Machine learning classification of single cell rna-seq across different types of cáncer.

Abstract

Human cancers are complex ecosystems composed of different types of cells. The diverse populations of co-existing cells within the same tumor that have genetic, functional, and environmental differences determine the tumor heterogeneity, which is one of the major challenges facing cancer diagnosis and treatment. The aim of this thesis was to apply different machine learning methods to classify single cell RNA-seq (scRNA-seq) samples across nine different types of cancer. We observed that T cells are the most abundant datasets in public repositories due to their important role in immunotherapies. For this reason, we performed an in-silico analysis from scRNA-seq data available in the Gene Expression Omnibus. A őrst approach was to analyze and characterize genetic T cell signatures from őve different types of cancer and apply dimensionality reduction and clus tering methods to identify subpopulations from malignant and non-malignant datasets. This analysis revealed that pathways related to immune response, metabolism and viral immunoregulation were observed exclusively in samples of malignant origin. A second approach was to perform two deep learning models to classify cells from nine different types of cancer, where the cells were grouped in the diversity of the cell state, giving us a new perspective in the different classes of tumors present in our dataset. Finally, we observed that working with unsupervised methods, our data help us understand the heterogeneity between tumors. Characterization of cellular diversity was associated with pathways that play a key role in tumor proliferation, progression, and regulation of the microenvironmental immune response.