Methods: We performed multi-steroid profiling by GC-MS, quantifying 34 steroid metabolites, in urine samples from 829 healthy controls and 178 untreated patients with inborn steroidogenic disorders. This cohort included patients with inborn deficiencies in the following enzymes: CYP21A2 (n=26), CYP11B1 (n=12), CYP17A1 (n=30), POR (n=37), HSD3B2 (n=22), and SRD5A2 (n=51). We assessed the diagnostic performance of conventional biochemical assessment employing 15 steroid precursor-to-product ratios, each historically established as indicative of a distinct steroidogenesis disorder. We compared this to the performance of our novel steroid metabolomics approach, which involved analysis of the GC-MS multi-steroid profiles by a custom-designed approach, Angle Learning Vector Quantization (ALVQ), which classifies samples by comparing similarity of their steroid metabolome to representative steroid metabolome prototypes for each enzyme deficiency.
Results: The conventional biochemical steroid ratio approach demonstrated acceptable sensitivity and specificity. However, the automated steroid metabolomics approach (ALVQ) performed significantly superior to this, particularly with regards to specificity. For differentiating patients from healthy controls, sensitivity and specificity of ALVQ were 97% and 98%, respectively. For differentiation of each pathogenic enzymatic defect, ALVQ performed superiorly, with sensitivity and specificity ranging between 95 and 100%.
Conclusion: We present a novel steroid metabolomics approach, able to automatically detect and differentiate six different inborn disorders of steroidogenesis, with improved performance when compared to reference standard metabolite ratios. Steroid metabolomics can expedite and standardise interpretation of complex urinary steroid metabolome data, making this technique more accessible to clinicians, and has excellent potential for implementation in routine clinical practice.