MicroRNA-Based Cancer Classification Using Feature Selection Wrapper

Abstract

Background: MicroRNAs (miRNAs) are a class of $\sim$22-nucleotide endogenous non-coding RNAs, having critical roles across various biological processes. It also performs a meaningful character in tumorigenesis. Therefore, the identification of differentially expressed miRNAs is a growing challenge. In this regard, our paper presents the approach for the selection of miRNA markers from high-throughput sequencing data with Least Absolute Shrinkage and Selection Operator (LASSO), Covariance Matrix Adaptation Evolution (CMA-ES), and inner classifiers. Results: The proposed method select features using regression analysis, evolutionary optimization, and inner classifiers, named after its underlying methods. LASSO is used as the dimensionality reduction method. CMA-ES optimizer is considered here to search the space of miRNA subsets, which yields the best performance concerning the inner classifiers. We have investigated several inner classifiers to choose the best-performing one in exchange for the given objective. Conclusions: The proposed miRNA selection task uses real, next-generation sequencing data from a United States-based consortium, The Cancer Genome Atlas (TCGA), and concerns miRNA expression levels in healthy and malignant tissues. Moreover, the emphasis is given here to determine the miRNAs with differential expression patterns. By doing so, the work of the proposed method is checked with other up-to-date methods as classification accuracy. We conclude by analyzing the selected, most relevant miRNAs in differentiation between sample types. These selected miRNAs are also been validated using different biological significance analyses. Our method reduces the number of miRNAs from several hundred to few, thereby facilitating a more target-oriented experimental validation.