicit arrow codes, graph-encoded datasets (node/edge representations), text-based encodings such as MechSMILES, and, for specialized applications, 3D-geometry and force-field information for transition-state sampling.

Mechanistic step encoding is standardized using arrow codes, SMARTS/SMIRKS templates, or "OrbChain" notation for electron flow. For example, elementary arrow-pushes are typically encoded as source atom and target atom indices, with associated bond change (e.g., Δbond = +1 for bond formation, Δbond = –1 for bond breaking) (Miller et al., 22 Apr 2025, Tavakoli et al., 2023). In graph-based ML, each mechanistic event is formalized as a set of edits Δ applied to a molecular graph Gₜ, updated as Gₜ₊₁ = APPLY(Gₜ, Δ) (Joung et al., 7 Mar 2024).

2. Major Public Mechanistic Reaction Datasets

Numerous datasets underpin recent advances in mechanistic ML:

a) Polar/Ionic Mechanisms:

• PMechDB: ~13,000 manually curated polar steps (nucleophilic substitution, proton transfer, acyl substitution, etc.) with explicit arrow-pushing, atom mapping, and class labels; combinatorially augmented with 48 million synthetic proton-transfer steps. Available as JSON/CSV (SMILES+arrow) (Miller et al., 22 Apr 2025).
• Large-scale mechanistic datasets from patent literature: Over 1.3 million full multi-step reaction pathways (5.8 million steps) constructed via elementary template imputation from Pistachio patents, yielding extensive coverage over 86 reaction classes (Joung et al., 7 Mar 2024).
• USPTO-based electron-flow datasets: e.g., MechUSPTO-31k, containing ~115,000 elementary steps and full mechanisms, allow benchmarking both stepwise and full-pathway mechanism models (Neukomm et al., 5 Dec 2025).

b) Radical Mechanisms:

• RMechDB: ~5,500 curated radical elementary steps annotated with half-arrow (fish-hook) notation, atom mapping, mechanistic class (hydrogen abstraction, radical addition, recombination, β-scission), and full SMILES (Tavakoli et al., 2023).

c) Organometallic and Transition Metal Catalysis:

• ReactMech: ~30,000 full reaction mechanisms (~105,000 elementary steps), spanning 67 mechanistic classes including 7 transition-metal (TM) reactions; explicit atom mapping and mass balance at every step, with templates encoded in SMARTS and detailed operation ("TMOp") class (Das et al., 19 Sep 2025).
• ReactAIvate: 100,000 annotated elementary steps from TM-catalyzed coupling chemistry (Suzuki, Buchwald–Hartwig, Kumada), with exhaustive step-type labels (oxidative addition, transmetallation, etc.) and atom-level reactivity masks (Hoque et al., 14 Jul 2024).

d) Physically-Resolved Path/Transition-State Datasets:

• Transition1x: 9.6 million DFT-evaluated molecular configurations (energies and forces) covering 10,073 minimum-energy paths for reactions up to seven heavy atoms. Unlike equilibrium-only datasets, Transition1x samples the full reaction coordinate via NEB trajectory, supplying geometries both at and near transition states (Schreiner et al., 2022).

e) Visual and Optical Benchmarks:

• SMiCRM: 453 PNG images of arrow-pushing diagrams, canonical SMILES, and SDF files, benchmarking optical chemical structure recognition (OCSR) under mechanistic graphical noise (Leung et al., 25 Jul 2024).

3. Mechanistic Annotation Schemes and Templates

Mechanical annotation of elementary steps consists of:

• Arrow-pushing codes: Explicit source and sink atoms (indexed), change in bond order (Δbond), and electron movement (single/double, for radical/polar). OrbChain and MechSMILES notations serve as canonical formats (Miller et al., 22 Apr 2025, Neukomm et al., 5 Dec 2025, Tavakoli et al., 2023).
• Template rules (SMARTS/SMIRKS): Templates define subgraph transformation patterns, e.g., for oxidative addition, deprotonation, nucleophilic addition, radical abstraction (Joung et al., 7 Mar 2024, Das et al., 19 Sep 2025).
• Step classes and taxonomy: Mechanistic classes are assigned at both the step level (e.g., acid–base deprotonation, transmetalation, β-scission) and sequence level, supporting stratified evaluation and class-specific ML tasks (Hoque et al., 14 Jul 2024, Tavakoli et al., 2023).
• Atom/atom mapping: Comprehensive atom mapping is mandatory, especially for ML models that require reactant–product correspondence and for tracking electron/bond flow across steps.

Data records typically incorporate explicit fields for reactant and product SMILES, arrow or template lists, class labels, mapping arrays, and, in some datasets, source references and mechanistic context (Miller et al., 22 Apr 2025, Tavakoli et al., 2023, Neukomm et al., 5 Dec 2025).

4. Dataset Construction, Curation, and Coverage

Mechanistic datasets arise via a combination of manual curation, expert-driven template annotation, and data-driven imputation:

• Manual curation: Hand-compiled reactive steps from textbooks and literature, especially for rare, organometallic, or atmospheric radical steps (e.g., RMechDB "core" subset, ReactMech TM cases) (Tavakoli et al., 2023, Das et al., 19 Sep 2025).
• Template-driven imputation: Algorithms apply expert-defined templates to patent or literature datasets, inserting intermediates and assigning mechanistic steps to recorded product transformations (Joung et al., 7 Mar 2024).
• Combinatorial augmentation: Synthetic expansion of coverage via systematic generation of plausible mechanistic steps (e.g., all acid/base pairs within pKₐ/threshold, as in PMechDB’s 48M-step augmentation) (Miller et al., 22 Apr 2025).
• Transition-state/path sampling: High-throughput NEB calculations on reaction pairs, sampling along the reaction coordinate and capturing off-equilibrium geometries frequently missing in equilibrium-based datasets (Schreiner et al., 2022).

