Multi-step reaction mechanisms move beyond single-step templates: graph-guided preference data lifts GRMP pathway accuracy to 48.1%

Multi-step mechanism prediction shifts from single elementary-step prediction to generating a whole molecular-state pathway, so the model must output an ordered intermediate sequence rather than only the final product. Earlier computational methods and language-model applications mostly targeted single elementary-step prediction or simplified multiple-choice tasks, and could not extrapolate to unseen reactions whose mechanistic logic departs from training patterns. On the ChemMechQA and oMe-Gold benchmarks, GRMP reaches pathway accuracy of 48.1% and 57.9%, the highest among the compared methods; on ChemMechQA it exceeds the strongest mechanistic baseline, Graph2SMILES at 33.5%, by 14.6 percentage points.

Pathway generation is guided by a chemical mechanism knowledge graph whose molecules and reaction templates are two distinct entity types, with edges labeling their connections. Common knowledge-graph augmentation of language models introduces topological noise when extracting multi-hop paths from dense chemical graphs, while shallow retrieval is too weak to govern chemical steps. The graph is built from 50,000 experimentally reported reactions in the Pistachio database, with intermediates imputed using 676 expert-annotated templates, and contains 247,832 entities, of which 247,156 are molecule nodes and 676 are template nodes.

Training runs in two stages: supervised fine-tuning on annotated pathways, then direct preference optimization on preference pairs built from graph convolutional network candidate pathways. A baseline that builds preference pairs only from reference pathways and incorrect predictions reaches just 5.5% pathway accuracy on the fixed 110-reaction sample from ChemMechQA-Hard. On the same sample, graph-guided preference data for direct preference optimization raises pathway accuracy to 27.2%, an absolute gain of 21.7 percentage points; against the supervised-fine-tuned predictor without language-model refinement, main product accuracy rises from 14.5% to 35.4% and pathway accuracy from 12.7% to 27.2%.

At inference, a relation retrieval module pulls relevant triplets from the knowledge graph as structural priors and hands them to a general-purpose language model to refine the candidate pathway. Without relation retrieval, the same configuration on ChemMechQA-Hard reaches 29.8% intermediate hit, 49.1% main product accuracy and 34.5% pathway accuracy. Adding relation retrieval with a template-hidden knowledge graph lifts these to 34.7%, 66.4% and 37.3%; with the standard knowledge graph, pathway accuracy is 36.4%.

The demonstrated scope is multi-step mechanism pathway prediction for organic reactions, evaluated on ChemMechQA, on oMe-Gold with 196 literature-verified pathways, and on 2,009 ChemMechQA-Hard reactions drawn from 11 held-out reaction classes and 24 unseen mechanism templates; ablations and human evaluation use a fixed 110-reaction sample from that set.

Generated pathways are positioned as credible mechanism candidates, and any route would require experimental validation before being treated as a discovered mechanism; the benefit of refinement varies with reaction difficulty and metric, with pathway accuracy dipping from 42.0% to 41.6% on medium-complexity reactions; preference thresholds that are too high or too low both lower pathway accuracy and main product accuracy.