Mechanistic Reaction Prediction via Discrete Flow Matching on Graph-Structured Electron Occupation
MAELLE predicts reactions by modeling electron rearrangements directly, rather than generating products or applying topology edits.
The paper frames reactant-to-product mapping as a continuous-time Markov chain over graph-structured electron occupation vectors. Its training path uses optimal transport to build interpretable intermediate edit moves without elementary-step annotations. On USPTO-480K, the authors report competitive performance against leading reaction prediction models, with stronger robustness in out-of-distribution tests for structural complexity and reaction type. They also say the learned electron flow can recover chemistry-aligned mechanisms and predict side products. ArXiv · AI/CL/LG's note
The paper frames reactant-to-product mapping as a continuous-time Markov chain over graph-structured electron occupation vectors. Its training path uses optimal transport to build interpretable intermediate edit moves without elementary-step annotations. On USPTO-480K, the authors report competitive performance against leading reaction prediction models, with stronger robustness in out-of-distribution tests for structural complexity and reaction type. They also say the learned electron flow can recover chemistry-aligned mechanisms and predict side products. ArXiv · AI/CL/LG's note
score 4