Cancer Classification using Microarray Gene Expression Data

Abstract

Accurate and early identification of the molecular nature of cancer is very important in prescribing an effective treatment for it. This has the potential to improve the quality of life for the patients and reduce the toxicity caused by general chemotherapy. Morpho- logical and clinical classification of tumors is effective only in the late stages of tumor development and does not capture molecular level variations. Gene-level data analysis can help in the early diagnosis and classification of cancers. newlineGene expression data made available by the Deoxyribose Nucleic Acid Microar- ray and Next-Generation Sequencing technologies allow the use of data analysis and machine learning techniques to identify useful patterns in this data. The focus of this thesis is on achieving an accurate classification of microarray gene expression profiles through innovations in the machine learning-based methods of feature selection, and classification. newlineOur original contributions to knowledge through this thesis are - the development of an adaptive ensemble feature selection technique with a new score metric, develop- ment of a scheme for globally optimized feature selection using simulated annealing and partial least squares regression, development of a multipurpose scheme for inference of gene regulatory networks and network-based feature selection, and development of an end-to-end optimized framework for accurate classification of gene expression data with few genes. newlineThe first contribution is a two-stage feature selection scheme using adaptive filter- ing of genes with a varying threshold at stage one, followed by an ensemble of feature selection methods at stage two. We identified a small number of relevant genes using this scheme and obtained a classification accuracy of about 100% for each of the six microarray cancer gene expression datasets with these genes. newlineThe second contribution is a feature selection scheme capable of selecting the globally-best features.