Machine Learning Strategies for Biomarker Discovery: A Senescence Case Study

Genetic biomarkers play a crucial role in genome-to-phenotype mapping. They are essential for annotating cell identities, assessing treatment effects, monitoring cell progression and development, and exploring cell-cell interactions. The wealth and complexity of noisy genomic data presents a formidable challenge to the discovery of biomarkers, necessitating the development of computa- tional approaches that can overcome experimental biases and sweep across vast genetic landscapes. Machine learning algorithms, which address both the scale of data and its inherent stochasticity, have emerged as a natural solution to these issues.

In this thesis, we explore machine learning strategies for biomarker discovery across different contexts. Our investigation is categorized into two primary themes; steady state and dynamic. In the steady state context our main objective is to identify biomarkers that differentiate conditions, cell types, and cell states within a given sample. We design computational methods that specifically target the static context across three learning settings, contingent upon the nature and availability of label information. These include supervised, weakly supervised, and unsupervised approaches. For the dynamic context, we are primarily concerned with the discovery of biomarkers that vary with time and explain the dynamics of the biological system being investigated. Here we focus on the problems of trajectory inference, as well as deep learning approaches for temporal graph learning. By applying these methods to senescence (a form of aging) and other real world datasets and diseases, we substantiate their practical value in unraveling the intricate relationship between genes and phenotypes