Designing Target-specific Data Sets for Regioselectivity Predictions on Complex Substrates

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Predicting regioselectivity in C(sp3)-H functionalization is crucial for synthetic chemistry.
• Current methods often rely on intuitive extrapolation from model substrates, limiting accuracy for complex molecules.

Purpose of the Study:

• To develop machine learning models for predicting C(sp3)-H functionalization regioselectivity.
• To investigate the efficacy of active learning strategies in curating efficient datasets for model training.
• To provide a quantitative, data-driven alternative to traditional reactivity prediction methods.

Main Methods:

• Curated a dataset for dioxirane oxidations from existing literature.
• Developed and compared various acquisition functions for active learning-based dataset selection.
• Leveraged predicted reactivity and model uncertainty in active learning acquisition functions.
• Experimentally validated the workflow on complex substrates and C-H radical borylation.

Main Results:

• Active learning acquisition functions incorporating reactivity and uncertainty outperformed similarity-based methods.
• Dataset curation using acquisition functions significantly reduced the number of required data points.
• Machine-designed, smaller datasets achieved accurate predictions where larger, randomly selected datasets failed.
• The developed workflow demonstrated applicability to predicting regioselectivity in arene C-H radical borylation.

Conclusions:

• Machine learning models, particularly when trained on active learning-curated datasets, offer a powerful and efficient approach to predicting C-H functionalization regioselectivity.
• This data-driven methodology provides a quantitative and reliable alternative to traditional, often less accurate, methods for complex molecules.
• The developed workflow streamlines the prediction process, reducing experimental effort and improving accuracy.