Mechanistic Reaction Prediction via Discrete Flow Matching on Graph-Structured Electron Occupation

Imagine you're trying to design a machine learning model that predicts the outcomes of chemical reactions. Traditionally, approaches either generate product molecules from scratch or modify existing molecular structures based on heuristics. These methods often struggle with accurately capturing the underlying electron dynamics, which are crucial for understanding how reactions actually occur. This limitation can lead to inaccurate predictions, especially in complex scenarios where the reaction pathways are not straightforward. This is what's called a failure in capturing the mechanistic details of reactions.

To address this, the authors propose a new method that models chemical reactions as discrete flow matching over electron occupation vectors. By framing the reactant-to-product mapping as a Continuous-time Markov Chain, they can represent the electron rearrangements more accurately. This approach leverages Optimal Transport to create interpretable edit trajectories without needing detailed annotations of elementary steps. Compared to existing models, MAELLE not only performs competitively on standard benchmarks but also shows robustness in out-of-distribution scenarios, maintaining accuracy where others falter. For builders, this means a more reliable tool for predicting chemical reactions that aligns better with established chemical knowledge and can even forecast side products, enhancing the utility of machine learning in chemistry.

MAELLE introduces a novel approach to modeling chemical reactions through electron flow matching.

The evaluation on the USPTO-480K benchmark and robustness tests provide solid evidence of performance.