Fuelling the Digital Chemistry Revolution with Language Models.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Synthetic Organic Chemistry

Background:

• The RXN for Chemistry project, initiated in 2017 by IBM Research, aimed to integrate data-driven methodologies into synthetic organic chemistry.
• Traditional chemical research often involves extensive manual experimentation and data analysis.
• The need for efficient, scalable, and digitized chemical research processes is paramount in modern scientific endeavors.

Purpose of the Study:

• To develop digital assets using machine learning (ML) for synthetic organic chemistry.
• To treat chemical reaction data as language, enabling prediction tasks analogous to language translation.
• To facilitate the digitalization of chemistry across various industrial sectors.

Main Methods:

• Developed language models by treating chemical reaction data as language records.
• Applied ML for tasks including forward reaction prediction, retrosynthesis, and reaction classification.
• Incorporated biochemical data for training models focused on sustainable chemical reactions.

Main Results:

• Successfully created language models for diverse applications: reaction prediction, retrosynthesis, atom-mapping, and procedure extraction.
• Enabled the inference of experimental protocols and programming of automated chemical laboratories.
• Demonstrated ease of model construction and enhancement via data augmentation with minimal human intervention.

Conclusions:

• The project successfully digitized synthetic organic chemistry through data-driven ML methodologies.
• The developed language models have been widely adopted, facilitating advancements in pharmaceuticals and chemical manufacturing.
• The approach promotes greener and more sustainable chemical reactions, contributing to the prestigious Sandmeyer Award in 2022.