Plasma Amino Acids in Individuals with Autism Spectrum Disorder: A Multivariate Statistical Analysis

Introduction: Autism spectrum disorder (ASD) is estimated to affect 1 in 68 children in the United States [1]. Currently, there are no established biomarkers for diagnosing ASD; rather, a diagnosis is typically determined after evaluation of an individual’s behavior by psychologists and specialized physicians. That being said, many biochemical markers have been explored with regards to their ability to distinguish individuals on the spectrum from their neurotypical (NT) peers. Studies investigating plasma amino acids for this purpose are well-represented in the literature, but are collectively inconclusive due to general disagreement among their results. Inconsistencies in these studies can typically be attributed to one of three challenges in study design: (1) small sample sizes, (2) inclusion of non-age-matched controls, and (3) a lack of fasting by study participants. However, the Nutrition/Diet Treatment Study at Arizona State University addressed all three of these challenges. The goal of the current work is to apply multivariate approaches to the analysis of these data to further explore the utility of plasma amino acids for classification of individuals as being NT or on the autism spectrum.

Materials and Methods: This work uses plasma amino acid measurements from the Nutrition/Diet Treatment Study, a 12-month nutritional and dietary intervention study at Arizona State University [2]. The data include concentrations of 21 amino acids and amino acid metabolites measured at baseline from 64 individuals with ASD and 49 NT individuals. Univariate comparison of measurements between ASD and NT cohorts was done using the two-sample Welch’s t-test (α = 0.05). Classification was performed using Fisher discriminant analysis (FDA) [3] and its nonlinear extension, kernel FDA (KFDA) [4]. These methods have previously been shown to be useful for multivariate analysis of biochemical data from individuals with ASD [2, 7]. Kernel density estimation [8] was used to construct the probability density functions (PDFs) of FDA/KFDA scores for the ASD and NT cohorts.

Results and Discussion: The univariate analysis indicated three variables to be significantly different between the ASD and NT cohorts. These were glutamate (p = 0.006), hydroxyproline (p = 0.018), and β-alanine (p = 0.047), with all three elevated in the ASD cohort. It should be noted that this assessment of significance does not account for multiple comparisons. As these variables were most likely to be useful for classification, FDA was performed while using these as inputs. While there is a significant difference in mean score between the two cohorts (p < 0.0001), these PDFs correspond to Type I and Type II error rates of 17% and 49%, respectively. Therefore, these results are not useful for classification purposes despite the significant difference in group means. Fitting a KFDA model with the same three inputs again yielded a significant univariate difference (p < 0.0001); however, the associated Type I and Type II error rates were 7.5% and 30%, respectively. Such large Type I/II errors from FDA and KFDA using the most significantly different measurements indicate that reliable classification will not be possible using these data.

Conclusions: Our results suggest that plasma amino acids are not a reliable biomarker for the diagnosis of ASD, which may explain the wide disagreement across studies in the literature. This is not to say that individual amino acids are not correlated with ASD pathophysiology, as supplementation with certain amino acids has been found to improve behavioral symptoms in individuals with ASD. However, amino acids have diverse roles in the body and their measurements in plasma alone do not appear to be a useful indicator of the occurrence of ASD.

References: [1] Christensen DL, MMWR Surveill Summ 2016, 65(3): 1-23. [2] Adams J, PLOS ONE 2017, 12(1): e0169526. [3] Fisher RA, Ann Eugenic 1936, 7(2): 179-188. [4] Mika S, Neural Netw Signal Process IX 1999, 41-48. [5] Wold S, Chemom Intell Lab Syst 2001, 58(2): 109-130. [6] Rosipal LJ, J Mach Learn Res 2001, 2: 97-123. [7] Howsmon DP, PLOS Comput Biol 2017, 13(3): e1005385. [8] Silverman B, Density Estimation for Statistics and Data Analysis (1986), New York: CRC Press.