[Paper Review] ReactAIvate: A Deep Learning Approach to Predicting Reaction Mechanisms and Unmasking Reactivity Hotspots

Ajnabiul Hoque, Manajit Das+2 more|arXiv (Cornell University)|Jul 14, 2024

Machine Learning in Materials Science4 citations

TL;DR

ReactAIvate is a novel interpretable graph neural network (GNN) model that predicts elementary reaction steps and identifies reactive atoms in transition-metal-catalyzed reactions with 96% accuracy in reactive atom detection and near-unity accuracy in step classification. It uses a dual-loss training scheme (graph-level and node-level) on a new dataset of seven mechanistic steps, enabling accurate, modular CRM prediction and robust out-of-distribution generalization for expert-guided reactivity analysis.

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.

Motivation & Objective

• To address the lack of comprehensive, mechanism-specific datasets for chemical reaction mechanisms (CRMs), especially for multi-step, transition-metal-catalyzed reactions.
• To overcome the limitations of sequence-to-sequence (Seq2Seq) models in CRM prediction, which are error-sensitive and prone to generating incoherent mechanisms due to single-character errors.
• To develop a model that not only classifies elementary steps but also identifies reactive atoms, offering chemically interpretable insights for experts.
• To enable modular extension to new reaction types by supporting out-of-distribution (OOD) class prediction with meaningful attention guidance.
• To provide a reliable, fast, and interpretable alternative to computationally expensive quantum mechanical (QM) methods for CRM prediction.

Proposed method

• The model employs a graph-attention-based GNN that processes molecular graphs of reactants, reagents, and catalysts as input, capturing both local and global structural features.
• It uses a dual-loss training objective: a graph-level loss for classifying the elementary reaction step (7 classes) and a node-level loss for identifying reactive atoms within each step.
• Attention mechanisms are explicitly guided by the node-level loss, ensuring that attention is concentrated on chemically relevant reactive centers, such as Pd centers and bond-breaking/forming sites.
• The model is trained on a newly curated dataset of 3 catalytic reaction families with 7 distinct elementary steps, including oxidative addition, transmetallation, and reductive elimination.
• For OOD generalization, the model assigns a dedicated 'no reaction' class (S8) and visualizes attention to guide experts on potential reactivity even when no reaction occurs.
• The framework enables full CRM reconstruction by sequentially predicting elementary steps from reactants and catalysts, avoiding the error propagation issues of SMILES-based sequence generation.

Figure 1: (a) A representative example of oxidative addition template, (b) complete workflow of our proposed ReactAIvate method, (c) process of reaction step classification and reactive atom identification using ReactAIvate

Experimental results

Research questions

• RQ1 Can a deep learning model accurately classify elementary reaction steps in multi-step catalytic mechanisms with high precision and interpretability?
• RQ2 How well can a GNN-based model identify reactive atoms within each elementary step, and does this improve mechanistic interpretability compared to standard sequence models?
• RQ3 To what extent can the model generalize to out-of-distribution (OOD) reaction pairs that are non-reactive or involve novel mechanisms?
• RQ4 Does the dual-loss training (graph-level + node-level) enhance attention localization on true reactive centers, aligning with chemist intuition?
• RQ5 Can the model serve as a reliable, modular tool for predicting full CRMs without relying on error-prone SMILES sequence generation?

Key findings

• The model achieves near-unity accuracy in classifying elementary reaction steps, demonstrating exceptional reliability in identifying the correct mechanistic transformation.
• The model attains 96% accuracy in identifying reactive atoms, with attention maps clearly highlighting key reactive centers such as Pd, Cl, and C atoms involved in bond cleavage or formation.
• Attention visualization confirms that the model focuses on chemically relevant regions, such as the Pd center and aryl chloride bond, validating its interpretability and alignment with expert reasoning.
• For OOD samples, the model correctly assigns the 'no reaction' class (S8) and still localizes attention on potential reactive sites, providing actionable insights for expert analysis.
• The model successfully reconstructs complete CRMs for complex reactions like Kumada coupling, correctly predicting all three steps (oxidative addition, transmetallation, reductive elimination) in sequence with accurate reactive atom identification.
• The model outperforms Seq2Seq baselines by avoiding catastrophic failure from single-character errors in SMILES generation, proving more robust for CRM prediction.

Figure 2: Effect of the inclusion of node-level loss in ReactAIvate demonstrated through attention visualization. The rightmost bar represents min-max rescaled attention values.