Effects of Metabolism Changes during Development on the Plasma Profile of ASD in Children

Background:  Current methods of autism diagnosis rely largely on behavioral testing that is time consuming, expensive, and usually performed around the age of 4 years. Prior work at Stemina showed that plasma metabolic profiles can be used to distinguish children with autism spectrum disorder (ASD) from typically developing (TD) children. With the aim of developing a diagnostic test, we studied metabolites in plasma from children ranging in age from 1-10 years, to identify potential age-related metabolic biomarkers of ASD. The outcome of these studies will define metabolic signatures capable of diagnosing autism at different stages of child development.

Objectives:  To identify metabolite profiles that can discriminate ASD from TD as children develop up to 10 years of age. Additionally, age dependent metabolic signatures will be used to further understand the impact of childhood development on the metabolic nature of ASD.

Methods: Plasma samples from 211 children (1-10 yrs; 97 ASD, 114 TD) were analyzed using 4 orthogonal Liquid Chromatography-High Resolution Mass Spectrometry (LC-HRMS) methods, in an untargeted metabolomic approach.   Patient samples were further divided into defined age ranges to examine metabolic changes associated with ASD over-time. For either the entire cohort, or age limited sample subsets, diagnostic-stratified random sampling was used to create a training set (75%) and an independent test set (25%). Univariate statistical tests were performed to select features differentially abundant (DA) between ASD and TD children. Classification models were built using the DA features and several multivariate supervised learning methods.  Finally, chemical structure identification of metabolites was performed using reference standards and MS/MS spectral matching.

Results: Using the entire cohort, a set of 76 statistically significant metabolic features (p-value < 0.05) differentiating ASD from TD were identified and ranked by their influence on class discrimination. Classification results were iteratively evaluated by varying the number of ranked features to identify the most predictive feature subsets.  Models were evaluated on the test set and classification accuracy for the best models was found to be 60-65%.  By contrast, when a subset of 77 samples (54 ASD, 23 TD) representing children with ages 2-4 years were evaluated using similar methods, the number of significant features comparing ASD to TD was 104, suggesting more pronounced differences  in the metabolism of ASD and TD in the younger age group.  Classification of these samples also yielded models exhibiting better classification accuracy, greater than 80%.  Confirmation of features of specific metabolism is ongoing but preliminary annotations include known biomarkers of autism as well as changes related to childhood development. 

Conclusions: This study demonstrates that metabolic variation in patients from a wide age range adds metabolic complexity; however a 2-3 year age range can provide a more predictive signature capable of classifying ASD from TD. In addition, the study further suggests that specific metabolites as biomarkers of autism can be measured and used over the course of childhood development allowing early diagnosis and subsequent therapy on a personalized basis.