Machine Learning Methods for Single Cell RNA-Sequencing Data to Improve Clinical Oncology

Abstract

Single-cell RNA sequencing (scRNA-seq) offers a detailed view of the cellular and phenotypic composition of healthy and diseased tissues. While machine learning (ML) methods are well-suited for the high-dimensional nature of scRNA-seq data, current computational tools face limitations, particularly when confronted with data from clinical oncology. This thesis presents the development and application of ML techniques for scRNA-seq data to address key computational challenges, with a focus on challenges in clinical oncology. It covers four key areas: identifying gene signatures and biomarkers in multiple myeloma, developing methods to account for somatic copy number variations in tumor samples, benchmarking large, pre-trained scRNA-seq foundation models, and creating a framework for predicting clinical outcomes using patient-level representations of single-cell data. Together, these studies aim to develop and evaluate novel ML algorithms for scRNA-seq data which can unlock actionable insights for personalized medicine.