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

Directed evolution (DE) mimics natural selection to improve the functions of a target protein. Machine learning (ML) has significantly streamlined DE by aiding in several steps, which includes identifying starting variants, generating diverse libraries and modeling sequence-fitness relationships. To date, the majority of ML-assisted DE (MLDE) approaches has relied predominantly on sequence information due to the challenges and cost of obtaining protein structure information. Here, we introduce a structure-augmented MLDE (saMLDE) approach for selecting high fitness variants from a library of Protein G B1 domain. We adopted and applied a zero-shot sequence-based prediction method (offering the potential to discover new insights without extensive training data) to select an initial training library of 96 variants for the saMLDE campaign. To leverage protein structure information, we used protein structure prediction with AlphaFold2 and molecular docking simulations performed with Rosetta FlexPepDock, resulting in structure-based features derived with an induced fit model. After three rounds of the saMLDE campaign, we demonstrated that saMLDE incorporating structural information gradually improves the average fitness scores and the precision of predicted binders. In addition, we found that the initial library selection with zero-shot subset selection methods significantly impacted the average fitness scores and precision, consequently influencing the overall directed evolutionary trajectories.