Simulating Directed Evolution In Silico with an Iterative, Physics-Based Workflow

Abstract

Abstract Directed evolution consistently delivers engineered enzymes that facilitate diverse biocatalytic transformations. Despite advances in automation and high-throughput workflows, our ability to rapidly engineer novel enzymes is limited by our inability to thoroughly sample the vast protein design space. Generative artificial intelligence (AI) methods have alleviated this challenge to some degree, due to their ability to propose mutations based on learned patterns from complex protein data sets. However, the hidden logic of these AI-based methods can complicate our ability to reveal and exploit fundamental details of these enzymes. To complement these hidden-logic models, we present a structure- and physics-based workflow that mimics directed evolution in silico. This Virtual Protein EngineeR (referred to as ViPER) aims to simulate directed evolution by leveraging computational modeling, molecular dynamics simulations, in silico mutagenesis, and structure-based property evaluation to iteratively engineer an enzyme toward a desired target. In this work, we present a retrospective analysis to demonstrate ViPER’s ability to efficiently navigate the complex protein–protein interfaces associated with the oligomerization of a ketoreductase (KRED) enzyme. In addition to successfully predicting >70% of sites that are crucial for toggling the oligomerization of this KRED enzyme (after only five rounds), this dynamics-based methodology identifies a distal mutation that was incorporated during experiment-based directed evolution. This work provides a novel foundation for computationally driven directed evolution efforts, opening the opportunity to simultaneously pursue and evaluate multiple evolutionary routes through a physics-based lens.