Molecular Transformer: A Model for Uncertainty-Calibrated Chemical Reaction Prediction.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Artificial Intelligence in Chemistry

Background:

• Organic synthesis is crucial in medicinal chemistry but predicting reaction outcomes (the forward problem) remains a significant challenge.
• Current methods often struggle with accuracy and require handcrafted rules, limiting their applicability.

Purpose of the Study:

• To develop a highly accurate and universally applicable computational model for predicting organic synthesis products.
• To address the unsolved forward problem in reaction planning using machine learning.

Main Methods:

• The study frames reaction prediction as a machine translation task, converting simplified molecular input line-entry system (SMILES) strings of reactants and reagents into product SMILES strings.
• A multihead attention-based model, termed Molecular Transformer, was employed to learn correlations between chemical motifs.

Main Results:

• The Molecular Transformer model achieved a top-1 accuracy exceeding 90% on a benchmark dataset, outperforming existing algorithms.
• The model demonstrated the ability to predict subtle chemical transformations without relying on handcrafted rules.
• It accurately estimates prediction uncertainty, with an 89% accuracy in classifying correct predictions.
• The model successfully handled inputs without a reactant-reagent split and incorporated stereochemistry.

Conclusions:

• The Molecular Transformer provides a powerful, data-driven solution for the organic synthesis forward problem.
• Its high accuracy, uncertainty estimation, and versatility make it a valuable tool for medicinal chemistry and drug discovery.
• The model's ability to generalize and handle complex inputs signifies a major step towards automated synthesis planning.