Untargeted metabolomics for metabolic diagnostic screening with automated data interpretation using a knowledge-based algorithm

Untargeted metabolomics may become a standard approach to address diagnostic requests, but, at present, data interpretation is very labor-intensive. To facilitate its implementation in metabolic diagnostic screening, we developed a method for automated data interpretation that preselects the most likely inborn errors of metabolism (IEM). The input parameters of the knowledge-based algorithm were (1) weight scores assigned to 268 unique metabolites for 119 different IEM based on literature and expert opinion, and (2) metabolite Z-scores and ranks based on direct-infusion high resolution mass spectrometry. The output was a ranked list of differential diagnoses (DD) per sample. The algorithm was first optimized using a training set of 110 dried blood spots (DBS) comprising 23 different IEM and 86 plasma samples comprising 21 different IEM. Further optimization was performed using a set of 96 DBS consisting of 53 different IEM. The diagnostic value was validated in a set of 115 plasma samples, which included 58 different IEM and resulted in the correct diagnosis being included in the DD of 72% of the samples, comprising 44 different IEM. The median length of the DD was 10 IEM, and the correct diagnosis ranked first in 37% of the samples. Here, we demonstrate the accuracy of the diagnostic algorithm in preselecting the most likely IEM, based on the untargeted metabolomics of a single sample. We show, as a proof of principle, that automated data interpretation has the potential to facilitate the implementation of untargeted metabolomics for metabolic diagnostic screening, and we provide suggestions for further optimization of the algorithm to improve diagnostic accuracy.