Deep Learning Models of Cellular Decision-Making Using Single-Cell Genomic Data

Abstract

Cellular decision-making, essential to regenerative medicine, disease research, and developmental biology, relies on complex molecular mechanisms that guide cells in responding to stimuli and committing to specific fates. This thesis introduces several deep learning methods to analyze single-cell RNA sequencing data, uncover regulatory programs driving these processes, and predict the outcomes of gene perturbations.

By applying representation learning and generative models, meaningful structures within high-dimensional data are identified, enabling tasks such as mapping cellular trajectories, reconstructing regulatory networks, and generating realistic synthetic data. Furthermore, integrating deep learning with dynamical systems theory enables the prediction of cellular decision timing and the identification of key regulatory genes involved in these processes. These methods enhance our understanding of gene activity dynamics, improve predictions of cellular behavior, and offer new avenues for progress in regenerative medicine, developmental biology, and disease research.