Use of relevancy and complementary information for discriminatory gene selection from high-dimensional cancer data

Abstract

With the advent of high-throughput technologies, life sciences are generating a huge 1 amount of biomolecular data. Global gene expression profiles provide a snapshot of all 2 the genes that are transcribed or not in a cell or in a tissue at a particular moment 3 under a particular condition. The high-dimensionality of such gene expression data 4 (i.e., very large number of features/genes analyzed in relatively much less number of 5 samples) makes it difficult to identify the key genes (biomarkers) that are truly and 6 more significantly attributing to a particular phenotype or condition, such as cancer or 7 disease, de novo. With the increase in the number of genes, simple feature selection 8 methods show poor performance for both selecting the effective and informative features 9 and capturing biological information. Addressing these issues, here we propose Mutual 10 information based Gene Selection method (M GS) for selecting informative genes and 11 two ranking methods based on frequency (M GS f ) and Random Forest (M GS rf ) for 12 ranking the selected genes. We tested our methods on four real gene expression datasets 13 derived from different studies on cancerous and normal samples. Our methods obtained 14 better classification rate with the datasets compared to recently reported methods. Our 15 methods could also detect the key relevant pathways with a causal relationship to the 16 phenotype. 17 25 indicator of a particular state). Identification of these informative genes is very 26 important for elucidating developmental and disease mechanisms, disease diagnosis, 27 February 22, 2020 1/15drug development, etc. Especially, for different cancer diseases, these informative genes 28 may be invaluable for the improvement of diagnosis, prognosis, and treatment. 29Usually, studies to generate cancer specific gene expression profiles comprise a small 30 number of control and patient samples in comparison to tens of thousands of genes 31 (high dimensionality of the data) in each sample where only a few numbers of genes are 32 responsible for a disease. From a large set of genes, identification of a subset that is 33 differently expressed in cancerous cells compared to the normal ones, is a challenging 34 task and is considered as NP hard or NP-complete [1]. Therefore, the feature/gene 35 selection methods can be a useful way to identify a subset of genes relevant to particular 36 cancer for better diagnosis and treatment. In this paper, we use the terms "gene" and 37 "feature" interchangeably. 38In bioinformatics, several gene selection methods have been proposed, particularly 39 for cancer data classification [2][3][4]. "Wrapper"and "Filter"are two popular categories of 40 feature selection methods [5] where wrapper methods are classifier dependent and filter 41 methods are classifier independent and their performance mainly depends on the 42 selection of a criterion. Wrapper based methods select the most discriminant subset of 43 features by minimizing the prediction error of a particular classifier [6]. Suppor...