The dissertation begins with an overview of current practices and limitations in the interpretation of variation in genes underlying inborn errors of metabolism and drug response, detailing experimental approaches to validating a variant as causative or pathogenic and summarizing advances in computational methods aiming to predict variant effect on protein function. We propose a vision for a genomic learning healthcare system (GLHS) that facilitates the translation of a patient’s genome into clinically actionable information for diagnostic and therapeutic purposes. After the overview, we present a rich set of experimental and computational approaches, which were developed and used to improve the functional prediction of genetic variants in OCTN2 for diagnosis of CTD. We functionally characterized 150 OCTN2 missense variants and found that 71% of variants had a significant effect on the uptake of carnitine. 25% of variants reduced transporter function to less than 20% of the wild-type OCTN2, a clinically meaningful threshold for CTD. We asked what was causing reduced function, and identified improper subcellular localization to be a major loss-of-function mechanism affecting 62% of variants. These data were then used in machine learning to build a protein-specific variant effect prediction model that accurately classified variants of OCTN2 as functional (>20%) or LOF (