A transfer learning approach for reaction discovery in small data situations using generative model.

Area of Science:

• Chemical Sciences
• Drug Discovery
• Sustainable Chemistry

Background:

• Reaction discovery is often heuristic, time-consuming, and resource-intensive.
• Fluorine-containing compounds, vital in pharmaceuticals, are synthesized via deoxyfluorination of alcohols.
• Machine learning (ML) offers potential for optimizing reaction discovery with limited data.

Purpose of the Study:

• To demonstrate the application of ML in discovering new chemical reactions for sustainable synthesis.
• To develop a deep generative model for exploring chemical space and identifying novel alcohol molecules.
• To enhance the efficiency of reaction discovery pipelines.

Main Methods:

• Utilized a recurrent neural network-based deep generative model.
• Trained the model on a small library of 37 alcohol molecules.
• Focused on generating novel, synthetically accessible, and higher-yielding alcohol candidates.

Main Results:

• The ML model effectively learned from a small dataset.
• Generated high-quality, novel alcohol molecules with potential for higher yields.
• Demonstrated the feasibility of using ML for exploring chemical space in reaction discovery.

Conclusions:

• Interdisciplinary approaches combining ML and chemistry can advance sustainable practices.
• This ML protocol shows promise for practical application in reaction discovery pipelines.
• The model's success with limited data highlights its utility for discovering valuable chemical entities.