Breaking barriers: A machine learning approach to efficiently explore the free energy surface of protein-surface systems

Understanding protein-surface interactions is crucial for the rational design of biotechnologies such as medical implants, biological sensors, and drug-delivery vehicles. Molecular dynamics (MD) simulations are excellent for this task due to their atomistic resolution, but the sizable free energy barriers between stable adsorbed protein conformations can limit sampling. To address this challenge, a framework developed by Wei et. al, titled Molecular Enhanced Sampling with Autoencoders (MESA), accelerates the sampling of protein conformations in isotropic environments (e.g., bulk solvent) using a machine learning architecture called an autoencoder. The model learns low-dimensional representations of the configurations sampled in MD simulations and directs future simulations to sample unexplored regions of the configuration space. A challenge in directly implementing this method to surface systems is that the model cannot distinguish conformations based on the orientation of the protein with the surface. In this work, we extend the MESA framework to efficiently sample adsorbed protein conformations through a modified training procedure and simulation protocol. We demonstrate the approach through studies of peptides on graphene and silica surfaces. We identify the various molecular driving forces governing peptide-surface interactions. Collectively, our method highlights the power of combining ML with molecular simulations in studying biomolecular systems near surfaces.