Modeling RNA secondary structure folding ensembles using SHAPE mapping data

Abstract

RNA secondary structure prediction is widely used for developing hypotheses about the structures of RNA sequences, and structure can provide insight about RNA function. The accuracy of structure prediction is known to be improved using experimental mapping data that provide information about the pairing status of single nucleotides, and these data can now be acquired for whole transcriptomes using high-throughput sequencing. Prior methods for using these experimental data focused on predicting structures for sequences assuming that they populate a single structure. Most RNAs populate multiple structures, however, where the ensemble of strands populates structures with different sets of canonical base pairs. The focus on modeling single structures has been a bottleneck for accurately modeling RNA structure. In this work, we introduce Rsample, an algorithm for using experimental data to predict more than one RNA structure for sequences that populate multiple structures at equilibrium. We demonstrate, using SHAPE mapping data, that we can accurately model RNA sequences that populate multiple structures, including the relative probabilities of those structures. This program is freely available as part of the RNAstructure software package.

Figure 1.

Overview of Rsample and experimental data. (A) The Rsample procedure. (B) The three conformations of a designed FMN riboswitch, color coded with SHAPE reactivity from Cordero et al. (43). The observed SHAPE reactivities are the weighted means of the reactivity of each structure. In the natural mixture in solution, the reactivity of each structure cannot be observed. (C) Distribution of normalized SHAPE reactivities for unpaired nucleotides, nucleotides paired at helix ends, and nucleotides paired and not at helix ends. The distributions show significant overlap and peak at a reactivity of 0, although the distribution for unpaired nucleotides skews farther towards higher reactivities. Nucleotides at the ends of helices also have some skew to higher reactivities than nucleotides in the interior of helices.

Figure 2.

Performance of methods for predicting the ratio of known conformations. Experimental ratios are compared to the predicted ratios for the method without using SHAPE restraints (No SHAPE), using our original method for including restraints (Pseudo ΔG) (21), and the method introduced here (Rsample). (A) The predicted centroid structures for the FMN riboswitch in the absence of ligand using Rsample. The three accepted structures are shown in Figure 1B. The remaining plots show the models for each sequence, FMN riboswitch (–FMN; B), FMN riboswitch (+FMN, C), HIV RRA (D), bistable sequence (E), 16S rRNA (F), ADD riboswitch (-adenine; G), ADD riboswitch (+adenine; H), and medloop (I). To illustrate the performance of the whole predicted ensemble predicted ratio for structures that were not close to any of the known conformations (‘other’) is also given.

Figure 3.

Performance of our method (Rsample) compared to other methods for predicting single conformation. The results are given as geometric mean of sensitivity and positive predictive value. RNAprobing predictions for 16S rRNA, 23S rRNA and its average over all sequences are not included because RNAprobing does not handle sequences longer than 1, 000 nucleotides.