Detection of micro-RNA/gene interactions involved in angiogenesis using machine learning techniques

Motivation: Angiogenesis is the process responsible for the growth of new blood vessels from existing ones. It is also associated with the development of cancer, as tumors need to be irrigated by blood vessels for growing. New cancer therapies appear that exploit angiogenesis inhibitors, also called angiostatic agents, to asphyxiate and starve the tumors. Better understanding the regulatory mechanisms that control angiogenesis is thus fundamental.
Recently, short non-coding RNA molecules, called micro-RNAs, have been discovered that are involved in post- transcriptional regulation of gene expressions. These molecules bind to RNA messengers following the base pairing rules, preventing them from being translated into proteins and/or tagging them for degradation. The main goal of this work is to use computational approaches to identify micro-RNAs involved in angiogenesis.
Method: In order to identify genes involved in angiogenesis, bovine endothelial cells were treated by a known angiogenesis inhibitor [1], prolactin 16K, and their gene expression profile was compared to the profile of untreated cells. The genes were then divided into three classes: up-regulated, down-regulated, and unaffected genes. The 3'UTR regions of these genes were then analysed by machine learning techniques.
Different approaches were considered. First, we described each gene by a vector of motif counts in their 3'UTR regions and used machine learning techniques to rank the motifs according to their relevance for separating the genes into the different classes. We considered successively motifs corresponding to the seeds of known micro- RNAs and also all possible motifs of a given length. To rank the motifs, we compared ensemble of decision trees and linear support vector machines. Second, we considered an approach called Segment and Combine that was proposed in [2]. Finally, we also carried out an exhaustive search of all motifs of a given length that satisfy some constraints on specificity and coverage with respect to a given gene category.
Results: The ability of the different approaches at identifying relevant motifs was first assessed on genes predicted to be the target of some known miRNAs. In this simple setting, most methods were able to identify the micro-RNA seed. The results obtained on the genes regulated by prolactin 16K are also very encouraging. We were able to identify one micro-RNA already known to play a role in angiogenesis and several motifs are predicted by different approaches as very specific of up- or down-regulation by prolactin 16K. Their relationship with known micro-RNAs is certainly worth exploring.
Conclusion: Machine learning approaches are promising techniques for the identification of micro-RNA/gene interactions. Future work will concern the application of the same kind of techniques on promoters for the identification of transcription factor binding sites.