Mechanistic Reaction Prediction via Discrete Flow Matching on Graph-Structured Electron Occupation
MAELLE models chemical reactions as discrete flow matching over electron occupation vectors, using a Continuous-time Markov Chain over graph-structured integer-valued electron occupation space. It generalizes discrete flow matching mixture paths to discrete electron rearrangements via Optimal Transport, producing mechanistically interpretable edit trajectories without elementary step annotations. MAELLE achieves competitive performance on USPTO-480K and is evaluated on two out-of-distribution settings.
MAELLE (Mechanistic Edit fLow-matching on eLectron rEarrangements) introduces a novel approach to reaction prediction by modeling reactions as transformations in electron space rather than direct molecular graph edits. It uses discrete flow matching over electron occupation vectors, formulated as a Continuous-time Markov Chain on graph-structured integer-valued electron occupation states. The method constructs intermediate edit trajectories by generalizing discrete flow matching mixture paths with Optimal Transport, yielding interpretable edit moves without requiring elementary step annotations. The model achieves competitive results on the USPTO-480K benchmark and is tested for robustness in two out-of-distribution settings.
The key technical contribution is the formulation of reaction prediction as discrete flow matching over electron occupation vectors, which shifts the modeling target from molecular topology to electron rearrangements. By using a Continuous-time Markov Chain on graph-structured integer-valued states and Optimal Transport to define mixture paths, MAELLE generates mechanistically interpretable edit trajectories. This approach avoids heuristic graph edits and does not require elementary step annotations, potentially improving generalization and interpretability.
This research could influence cheminformatics and drug discovery by providing more interpretable and robust reaction prediction models. The focus on electron-level transformations may lead to better handling of complex reactions and out-of-distribution scenarios, which are common in real-world chemical synthesis planning.
Improved reaction prediction can accelerate drug discovery and materials science by reducing the need for expensive laboratory experiments. MAELLE's interpretability and robustness may lower barriers to adoption in pharmaceutical and chemical companies, potentially leading to cost savings and faster development cycles.
Future work may explore scaling MAELLE to larger reaction datasets, integrating it with retrosynthesis planning tools, and validating its out-of-distribution robustness in practical chemistry workflows. The approach could also inspire similar electron-space modeling in other molecular property prediction tasks.