Augmentation of Structure Information to the Sequence-Based Machine Learning-Assisted Directed Protein Evolution

Directed evolution (DE) adapts the evolutionary process of the nature to improve the functions of a target protein. Machine learning (ML) has contributed to many steps in DE to identify starting variants, to generate a pool of variants, to map sequence to fitness and to optimize sequence-function models. To date, the majority of ML-assisted DE (MLDE) approaches has utilized exclusively sequence information for their DE campaigns due to the challenges and cost associated with generating protein structure information. Here, we examine two examples of a structure-informed MLDE approach to select high fitness variants from a library of Protein G B1 domain. We adapted and applied the zero-shot sequence prediction to select an initial training set of 96 variants for our MLDE campaign. To generate structure-based input features for use in ML model training, we leveraged protein structure prediction with AlphaFold2 and molecular docking with Rosetta FlexPepDock. After three rounds of MLDE campaign, we demonstrated that leveraging structure information has the potential to improve both the average and the maximum fitness scores of predicted variants when compared to predictions made with sequence information alone. In addition, we found that zero-shot sampling methods could significantly affect the average fitness scores, the maximum fitness score and the number of predicted variants.