Machine learning for computational phenotyping and genetic discovery

Unsupervised phenotyping of medical images to better characterize the genetics of complex diseases

Medical images are a rich source of phenotypic data but typically require specific and expensive annotations of a given phenotype to utilize. Supervised approaches to “phenotyping images” therefore have both a time and information bottleneck: It is difficult to obtain annotations and they only quantify a subset of the possible phenotypes present in each image acquisition. In this seminar, I will discuss methods my group are using to perform self-supervised and unsupervised phenotyping of histology and population-based MRI data. Our primary goal is to extract phenotypes from images and combine them with clinical and genetic data to better understand the mechanisms of complex disease. I will discuss strategies we are using to provide interpretability of such unsupervised phenotyping methods and outline how we can use such methods to discover novel genetic variation associated with image derived phenotypes.

Dr Craig A. Glastonbury holds a PhD in computational biology from King’s College London (2013-2017) where he focused on mapping tissue specific eQTLs in multiple human tissues. After his PhD Craig worked as a Postdoctoral Fellow for 2 years in the lab of Cecilia Lindgren. Craig spent 4 years after his postdoc as a lead machine learning researcher at the drug discovery company, BenevolentAI (2019-2022), focusing on how to use human genetics for target discovery and ML methods for patient stratification. Craig’s research has focused on multiple aspects of computational biology, with a key focus on histopathology imaging using machine learning and human genetics. The Glastonbury Group at Human Technopole aims to combine these expertise, by studying how genetic variation influences quantifiable phenotypes extract from a whole range of biomedical imaging modalities. Craig serves on the organising committee of the International Common Disease Alliance (ICDA) and also as a guest associate editor for machine learning at AHA Circulation.