Towards the Development of a Diagnostic Test for Autism Spectrum Disorder: Big Data Meets Metabolomics

Abstract

Autism spectrum disorder (ASD) is defined as a neurodevelopmental disorder which results in impairments in social communications and interactions as well as repetitive behaviors. Despite current estimates showing that approximately 2.2% of children are affected in the United States, relatively little about ASD pathophysiology is known in part due to the highly heterogenous presentation of the disorder. Given the limited knowledge into the biological mechanisms governing its etiology, the diagnosis of ASD is performed exclusively based on an individual’s behavior assessed by a clinician through psychometric tools. Although there is no readily available biochemical test for ASD diagnosis, multivariate statistical methods show considerable potential for effectively leveraging multiple biochemical measurements for classification and characterization purposes. In this work, markers associated with the folate dependent one-carbon metabolism and transulfuration (FOCM/TS) pathways analyzed via both Fisher Discriminant Analysis and Support Vector Machine showed strong capability to distinguish between ASD and TD cohorts. Furthermore, using Kernel Partial Least Squares regression it was possible to assess some degree of behavioral severity from metabolomic data. While the results presented need to be replicated in independent future studies, they represent a promising avenue for uncovering clinically relevant ASD biomarkers.

1. INTRODUCTION

Autism Spectrum Disorder (ASD) is defined as a neurodevelopmental condition that is characterized by difficulty in communication, social interaction, and restricted repetitive behaviors. It is currently estimated that 1 in 44 children in the United States are affected by ASD.^[1] Although a broad range of different genes and metabolic pathways have been associated or implicated in ASD pathophysiology, the biological underpinnings of ASD etiology continues to be difficult to ascertain. Complicating research into ASD is the highly heterogeneous nature of the condition, which is associated in 95% of cases with co-occurring conditions.^[2] Subsequently, individuals with autism are distinguished primarily by their behavior, and an ASD diagnosis is performed exclusively based on an individual’s evaluation by a clinician through psychometric tools.

One of the common misleading conclusions that stems from the behavioral criteria used to diagnose ASD is that only an analysis of the brain and the neurological system is required to characterize its manifestation more quantitatively. While there are considerable effects on the brain, ASD has been shown to have an impact on many other physiological systems. Notably, differences observed among individuals with ASD include perturbations of mitochondrial metabolism, the gut-brain axis, handling of oxidative stress and immune response. ^[3–6] The metabolomic implications of these differences have begun to be more thoroughly investigated for insight into both the underlying mechanisms that give rise to behavioral characteristics associated with ASD, as well as the potential for the development of a biochemical test for predicting ASD diagnosis.

A biochemical approach towards ASD diagnosis holds considerable promise to promote earlier and more widespread opportunities for support. Many social and communication skills do not emerge until after the first 12 months. Repetitive behaviors that serve as hallmarks of the condition, may not emerge until after other less pronounced traits, thereby missing a potential window for earlier diagnosis when relying on psychometric evaluations.^[7] It has been demonstrated that it is possible to achieve a stable diagnosis of ASD as early as at 14 months.^[8] However, the median age of ASD diagnosis in the United States is estimated to be 52 months^[9]. Benefits of earlier diagnosis for both children and caregivers underscore the imperative to developing more robust ways of clinically screening for ASD.

Comprehensive behavioral early intervention techniques for ASD such as Applied Behavioral Analysis (ABA) and the Early Start Denver Model (ESDM) have been found to be most effective when initiated in early childhood.^[10,11] Despite the benefits of earlier diagnosis, accessibility to clinical professionals for diagnosis and the heterogeneity of ASD manifestation have contributed to a delayed age of diagnosis for many children.^[12] Disparities in average age of diagnosis by socioeconomic status and race/ethnicity also suggest limitations in access to services and treatment. ^[13,14] Thus, a less subjective approach towards diagnosis can provide a useful toolset for medical providers to bridge the gap between the age of stable onset and the age of diagnosis.

The heterogeneity of individuals with ASD has been one of the perennial challenges with uncovering potential biomarkers that can effectively provide a reliable predictor of ASD diagnosis. However, the use of multivariate statistical techniques on metabolomic data has shown promise in allaying this challenge. Metabolites derived from several physiological systems have been examined for their capacity to reliably predict an ASD diagnosis as well as symptom severity. Blood, urine, and fecal derived metabolites and biochemical compounds have been examined for their potential role in differentiating children with an ASD diagnosis and their typically developing peers (TD).^[15–21] Using folate-dependent one-carbon metabolism (FOCM) and transsulfuration (TS) metabolites found in blood plasma, it was possible to correctly classify between ASD and TD cohorts with a specificity of 96.1% and sensitivity of 97.6%, following leave-one-out cross validation^[15]. Using an independently collected data set, an 88% accuracy using FOCM/TS derived model panels was attained.^[22]

The relationship between ASD behavioral severity and metabolomics has also been a promising area of exploration with multivariate statistical analysis. Howsmon et al.^[15] achieved an R² of 0.45 between metabolomic and behavioral data after cross-validation using FOCM/TS metabolites. In another study, metabolites related to the FOCM/TS pathway, glutathione and SAM have been shown to be correlated with ASD severity in blood plasma.^[23] Analysis of urinary metabolites in a case-control study (n=57) have similarly shown that specific metabolites such as adipic acid, palmitic acid, and 3-(3-hydroxyphenyl)-3-hydroxypropanoic were correlated with behavioral symptom severity.^[24] A nutritional and metabolomic study also showed some degree of correlation between groups of vitamins, minerals and plasma amino acids with behavioral symptom severity.^[25] By uncovering multivariate relationships between metabolomics and behavior, it may be possible to better understand mechanisms of ASD etiology.

