Leveraging Language Model Multitasking To Predict C-H Borylation Selectivity.

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Medicinal Chemistry

Background:

• C-H borylation is crucial for synthesizing pharmaceutical lead candidates due to versatile downstream coupling reactions.
• Regioselectivity prediction in C-H borylation, particularly for complex drug-like molecules with heterocycles, remains a significant challenge.

Purpose of the Study:

• To evaluate the efficacy of a language model, T5Chem, for predicting C-H borylation reaction outcomes.
• To assess the model's performance in both product generation and site reactivity classification for borylation reactions.

Main Methods:

• Utilized a dataset of borylation reactions sourced from Reaxys.
• Fine-tuned a T5Chem multitask language model, originally trained on USPTO_500_MT patent data.
• Evaluated the model's ability to predict reaction products and classify reactive C-H bonds.

Main Results:

• The fine-tuned T5Chem model achieved 79% accuracy in generating the correct C-H borylation product.
• The model demonstrated high performance in classifying reactive aromatic C-H bonds, with 95% accuracy and 88% positive predictive value.
• Performance surpassed that of specialized graph-based neural networks.

Conclusions:

• Language models, specifically T5Chem, offer a powerful and accurate approach for predicting C-H borylation regioselectivity.
• This computational strategy can aid in the efficient design and synthesis of pharmaceutical compounds.
• The model's predictive capabilities exceed current graph-based methods, paving the way for improved reaction prediction tools.