Machine learning predictions of the Hessian matrix for peptides chains and small proteins

Abstract

Molecular Hessians have a key role in describing molecular vibrations, trajectories and optimization paths. An explicit calculation of them through standard quantum mechanical methods can be computationally expensive for medium-large systems, and in many applications not even needed. Machine learning methods can be a shortcut to tackle efficiently the computational difficulties. This paper will present a ML model able to predict the Hessian matrix of biological system made of thousands of atoms. The method, based on learning the Hessian in internal coordinates is intrinsically invariant to molecular rotations and translations, and has a very good scaling with the systems' size. The training was performed on a dataset of simple aminoacids, as they constitute the building blocks of larger proteins. From the predicted Hessian matrix it is possible to calculate thermochemical properties within the harmonic approximations, among them enthalpies, entropies, Gibbs' free energies, and zero point vibrational energies.