Machine-learning to create predictive models of genetic regulation

Genetic regulatory networks control critical biological processes, from development and cellular differentiation to the response to environmental cues and stress. In eukarytotes however the networks are substantially unknown and difficult to predict, since they may differ substantially in different cell types and because enhancer or supressor activities for individual genes may be driven from distal genomic regions. In recent years there has been an explosion in high throughput data generation relevant to genetic regulation. This has been primarily based on next generation DNA sequencing, and includes DNase/ATAC-seq, ChIP-seq for chromatin modification and transcription factor binding, RNA-seq for gene expression measurement and increasingly 3D chromosome structure data. This project aims to apply machine learning methods to these large, heterogeneous data sets to predict genetic regulatory networks. It follows on from recent work in the group, where we have developed penalized regression methods integrate DNase-seq, ChIP-seq and RNA-seq data into regulatory models for individual genes. It will employ methods such as neural networks and support vector machines, where we have existing expertise, and an important aspect of the project will be assessment of the utility of the prediction models in analysing the biological consequences of regulatory (non-coding) DNA changes, for instance in genome wide association studies of disease phenotypes and the interpretation of mutational data, for instance from cancer cells.