Abstract

Recent developments and applications of machine learning to physical systems have led to significant advances in the construction of coarse-grained force fields for efficient simulation and sampling [1]. Yet, transferability and extrapolation between different systems of interest remain an outstanding limitation for machine-learned models. Using force matching, a bottom-up coarse-graining approach, and a database of chemically diverse peptides, we present a coarse-grained force field that is transferable across protein sequences enabling us to explore their conformational landscape. Our model, based on a graph neural network architecture, is validated/tested against all-atom simulations of unseen proteins [2].

[1] Wang, J. et al. Machine Learning of Coarse-Grained Molecular Dynamics Force Fields. ACS Cent. Sci. 5, 755–767 (2019).

[2] Lindorff-Larsen, K. et al. How fast-folding proteins fold. Science (80-. ). 334, 517–520 (2011).