Inferring Cell State In Disease Progression From Gene Expression Data Using Tractable Deep Probabilistic Models

Ava Khamseh, ava.khamseh@ed.ac.uk
Antonio Vergari, avergari@exseed.ed.ac.uk
Sjoerd Beentjes, Sjoerd.Beentjes@ed.ac.uk

In this cross-disciplinary project spanning machine learning, statistics, and biomedicine, the student will study a large class of tractable probabilistic models (TPMs), including probabilistic circuits (PCs) and normalising flows (NFs). PCs and NFs are expressive deep generative models as they encode several layers of latent variables into large graphs with potentially millions of connections and parameters. These models allow for exact probabilistic inference in time linear in the size of the model, which is necessary when dealing with thousands of gene expression variables across multiple cell types and conditions. The biomedical aim of this project is to identify and compare cell (sub)types and states in normal and disease tissue at large-scale, e.g. human cell atlases involving hundreds to millions of cells across multiple tissues and/or disease progression. This project is suitable for a candidate with a strong mathematical and/or machine learning background, as well as biomedical motivation. The project is a cross-disciplinary project, and will be supervised across 3 institutes: Institute of Genetics and Cancer and School of Informatics (Khamseh [quantitative biology]), School of Informatics (Vergari [deep probabilistic models]) and School of Mathematics (Beentjes [mathematical statistics]).