Evidence of distal regulations orchestrated by RNAs initiating at short tandem repeats

Short Tandem Repeats (STRs), also called microsatellites, correspond to tandemly repeated short DNA motifs (1 to 6 bp) and are one of the most polymorphic and abundant repetitive elements in the human genome. Variations of their length (i.e. number of consecutive repeats) have been implicated in gene expression regulation (termed expression(e)STRs). Using cap-trapping followed by long read sequencing, we discovered that STRs can host transcription start sites, the presence of which depends mainly on STR flanking sequences. Here, we investigate the effect of SNPs located in these sequences and ask whether STR-initiating RNAs have regulatory potential. First, we develop fully interpretable deep learning-based models, called Modular Neural Networks, able to predict, for each STR class, the level of RNAs using 101bp-long sequences encompassing STRs. Analysis of MNN filters allows us to identify multiple regulatory elements and candidate transcription factors. Second, leveraging genome sequencing and gene expression data from the Genotype-Tissue Expression project, we use the output of our models to link the levels of STR-initiating RNAs to the expression of nearby genes. We identify 14,340 significant associations (coined RNA(r)STRs) and illustrate how this novel resource can help interpret non-coding variants associated with complex traits and diseases. Third, we unveil an intricate transcriptional interplay between STR-initiating RNAs and Alu repeats that may couple their regulatory actions, extending both the nature and the functional importance of non-coding transcription and shedding new light on the complexity of distal regulations orchestrated by repeated sequences.