Accelerated Chemical Reaction Optimization Using Multi-Task Learning.

Area of Science:

• Medicinal Chemistry
• Chemical Synthesis
• Machine Learning

Background:

• C-H bond functionalization is crucial for drug discovery but challenging with polar groups.
• Bayesian optimization (BO) has shown promise for self-optimizing chemical reactions.
• Previous BO methods lacked prior reaction knowledge.

Purpose of the Study:

• To explore multitask Bayesian optimization (MTBO) for accelerating new reaction optimization using historical data.
• To apply MTBO to real-world medicinal chemistry problems, specifically optimizing pharmaceutical intermediates.
• To demonstrate MTBO's efficiency and cost-saving potential compared to traditional methods.

Main Methods:

• Utilized multitask Bayesian optimization (MTBO) with historical reaction data for *in silico* studies.
• Implemented MTBO on an autonomous flow-based reactor platform for medicinal chemistry applications.
• Tested MTBO on C-H activation reactions with diverse substrates.

Main Results:

• MTBO successfully accelerated the optimization of new reactions by leveraging past experimental data.
• The methodology efficiently optimized yields of pharmaceutical intermediates using flow chemistry.
• MTBO identified optimal conditions for unseen C-H activation reactions, proving its versatility.

Conclusions:

• MTBO is an effective tool for accelerating reaction optimization in medicinal chemistry.
• This approach represents a significant advancement in utilizing data and machine learning for chemical synthesis.
• MTBO offers potential for substantial cost reductions in process optimization.