Interpretable Deep Learning for Polar Mechanistic Reaction Prediction

TL;DR

The paper tackles interpretability in chemical reaction prediction by moving from black-box product outcomes to mechanistic, polar elementary steps using the PMechDB dataset. It introduces PMechRP, a hybrid system that fuses a $5$-ensemble Chemformer with a two-step Siamese framework to enforce arrow-pushing plausibility and filter alchemical outputs. Augmenting training with ~48.8M combinatorial proton-transfer reactions and benchmarking against a human pathway dataset enables robust evaluation of forward prediction and pathway recovery. The results show top-10 accuracies near the mid-to-high 90s and pathway target recovery around $84.9\%$, demonstrating a practical path toward interpretable, mechanism-aware reaction prediction.

Abstract

Accurately predicting chemical reactions is essential for driving innovation in synthetic chemistry, with broad applications in medicine, manufacturing, and agriculture. At the same time, reaction prediction is a complex problem which can be both time-consuming and resource-intensive for chemists to solve. Deep learning methods offer an appealing solution by enabling high-throughput reaction prediction. However, many existing models are trained on the US Patent Office dataset and treat reactions as overall transformations: mapping reactants directly to products with limited interpretability or mechanistic insight. To address this, we introduce PMechRP (Polar Mechanistic Reaction Predictor), a system that trains machine learning models on the PMechDB dataset, which represents reactions as polar elementary steps that capture electron flow and mechanistic detail. To further expand model coverage and improve generalization, we augment PMechDB with a diverse set of combinatorially generated reactions. We train and compare a range of machine learning models, including transformer-based, graph-based, and two-step siamese architectures. Our best-performing approach was a hybrid model, which combines a 5-ensemble of Chemformer models with a two-step Siamese framework to leverage the accuracy of transformer architectures, while filtering away "alchemical" products using the two-step network predictions. For evaluation, we use a test split of the PMechDB dataset and additionally curate a human benchmark dataset consisting of complete mechanistic pathways extracted from an organic chemistry textbook. Our hybrid model achieves a top-10 accuracy of 94.9% on the PMechDB test set and a target recovery rate of 84.9% on the pathway dataset.

Figures (9)

• Figure 1: Example of an overall transformation vs an elementary step approach. This is a the final reaction step in the synthesis of enzalutamide, a drug used to treat prostate cancer that generates over $6 billion a year in revenue zhou2017improved.
• Figure 2: A side reaction occurring at an intermediate step in the synthesis of the autoimmune drug Deucravacitinib, generated unwanted side products due to competing addition of chloride anion to the key NO$_{2}^+$ intermediate. This led to a decrease in overall purity of the products treitler2022development.
• Figure 3: The first pathway generated for mesylation of an alcohol with methanesulfonyl chloride.
• Figure 4: The first pathway generated for an aldol condensation under acidic conditions. As with most reactions involving proton transfers, other mechanistic variations are plausible. For example, the formation of the oxonium and enol in step 2 could have been depicted as a two-step process (e.g., step 2a and 2b, not shown) using the bisulfate anion to affect the proton transfer.
• Figure 5: Pipeline for combinatorial reaction generation.