Machine Learning Prediction of DNA Charge Transport.

First-principles calculations of charge transfer in DNA molecules are computationally expensive given that conducting charge carriers interact with intra- and intermolecular atomic motion. Screening sequences, for example, to identify excellent electrical conductors, is challenging even when adopting coarse-grained models and effective computational schemes that do not explicitly describe atomic dynamics. We present a machine learning (ML) model that allows the inexpensive prediction of the electrical conductance of millions of long double-stranded DNA (dsDNA) sequences, reducing computational costs by orders of magnitude. The algorithm is trained on short DNA nanojunctions with n = 3-7 base pairs. The electrical conductance of the training set is computed with a quantum scattering method, which captures charge-nuclei scattering processes. We demonstrate that the ML method accurately predicts the electrical conductance of varied dsDNA junctions tracing different transport mechanisms: coherent (short-range) quantum tunneling, on-resonance (ballistic) transport, and incoherent site-to-site hopping. Furthermore, the ML approach supports physical observations that clusters of nucleotides regulate DNA transport behavior. The input features tested in this work could be used in other ML studies of charge transport in complex polymers in the search for promising electronic and thermoelectric materials.