Learning from multi-omics data of cancer

Abstract

Analysing multiple types of omics data is a keystone methodology in biomedical research nowadays. With the significant advances of high-throughput experimental technologies, an enormous amount of omics data with extremely high dimensionality are generated every day at an unprecedented speed, which leaves a massive data gold ore waiting to be mined. However, it is difficult for traditional bioinformatics methods to deal with the high dimensionality and enormous data amount. The rapid development of machine learning, especially the deep learning methodology, has dramatically revolutionised fields like natural language processing and computer vision over the past decade. Deep learning has shown great success in decoding high-dimensional data like images, which makes it promising to adopt this cutting-edge technology to the analysis of multi-omics data. In this thesis, we propose a comprehensive toolbox for deep learning-based multi-omics data analysis named OmiSuite to decode high-dimensional multi-omics data and unveil the correlation between the phenotype profile and different types of omics profiles. OmiSuite is comprised of four components: OmiVAE, OmiEmbed, XOmiVAE, and OmiTrans. Among them, OmiVAE is one of the first endeavours to decode high-dimensional multi-omics data using variational autoencoders for pan-cancer classification. OmiEmbed is a unified multi-task deep learning framework for multi-omics data, supporting multi-omics integration, dimensionality reduction, omics embedding, tumour type classification, phenotypic feature reconstruction, survival prediction, and multi-task learning for aforementioned tasks. XOmiVAE is the explainable upgrade of OmiVAE, which can provide the contribution score of each molecular feature and each latent dimension for each phenotype prediction. OmiTrans is the first generative adversarial networks-based omics-to-omics translation framework, which ushered in a brand-new research topic with a promising vision. These four components of OmiSuite created a unified ecosystem for high-dimensional multi-omics analysis, which covered almost every aspect of this field, benefited follow-up studies and led to an upsurge of research in this field.