A machine learning model to predict quantum-accurate electron densities for DNA: applications for X-ray crystal structure refinement

Abstract

We developed a machine learning model based on a Euclidean neural network framework to study large biological macromolecules such as DNA. The machine learning model bypasses computationally demanding quantum simulations to predict electron densities of up to 99% accuracy. We show that the machine learning model extrapolates to larger system sizes with negligible loss of accuracy by comparing model predictions to quantum calculations for DNA structures of increasing chain length. We apply the model to produce electron densities for various experimental DNA crystal structures typically considered to be too large for conventional calculations.