Principles for Understanding the Accuracy of SHAPE-Directed RNA Structure Modeling

Abstract

Accurate RNA structure modeling is an important, incompletely solved, challenge. Single-nucleotide resolution SHAPE (selective 2'-hydroxyl acylation analyzed by primer extension) yields an experimental measurement of local nucleotide flexibility that can be incorporated as pseudo-free energy change constraints to direct secondary structure predictions. Prior work from our laboratory has emphasized both the overall accuracy of this approach and the need for nuanced interpretation of some apparent discrepancies between modeled and accepted structures. Recent studies by Das and colleagues [Kladwang et al., Biochemistry 50:8049 (2011) and Nat. Chem. 3:954 (2011)], focused on analyzing six small RNAs, yielded poorer RNA secondary structure predictions than expected based on prior benchmarking efforts. To understand the features that led to these divergent results, we re-examined four RNAs yielding the poorest results in this recent work – tRNAPhe, the adenine and cyclic-di-GMP riboswitches, and 5S rRNA. Most of the errors reported by Das and colleagues reflected non-standard experiment and data processing choices, and selective scoring rules. For two RNAs, tRNAPhe and the adenine riboswitch, secondary structure predictions are nearly perfect if no experimental information is included but were rendered inaccurate by the Das and colleagues SHAPE data. When best practices were used, single-sequence SHAPE-directed secondary structure modeling recovered ~93% of individual base pairs and greater than 90% of helices in the four RNAs, essentially indistinguishable from the mutate-and-map approach with the exception of a single helix in the 5S rRNA. The field of experimentally-directed RNA secondary structure prediction is entering a phase focused on the most difficult prediction challenges. We outline five constructive principles for guiding this field forward.

RNA secondary structure modeling at consistent high accuracy using differential SHAPE

“…These included six riboswitch aptamer domains that require ligand binding to fold into their accepted structures (the TPP, adenine, glycine, cyclic-di-GMP, M-Box, and lysine riboswitches); four RNAs longer than 300 nucleotides (nt), including several domains of the Escherichia coli 16S and 23S ribosomal RNAs; four pseudoknot-containing RNAs; and every other RNA of which we are aware that contains up to one pseudoknot for which the single-reagent 1M7 modeling accuracy is <90% (Cordero et al 2012;Hajdin et al 2013;Leonard et al 2013; (Steen et al 2012). We therefore used a windowed scaling algorithm to locally normalize NMIA and 1M6 SHAPE profiles to each other (see Materials and Methods) and then subtracted the normalized profiles to generate differential SHAPE reactivity traces (Fig.…”

“…SHAPE has been used to create nucleotide-resolution models for the viral genomes of HIV-1 (Watts et al 2009) and STMV (Archer et al 2013) and to analyze conformational changes in HIV-1 (Wilkinson et al 2008) and the Moloney murine leukemia virus (Grohman et al 2013). Although SHAPE-directed folding yields near-perfect models for many RNAs, there remain a few RNAs whose structures are difficult to recover using a single structure probing experiment (Cordero et al 2012;Leonard et al 2013). These "hard" RNAs are modeled with sensitivities in the 75%-85% range.…”

RNA structure through multidimensional chemical mapping

“…For example, reanalysis of a model based on selective 2´-OH acylation by primer extension (SHAPE) of the 9173-nucleotide HIV-1 RNA genome (Watts et al 2009) suggested that more than half of the presented helices were not well-determined (Kladwang et al 2011c), and subsequent work, including both experimental and computational improvements, have significantly revised these uncertain regions (Pollom et al 2013;Siegfried et al 2014;Sukosd et al 2015). The debate over whether these methods produce acceptable structure accuracies continues (Deigan et al 2009;Eddy, 2014;Kladwang et al 2011c;Leonard et al 2013;Rice et al 2014;Sukosd et al 2013;Tian et al 2014) and will not be reviewed in detail here. There is general agreement, however, on some points.…”

“…These experiments raise the prospect of nucleotide-resolution structural portraits of all RNAs being transcribed in an organism -the 'RNA structurome'. Nevertheless, when tested through independent experiments, de novo models derived from chemical mapping and computational modeling have not always given consistently accurate structures, even on small domains folded into well-defined states and probed in vitro (Deigan et al 2009;Kladwang et al 2011c;Leonard et al 2013;Tian et al 2014). These issues can be traced to the poor information content of chemical mapping measurements, which typically give single or few measurements per nucleotide, compared with high-resolution technologies such as crystallography, NMR, or cryo-electron microscopy, which can return datasets with ten or more measurements per nucleotide.…”

Capillary electrophoresis based on nucleic acid detection for diagnosing human infectious disease

“…Most technologies for determining RNA structure are unable to analyze RNA from a small number of biological samples. The highly sensitive selective 2′-hydroxyl acylation analyzed by primer extension (SHAPE) [98][99][100][101] combined with high resolution CE could resolve this problem and quantitatively analyze RNA structure to solve RNA structurerelated function problems. Thus, Jacob et al [102] developed SHAPE-CE to structurally analyze genomic RNA from the xenotropic murine leukemia virus-related virus (XMRV) (Figure 13).…”