The group of Mitsuo UMEZU first prepared a small number of variants of green fluorescent protein (GFP) by conventional random mutagenesis and measured their fluorescent properties in ordert o acquire learning data for artificial intelligence. Next, by Bayesian optimization, an artificial intelligence techniques, they predicted which mutation if introduced would convey the desired function. This made it possible to propose a small variant group („smart hot library“) which abundantly should contain protein variants having the desired function and which could easily be tested. In this study, they succeeded to change GFP to YFP (yellow fluorescent protein) and find new YFPs with higher fluorescence intensity at longer wavelength compared to known YFPs. About 3% of the new YFPs was not contained in a library prepared by conventional random mutagenesis, but about 70% of all YFPs in the AI-supported library had improved properties.

Tohoku University news release, August 31, 2018