New insights from cluster analysis methods for RNA secondary structure prediction

Abstract

A widening gap exists between the best practices for RNA secondary structure prediction developed by computational researchers and the methods used in practice by experimentalists. Minimum free energy (MFE) predictions, although broadly used, are outperformed by methods which sample from the Boltzmann distribution and data mine the results. In particular, moving beyond the single structure prediction paradigm yields substantial gains in accuracy. Furthermore, the largest improvements in accuracy and precision come from viewing secondary structures not at the base pair level but at lower granularity/higher abstraction. This suggests that random errors affecting precision and systematic ones affecting accuracy are both reduced by this “fuzzier” view of secondary structures. Thus experimentalists who are willing to adopt a more rigorous, multilayered approach to secondary structure prediction by iterating through these levels of granularity will be much better able to capture fundamental aspects of RNA base pairing.

Statistical modeling of RNA structure profiling experiments enables parsimonious reconstruction of structure landscapes

“…We statistically sampled 1000 secondary structures for each allele sequence, removed duplicates, and combined samples into one set of >100 structures. To account for inter-sample variation at the base-pair level, which is characteristic of statistical samples 24 , 26 , we considered 10 independent samples in all analyses. Although DMS-MaPseq can report multiple modifications per read by way of mutational profiling, it is also possible to generate an equivalent truncation data set in silico, where each read reports up to one modification (see Methods for details and Supplementary Fig.…”

“…Otherwise, a two-step approach may be more appropriate, where LASSO is applied to limit the number of selected structures, followed by NNLS for accurate abundance estimation 54 . Also, one may cluster either the input structures or the structures selected by NNLS using methods specialized to RNA structure 4 , 24 , 26 . However, clustering might eliminate features that may be informative to downstream analysis.…”

“…Other than M 2 -REEFFIT and RING-MaP 22 , 23 , current approaches entail large-scale data-directed sampling of suboptimal secondary structures, which interprets SP data as the signature of one structure that dominates in solution 17 . Furthermore, generated secondary structures are often subject to limitations of thermodynamic models 21 and also warrant lossy post-processing via clustering and/or dimensionality reduction to mine a simplified view of key structures 24 – 26 . While this paper was under review, another group reported a thermodynamics-based method, Rsample, which employs data-directed suboptimal sampling followed by large-scale clustering, yet it explicitly considers multiple structures as the sources of the data 27 .…”

PATTERNA: transcriptome-wide search for functional RNA elements via structural data signatures

“…Overall, these results are not surprising as an ensemble-based approach considers many competing alternative structures while an MFE approach encapsulates structural dynamics into a single structure [ 69 ]. Nevertheless, our results also indicate that PATTERNA performs surprisingly well, given that no sequence information is used, and can even outperform thermodynamics modeling on occasions.…”

“…In addition, long RNAs (>1000 nt) generally require deeper sampling of the ensemble, compared to short RNAs, to ensure that all biologically relevant structures are represented in the sampled pool and that sample frequencies reliably approximate their theoretical counterparts. Therefore, this approach is typically reserved for small-scale studies of small RNAs [ 69 ], or alternatively, for small targeted regions within transcripts (e.g., 100 nt long), as these are easily folded computationally. MFE prediction suffers a similar drawback given its complexity is also (Additional file 2 : Figure S3) and, while faster compared to ensemble sampling, it would still require about a month for a typical transcriptome-wide data set (Additional file 2 : Table S7).…”