Development of automated computational methods for the redesign of protein dynamics using biomolecular simulations and machine learning

Abstract

Proteins are responsible for almost all biological mechanisms, and their three-dimensional structures and dynamics define their function. In recent years, outstanding advances have been made in protein design. However, redesigning protein dynamics to achieve desired properties or states remains a significant challenge in computational protein design. This thesis addresses this gap by introducing three novel toolkits—MDSubSampler, MDAutoMut, and MDAutoPredict—developed to integrate biomolecular simulations with machine learning for the automated redesign of protein dynamics. MDSubSampler is designed to preprocess and a posteriori subsample molecular dynamics simulations, preserving critical dynamic information while reducing noise and data complexity. Its application demonstrates effective noise reduction and compatibility with machine learning workflows, validated using adenylate kinase as a model system. MDAutoMut automates mutation generation, simulation, and analysis, facilitating systematic identification of mutations that have a desired impact on protein dynamics. This toolkit successfully identifies mutations on adenylate kinase structure shifting dynamics towards a closed conformation, validated by literature benchmarks. MDAutoPredict extends the workflow by using machine learning models to predict conformational states from molecular dynamics data, offering an adaptable framework for dynamic state prediction. These contributions represent an advance in computational protein design, providing scalable, automated solutions for mutation engineering and dynamic prediction. The toolkits are modular, extensible, and integrated with well-consolidated libraries, ensuring broad applicability across protein engineering challenges. This research highlights the potential of combining biomolecular simulations with machine learning to redesign protein dynamics and sets the stage for future innovations in computational biology.