The discriminant power of RNA features for pre-miRNA recognition

I.O.N. Lopes, A. Schliep and A.P.L.F. Carvalho

Technical report, Oct 2013. arXiv:1312.5778.

Computational discovery of microRNAs (miRNA) is based on pre-determined sets of features from miRNA precursors (pre-miRNA). These feature sets used by current tools for pre-miRNA recognition differ in construction and dimension. Some feature sets are composed of sequence-structure patterns commonly found in pre-miRNAs, while others are a combination of more sophisticated RNA features. Current tools achieve similar predictive performance even though the feature sets used and their computational cost differ widely. In this work, we analyze the discriminant power of seven feature sets, which are used in six pre-miRNA prediction tools. The analysis is based on the classification performance achieved with these feature sets for the training algorithms used in these tools. We also evaluate feature discrimination through the F-score and feature importance in the induction of random forests. More diverse feature sets produce classifiers with significantly higher classification performance compared to feature sets composed only of sequence-structure patterns. However, small or non-significant differences were found among the estimated classification performances of classifiers induced using sets with diversification of features, despite the wide differences in their dimension. Based on these results, we applied a feature selection method to reduce the computational cost of computing the feature set, while maintaining discriminant power. We obtained a lower-dimensional feature set, which achieved a sensitivity of 90% and a specificity of 95%. Our feature set achieves a sensitivity and specificity within 0.1% of the maximal values obtained with any feature set while it is 34x faster to compute. Even compared to another feature set, which is the computationally least expensive feature set of those from the literature which perform within 0.1% of the maximal values, it is 34x faster to compute.

A reprint is available as PDF.