Data-driven analysis of the number of Lennard\u2013Jones types needed in a force field

Michael Schauperl,Sophie M Kantonen,Michael K Gilson,Lee-Ping Wang

doi:10.1038/s42004-020-00395-w

Michael Schauperl, Sophie M Kantonen + Show 2 more

Open Access

https://doi.org/10.1038/s42004-020-00395-w

Copy DOI

Abstract

Force fields used in molecular simulations contain numerical parameters, such as Lennard–Jones (LJ) parameters, which are assigned to the atoms in a molecule based on a classification of their chemical environments. The number of classes, or types, should be no more than needed to maximize agreement with experiment, as parsimony avoids overfitting and simplifies parameter optimization. However, types have historically been crafted based largely on chemical intuition, so current force fields may contain more types than needed. In this study, we seek the minimum number of LJ parameter types needed to represent the key properties of organic liquids. We find that highly competitive force field accuracy is obtained with minimalist sets of LJ types; e.g., two H types and one type apiece for C, O, and N atoms. We also find that the fitness surface has multiple minima, which can lead to local trapping of the optimizer.

Highlights

We report on the training- and test-set performance of all of the LJ typing models depicted in Fig. 1, using RESP1 charges and both training set/test set splits
The central finding of this study is that highly competitive force field accuracy can be obtained with minimalist sets of LJ types
Merely splitting apolar from polar hydrogens, while using one type apiece for carbon, oxygen, and nitrogen, yields a model that meets or exceeds the accuracy of SmirFF.[7] and GAFF1.8 for the present test sets. This data-driven observation arguably challenges a chemical intuition that distinct LJ types are needed for atoms in distinct functional groups

Summary

Introduction

For the second training/test-set split (see Methods section), the accuracy of the HCON model (5.46%, 15.28%, 49.3%) is again slightly lower than that of SmirFF.[7] (3.89%, 13.50%, 47.5%) Splitting the single H type in HC3ON, HCO3N, and HCON2 into polar and apolar types to yield models H2C3ON, H2CO3N, and H2CON2, respectively, yields consistently lower values of the objective function for both test sets

Results

Conclusion