Multivariate Analysis of Metabolomic and Nutritional Profiles among Children with Autism Spectrum Disorder

Abstract

There have been promising results regarding the capability of statistical and machine-learning techniques to offer insight into unique metabolomic patterns observed in ASD. This work re-examines a comparative study contrasting metabolomic and nutrient measurements of children with ASD (n = 55) against their typically developing (TD) peers (n = 44) through a multivariate statistical lens. Hypothesis testing, receiver characteristic curve assessment, and correlation analysis were consistent with prior work and served to underscore prominent areas where metabolomic and nutritional profiles between the groups diverged. Improved univariate analysis revealed 46 nutritional/metabolic differences that were significantly different between ASD and TD groups, with individual areas under the receiver operator curve (AUROC) scores of 0.6-0.9. Many of the significant measurements had correlations with many others, forming two integrated networks of interrelated metabolic differences in ASD. The TD group had 189 significant correlation pairs between metabolites, vs. only 106 for the ASD group, calling attention to underlying differences in metabolic processes. Furthermore, multivariate techniques identified potential biomarker panels with up to six metabolites that were able to attain a predictive accuracy of up to 98% for discriminating between ASD and TD, following cross-validation. Assessing all optimized multivariate models demonstrated concordance with prior physiological pathways identified in the literature, with some of the most important metabolites for discriminating ASD and TD being sulfate, the transsulfuration pathway, uridine (methylation biomarker), and beta-amino isobutyrate (regulator of carbohydrate and lipid metabolism).

Keywords: ASD; Fisher discriminant analysis; SVM; machine learning; metabolomics; multivariate statistics.

Figure A1

Univariate distribution for free sulfate in plasma, which was the metabolite that had the highest AUROC (0.90).

Figure A2

Boxplots of the FDA scores for both the 5-marker and 6-marker optimized model based upon cross-validated AUROC value. Each box represents scores that fall between the 25th and 75th percentile for that respective set of scores.

Figure 1

Univariate hypothesis test selection paradigm. Each sample set was examined for both its variance and distribution to select the appropriate parametric or nonparametric test.

Figure 2

Sankey diagram showing the biochemical and xenobiotic measurements that served as the inputs to the hypothesis testing protocol. Measurements that had a p-value greater than 0.05 or a false discovery rate greater than 0.10 were deemed to not be significantly different (n.s.) between the ASD and TD groups. The measurements that were determined to be significant were used in the development of the FDA and SVM models.

Figure 3

Correlation network between significant biochemical and xenobiotic compounds in the TD cohort (strength of the correlation is visualized by the line thickness, positive correlations are in blue and negative correlations are in red). In order for a relationship to be deemed significant, the correlation coefficient had to be greater than 0.35, FDR less than 0.10, and the p-value less than 0.05. In total, 378 significant correlations were observed that met these criteria. NADP and total sulfate had the greatest number of relationships, with 19 significant relationships. Only those relationships with r > 0.40 are presented (see Table A1 for details).

Figure 4

Correlation network between significant biochemical and xenobiotic compounds in the ASD cohort (strength of the correlation is visualized by the line thickness, positive correlations are in blue and negative correlations are in red). In order for a relationship to be deemed significant, the correlation coefficient had to be greater than 0.35, FDR less than 0.10, and the p-value less than 0.05. In total, 212 significant correlations (106 pairs) were observed. Acetylcholine had the greatest number of relationships, with 14 significant relationships (see Table A2 for details).

Figure 5

Marker prevalence among the top-1000 FDA 5-marker models as judged by their performance on the test set. Among the most prominent potential biomarkers are free sulfate, uridine, and beta-amino isobutyrate (highlighted in red). Each of these was present in more than 75% of the top models. Free sulfate in particular was present in every single top model.

Figure 6

Marker prevalence among the top 1000 4-marker FDA models as judged by their performance on the test set, with both total and free sulfate excluded. Due to the predominance of sulfate in model panels, models with other constituents were explored by conducting the FDA analysis with these two metabolites excluded. The metabolites observed to be most prevalent in the resulting models were highlighted in red and include (A) glutathione present in 43.3% (B) uridine present in 74.7%, and (C) homocystine + homocysteine present in 32.1% of models.

Figure 7

Marker prevalence among the top-1000 5-marker SVM models as judged by their performance on the test set. Among the most prominent potential biomarkers are (A) free sulfate in serum, (B) uridine, (C) tryptophan, (D) beta-amino isobutyrate, and (E) copper in whole blood.