Learning coarse-grained force fields for fibrogenesis modeling

Abstract

Fine-grained molecular dynamics (FGMD) can examine the dynamical and structural details of many complex systems including those found in biology, and can mitigate the many deficiencies of in vitro experiments. However these advantages come at the cost of prohibitive computing expense. Coarsening molecular structure and dynamics can dramatically diminish the computing complexity of simulations, while retaining modeling accuracies, and have become the state-of-the-art alternative to FGMD studies. Here we formulate the coarsening operator for application to fine-grained coordinates and develop the corresponding coarsening operator for forces to mathematically facilitate the transformation. In an exemplary application to studying the formation of fibrin networks of a few hundreds of soluble protein fibrinogens, we leverage multiscale molecular dynamics and machine learning to develop a physics-informed parameter learning (PIPL) framework with the learnable force parameters in physics priors as the learning target. While its adaptivity or performance on more general macromolecules remains untested, we show its performances, measured in terms of speed and accuracy, and demonstrate that simulations with the learned CGMD force field achieve higher accuracies and higher speeds than with existing methods. The resulting predictions of our model, additionally, are validated by in vitro observations.