Predicting the Effects of Genetic Medicines Using Transfer Learning
Genetic medicines promise the ability to precisely target the root causes of disease. At Deep Genomics, we are developing machine learning systems to predict the properties of these medicines, including activity and safety. A fundamental problem in doing so is that large collections of therapeutic data is infeasible to collect. Using transfer learning allows us to fuse large amounts of inexpensive biology data with small amounts of therapeutic data. I will discuss how we have successfully used transfer learning to predict the on-target activity of genetic medicines, enabling us to test five times fewer compounds for some of our targets.
Amit Deshwar is the Director of Predictive Systems at Deep Genomics. His doctoral work was at the University of Toronto under Quaid Morris using Bayesian Non-parametrics to study intra-tumour heterogeneity and evolution. He is a Vanier Scholar and former Junior Fellow at Massey College. Previously he worked at Google, started two companies and obtained undergraduate degrees in Software Engineering and Psychology from the University of Calgary.