Quantitative prediction of variant effects on alternative splicing using endogenous pre-messenger RNA structure probing

Abstract

Splicing is a highly regulated process that depends on numerous factors. It is particularly challenging to quantitatively predict how a mutation will affect precursor messenger RNA (mRNA) structure and the subsequent functional consequences. Here we use a novel Mutational Profiling (-MaP) methodology to obtain highly reproducible endogenous precursor and mature mRNA structural probing data in vivo. We use these data to estimate Boltzmann suboptimal ensembles, and predict the structural consequences of mutations on precursor mRNA structure. Together with a structural analysis of recent cryo-EM spliceosome structures at different stages of the splicing cycle, we determined that the footprint of the Bact complex on precursor mRNA is best able to predict splicing outcomes for exon 10 inclusion of the alternatively spliced MAPT gene. However, structure alone only achieves 74% accuracy. We therefore developed a β-regression weighting framework that incorporates splice site strength, structure and exonic/intronic splicing regulatory elements which together achieves 90% accuracy for 47 known and six newly discovered splice-altering variants. This combined experimental/computational framework represents a path forward for accurate prediction of splicing related disease-causing variants.

Computational Approaches for RNA Structure Ensemble Deconvolution from Structure Probing Data

“…Numerous popular techniques have been used, including Principal Component Analysis (PCA), 65,[75][76] multidimensional scaling (MDS), [73][74] and t-SNE. [77][78] Method details. Analysis starts with data in the form of reactivities, which are input to datadirected partition function and stochastic sampling routines implemented in popular structure prediction software packages.…”

“…It is often used in tandem with the so-called base pair distance between structures, defined as the number of base pairs in either structure but not in both. 73 Other representations and distance metrics have also been used, e.g., nucleotide paired/unpaired states, 65,76 or counts of small structural features with/without their positions, 74,77 jointly with Euclidean distance. In terms of sequence lengths to which this approach has been applied, to the best of our knowledge, the longest is 554 nt.…”

“…Thus, it can be used as the datadirected sampling module in 'sample and cluster' analyses. 77 The other methods assign abundances to previously-obtained structures. In doing so, they build on an earlier idea, termed 'sample and select', which was introduced to predict a single structure from sequence and data.…”

“…To predict reactivities based on structures in the first sample, Rsample employs a statistical empirical model of reactivity data, which is typically compiled from several thousands of reactivities observed for structured RNAs with known structures. 77,[81][82]25,83 The model captures the statistical behavior of reactivities, depending on a nucleotide's structural context, that is, whether it is unpaired, paired at a helix end, or paired inside a helix. 81 For each nucleotide in each structure, a value is drawn according to its structural context, and reactivities are predicted by averaging over all structures.…”

“…This strategy allows Rsample to account for the high variability and ambiguity often characteristic of probing data, in particular the observation that unpaired nucleotides can often be unreactive to SHAPE or DMS. 13,77,81 Method details. Rsample takes probing data and the RNA sequence as input.…”