Autism spectrum disorder (ASD) is a neurodevelopmental disorder with no clinical biomarker. The aims of this study were to characterize a metabolic signature of ASD and to evaluate multiplatform analytical methodologies in order to develop predictive tools for diagnosis and disease follow-up. Urine samples were analyzed using (1)H and (1)H-(13)C NMR-based approaches and LC-HRMS-based approaches (ESI+ and ESI- on HILIC and C18 chromatography columns). Data tables obtained from the six analytical modalities on a training set of 46 urine samples (22 autistic children and 24 controls) were processed by multivariate analysis (orthogonal partial least-squares discriminant analysis, OPLS-DA). The predictions from each of these OPLS-DA models were then evaluated using a prediction set of 16 samples (8 autistic children and 8 controls) and receiver operating characteristic curves. Thereafter, a data fusion block-scaling OPLS-DA model was generated from the 6 best models obtained for each modality. This fused OPLS-DA model showed an enhanced performance (R(2)Y(cum) = 0.88, Q(2)(cum) = 0.75) compared to each analytical modality model, as well as a better predictive capacity (AUC = 0.91, p-value = 0.006). Metabolites that are most significantly different between autistic and control children (p < 0.05) are indoxyl sulfate, N-α-acetyl-l-arginine, methyl guanidine, and phenylacetylglutamine. This multimodality approach has the potential to contribute to find robust biomarkers and characterize a metabolic phenotype of the ASD population.
Keywords: ASD; LC−HRMS; NMR; autism spectrum disorder; data fusion; metabolomics.