Research map / Graph learning / Graph self-supervised learning
Reaction Graph: Towards Reaction-Level Modeling for Chemical Reactions with 3D Structures
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Reaction Graph, a unified graph of reactants and products with reaction edges from atom mapping and rotation/translation-invariant 3D edges (bond length and angular edges), evaluated on condition, yield, and classification prediction.
Problem
Existing graph-based models for chemical reactions encode reactants and products separately and ignore 3D structure, which limits reaction-level modeling of condition, yield, and type prediction.
Approach
Extends the molecular graph into a single reaction graph by adding reaction edges linking mapped atoms in reactants and products, plus an invariant 3D encoding using RBF-embedded bond lengths and an angular edge that forms triangles representing bond angles, processed by an MPNN with Set2Set aggregation.
Key results
- On USPTO-Condition, RG reaches top-1 accuracy 0.325 vs 0.269 for Parrot and 0.198 for D-MPNN.
- On USPTO-TPL, RG reaches ACC 0.999 and reduces classification error by 66.6% relative to T5Chem.
- On yield prediction, RG achieves R2 0.68 on B-H-4 and 0.216 on Subgram, the highest on six of eight datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Reaction Graph instead of Parrot | reaction condition prediction · USPTO-Condition | Top-1 accuracy: 0.269* -> 0.325 | Helped |
| use Reaction Graph instead of Parrot | reaction condition prediction · Pistachio-Condition | Top-1 accuracy: 0.350 -> 0.392 | Helped |
| use Reaction Graph instead of the strongest listed baseline | reaction yield prediction · B-H-1 | R2: 0.84 -> 0.80 | Hurt |
| use Reaction Graph instead of the strongest listed baseline | reaction yield prediction · B-H-3 | R2: 0.65 -> 0.76 | Helped |
| use Reaction Graph instead of the strongest listed baseline | reaction yield prediction · B-H-4 | R2: 0.56 -> 0.68 | Helped |
| use Reaction Graph instead of the strongest listed baseline | reaction yield prediction · Gram | R2: 0.130 -> 0.129 | Hurt |
| use Reaction Graph instead of the strongest listed baseline | reaction yield prediction · Subgram | R2: 0.206 -> 0.216 | Helped |
| use Reaction Graph instead of the strongest listed baseline | reaction classification · USPTO-TPL | accuracy: 0.997 -> 0.999 | Helped |
| use Reaction Graph instead of the strongest listed baseline | reaction classification · Pistachio-Type | accuracy: 0.982 -> 0.987 | Helped |
| add 3D information (ablation) | Reaction Graph for reaction condition prediction · USPTO-Condition | accuracy: 0.3133 -> 0.3246 | Helped |
| add angular edge length rather than directly using bond angle (ablation) | Reaction Graph with bond edge length · USPTO-Condition | accuracy: 0.3179 -> 0.3246 | Helped |
| add reaction information and 3D structure (ablation) | Reaction Graph · USPTO-Condition | accuracy: +6.4% | Helped |
Surprising findings
- Angular edge length outperformed directly using bond angles as features, and was cheaper at inference.
- Yield prediction was restricted to reaction information only because 3D priors were judged hard to learn on noisy small datasets.
Limitations
- Performance depends on quality of 3D coordinates from MMFF94; inaccurate coordinates degrade results.
- Yield datasets are noisy, so 3D information was not tested for yield prediction.
- Relies on atom mapping from RXNMapper, whose errors propagate into reaction edges.
Experimental setup
Datasets: USPTO-Condition Pistachio-Condition Buchwald-Hartwig Suzuki-Miyaura USPTO-Yield (Gram, Subgram) USPTO-TPL Pistachio-Type LvG dataset from USPTO
Benchmarks: USPTO-Condition Pistachio-Condition Buchwald-Hartwig Suzuki-Miyaura Gram Subgram USPTO-TPL Pistachio-Type
Metrics: top-k accuracy R2 likelihood log variance ACC CEN MCC macro F1
Baselines: CRM Parrot AR-GCN CIMG D-MPNN Rxn Hypergraph DRFP Yield-Bert Egret UGNN RXNFP T5Chem
Builds on: MPNN Set2Set Rxn Hypergraph D-MPNN RXNMapper
Closest papers by meaning
- Learning Condensed Graph via Differentiable Atom Mapping for Reaction Yield Prediction (ICML 2025 · similarity 0.88)
- Reaction Prediction via Interaction Modeling of Symmetric Difference Shingle Sets (NeurIPS 2025 · similarity 0.88)
- Uncertainty-Aware Yield Prediction with Multimodal Molecular Features (AAAI 2024 · similarity 0.86)
- GDiffRetro: Retrosynthesis Prediction with Dual Graph Enhanced Molecular Representation and Diffusion Generation (AAAI 2025 · similarity 0.86)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.85)
- ReactGPT: Understanding of Chemical Reactions via In-Context Tuning (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.