Sequence-based deep learning models for prioritisation of disease-associated non-coding variants in cardiovascular disease.

The vast majority of genetic variants linked to common cardiovascular diseases are found in intronic or intergenic regions of the genome and do not exhibit an obvious mechanism by which they influence disease.

Our previous work on one disease (hypertension) and one tissue (kidney) found that roughly half of these variants exert a causal influence on disease risk through regulation of gene expression, alternative splicing or DNA methylation.

In this project we intend to expand our work to all human tissues of relevance to cardiovascular disease (i.e. vascular, heart or adipose tissue) using a deep-learning approach to capture the DNA-sequence features relevant to gene expression and alternative splicing.

A panel of known deep-learning architectures will be implemented and predictive performance will be assessed against a held-out set of genes (across all tissues).

The project will make extensive use of data-augmentation methods, via resampling and down-sampling of the primary RNA-sequencing input data, to improve the optimisation of neural network weights, without generation of additional data.

An optimal deep-learning model will be chosen based on independent validation of known variant effects on gene expression (eQTLs) and alternative splicing (sQTLs).

The final model will then be dissected using in-silico mutagenesis (input of synthetic DNA sequences) to identify the most relevant regulatory sequences per tissue and these will then be combined with the results of genome-wide association studies to create an encyclopaedia of regulatory sequences that are tissue and disease specific.