RXN

RXN employs the Molecular Transformer to predict retrosynthetic routes and chemical reactions, supporting digital chemistry tasks such as reaction prediction and experimental protocol conversion.

Key Features:

• Molecular Transformer Architecture: The core model is an extension of the Molecular Transformer that processes chemical reactions for forward and retrosynthetic prediction.
• Hyper-Graph Exploration Strategy: A hyper-graph exploration strategy dynamically constructs and expands hypergraphs for automatic retrosynthesis route planning.
• Bayesian-like Filtering: Bayesian-like probabilities are used to filter and prioritize proposed retrosynthetic pathways.
• Single-Step Retrosynthetic Model: A single-step retrosynthetic model predicts reactants, reagents, solvents, and catalysts for each transformation step.
• Evaluation Metrics: Model performance is evaluated using coverage, class diversity, round-trip accuracy, and Jensen-Shannon divergence focusing on forward prediction and reaction classification.

Scientific Applications:

• Retrosynthesis Pathway Prediction: Predicts retrosynthetic pathways to support synthetic route planning.
• Chemical Reaction Prediction: Predicts reaction outcomes to aid mechanism analysis and experimental optimization.
• Experimental Protocol Conversion: Converts experimental protocols into actionable sequences for laboratory workflows and reproducibility.
• Automation of Chemical Syntheses: Automates compilation and execution steps for chemical syntheses.

Methodology:

Uses a transformer-based (Molecular Transformer) architecture for chemical data processing; applies a hyper-graph exploration strategy with dynamic graph construction and Bayesian-like probabilistic filtering; implements a single-step retrosynthetic prediction model; and assesses performance using coverage, class diversity, round-trip accuracy, Jensen-Shannon divergence and examples drawn from literature and academic exams.