A Systematic Evaluation of Feature Selection and Classification Algorithms Using Simulated and Real miRNA Sequencing Data

Sequencing is widely used to discover associations between microRNAs (miRNAs) and diseases. However, the negative binomial distribution (NB) and high dimensionality of data obtained using sequencing can lead to low-power results and low reproducibility. Several statistical learning algorithms have been proposed to address sequencing data, and although evaluation of these methods is essential, such studies are relatively rare. The performance of seven feature selection (FS) algorithms, including baySeq, DESeq, edgeR, the rank sum test, lasso, particle swarm optimistic decision tree, and random forest (RF), was compared by simulation under different conditions based on the difference of the mean, the dispersion parameter of the NB, and the signal to noise ratio. Real data were used to evaluate the performance of RF, logistic regression, and support vector machine. Based on the simulation and real data, we discuss the behaviour of the FS and classification algorithms. The Apriori algorithm identified frequent item sets (mir-133a, mir-133b, mir-183, mir-937, and mir-96) from among the deregulated miRNAs of six datasets from The Cancer Genomics Atlas. Taking these findings altogether and considering computational memory requirements, we propose a strategy that combines edgeR and DESeq for large sample sizes.