SMOTE-LASSO-DeepNet Framework for Cancer Subtyping from Gene Expression Data

Cancer subtyping from gene expression data is trending research in the field of bioinformatics. Classification of gene expression data is a challenging task due to the small number of samples and large number of features involved. The problem is further complicated due to the strong class imbalance issue prevalent in gene expression datasets. The challenge here is to find an end-to-end machine learning solution to classify cancer subtypes from small sample, high-dimensional, imbalanced gene expression datasets. In this study, we propose a SMOTE-LASSO-DeepNet framework for the identification of cancer subtypes from gene expression data. The proposed framework balances the training set using SMOTE, and then finds the most informative genes using LASSO. The balanced and pruned training set is then applied as input to a deep neural network (DeepNet) with four hidden layers having 512, 256, 128 and 64 neurons respectively. We tested our framework on four different cancer gene expression datasets: Leukemia, Lung cancer, Brain cancer and Breast cancer. It is observed from the results that our proposed SMOTE-LASSO-DeepNet framework performs consistently best as compared to the existing methods.