The Limitations of Existing Approaches in Improving MicroRNA Target Prediction Accuracy.
Bioinformatics in MicroRNA Research
Methods in molecular biology, January 2017
Rasiah Loganantharaj Ph.D., Thomas A. Randall, Rasiah Loganantharaj, Loganantharaj, Rasiah, Randall, Thomas A.
Jingshan Huang, Glen M. Borchert, Dejing Dou, Jun (Luke) Huan, Wenjun Lan, Ming Tan, Bin Wu
MicroRNAs (miRNAs) are small (18-24 nt) endogenous RNAs found across diverse phyla involved in posttranscriptional regulation, primarily downregulation of mRNAs. Experimentally determining miRNA-mRNA interactions can be expensive and time-consuming, making the accurate computational prediction of miRNA targets a high priority. Since miRNA-mRNA base pairing in mammals is not perfectly complementary and only a fraction of the identified motifs are real binding sites, accurately predicting miRNA targets remains challenging. The limitations and bottlenecks of existing algorithms and approaches are discussed in this chapter.A new miRNA-mRNA interaction algorithm was implemented in Python (TargetFind) to capture three different modes of association and to maximize detection sensitivity to around 95% for mouse (mm9) and human (hg19) reference data. For human (hg19) data, the prediction accuracy with any one feature among evolutionarily conserved score, multiple targets in a UTR or changes in free energy varied within a close range from 63.5% to 66%. When the results of these features are combined with majority voting, the expected prediction accuracy increases to 69.5%. When all three features are used together, the average best prediction accuracy with tenfold cross validation from the classifiers naïve Bayes, support vector machine, artificial neural network, and decision tree were, respectively, 66.5%, 67.1%, 69%, and 68.4%. The results reveal the advantages and limitations of these approaches.When comparing different sets of features on their strength in predicting true hg19 targets, evolutionarily conserved score slightly outperformed all other features based on thermostability, and target multiplicity. The sophisticated supervised learning algorithms did not improve the prediction accuracy significantly compared to a simple threshold based approach on conservation score or combining the results of each feature with majority agreements. The targets from randomly generated UTRs behaved similar to that of noninteracting pairs with respect to changes in free energy. Availability of additional experimental data describing noninteracting pairs will advance our understanding of the characteristics and the factors positively and negatively influencing these interactions.
|Members of the public||1||100%|
|Readers by professional status||Count||As %|
|Student > Master||2||20%|
|Student > Ph. D. Student||2||20%|
|Student > Bachelor||1||10%|
|Readers by discipline||Count||As %|
|Biochemistry, Genetics and Molecular Biology||2||20%|
|Agricultural and Biological Sciences||1||10%|