Principal Investigator: Dr. Frank Wendt

Affiliation: University of Toronto

Start Year: 2022

Investigative leads in forensic casework often rely on genetic information associated with externally visible characteristics such as eye colour and hair colour. These traits are predicted using a small set of positions in the genome that, when occurring in the correct pattern, are predictive of blue eyes versus brown eyes, for example. Recent large genetic studies support the notion that these traits are far more complex than existing commercial products are capable of predicting.

Furthermore, the pattern-matching used by commercial products may inadvertently detect off-target features of a person such as ancestry and even cancer risk. This project will uncover where biases exist in commercial algorithms that use genetic data to predict eye colour, hair colour, etc. and then build new prediction algorithms that are rigorously tested to avoid/minimize off-target bias. This work will (i) inform best practices for introducing low-bias assays into the legal space, (ii) determine the ability of commercial assays to inform cancer, eczema, and psoriasis risk resulting from similar biological processes, and (iii) minimize the off-target effects of commercial algorithms towards maximizing public wellbeing and minimizing the risk of undesired prediction outcomes.

CanPath data for externally visible characteristics will be used to assess the ability of different machine learning algorithms to predict the target trait independent of ancestry confounding. Follow-up studies in CanPath data will discover regions of the genome associated with externally visible characteristics such that better algorithms can be developed and tested in external cohorts.