Beyond Structure: Advancing Protein Biophysics Through Molecular Dynamics Simulations and Artificial Intelligence
ABSTRACT
The major paradigm in protein biophysics asserts that “structure determines function,” whereby proteins spontaneously fold into well-structured nanomachines that perform specific mechanical and chemical functions. Consequently, predicting the three-dimensional structure of folded domains for a given protein sequence has been a major challenge in computational biophysics. Last year, the AI-based AlphaFold method won the Nobel Prize by solving the conventional protein folding problem. However, this structure-centric view has been challenged by recent discoveries of intrinsically disordered proteins (IDPs), which perform their functions through the collective behavior of disordered ensembles, without any ordered structure. Bioinformatics studies indicate that approximately half of the human proteome contains IDP domains. Despite their importance, our understanding of IDPs remains significantly limited due to their dynamic and disordered nature, which renders them experimentally elusive and challenging for AI model training. In this presentation, I will demonstrate that computer simulation based on the fundamental principles of physics—specifically, molecular dynamics (MD) simulation—represents the most efficient method for the molecular characterization of IDPs. First, I will assess the state-of-the-art models of MD simulations, highlighting their significant limitations in accuracy across various model systems. Subsequently, I will propose a novel strategy to enhance the accuracy of MD simulations based on thermodynamic principles. Finally, I will illustrate that MD simulations utilizing the newly optimized physics model can achieve the thermodynamic accuracy limit, aligning simulation results with experimental counterparts. Specifically, I will showcase remarkably realistic simulation results for several important categories of biophysical systems, including protein folding, sampling of IDP ensembles, phase separation of IDPs, target searching of DNA-binding proteins, and biological membranes. Further, I will demonstrate that a diffusion-based AI model, trained on simulation data, can reconstruct the entire free energy landscape. These advancements indicate that MD simulations have finally reached the realm of super-accuracy, potentially providing training data for systems that are experimentally inaccessible.
