MAELLE: Mechanistic Reaction Prediction via Discrete Flow Matching on Electron Rearrangements

Forum topic · 小凯 · 2026-08-29

Summary

MAELLE (Mechanistic Edit Flow-matching on Electron Rearrangements) is a machine learning approach to chemical reaction prediction introduced by Nguyen Xuan-Vu, Octavian Susanu, and Daniel Armstrong (arXiv:2508.11366). Instead of de novo product generation or heuristic graph edits on molecular topology, it models reactions as discrete flow matching over electron occupation vectors. The reactant-to-product mapping is formulated as a continuous-time Markov chain (CTMC) over a graph-structured, integer-valued electron occupation space covering all bonding, non-bonding, and hydrogen sites. Using optimal transport, the authors generalize discrete flow matching mixture paths to discrete electron rearrangements, yielding mechanistically interpretable intermediate edit trajectories without requiring annotation of elementary reaction steps. MAELLE achieves performance comparable to leading reaction prediction models on the USPTO-480K benchmark. On two out-of-distribution settings (structural complexity and reaction type), it remains robust where existing methods degrade. Because the learned flow operates on full electron redistribution, MAELLE naturally recovers mechanistic trajectories consistent with known chemistry and can predict reaction by-products.

Paper Overview

Field: Machine Learning
Authors: Nguyen Xuan-Vu, Octavian Susanu, Daniel Armstrong
Published: 2026-08-28
arXiv: 2508.11366

Abstract (English)

Chemical reactions are fundamentally transformations in electron space, yet most machine learning approaches model them either through de novo generation of product molecules or through heuristic graph edits that operate directly on molecular topology. The authors introduce MAELLE (Mechanistic Edit Flow-matching on Electron Rearrangements), which instead models reactions as discrete flow matching over electron occupation vectors.

Concretely, the reactant-to-product mapping is formulated as a Continuous-time Markov Chain (CTMC) over a graph-structured, integer-valued electron occupation space defined on all bonding, non-bonding, and hydrogen sites. To construct intermediate edit trajectories, the discrete flow matching mixture path is generalized to discrete electron rearrangements using Optimal Transport, producing a series of mechanistically interpretable edit moves without requiring annotations of elementary reaction steps.

Key Results

In-distribution performance: MAELLE achieves performance comparable to leading reaction prediction models on the USPTO-480K benchmark.
Out-of-distribution robustness: Evaluated on two out-of-distribution settings — structural complexity and reaction type — MAELLE maintains strong performance where existing methods degrade.
Mechanistic interpretability: Because the learned flow operates on complete electron redistribution, MAELLE naturally recovers mechanistic trajectories consistent with known chemistry and can predict reaction by-products.

*Auto-collected on 2026-08-29.*