Characterizing RNA Tetramer Conformational Landscape Using Explainable Machine Learning

The conformational flexibility of RNA molecules enables them to play vital physiological roles, including carrying genetic information, catalyzing reactions, and forming organelles. However, this structural diversity complicates the quantitative sampling of their conformational landscape, even for simple single-stranded RNA tetramers. We show that combining explainable artificial intelligence (XAI) with enhanced sampling algorithms can effectively explore the complex free energy landscapes of RNA tetramers. Our simulations capture key conformational states, such as stacked, intercalated, nucleobase-flipped, and random coil structures, while reproducing unbiased populations with much less computational effort than conventional molecular dynamics. This data-driven approach distinguishes several metastable states that are often indistinguishable in standard analysis. Additionally, our interpretable machine learning framework identifies key torsion angles driving slow transitions and those responsible for unphysical intercalated structures, paving the way for improvements in nucleic acid force fields.