Alternative splicing impacts microRNA regulation within coding regions

Abstract

MicroRNAs (miRNAs) are small non-coding RNA molecules that bind to target sites in different gene regions and regulate post-transcriptional gene expression. Approximately 95% of human multi-exon genes can be spliced alternatively, which enables the production of functionally diverse transcripts and proteins from a single gene. Through alternative splicing, transcripts might lose the exon with the miRNA target site and become unresponsive to miRNA regulation. To check this hypothesis, we studied the role of miRNA target sites in both coding and non-coding regions using six cancer data sets from The Cancer Genome Atlas (TCGA) and Parkinson's disease data from PPMI. First, we predicted miRNA target sites on mRNAs from their sequence using TarPmiR. To check whether alternative splicing interferes with this regulation, we trained linear regression models to predict miRNA expression from transcript expression. Using nested models, we compared the predictive power of transcripts with miRNA target sites in the coding regions to that of transcripts without target sites. Models containing transcripts with target sites perform significantly better. We conclude that alternative splicing does interfere with miRNA regulation by skipping exons with miRNA target sites within the coding region.

Graphical Abstract

Figure 1.

Analysis pipeline using miRNA and mRNA expression and sequence data. (1) TarPmiR target site prediction, (2) categorization in non-binding vs. binding transcripts, (3) filtering (A) for expression and variance above the chosen thresholds (see Materials and Methods), (B) alternatively spliced genes, (C) negative correlation of miRNA and gene expression, (4) per miRNA–gene pair nested linear regression: non-binding transcript regression and all transcript regression, (5) subsampling of nested models, (6) subsampling and label randomization of nested models, (7) likelihood ratio test between nested model pairs. The pipeline was run for the three settings ALLT (all transcripts), TNBN (transcripts not binding in non-coding region) and TBN (transcripts binding in non-coding region) separately from step 2) on.

Figure 2.

(A) Transcripts are divided into four different transcript types: (A) non-binding transcripts, (B) transcripts with target sites only in the coding region, (C) transcripts with target sites only in the non-coding region (3′-UTR or 5′-UTR), (D) transcripts with target sites in both the coding and non-coding region B Structure of the nested miRNA–gene-level linear regression models. The full model is trained on: all transcripts (ALLT), only transcripts without target sites in non-coding region (TNBN), and only transcripts with target sites in non-coding region (TBN). Accordingly, the reduced model is only trained on a subset of the transcripts without target sites in the investigated region.

Figure 3.

Categorization of predicted TarPmiR target sites before any filtering (A) in target sites with binding probability between 50% and 80% and above 80% and (B) in miRNA-transcript pairs based on target region (coding/non-coding region).

Figure 4.

The ratio of models with statistically significant (<0.05) corrected p-values of the likelihood ratio test statistic calculated between nested regression models is shown as the dashed green line. To estimate the distribution, the ratio was calculated 1000 times for random subsets of miRNA–gene pairs (green histogram) and to estimate the impact of alternative splicing, the ratio was calculated 1000 times for random subsets of miRNA–gene pairs while randomizing the transcript binding labels within a gene (yellow histogram). Dist describes the difference between the average ratio of models based on subsampled real miRNA–gene pairs and after randomizing the transcript category labels. This is shown separately for diseases Brain lower grade glioma (LGG) and Kidney chromophobe carcinoma (KICH) for settings ALLT, TNBN and TBN.

Figure 5.

The ratio of models with statistically significant (<0.05) corrected p-values of the likelihood ratio test statistic calculated between nested regression models is shown as the dashed green line. To estimate the distribution, the ratio was calculated 1000 times for random subsets of miRNA–gene pairs (green histogram) and to estimate the impact of alternative splicing, the ratio was calculated 1000 times for random subsets of miRNA–gene pairs while randomizing the transcript binding labels within a gene (yellow histogram). Dist describes the difference between the average ratio of models based on subsampled real miRNA–gene pairs and after randomizing the transcript category labels. This is shown for Parkinson disease for settings ALLT, TNBN and TBN.

Figure 6.

The ratio of models with statistically significant (<0.05) corrected P-values of the likelihood ratio test statistic calculated between nested regression models is shown as the dashed green line. To estimate the distribution, the ratio was calculated 1000 times for random subsets of miRNA–gene pairs (green histogram) and to estimate the impact of alternative splicing, the ratio was calculated 1000 times for random subsets of miRNA–gene pairs while randomizing the transcript binding labels within a gene (yellow histogram). Dist describes the difference between the average ratio of models based on subsampled real miRNA–gene pairs and after randomizing the transcript category labels. This is shown for Kidney chromophobe carcinoma (KICH) for setting TNBN with the addition of the most correlated miRNA as a covariate.