Aprendizagem de Máquina em Análise de Expressão Gênica: Classificação e Seleção de Genes Relevantes em Câncer

Abstract

The selection of gene biomarkers can benefit the clinical management of patients and provide insights into the understanding of disease. However, the analysis of gene expression data is still a challenge, due to: curse of dimensionality (high number of features and small number of samples); presence of non-linear complex relationships; high background noise; and difficulty in analyzing datasets together (with different noise and scales). This work aims to analyze Deep and Shallow Machine Learning methods for sample classification and selection of relevant genes in microarray datasets from different tissues. We explore the following methods: Convolutional Neural Networks (CNNs); Vision Transformers; Generative Adversarial Network (GAN); and Multilayer Perceptron (MLP). We propose the adoption of the Transcriptogram method for analyzing gene interaction networks and obtaining an ordering of genes that can be explored by CNNs. Furthermore, we propose an approach, called MLPEns, which exploits dropout to treat a Neural Network model as an ensemble of models, aiming to take advantage of the high generalization capacity of the ensembles. For the selection of the relevant genes, we analyze: coefficients of Support Vector Machines (SVMs); Boruta Shap; and weights of MLPEs. Our results demonstrate that, contrary to the recent trend, some Shallow Learning methods (such as SVM Linear L2 and Ridge Regression) provide results that are statistically equivalent to our best Deep Learning approach and, in addition, have shorter execution times and greater simplicity in interpretability of results. Therefore, we perceive these methods as the best options for analyzing this type of data. The approach of employing the Transcriptogram has improved CNN accuracy; and MLPEns obtained results statistically equivalent to the best approaches analyzed.