Abstract
A chemical reaction mechanism (CRM) is a sequence of molecular-level events involving bond-breaking/forming processes, generating transient intermediates along the reaction pathway as reactants transform into products. Understanding such mechanisms is crucial for designing and discovering new reactions. One of the currently available methods to probe CRMs is quantum mechanical (QM) computations. The resource-intensive nature of QM methods and the scarcity of mechanism-based datasets motivated us to develop reliable ML models for predicting mechanisms. In this study, we created a comprehensive dataset with seven distinct classes, each representing uniquely characterized elementary steps. Subsequently, we developed an interpretable attention-based GNN that achieved near-unity and 96% accuracy, respectively for reaction step classification and the prediction of reactive atoms in each such step, capturing interactions between the broader reaction context and local active regions. The near-perfect classification enables accurate prediction of both individual events and the entire CRM, mitigating potential drawbacks of Seq2Seq approaches, where a wrongly predicted character leads to incoherent CRM identification. In addition to interpretability, our model adeptly identifies key atom(s) even from out-of-distribution classes. This generalizabilty allows for the inclusion of new reaction types in a modular fashion, thus will be of value to experts for understanding the reactivity of new molecules.
Introduction
Understanding Chemical Reaction Mechanisms (CRMs)—the step-by-step sequence of bond-breaking and bond-forming events—is fundamental to catalyst design and process optimization. Traditionally, researchers have relied on Quantum Mechanical (QM) simulations to probe these pathways. However, QM methods are computationally expensive and slow. While machine learning has advanced "forward synthesis" (predicting final products), many existing models treat reactions as "black boxes." Sequence-to-sequence (Seq2Seq) models using SMILES strings often suffer from "hallucinations" where minor character errors produce chemically invalid molecules. Furthermore, these models typically struggle with out-of-distribution (OOD) samples that differ structurally from the training data.
The ReactAIvate Framework
To address these challenges, we developed ReactAIvate, an interpretable attention-based Graph Neural Network (GNN) that predicts full reaction mechanisms by decomposing them into individual elementary steps. Unlike string-based models, ReactAIvate treats molecules as graphs and performs two concurrent tasks: Reaction Step Classification (RSC) and Reactive Atom Identification (RAI).
The model utilizes a "supernode" for graph-level classification (RSC) to identify which of seven fundamental mechanistic steps (e.g., oxidative addition, reductive elimination) is occurring. Simultaneously, individual node embeddings are used to pinpoint the specific atoms where bonds are breaking or forming (RAI). By predicting the type of reaction and then applying a corresponding chemical template, ReactAIvate ensures that generated intermediates are always chemically valid.
Data and Training
We curated a first-of-its-kind dataset comprising 100,000 elementary mechanistic steps derived from three major transition metal-catalyzed reaction families: Suzuki-Miyaura (SMC), Buchwald-Hartwig (BHA), and Kumada (KC) couplings. To test generalization, we also developed an OOD test set of 1,000 samples with structural components entirely absent from the training data.
A key technical innovation is the use of a weighted Binary Cross-Entropy loss. Since most atoms in a molecule are non-reactive, this weighting prevents the model from becoming biased toward "non-reactive" predictions, ensuring high sensitivity for the RAI task.
Results and Interpretability
ReactAIvate achieved near-perfect performance on in-distribution data, with 100% accuracy in reaction step classification and ~96% accuracy in identifying reactive atoms. Most notably, the model demonstrated exceptional robustness on OOD samples, maintaining a 95.7% accuracy for CRM prediction. In comparison, state-of-the-art baselines like T5Chem and traditional Transformers saw their accuracy plummet to less than 1% on the same OOD tasks.
The inclusion of the RAI task significantly enhances the model's interpretability. As shown in the figure above, forcing the model to identify reactive atoms causes the GNN's attention to center around the actual reaction sites, mirroring the logic a human chemist would use. This modular generalization allows the model to identify reactive centers even in novel molecules it has never encountered before.
Conclusion
ReactAIvate offers a lightweight (0.87M parameters) and efficient alternative to traditional CRM prediction methods. By combining rule-based templates with deep learning-based atom identification, it bypasses the syntax errors of Seq2Seq models while providing chemically grounded insights. Future work will focus on expanding the dataset to include a broader range of organic reactions and moving toward dynamic product generation without fixed templates.
Summary updated