miRNATIP: A self-organizing map based miRNA-Target Interactions Predictor

The miRNA target prediction method is composed of four steps. In the first step, a set of miRNAs is used for the training of a SOM. In detail, in this step we consider only the 8-mer miRNA seeds, because it has been demonstrated the seed is mainly responsible of the miRNA target binding. Each miRNA seed is then converted in a 4X8 (nucleotide symbols vs. seed length) Position Weight Matrix (PWM), according to the representation used for motifs in biological sequences. At this point, a SOM is trained with this set of PWMs. The second step is the projection of a mRNA sequence over the trained SOM. For this reason, we extracted all the 8-length mRNA fragments through a 8-mer sliding window. This way, we obtained a set of 4x8 PWMs that can be projected over the trained SOM. The result of this step is, for each neural unit (cluster), a list of miRNA_seed-mRNA_fragment. Each cluster can be considered as a preliminary list of predicted miRNAs-mRNAs interactions. In the third step, we filtered these putative interactions considering the remaining part of the miRNA sequences (miRNA tail). For each miRNA_seed-mRNA_fragment interaction, we considered respectively the miRNA tail and the mRNA sequence of the same length of miRNA tail, next to the predicted mRNA_fragment. Then we computed a dissimilarity measure based on normalised euclidean distance between the PWM representations of those two sequences, and we retained only the couples of miRNA-mRNA interactions whose distance is below a certain threshold. In order to take into account also the presence of possible bulge loops between the 8-mer seed and the tail of the miRNA, an offset for the selection of the mRNA fragment corresponding to the miRNA tail is used. Finally, in the fourth step another filtering to the miRNA-mRNA interaction list is performed, by computing the free-energy of the miRNA-target site duplex.