Advances in RNA contact prediction: a benchmark evaluation of computational methods

Abstract Ribonucleic Acid (RNA) contact prediction holds great significance for modeling RNA 3D structures and further understanding RNA biological functions. The rapid growth of RNA sequencing data has driven the development of diverse computational methods for RNA contact prediction, and a benchmark evaluation of these methods remains essential. In this work, we first classified RNA contact prediction methods into statistical inference-based and neural network-based ones. We then evaluated eight state-of-the-art methods on three test sets: a sequence-diverse set, a structurally non-redundant set and a CASP RNA targets set. Our evaluation shows that for identifying non-local and long-range contacts, neural network-based methods outperform statistical inference-based ones, with SPOT-RNA-2D achieving the best performance, followed by CoCoNet and RNAcontact. However, for identifying the long-range tertiary contacts, which are vital for stabilizing RNA tertiary structure, statistical inference-based methods exhibit superior performance with GREMLIN emerging as the top performer. This work provides a comprehensive benchmarking of RNA contact prediction methods, highlighting their strengths and limitations to guide further methodological improvements and applications in RNA structure modeling.