Vast-scale linear mixed modelling genetic discovery approaches for genome- and exome-wide association analyses to enable therapeutic target validation

Large-scale publicly available datasets, such as the UK Biobank (n=500,000 participants), which combine genome-wide genotyping and exome sequencing data with linkage to detailed phenotype measurement and electronic healthcare records have the opportunity to transform human genetic discovery analyses. Such datasets are transformative both in their scale and in the depth and diversity of quantitative and disease phenotypes available, and raised a strong interest both in the academia and the industry. In this regard, we have identified partners in Target Sciences (TSci) at GlaxoSmithKline (GSK), a leading team in the application of genetics in drug target discovery and validation. They have previously shown that drugs developed against targets with genetic support for the proposed disease are more likely to reach approval (PMID: 26121088), have used existing GWAS results to search for drug repurposing opportunities (PMID: 22491277) and to develop databases of gene-disease pairs to inform target discovery and validation decisions (PMID: 27899665, 28472345), and have used other biobank samples to influence selection of cardiovascular endpoints (PMID: 26791069) and search for drug repurposing opportunities (PMID: 27301456). GSK have previously performed large-scale targeted sequencing studies (PMID: 22604722) and recently funded exome sequencing of 50,000 participants in UK Biobank, with the aim of further supporting drug target discovery and validation. A major aim at GSK is to use UK Biobank data to conduct phenome-wide association studies (PheWAS), for variants known or predicted to affect gene function for drug targets of interest. The approach currently used is to test each single variant against thousands of disease traits, in the subset of unrelated individuals. However, this approach needs to be improved to distinguish between associations where the drug target variants are likely causal, from associations where the drug target variants are merely correlated (in linkage disequilibrium).

Testing all variants (potentially thousands) in order to fine map in the genomic context of each association of interest is inefficient. A preferable approach is to conduct PheWAS and fine mapping in genomic context, by querying a database of genome-wide association results for all diseases and phenotypes of interest. To maximize discovery power and fine mapping resolution, it is preferable to populate this database with results calculated using in the largest possible sample size. However, an almost inevitable consequence of increasing sample sizes from human populations, is that a larger fraction of participants are related to other participants in the sample. Traditional approaches, such as removing one participant from each related pair, may lead to the removal of a significant proportion of participants from the analysis with consequent loss of statistical power. An alternative approach is using mixed linear model approaches to correct for population structure. However, these approaches require the development of new software tools to deal with large sample sizes, variants and numbers of phenotypes. However, GSK TSci scientists lack the technical expertise required to implement efficient mixed model association testing at the scale required, so this joint project is aimed to collaborate with them to develop the required methods to populate the database. Our work has the opportunity to be impactful on drug discovery and development.

To address the objectives of the fellowship, we will further develop DISSECT (PMID: 26657010). This is a software tool developed within the group, which was designed to overcome the compute and memory limitations of single compute nodes by taking advantage of the aggregate power of the thousands of processor cores and large distributed memory available on supercomputers or large compute clusters. For this purpose, DISSECT distributes the available data over the multiple nodes. At any given time, each node has access to only a small portion of the data on which it performs local computations. When the algorithm requires access to blocks of data currently held on other nodes, the nodes communicate to coordinate data redistribution. This approach provides access to much larger computational resources for a single analysis (i.e. increases the scalability) than standard tools that can only use the resources of a single compute node for each analysis, even when running on similar computer clusters environments. In addition, using as a basis our current development, we will further develop, evaluate and implement previous approaches (PMID: 25642633, 21465547) that propose to perform approximations to reduce the computational cost of fitting these models on large datasets, and find a balance between speed, accuracy, and computation requirements. The proposed analyses will be run on Tier-1 and Tier-2 High Performance Computing Centres such as ARCHER (https://www.archer.ac.uk) and CIRRUS (http://www.cirrus.ac.uk).