Protein Language Models in Directed Evolution

The dominant paradigms for integrating machine-learning into protein engineering are de novo protein design and guided directed evolution. Guiding directed evolution requires a model of protein fitness, but most models are only evaluated in silico on datasets comprising few mutations. Due to the limited number of mutations in these datasets, it is unclear how well these models can guide directed evolution efforts. We demonstrate in vitro how zero-shot and few-shot protein language models of fitness can be used to guide two rounds of directed evolution with simulated annealing. Our few-shot simulated annealing approach recommended enzyme variants with 1.62 x improved PET degradation over 72 h period, outperforming the top engineered variant from the literature, which was 1.40 x fitter than wild-type. In the second round, 240 in vitro examples were used for training, 32 homologous sequences were used for evolutionary context and 176 variants were evaluated for improved PET degradation, achieving a hit-rate of 39 % of variants fitter than wild-type.