Fig. 2: ML-accelerated protein sequence optimization.

a An overview of our sequence space search strategy. We first initialize the search by designing a diverse set of sequences that broadly sample the landscape. We then iterate through multiple design-test-learn cycles to efficiently understand and optimize in vivo fatty alcohol production. b Sequence space visualization over ten rounds of UCB optimization (each round is shown as a different color). The three parent enzymes are found at the vertices of this chimeric sequence space and all chimeras fall within the parents’ envelope. The UCB optimization started by broadly sampling the landscape, but quickly converged on highly active regions. c The in vivo fatty alcohol titers over the course of the sequence optimization. Each point depicts an individual sequence’s mean fatty alcohol production in the sequence optimization phase and the horizontal gray bars represent the average titer during that round of sequence optimization. The mean, standard deviation and number of replicates, where n is equal to the number of cultures analyzed (each one from an individual colony), are shown in Supplementary Table 5. Source data underlying Fig. 2c are provided as a Source Data file.