Deep Learning for Single-cell Omics Data Analysis

Abstract

Single-cell omics sequencing technologies have provided unprecedented opportunities

in characterising biological systems at the omics scale, leading to systematic

understanding of various biological processes, such as stem cell fate decisions and

disease progression. Machine learning models have been at the centre for analysing

single-cell omics data and inferring biological knowledge. Owing to the high dimensionality

(e.g. large number of profiled genes) and highly complex nature of

single-cell omics data, feature selection, a class of machine learning techniques that

select a subset of informative features, is essential for handling single-cell omics

data for downstream analyses.

In this thesis, I first summarise machine learning and feature selection techniques,

and their development and application to single-cell omics data based my

recently published review. I then apply and evaluate deep learning-based feature

selection methods for selecting cell type-specific genes for classifying cells in

single-cell transcriptomics data. I demonstrate the utility and advantages of deep

learning-based feature selection methods for cell type classification by creating

single-cell transcriptomics datasets sampled from Tabula Muris and Tabula Sapiens

atlases, with various data properties. Next, I incorporate the deep learningbased

feature selection model in a multi-task learning framework for analysing

single-cell multimodal omics data that simultaneously profile multiple molecular

features of individual cells. I show that our multi-task deep learning framework is

capable of performing data simulation, data augmentation, cell type classification,

and feature selection for various single-cell multimodal omics data types.