Predicting the Evolutionary and Medical Significance of Human Genetic Variations with Machine Learning

• Main Author: Saha Mandal, Arnab
• Format: Article in Journal/Newspaper
• Language: unknown
• Published: Cumming School of Medicine 2019
• Subjects: bioinformatics, genomics, machine learning, genomic variants, classification, pathogenic, benign, rare disease, genetics; Bioinformatics; FOS Computer and information sciences; Genetics; FOS Biological sciences; Artificial Intelligence; Computer Science; taiga
• Online Access: https://dx.doi.org/10.11575/prism/36479; https://prism.ucalgary.ca/handle/1880/110303

The advent of inexpensive and high-throughput genome sequencing technologies has facilitated the acquisition of patient exome and genome sequences at a vast scale. One of the primary challenges of such data is its functional interpretation, and specifically, the ability to distinguish functionally important, deleterious, and pathogenic variants from neutral or benign variants (“variant impact prediction” or VIP). Over the last two decades, many approaches have been proposed for VIP, which utilize data from patterns of evolutionary conservation, population genomics, protein structures and other sources to inform machine learning classification algorithms. However, existing approaches are fraught with limitations, especially when they are trained on databases of putatively pathogenic variants that may have been identified with reference to existing prediction methods (a type of ‘circularity’). This dissertation identifies shortcomings of existing variant impact prediction methods and discusses how they can be better understood (Chapter 1). Approaches to overcome these shortcomings are presented (Chapter 2), and a new method, TAIGA (Transformation and Integration of Genomic Annotations), is developed. The utility of this method and its accompanying refinements are evaluated (Chapter 3) and later scrutinized (Chapter 4). As part of this work, I have produced TAIGA scores for all protein coding positions of the human genome, and I show these have substantially superior performance in distinguishing known pathogenic variations from neutral variations in a number of high-quality datasets. Variant prediction scores from TAIGA are later integrated with clinical information from human phenotypes (Chapter 5) and this extension demonstrated the highest sensitivity and smallest candidate gene search space over a large set of rare genetic disorders. It is my hope that TAIGA will aide clinicians and researchers alike in the new era of personalized genomic medicine in which we find ourselves.