Researchers at iMed.ULisboa, Instituto de Medicina Molecular (iMM), University of Cambridge and the Massachusetts Institute of Technology (MIT) have developed an algorithm harnessing adaptive artificial intelligence to iteratively optimize synthesis protocols in chemistry.
The LabMate.ML software was conceived to help bench chemists prioritize the next experiment in a statistically motivated manner, while generating new reactivity insights that would be overlooked by expert intuition, even in small datasets.
The study “Adaptive Optimization of Chemical Reactions with Minimal Experimental Information” was published on November 11th at Cell Reports Physical Science from the prestigious Cell Press journal family. The lead author, iMed.ULisboa researcher Tiago Rodrigues, mentions “we provided initial proof-of-concept that machine learning methods have wide applicability in chemical syntheses beyond big data, which is inexistent or hardly found in real-world scenarios. The algorithm we report runs on personal computers and uses active learning, with a full and autonomous self-optimization cycle to provide predictions on the most promising experiment to carry out next.” As detailed in the study, the algorithm is not only efficient in low data regimes, but also highly competitive with expert intuition. In a series of double blind tests, LabMate.ML outperformed PhD-level chemists in the identification of optimized synthesis protocols.
Tiago Rodrigues expects that this and related technologies “can be fully integrated in closed loop cycles to free up human brainpower to other tasks, reduce the number of redundant or non-informative experiments and feedstocks”. The researchers are now further improving and translating the technology in collaboration with an international company in the pharmaceutical space.