Achieving Atomic Scale Resolution of Metastable Polymers in Solution using Machine Learning

Single-stranded DNA (ssDNA) is a metastable biopolymer that plays an important role in biological processes and has shown promise for applications in medicine and DNA nanotechnology. Understanding the structure of ssDNA in solution can provide a framework on how to control and manipulate the self-assembly and structure of ssDNA-based materials. Typically, experimental methods such as Small Angle X-ray Scattering (SAXS) are used to resolve conformational distributions of ssDNA; however, the low resolution does not provide enough information on structural details. We have developed a new method that utilizes molecular dynamic (MD) simulations in conjuncture with machine learning (ML) that can obtain converged structures of metastable polymers. This study specifically performs MD simulations on ssDNA and evaluates the conformational landscape with respect to SAXS. ML methods are first used to obtain a comprehensive collection of possible ssDNA structures through simulation and subsequently used to optimize which group of individual chains closely match experimental SAXS curves. Ultimately, this process lets us break down the ensemble average embedded in SAXS and find the most probable group of metastable conformations in solution with structures defined at the atomic scale.