4 Year MRC PhD Programme: Predicting drug outcomes in diabetes using deep learning

Precision therapeutics aims to identify what patient characteristics predict the best treatment for an individual, from simple clinical features to molecular signals conveyed by genomic, metabolomic and other –omic data (1).

Within diabetes, some patients respond well to a treatment, others do not; some progress rapidly to require insulin treatment, others can manage on one tablet for many years. The response to a therapeutic intervention depends on the physiological state of an individual at the time of treatment. The physiological state may be better understood by more careful molecular phenotyping measured at a point in time. In addition, we propose that the longitudinal exposures in the months or years prior to a drug intervention will influence this physiological state and subsequent drug response.

This studentship will apply artificial intelligence approaches to improve predictive models of drug outcomes in diabetes. Conventional bootstrap analysis based on logistic regression with feature selection, supervised machine learning and conventional statistics will provide a baseline. We shall then explore deep learning (2) using the NVIDIA GPU facilities in Computing. We envisage investigating unsupervised learning to discover embedded data representations, and recurrent deep networks to model longitudinal data (see (3) and https://arxiv.org/pdf/1511.03677.pdf).

We will use the GoDARTS study, containing extensive, well characterized clinical and molecular phenotypic data on 10,000 diabetic patients in Tayside, with >20 years follow-up captured via the electronic medical record. In addition we will have access to data on 270K diabetics in Scotland and 400,000 in India (via the recently awarded £7M NIHR Unit in global outcomes in diabetes).

The student will join a rich, interdisciplinary environment across the Schools of Medicine and Science and Engineering, receiving training in both medical and computational topics. To maximize interdisciplinary learning, the student will be co-located with cognate researchers and data scientists, with space both in the VAMPIRE/CVIP laboratory in Computing (Queen Mother Building) and the new purpose-built space planned in an expanded Farr@Dundee/Health Informatics Centre housing researchers from a spectrum of disciplines.