Unified Deep Learning Model for Multitask Reaction Predictions with Explanation.

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Organic synthesis

Background:

• Traditional machine learning models for organic chemistry synthesis are often task-specific.
• Developing unified models can improve efficiency and robustness in reaction prediction.

Purpose of the Study:

• To develop a unified deep learning model, T5Chem, for various chemical reaction prediction tasks.
• To introduce a new unified multitask reaction prediction dataset, USPTO_500_MT.
• To enhance the interpretability of deep learning models in chemistry.

Main Methods:

• Adapted the "Text-to-Text Transfer Transformer" (T5) framework from natural language processing.
• Utilized self-supervised pretraining with PubChem molecules.
• Introduced the USPTO_500_MT dataset for training and testing five reaction tasks.
• Employed SHAP (SHapley Additive exPlanations) for model interpretability.

Main Results:

• T5Chem achieved state-of-the-art performance on four distinct reaction prediction tasks (classification, forward prediction, retrosynthesis, yield prediction).
• Models trained on multiple tasks demonstrated increased robustness and benefited from cross-task learning.
• The new USPTO_500_MT dataset facilitates multitask learning for reaction prediction and reagent suggestion.

Conclusions:

• T5Chem offers a unified and robust approach to machine learning-assisted organic synthesis.
• Multitask learning enhances model performance and generalization.
• SHAP provides valuable insights into T5Chem's predictions, demystifying deep learning in chemistry.