Characterizing RNA Tetramer Conformational Landscape Using Explainable Machine Learning

We present a simulation framework that combines explainable artificial intelligence (XAI) with the on-the-fly probability enhanced sampling (OPES) algorithm to efficiently sample the complex free energy landscapes of RNA tetramers. Our simulations effectively capture key conformational states—including stacked, intercalated, nucleobase-flipped, and random coil structures—while accurately reproducing the unbiased populations of these states with two orders of magnitude less computational effort compared to conventional molecular dynamics. This approach distinguishes, in an entirely data-driven manner, the structural ensembles of several metastable states that are nearly indistinguishable when using standard metrics in RNA simulation literature. Using explainable machine learning, we can also identify, without incurring additional computational costs, the key torsion angles of the RNA molecule that drive these slow transitions. This built-in interpretability in our machine learning model allows us to pinpoint which backbone dihedral angles contribute to the formation of unphysical intercalated structures in conventional classical force fields, paving the way for data-driven improvements in nucleic acid force fields.