Cross-talk between RNA secondary and three-dimensional structure prediction: a comprehensive study

In recent years, various computational methods have been developed to predict the three-dimensional (3D) structures of RNAs. Due to its hierarchical folding property, RNA secondary (2D) structure is usually used as input for 3D structure prediction to improve accuracy and efficiency. However, the extent to which the accuracy of input 2D structure affects the performance of 3D structure prediction remains to be further investigated. Additionally, whether and how the input base-pairing interactions are modified during the 3D structure modeling process is another question worth exploring. To address these issues, here we comprehensively benchmark five representative 3D structure prediction models on extensive datasets, using 2D structures of varied accuracies as input. Our results indicate that there is a pervasive cross-talk between RNA 2D and 3D structure predictions, where the performance dependence of 3D structure prediction on the accuracy of input 2D structure is closely associated with the 3D models ability to modify the input base-pairing interactions during structure modeling. Furthermore, we also observed that RNA 3D structure prediction performance is more sensitive to the occurrence of false positive base pairs in the input 2D structure than to true positive base pairs, suggesting a worthy direction to further improve the model performance. Author summaryThree-dimensional (3D) structural modeling of RNAs with large sizes and complex topologies remains challenging despite the availability of (predicted) 2D structures as constraints. Exploring the potential cross-talk between 2D and 3D structure predictions is a worthy direction to improve the performance of RNA structure modeling. In this study, we found that all tested popular RNA 3D structure prediction models were able to modify the original base pairing interactions contained in the input 2D structure during the 3D modeling process. Especially, all of these models presented increased F1-score for the optimal combination of RNA 2D and 3D structure prediction models. The results suggest that a worthwhile direction to further improve the performance of RNA 3D structure prediction is to minimize the incidence of incorrectly predicted base pairing interactions during modeling process without compromising or even improving the presence of correct interactions in the input 2D structure.