Seminar Series in Quantitative Life Sciences and Medicine

"Developing machine learning models for phenotype prediction vs disease mechanism detection. A play in 3 acts."

Anna Goldenberg (University of Toronto)
Tuesday December 4, 12-1pm
McIntyre Building, Room 1027

Abstract: There is a great potential for machine learning to contribute to understanding complex human diseases and clinical decision making. Rapidly evolving biotechnologies are making it progressively easier to collect multiple and diverse genome-scale datasets to address clinical and biological questions. Much of the work is driven by a great human propensity to explain the unknown. We have to be careful, however, in trying to explain models applied to biological data that were built for purposes other than data explanation, such as models that were built purely to predict a specific phenotype. In this talk, if time allows, I will talk about 3 different applications. The first model predicts whether a child with a TP53 mutation is likely to get cancer before the age of 6 using regularized regression; the second model predicts drug response using variational autoencoders; the third is a graphical model that was built specifically to identify disease mechanisms. In each case, I will highlight our attempts to explain the biology behind the predictions and the perils of doing so.