Learning Disease Patterns from High-Throughput Genomic Profiles: Why Is It So Challenging?

Abstract

In the 20^th century, genetic scientists anticipated that shortly after availability of the whole-genome profiling technologies, the patterns of complex diseases would be decoded easily. However, we recently found it extremely difficult to predict women’s susceptibility to breast cancer based on their germline genomic profiles and achieved an accuracy of 59.55% over the baseline of 51.52% after applying a wide variety of biologically-naïve and biologically-informed feature selection and supervised learning methods. By contrast, in a separate study, we showed that we can utilize these genomic profiles to accurately predict ancestral origins of individuals. While there are biomedical explanations of accurate predictability of an individual’s ancestral roots and poor predictability of her susceptibility to breast cancer, my research attempts to utilize the computational learning theory framework to explain what concepts are learnable, based on the three common characteristics of biomedical datasets: the high dimensionality, the label heterogeneity, and the noise.