Prediction of the multivariate profile of complications of Type 1 diabetes from genotypes and phenotypic biomarker

Long-term complications of diabetes include nephropathy, retinopathy, neuropathy and vascular disease. Learning to predict these complications is important for clinical practice, for stratifying patients for drug trials, for identifying biomarkers that can be used as surrogate end-points, and for discovery of new therapeutic targets. Although there is evidence that the risks of these complications are under strong genetic influence, only a few associations of complications with genetic variants have been identified. Studies so far have tested associations only with single complications such as nephropathy or retinopathy, not at multivariate complications.
The Scottish Diabetes Research Network Type 1 Bioresource is a cohort of 6116 people aged over 16 years with a clinical diagnosis of Type 1 diabetes, linked to electronic health records, on whom genome-wide genotype and biomarker profiles have been measured. Measurement of additional biomarker panels is in progress. This makes it possible to construct predictive models of the entire profile of diabetic complications.
The principal aim of the project is to develop a predictive model for multiple diabetic complications, using all available genotypic and phenotypic biomarkers. We have developed these methods for single outcomes, but extending these to joint modelling of multiple outcomes presents a considerable challenge. In principle this type of problem is tractable, using recently developed software tools for Bayesian statistical modelling of high-dimensional data.
Subsidiary aim is to discover pathways underlying susceptibility to these complications. One approach to this is to construct genotypic predictors of intermediate variables of interest and test them for association with the outcome. This can be extended to a "mendelian randomization" hypothesis test which uses the genotypic scores as "instruments" that perturb the intermediate variable. We have developed a platform (GENOSCORES) for this type of analysis.
To sum up, our predictive modelling approaches will range from standard penalized regression methods such as LASSO to cutting-edge Bayesian methods using Markov chain Monte Carlo and variational approximations, implemented in programs such as Stan and Edward. We will use individual-level and summary-level data from genome-wide association studies for prediction of univariate and multivariate outcomes, allowing for relatedness.