Untargeted urine metabolomics and machine learning provide potential metabolic signatures in children with autism spectrum disorder.

Abstract

Complementary to traditional biostatistics, the integration of untargeted urine metabolomic profiling with Machine Learning (ML) has the potential to unveil metabolic profiles crucial for understanding diseases. However, the application of this approach in autism remains underexplored. Our objective was to delve into the metabolic profiles of autism utilizing a comprehensive untargeted metabolomics platform coupled with ML. Untargeted metabolomics quantification (UHPLC/Q-TOF-MS) was performed for urine analysis. Feature selection was conducted using Lasso regression, and logistic regression, support vector machine, random forest, and extreme gradient boosting were utilized for significance stratification. Pathway enrichment analysis was performed to identify metabolic pathways associated with autism. A total of 52 autistic children and 40 typically developing children were enrolled. Lasso regression identified ninety-two urinary metabolites that significantly differed between the two groups. Distinct metabolites, such as prostaglandin E2, phosphonic acid, lysine, threonine, and phenylalanine, were revealed to be associated with autism through the application of four different ML methods (p<0.05). The alterations observed in the phosphatidylinositol and inositol phosphate metabolism pathways were linked to the pathophysiology of autism (p<0.05). Significant urinary metabolites, including prostaglandin E2, phosphonic acid, lysine, threonine, and phenylalanine, exhibit associations with autism. Additionally, the involvement of the phosphatidylinositol and inositol phosphate pathways suggests their potential role in the pathophysiology of autism.