RNAs not only transcribe instructions from DNA and transport them to ribosomes in the cytoplasm, but they also act as catalysts of biochemical reactions and as key regulators of gene expression and other biological processes. Consequently, understanding RNA structures is just as crucial as understanding protein structures. However, RNA structures still remain underrepresented because of the combined challenges of difficulties in RNA crystallization, high costs associated with ¹³C,¹⁵N-labeled nucleotides for determining RNA structures by NMR, and limited access to high-end cryoelectron microscopy. Thus, various computational methods have been developed to predict RNA 3D structures. For a computational method to reliably predict an RNA 3D structure, a reliable energy function is essential for evaluating the predicted structures. Currently available energy functions show a limited performance for realistic test datasets. Incorporation of machine learning–based energy functions, including RNA 3D convolutional neural networks and  atomic rotationally equivariant scorer, showed an improved performance in evaluating structures.  

An all-atom statistical potential (rsRNASP) and a coarse-grained version (cgRNASP) have demonstrated superior performance compared to traditional energy functions, because of their ability to distinguish between short- and long-range interactions on the basis of residue separation. In a recent study, the authors extended this framework by incorporating both torsional angle–dependent and all-atom distance–dependent energy terms with residue separation for RNA 3D structure evaluation, resulting in the development of rsRNASP1.

This new function integrates multiple dihedral-dependent parameters—including six backbone torsion angles, three sugar ring torsion angles, and one base rotation angle—together with distance-dependent contributions. Benchmarking against extensive datasets revealed that rsRNASP1 consistently outperforms other existing energy functions. Although there remains room for further refinement, rsRNASP1 already offers significantly improved accuracy in evaluating computationally predicted RNA 3D structures.