Machine learning assesses multibody terms in a coarse-grained model

Coarse-grained models, which are less detailed than atomistic models, are used to study important complex molecular systems. Multibody terms improve the accuracy of these simplified models, but there is no general, systematic and computationally feasible approach for including multibody terms in their effective potential energies.

Wang et al. constructed a neural network architecture for a coarse-grained model’s effective potential energy that resembles a multibody energy expansion. Machine learning allowed them to examine how multibody terms contributed to the model’s effective potential energy. They applied their approach to a small protein called chignolin.

The authors were surprised to find the effective potential energy for the model of this protein needed five body terms to match the free energy landscape of the reference atomistic model. Previous models usually used up to three body or four body terms, but these results suggest that is not enough for a coarse-grained model to reproduce the free energy of an atomistic model. More than five body terms did not demonstrate further improvement.

“These results highlight the importance of machine learning for these kinds of problems,” said author Cecilia Clementi. “It was not possible to do this without neural networks, because these potentials are very complicated. At the moment, we don’t have enough knowledge to represent with analytical expressions the effective coarse-grained potential energy that can reproduce the free energy of an atomistic system.”

The authors’ machine learning approach will help improve coarse-grained models, as well as increase understanding of their energy landscapes. The authors are currently working on expanding their approach to larger molecular systems beyond chignolin.

Source: “Multi-body effects in a coarse-grained protein force field,” by Jiang Wang, Nicholas Charron, Brooke Husic, Simon Olsson, Frank Noé, and Cecilia Clementi, Journal of Chemical Physics (2021). The article can be accessed at https://doi.org/10.1063/5.0041022 .