Our next Lunch & Learn will be held Thursday, February 6, in the Interaction Zone. Lunch will be served at 12:00 noon followed by a talk from 12:30 to 1:30 pm.
The speaker for this event will be Lan Yang from the Gumbart Group
Title: Machine Learning Collective Variables for Protein Unfolding: Enhanced All-Atom Sampling Driven by Coarse-Graining Data
Abstract: Autotransporters play a key role in Gram-negative bacterial infections by facilitating the secretion of their passenger domains across the outer membrane without requiring chemical energy. Although the mechanism of this secretion is not fully understood, studies suggest that it may be closely linked to the folding of the passenger domain in the extracellular space. In this study, we use pertactin, a virulence factor of Bordetella pertussis, as a model protein to investigate the physical principles underlying passenger-domain folding. Pertactin has a structure composed of a β-helix with 18 turns. Using all-atom molecular dynamics (MD) simulations and coarse-grained models, we sample the unfolding processes of the approximately 100 residues at both the N- and C-termini of pertactin. Since protein unfolding is a rare event, even small proteins with only a few dozen residues require simulations spanning hundreds of microseconds, making all-atom simulations extremely challenging. To address this, we employed various enhanced sampling techniques in our all-atom simulations, including unbiased methods such as adaptive sampling and weighted ensemble as well as biased methods like Gaussian accelerated MD. For many of these approaches, the choice of CVs is crucial, especially for unbiased methods that rely on sampling along collective variables (CVs). We trained machine-learned CVs (ML CVs) using simulation trajectories obtained from coarse-grained models and used these ML CVs to guide all-atom enhanced sampling. Compared to traditional CVs based on physicochemical intuition, our ML CVs significantly improved sampling efficiency, allowing us to observe unfolding conformational changes within just tens of microseconds. These trajectories were then used to construct Markov State Models (MSMs), which provide detailed insights into intermediate states and transition rates, allowing us to better understand the kinetic aspects underlying pertactin’s folding process.
