18.4: MicroRNA Genes and Targets

MiRNA Gene Discovery

MiRNAs are post-transcriptional regulators that bind to mRNAs to silence a gene. They are an extremely important regulator in development. These are formed when a miRNA gene is transcribed from the genome. The resulting strand forms a hairpin at some point. This is processed, trimmed and exported to the cyto- plasm. Then, another protein trims the hairpin and one half is incorporated into a RISK complex. By doing this, it is able to tell the RISK complex where to bind, which determines which gene is turned off. The second strand is usually discarded. It is a computational problem to determine which strand is which. The computational problem here is how to find the genes which correspond to these miRNAs.

The first problem is finding hairpins. Simply folding the genome produces approximately 760,000 hairpins, but there are only 60 to 200 true miRNAs. Thus we need methods to help improve specificity. Structural features, including folding energy, loops (number, symmetry), hairpin length and symmetry, substructures and pairings, can be considered, however, this only increases specificity by a factor of 40. Thus structure alone cannot predict miRNAs. Evolutionary signatures can also be considered. MiRNA show characteristic conservation properties. Hairpins consist of a loop, two arms and flanking regions. In most RNA, the loop is the most well conserved due to the fact that it is used in binding. In miRNA, however, the arms are more conserved because they determine where the RISK complex will bind. This increases specificity by a factor of 300. Both these structural features and conservation properties can be combined to better predict potential miRNAs.

These features are combined using machine learning, specifically random forests. This produces many weak classifiers (decision trees) on subsets of positives and negatives. Each tree then votes on the final classification of a given miRNA. Using this technique allows us to reach the desired sensitivity (increased by 4,500 fold).

Validating Discovered MiRNAs

Discovered miRNAs can be validated by comparing to known miRNAs. An example given in class shows that 81% of discovered miRNAs were already known to exist, which shows that these methods perform well. The putative miRNAs have yet to be tested, however this can be difficult to do as testing is done by cloning.

Region specificity is another method for validating miRNAs. In the background, hairpins are fairly evenly distributed between introns, exons, intergenic regions, and repeats and transposons. Increasing confidence in predictions causes almost all miRNAs to fall in introns and intergenic regions, as expected. These predictions also match sequencing reads.

This also produced some genomic properties typical of miRNAs. They have a preference for transcribed strand. This allows them to piggyback in intron of real gene, and thus not require a separate transcription. They also clustering with known and predicted miRNAs. This indicates that they are in the same family and have a common orgin.

MiRNA’s 5’ End Identification

The first seven bases determine where an miRNA binds, thus it is important to know exactly where clevage occurs. If this clevage point is wrong by even two bases, the miRNA will be predicted to bind to a completely different gene. These clevage points can be discovered computationally by searching for highly conserved 7-mers which could be targets. These 7-mers also correlate to a lack of anti-targets in ubiquitously expressed genes. Using these features, structural features and conservational features, it is possible to take a machine learning approach (SVMs) to predict clevage site. Some miRNAs have no single high scoring position, and these also show imprecise processing in the cell. If the star sequence is highly scored, then it tends to be more expressed in the cell also.

Functional Motifs in Coding Regions

Each motif type has distinct signatures. DNA is strand symmetric, RNA is strand-specific and frame- invariant, and Protein is strand-specific and frame-biased. This frame-invariance can be used as a signature. Each frame can then be evaluated separately. Motifs due to di-codon usage biases are conserved in only one frame offset while motifs due to RNA-level regulation are conserved in all three frame offsets. This allows the ability to distinguish overlapping pressures.