Enabling late-stage drug diversification by high-throughput experimentation with geometric deep learning.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Chemical Engineering

Background:

• Late-stage functionalization (LSF) is crucial for optimizing drug properties.
• The chemical complexity of drug molecules presents significant challenges for LSF.
• Efficient diversification strategies are needed to accelerate drug discovery.

Purpose of the Study:

• To develop a novel LSF platform integrating geometric deep learning and high-throughput experimentation.
• To accurately predict reaction yields and regioselectivity for borylation reactions.
• To identify structural diversification opportunities in diverse commercial drug molecules.

Main Methods:

• Development of a geometric deep learning computational model for reaction prediction.
• High-throughput reaction screening to validate computational predictions.
• Application of the platform to a diverse set of 23 commercial drug molecules.
• Quantification of steric and electronic influences on model performance.

Main Results:

• The computational model predicted reaction yields with a 4-5% mean absolute error.
• Reactivity prediction achieved 92% balanced accuracy for known substrates and 67% for unknown substrates.
• Regioselectivity of major products was captured with a 67% F-score.
• Numerous structural diversification opportunities were identified across 23 drug molecules.

Conclusions:

• The developed platform effectively integrates deep learning and high-throughput experimentation for LSF.
• The platform enables efficient identification of diversification strategies for drug candidates.
• A user-friendly reaction format facilitates seamless integration of computational and experimental approaches.