Applications of machine-learned electron densities of nucleic acids.

Abstract

The efficient prediction of large-scale electron densities of biological macromolecules and molecular arrays could lead to breakthroughs in quantum chemistry, structural biology, and molecular design. This work leverages a novel machine learning (ML) algorithm based on Euclidean neural networks to learn the electron densities of arbitrary sequences of DNA by performing many smaller ab initio calculations of the component base-pair steps of DNA. These electron density predictions are currently accurate to around 1% error for arbitrary sequences of DNA, and are easily size-extensible unlike traditional quantum calculations. Using this machine learning algorithm, electron density predictions of large-scale DNA structures consisting of tens-of-thousands to hundreds-of-thousands of electrons can be performed in minutes. Several applications of these large-scale machine-learned electron densities are of interest. We focus here on the accurate calculation of energies and forces from the machine-learned electron densities, allowing for dynamical propagation of the electronic system, termed ab initio molecular dynamics (AIMD). Calculation of accurate forces directly from the machine-learned electron densities is not a trivial step, as specialized basis set expansions are required to capture this information. These machine-learned electron densities would allow for AIMD simulations of systems that are much larger than can be propagated using traditional quantum calculations, leading to potential applications in biomolecular binding and structural prediction of macromolecules.