Active machine learning for reaction condition optimization

November 11, 2020

In a collaboration with the University of Cambridge and the University of Lisbon, we have developed a new software tool ”LabMate.ML” that can optimize organic synthesis conditions through active machine learning.

The tool requires only 5-10 data points as training data and use this data to suggest a new experimental protocol. The success of this protocol is then feedback into the algorithm, and the adaptive learning algorithm will incorporates this knowledge to suggest a further improved protocol.

We tested the tool in nine prospective use cases, optimizing small-molecule, glyco, or protein chemistry. In all cases, the tool found suitable conditions using only 1-10 additional experiments. To contextualize this performance, we asked multiple PhD level chemists to optimize reaction conditions and found that they needed at least as many experiments as our software to find suitable conditions.

Through employing random forest models, the predictive reasoning of our machine learning tool can be easily analyzed by quantifying the importance of different parameters. We found that our tool can rapidly learn known chemical constraints, but also learned novel relationships between the parameters that defied the intuition of dozens of PhD level chemists.

Using AI to optimize reaction conditions is a flourishing field of research currently, with many successful developments from academia and industry. Many of these applications use big data from the chemical and patent literature, or employ advanced laboratory automation and flow chemistry setups. In contrast, our software can be executed on a desktop computer and with extremely limited data.

We hope that this algorithm and extensions thereof can help to streamline laborious chemistry optimization and thereby free human labor for more creative tasks. Additionally, we expect that this concept is applicable to other optimization problems in biology, chemistry, and engineering.

Reker, D., Hoyt, E. A., Bernardes, G. J., & Rodrigues, T. (2020). Adaptive Optimization of Chemical Reactions with Minimal Experimental Information. Cell Reports Physical Science, 100247.

Update on November 12th:

ChemistryWorld has written a short perspective on our article.