基于图结构电子占据离散流匹配的机理反应预测
文章背景与核心概要
化学反应的本质是电子空间的转化，然而现有的绝大多数机器学习模型主要依赖启发式图编辑或产物分子的从头生成（de novo generation）。本文介绍了 MAELLE（MechAnistic Edit fLow-matching on e*Lectron rE*arrangements），这是一种全新的框架，它将化学反应构想为图结构电子占据向量上的离散流匹配过程。
该方法的核心亮点包括：通过连续时间马尔可夫链（CTMC）在整数值电子占据空间（涵盖成键、非成键和氢位点）上建模反应物到产物的映射；利用最优传输将离散流匹配混合路径推广至电子重排，无需显式的基元步骤标注；在 USPTO-480K 基准测试中展现出卓越的分布外（OOD）鲁棒性与准确性；并且由于学习到的流直接作用于全局电子再分配，MAELLE 能够自然地重建符合已知化学机理的反应轨迹并成功预测副产物。
Summary
Chemical reactions are fundamentally transformations in electron space, yet most machine learning models rely on heuristic graph edits or the de novo generation of product molecules. This paper introduces MAELLE (MechAnistic Edit fLow-matching on e*Lectron rE*arrangements), a novel framework that frames chemical reactions as discrete flow matching over graph-structured electron occupation vectors.
Key highlights of the approach include: * Continuous-time Markov Chain (CTMC): Models reactant-to-product mapping as a CTMC over integer-valued electron occupation spaces (covering bonding, non-bonding, and hydrogen sites). * Optimal Transport Mixures: Generalizes discrete flow matching mixture paths to electron rearrangements without requiring explicit elementary step annotations. * Robust Out-of-Distribution (OOD) Performance: Evaluated on the USPTO-480K benchmark, MAELLE maintains high accuracy and strong robustness even when existing methods degrade in structural complexity and alternative reaction types. * Mechanistic Interpretability: Because the learned flow operates across the full electron redistribution, MAELLE naturally reconstructs reaction trajectories aligned with known chemical mechanisms and successfully predicts reaction side products.
化学反应的本质是电子空间的转化，然而大多数机器学习模型要么通过产物分子的从头生成来对其进行建模，要么通过直接对分子拓扑进行操作的启发式图编辑来建模。
Chemical reactions are fundamentally transformations in electron space, yet most machine learning approaches model them either through \textit{de novo} generation of product molecules or through heuristic graph edits that operate directly on molecular topology.
我们引入了 MAELLE（MechAnistic Edit fLow-matching on e*Lectron rE*arrangements），该方法转而将反应建模为电子占据向量上的离散流匹配。
We introduce MAELLE (\textbf{M}ech\textbf{A}nistic \textbf{E}dit f\textbf{L}ow-matching on e\textbf{L}ectron r\textbf{E}arrangements), which instead models reactions as discrete flow matching over electron occupation vectors.
具体而言，我们将反应物到产物的映射构建为在所有成键、非成键和氢位点上定义的图结构整数值电子占据空间上的连续时间马尔可夫链（CTMC）。
Concretely, we formulate the reactant-to-product mapping as a Continuous-time Markov Chain (CTMC) over the graph-structured integer-valued electron occupation space defined on all bonding, non-bonding, and hydrogen sites.
为了构建中间编辑轨迹，我们利用最优传输将离散流匹配混合路径推广至离散的电子重排，从而在不需要基元步骤标注的情况下产生一系列具有机理可解释性的编辑动作。
To construct the intermediate edit trajectories, we generalize the discrete flow matching mixture path to discrete electron rearrangements using Optimal Transport, yielding a sequence of mechanistically interpretable edit moves without requiring elementary step annotations.
与领先的反应预测模型相比，MAELLE 在 USPTO-480K 基准测试上取得了具有竞争力的性能。
MAELLE achieves competitive performance on the USPTO-480K benchmark compared with leading reaction prediction models.
除了分布内准确率外，我们还评估了结构复杂度和反应类型这两种分布外（OOD）设置下的鲁棒性，发现 MAELLE 在现有方法性能下降的地方仍能保持强大的性能。
Beyond in-distribution accuracy, we evaluate robustness across two out-of-distribution settings - structural complexity and reaction type - and find that MAELLE maintains strong performance where existing methods degrade.
最后，由于学习到的流操作覆盖了完整的电子再分配，MAELLE 能够自然地恢复符合已知化学规律的机理轨迹，并能预测反应的副产物。
Finally, because the learned flow operates over the full electron redistribution, MAELLE naturally recovers mechanistic trajectories that align with known chemistry and can predict side products of a reaction.
Submission History
- [v1] Thu, 27 Aug 2026 17:50:44 UTC (1,078 KB)