Coverage varies: large patent-derived datasets achieve breadth but may miss rare mechanisms or explicit TS characterization; curated/augmented sets fill specialist gaps (e.g., radical/organometallic steps, transition states). Mechanistic class balance, template diversity, and atom-level detail are documented for each resource (Joung et al., 7 Mar 2024, Miller et al., 22 Apr 2025, Das et al., 19 Sep 2025, Tavakoli et al., 2023, Schreiner et al., 2022).

5. Machine Learning Applications and Benchmarking

Mechanistic datasets serve as the foundation for a spectrum of ML tasks:

• Mechanism prediction: Stepwise sequence modeling (recapitulating entire mechanistic pathways), e.g., Transformer-based or graph-edit models trained to propose all intermediates from reactant to product (Joung et al., 7 Mar 2024, Das et al., 19 Sep 2025, Chen et al., 13 Mar 2025).
• Elementary step classification: Predicting the next elementary mechanistic move from a given state, as in ReactAIvate GNN or RMechRP for radical steps (Hoque et al., 14 Jul 2024, Tavakoli et al., 2023).
• Reactive-site identification: Pinpointing reactive atoms/bonds for proposed transformations (node-level classification) (Hoque et al., 14 Jul 2024, Tavakoli et al., 2023).
• Byproduct/impurity pathway inference: Explicit enumeration of competing mechanistic branches and byproducts, critical for process chemistry and impurity control (Joung et al., 7 Mar 2024).
• Interpretability and template extraction: Models trained on mechanistic datasets readily yield physically interpretable explanations for each step, including catalyst and spectator identification (Neukomm et al., 5 Dec 2025, Miller et al., 22 Apr 2025).
• Benchmark metrics: Top-k accuracy at the elementary step and full-mechanism levels, OOD (out-of-distribution) generalization, atom/bond conservation, sequence-rank strictness, and pathway recovery are established metrics (see Table below for example benchmark results):

Dataset	Task	Top-1 Accuracy	Top-3 / Top-5	Mechanism Retrieval
ReactMech	Step prediction	98.98 %	n.r.	CRM: 95.94 %
PMechDB	Step prediction	94.9 % (Top-10)	n.r.	Pathway: 84.9 %
RMechDB	Mechanism ranking	62–64 %	74–79 %	96–97 % (Top 10)
MechUSPTO	Elementary step (LM)	95.72 %	96.56 %	73 % (Beam 1)

(Das et al., 19 Sep 2025, Miller et al., 22 Apr 2025, Tavakoli et al., 2023, Neukomm et al., 5 Dec 2025).

6. Current Limitations and Future Directions

Mechanistic datasets, while transformative for reaction informatics, present challenges:

• Scale and diversity: Some datasets (e.g., PMechDB, RMechDB) are small by ML standards and may underrepresent rare or exotic mechanisms; combinatorial expansion addresses this but may be limited to specific step types (Miller et al., 22 Apr 2025, Tavakoli et al., 2023).
• Direct arrow-pushing prediction: Sequence-to-sequence models trained on SMILES alone have limited intrinsic mechanistic interpretability; proposed solutions include including arrow codes directly in model vocabularies or using hybrid models (Miller et al., 22 Apr 2025, Neukomm et al., 5 Dec 2025).
• Transition-state energies/3D geometry: Most datasets are 2D/graph-based and lack direct energetic information; Transition1x addresses this via dense DFT sampling, but expansion to broader chemistry and mechanism types is needed (Schreiner et al., 2022).
• Coverage of non-polar/radical/pericyclic mechanisms: Expansion to non-polar, pericyclic, and enzymatic classes remains an ongoing effort (Miller et al., 22 Apr 2025).
• Standardization: Uniform formats for arrow encoding, atom mapping, template language, and metadata are being developed (e.g., MechSMILES, JSON schema), with efforts to provide APIs for community enrichment (Neukomm et al., 5 Dec 2025, Miller et al., 22 Apr 2025).

Planned extensions include increased automation for curation (mining literature, patents, or quantum-chemical outputs), standardization of step-class and arrow encoding, and integration of 3D structure with mechanistic templates (Miller et al., 22 Apr 2025, Neukomm et al., 5 Dec 2025, Schreiner et al., 2022).

7. Utility in Physical Modeling, Data Analysis, and Machine Learning

Mechanistic reaction datasets facilitate:

• Mechanistically grounded ML: Models trained on these datasets predict full reaction pathways and provide interpretable, stepwise explanations, byproduct/catalyst assignments, and atom-resolved mappings (Joung et al., 7 Mar 2024, Chen et al., 13 Mar 2025, Das et al., 19 Sep 2025).
• Physical property regression: Through unified frameworks such as the Hammett equation, mechanistic datasets enable interpretable, transferable models of substitution effects, including Δ-ML corrections for activation barriers (Bragato et al., 2020).
• Force field and path-sampling ML: Transition1x’s design, with extensive sampling along reaction coordinates, supports the direct training of ML force fields valid far from equilibrium and in transition-state regions (Schreiner et al., 2022).
• CASP validation and mechanistic search: Mechanism datasets enable post hoc plausibility checks and synthesis planning by reconstructing electron-flow-consistent pathways for proposed transformations (Neukomm et al., 5 Dec 2025).
• Benchmarking and stress-testing: Image-based datasets like SMiCRM provide direct tests of OCSR and image-to-structure tools on diagrams containing curved arrows and partial charges (Leung et al., 25 Jul 2024).

These datasets, by providing mechanistic granularity and comprehensive annotation, underpin the development of next-generation interpretable AI frameworks and mechanistically faithful physical-chemical models.