Neural potentials of proteins extrapolate beyond training data

22 December 2022, Version 1
This content is a preprint and has not undergone peer review at the time of posting.


We evaluate neural network coarse-grained force fields compared to traditional CG molecular mechanics force fields. We conclude neural network force fields are able to extrapolate and sample from unseen regions of the free energy surface free energy surfaces when trained with limited data. Our results come from 66 trained force fields trained on different combinations of clustered free energy surfaces across three proteins. We used total variation similarity as our metric, which assesses agreement between free energy surfaces of force fields. Additionally, force matching error was found to only be weakly correlated with a force field's ability to reconstruct the correct free energy surface. These conclusions support the common hypothesis that constructing force fields on one region of the protein free energy surface can extrapolate well.


molecular dynamics
neural force field

Supplementary materials

Supporting Information
Additional details on model training hyperparameters


Comments are not moderated before they are posted, but they can be removed by the site moderators if they are found to be in contravention of our Commenting Policy [opens in a new tab] - please read this policy before you post. Comments should be used for scholarly discussion of the content in question. You can find more information about how to use the commenting feature here [opens in a new tab] .
This site is protected by reCAPTCHA and the Google Privacy Policy [opens in a new tab] and Terms of Service [opens in a new tab] apply.