Assessment of Embedding Schemes in a Hybrid Machine Learning/Classical Potentials (ML/MM) Approach

Abstract

Machine learning (ML) methods have reached high accuracy levels for theprediction of in vacuo molecular properties. However, the simulation of large systems solelythrough ML methods (such as those based on neural network potentials) is still a challenge. Inthis context, one of the most promising frameworks for integrating ML schemes in thesimulation of complex molecular systems is the so-called ML/MM methods. These multiscaleapproaches combine ML methods with classical force fields (MM), in the same spirit as thesuccessful hybrid quantum mechanics−molecular mechanics methods (QM/MM). The keyissue for such ML/MM methods is an adequate description of the coupling between the regionof the system described by ML and the region described at the MM level. In the context ofQM/MM schemes, the main ingredient of the interaction is electrostatic, and the state of theart is the so-called electrostatic-embedding. In this study, we analyze the quality of simplermechanical embedding-based approaches, specifically focusing on their application within anML/MM framework utilizing atomic partial charges derived in vacuo. Taking as referenceelectrostatic embedding calculations performed at a QM(DFT)/MM level, we explore different atomic charges schemes, as well as apolarization correction computed using atomic polarizabilites. Our benchmark data set comprises a set of about 80k small organicstructures from the ANI-1x and ANI-2x databases, solvated in water. The results suggest that the minimal basis iterative stockholderatomic charges yield the best agreement with the reference coupling energy. Remarkable enhancements are achieved by including asimple polarization correction.