Target Prediction Algorithms and Bioinformatics Resources for miRNA Studies - MicroRNAs as Tools in Biopharmaceutical Production - page 37

associate a mature miRNA (as opposed to mature miRNA pattern) those areas with
low pattern aggregation are discarded and the remaining target islands are aligned
to the miRNA sequences in a straightforward manner followed by thermodynamic
assessment of binding via the Vienna package after insertion of a linker to form an
in-silico hetroduplex.

3.3.8 mirWalk

mirWalk (Dweep et al. 2011 ) attempts to locate putative miRNA-mRNA interactions
not only in 3 UTR but in any region of the mRNA. Once Watson-Crick base pairing
across 7 nucleotides is located the miRNA seed is extended until a mismatch is found.
For each miRNA the targets matched are categorised in terms of their position within
the transcript (promoter region, 5 UTR, coding sequence, 3 UTR or mitochondrial
genes) with a p-value reported for each match based on a random seed match. The
mirWalk system allows user friendly overlap with several other prediction algorithms
as well as a collection of validated targets.

3.3.9 Machine Learning

Several approaches have recently emerged that utilise features from known miRNA-
mRNA interactions. Such methods aim to develop a statistical model to predict novel
interactions from the characteristics of those that have already been discovered. One
such approach was developed by Kim and co-workers (Kim et al. 2006 ). A sup-
port vector machine approach (SVM) algorithm was applied to predict targets for
multiple species. The data utilised to train the algorithm on features selected from
a combination of sequence composition and binding position along with thermody-
namic calculations. miRSVR is an example of a similar approach where a model was
constructed from 57 features derived using a two-class Natrive Bayes (NBmiRTar)
(Yousef et al. 2007 ) (an artificial negative target dataset was generated for training).
The model was integrated with the predicted targets of miRanda demonstrating how
machine learning and target prediction strategies can be used together. Several other
machine learning approaches have also been described using a variety of compet-
ing algorithms including random forest prediction (Jiang et al. 2007 ), and in some
instances a combination of different machine learning algorithms (Yan et al. 2007 ).

3.3.10 Combining in-silico Prediction with Experimental Data

One recently developed approach, HOCTAR (Host Gene Oppositely correlated Tar-
gets) (Gennarino et al. 2009 ) utilises gene expression analysis to search for inversely
correlated target genes. The creators of the algorithm compiled at list of predicted tar-
gets outputted by TargetScan, miRanda and PicTar. The intergenic miRNAs within