Enhancing Multi-Step Chemical Mechanism Prediction with Graph-Guided Preference Data
Organizations: Shanghai Jiao Tong University Shanghai, China
Abstract
Predicting multi-step chemical reaction mechanisms requires identifying plausible intermediates and correctly ordering the elementary reaction steps. Existing computational methods focus on single-step elementary reaction prediction, limiting their ability to generalize to unseen reaction spaces. We introduce a graph-guided framework for multi-step chemical mechanism prediction that leverages a chemical mechanism knowledge graph (ChemMechKG) to guide large language models (LLMs). To improve pathway prediction, we develop a graph-aligned reaction mechanism predictor (GRMP) that uses graph neural networks to generate synthetic candidate pathways for pathway optimization. During inference, we design a relation retrieval (RR) module to extract relevant graph evidence from the KG for pathway refinement. Experiments show that GRMP achieves the highest pathway accuracy among all compared methods on ChemMechQA and oMe-Gold, reaching 48.1% and 57.9%, respectively. On a fixed sample of 110 out-of-distribution reactions from ChemMechQA-Hard, GRMP achieves 36.4% top-1 pathway accuracy. LLM-as-Judge and human assessment further demonstrate its ability to generalize to unseen reactions.
Figures & tables
| Method | ChemMechQA | oMe-Gold | ||||||
| Validity (%) | MPA (%) | I-hit (%) | PA (%) | Validity (%) | MPA (%) | I-hit (%) | PA (%) | |
| General LLMs for mechanism reasoning | ||||||||
| GPT-5.2 ( Singh et al., 2025 ) | 70.3 ±0.09 | 10.0 ±0.09 | 8.1 ±0.51 | 5.0 ±0.22 | 90.1 | 23.1 | 29.1 | 23.6 |
| GPT-5.2 + Few-shot ( Wei et al., 2022 ) | 78.9 ±0.85 | 13.1 ±0.27 | 13.0 ±0.00 | 0.0 ±0.35 | 90.6 | 17.8 | 11.3 | 18.3 |
| Gemini-3.1-Pro ( Team et al., 2023 ) | 90.1 ±0.30 | 52.2 ±0.20 | 7.3 ±0.31 | 10.0 ±0.27 | 90.5 | 57.8 | 37.9 | 34.9 |
| DeepSeek-V4-Pro ( DeepSeek-AI, 2026 ) | 95.0 ±0.24 | 48.0 ±0.30 | 12.9 ±1.01 | 5.0 ±0.40 | 87.6 | 45.3 | 25.0 | 22.5 |
| Method | MPA (%) | I-hit (%) | PA (%) | SR (%) |
| Easy Complexity | ||||
| FlowER | 63.6 | 57.0 | 53.5 | 60.5 |
| Graph2SMILES | 73.4 | 64.2 | 62.1 | 67.4 |
| 64.2 | 40.9 | 32.4 | 62.3 | |
| GRMP | 70.8 | 71.4 | 57.2 | 67.5 |
| Medium Complexity | ||||
| Configuration | Validity (%) | I-hit (%) | MPA (%) | PA (%) | ||||||||
| (Components of methods) | Top-1 | Top-3 | Top-5 | Top-1 | Top-3 | Top-5 | Top-1 | Top-3 | Top-5 | Top-1 | Top-3 | Top-5 |
| Mechanistic models | ||||||||||||
| FlowER | 100.0 | 100.0 | 100.0 | 18.6 | 19.0 | 20.2 | 14.6 | 14.6 | 14.6 | 10.9 | 10.9 | 10.9 |
| Graph2SMILES | 89.1 | 90.9 | 90.9 | 15.7 | 17.7 | 18.2 | 12.7 | 14.6 | 14.6 | 7.3 | 9.1 | 9.1 |
| Ablation of general LLMs (SFT-only predictors; w/o DPO, w/o RR) | ||||||||||||
| SFT-only predictor: Llama-3.1-8B | ||||||||||||
| Dataset | #Reactions | #Reaction Types | #Mech. types/ subtypes | MI? | NL? | MP? | Cond? |
| Tavakoli et al. (2023b) | 5k+ | – | 7 | ✓ | ✗ | ✗ | ✗ |
| Chen et al. (2024b) | 31k | 86 | 13 | ✗ | ✗ | ✓ | ✗ |
| Joung et al. (2024) | 192k | 86 | 13 | ✓ | ✗ | ✗ | ✓ |
| Hao et al. (2026) | 275 | 86 | 13 | ✗ | ✓ | ✗ | ✗ |
| Xu et al. (2025) | 2.8k | – | 8 | ✓ | ✓ | ✓ | ✗ |
| ChemMechQA | 50k | 182 | 259 | ✓ | ✓ | ✓ | ✓ |
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
| Statistic | Count |
| Entities | 247,832 |
| Molecule nodes | 247,156 |
| Template nodes | 676 |
| Triples | 422,160 |
| Reaction types | 182 |
| Mechanism types | 13 |
| Dataset | Split | Mol. | Rxns. | Steps | Types |
| oMeBench | Train | 12,684 | 2,508 | 10,619 | – |
| Test | 1,151 | 196 | 722 | – | |
| ChemMechQA | Train | 197,759 | 40,000 | 168,545 | 167 |
| Val. | 25,162 | 5,000 | 21,354 | 132 | |
| Test | 24,981 | 5,000 | 21,181 | 132 | |
| ChemMechQA-Hard | OOD Test | 8,008 | 2,009 | 6,361 | 11 |
| Category | Space |
| Mechanism Type | Pericyclic, Rearrangement, Radical, Electron transfer, Substitution, Cleavage, Addition, Elimination, Coordination, Proton transfer / Dissociation, Organometallic, Activation, Condensation |
| Mechanism Subtype | Deprotonation, Proton transfer, Elimination, Neutralization of protonated ether / Addition of alcohol under basic pKa, Nucleophilic Addition, Proton exchange, Amine addition, Halide leaves, Alkoxide leaves, Hydroxide addition, … |
| Hard Reaction Type | Appel bromination (index 5), Carboxylic acid + sulfonamide condensation (index 6), Diels-Alder cycloaddition (index 9), Hydroxy to bromo (index 1), Menshutkin reaction (index 8), N-Bn deprotection (index 11), Negishi coupling (index 2), Prilezhaev epoxidation (index 10), Staudinger reduction (index 4), Transamidation (index 3), Urea Schotten-Baumann (index 7) |