Deep learning approaches for an integrated study of DNA sequence, epigenome and chromatin architecture

Thesis

In the era of genome sequencing, it has become clear that interpreting sequence variation in the non-coding genome represents a critical bottleneck for translating genomic insights to benefits for human health. The non-coding genome harbours the complex sequence patterns that encode regulatory elements and their intricate interplay over large genomic distances, ensuring precise gene regulation throughout time and cell differentiation. This complex sequence grammar is challenging for humans to comprehend and the experiments required to probe it are expensive and time-consuming. Machine learning approaches offer a powerful tool for studying the non-coding genome in silico.

In this thesis work, we explored the use of deep neural networks for the purpose of interpreting genomic variations in the non-coding genome. We re-implemented, optimised and evaluated proposed convolutional neural networks models for predicting chromatin features from DNA sequence and utilise them in an integrated pipeline for dissecting non-coding sequence variants identified by Genome-Wide Association Studies (GWAS). We applied this pipeline to variants associated with red blood cell traits, demonstrating how neural networks can facilitate the interpretation of loss and gain of function variants in regulatory elements.

Although chromatin feature networks evaluate variation in larger sequence context than previously possible, they still only offer a highly local perspective ignoring chromatin architecture and the distal interplay of regulatory elements. To address this limitation, we developed a novel neural network framework (deepC) using dilated convolutions and transfer learning to predict chromatin interactions from megabase scale DNA sequence. We validated deepC predictions using chromosome conformation assays at high sensitivity. Finally, we demonstrated how deepC can be used to dissect the sequence determinants of chromatin folding and to predict the impact of structural and single nucleotide variation on genome architecture. Together this thesis evaluates, utilises and expands the deep learning toolbox for genomics and our capabilities of interpreting non-coding genomic variation.

Authors: Schwessinger, R

• DOI: 10.5287/ora-ao1eok4n2
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: Functional Genomics; Deep Learning; Genomics; Machine Learning; Chromatin Architecture; Gene Regulation
• Subjects: Functional genomics; Machine learning