Many phenotypic traits – observable qualities like height or disease susceptibility -- have a strong genetic basis. Models that can predict traits from genetic data have applications in fields ranging from healthcare and agriculture to forensics. Because of the genome's large size and high sequencing cost, models that predict traits often have more features (genomic sites) than training observations (sequenced individuals) and are thus prone to low accuracy from overfitting. In particular, it can be difficult to distinguish the few truly causal features from the large number of benign but correlated genetic, environmental, or cultural differences that exist between populations.
In this episode, we describe the challenges posed by population structure and a computational approach developed by Microsoft Research to address the problem: Factored Spectrally-Transformed Linear Mixed Models (FaST-LMM). We provide an online walkthrough demonstrating how their method can be applied to real genomic data and simulated phenotypes in Python on a Linux Data Science Virtual Machine.