Mechanism-Aware Deep Learning for Polar Reaction Prediction.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Accurate chemical reaction prediction is vital for synthetic chemistry innovation across various industries.
• Current deep learning models often lack mechanistic insight, treating reactions as simple input-output transformations.
• Existing models are typically trained on limited datasets, hindering generalization.

Purpose of the Study:

• To develop advanced deep learning models for accurate and mechanistically detailed chemical reaction prediction.
• To improve the generalization capabilities of reaction prediction models by augmenting training data.
• To provide interpretable predictions and mechanistic insights beyond simple product identification.

Main Methods:

• Introduction of PMechRP (Polar Mechanistic Reaction Predictor) trained on PMechDB, a dataset of polar elementary steps.
• Augmentation of PMechDB with combinatorially generated reactions to enhance model coverage.
• Development of ArrowFinder for direct prediction of arrow-pushing mechanisms.
• Implementation of hybrid pipelines combining transformer (Chemformer) and Siamese architectures.

Main Results:

• A hybrid pipeline combining Chemformer ensembles and a two-stage Siamese network achieved high predictive accuracy.
• The developed models successfully filtered "alchemical" products and provided mechanistic annotations.
• Performance was validated across multiple benchmarks, including PMechDB, USPTO, and a human-curated textbook dataset.

Conclusions:

• The developed PMechRP and ArrowFinder models significantly advance the state-of-the-art in chemical reaction prediction.
• The hybrid approach offers accurate, interpretable, and mechanistically detailed predictions.
• This work paves the way for more efficient and insightful synthetic chemistry research.