Re-examining the work done by Howsmon et al.^[15] this work focuses on approaches to potential biomarker discovery, practical relevance for ASD characterization and assesses the relationship between symptom severity and metabolic processes.

2. MATERIALS AND METHODS:

This work utilizes data that was collected from the Arkansas Children’s Hospital Research Institute’s autism IMAGE case-control study.^[26] This study involves data collected from 159 participants, 83 with ASD and 76 age-matched control children. The inclusion criteria required participants to be between the ages of 3–10. Furthermore, participants grouped as belonging to the ASD cohort were required to meet the diagnostic criteria laid out by the DSM-IV, the Autism Diagnostic Observation Schedule (ADOS), and/or the Childhood Autism Rating Scales (CARS> 30). Plasma samples were collected from all participants from which a panel of transmethylation and transsulfuration metabolites were derived. Additionally details regarding measurement techniques and variables examined are provided by Melnyk et al.^[26] Behavioral information for 55 of the ASD participants was also collected using the Vineland Adaptive Behavior Composite. This metric assesses communication, daily living skills, socialization, motor skills, and maladaptive behavior.

Melnyk et al.^[26] performed univariate analysis by comparing all measurements collected between the ASD and TD cohorts using the Student’s t test. An extension of this analysis was performed here by determining hypothesis testing based on sample distribution. Either Welch’s test or Student’s t test were utilized for comparing two parametric groups, while the Mann-Whitney u-test was performed when the normality assumptions were not met. To account for the multiple hypothesis testing problem, the false discovery rate (FDR) was determined for each measurement deemed to be statistically significant between the ASD and TD cohorts. A leave-one-out approach was used to determine the FDR, in which hypothesis testing was performed on every subset comparing two groups with one sample excluded. The proportion of null hypothesis across all subsets that were deemed to not be rejected corresponded to the FDR. Findings were deemed significant when the p-value was less than 0.05 and the FDR <0.10. Additionally, the receiver operator curve was individually assessed for each measurement.

Additionally, the correlations between all pairs of 24 measurements were assessed. Correlations were initially screened to be significant if the Pearson correlation coefficient between them was greater in magnitude than 0.40 and had a p-value of less than 0.05. Those correlations that were able to attain a p-value of 0.05 were subsequently subjected to leave-one-out FDR testing. Those that were able to achieve an FDR value less than 0.10 were deemed statistically significant. The Pearson correlation coefficients between all biochemical measured quantities were also determined with respect to their relationship with behavioral data. Correlation analysis was performed between all biochemical measurements and the Vineland Adaptive Behavioral Composite (VABC) for the 55 individuals with ASD where VABC scores were available.

Fishers Discriminant Analysis (FDA) was utilized to determine the degree to which it was possible to distinguish the ASD and TD cohorts using their respective metabolite measurements. This method is defined as a dimensionality reduction technique that seeks to separate n different classes by finding a projection where such differences are maximized.^[27] The objective function for this technique simultaneously seeks to maximize between class scatter (S_B) and minimize within class scatter (S_w). The performance of different linear classifiers against the resulting FDA scores can subsequently be used to examine the tradeoff of specificity and sensitivity for distinguishing between two groups. The goal of this technique is thus to determine n − 1 vectors (w) that maximize the function:

$$J\left(W\right)=\frac{W^T{S}_{B}W}{W^T S_W W}$$ (1)

The solution of this optimization problem corresponds to eigenvectors related to the k-1 largest generalized eigenvalues of S_w⁻¹ S_B.

An alternative approach for classifying between the ASD and TD groups was Support Vector Machine (SVM). By performing analysis with multiple techniques, the repeatability and consistency of the findings could be ascertained. SVM is defined as a supervised machine learning algorithm which is used to find an optimum hyperplane separator between two classes. This technique operates by utilizing the data points in multidimensional space that are closest to the opposite group known as support vectors.^[28] The distance between support vectors of opposite classes determines the location of the optimal hyperplane that maximizes separation between both classes.

Given data that are not separable, the SVM-soft constraint approach is utilized. This technique introduces a slack term that allows for some data values that are not categorized accurately by the margin of the hyperplane, but penalizes instances in which this occurs. The objective function for a weight vector given data x labelled into two categories y_i = {−1,1}, bias b, slack variable ξ and regularization parameter C:

$$\begin{gathered} min\frac{1}{2}{\Vert W\Vert }^2+C\sum _i{\xi }_i \\ s.t y_i\left(w*x_i+b\right)\ge 1-\xi ,\forall x_i \end{gathered}$$ (2)

In order to ensure robustness of the findings, cross-validation was performed on SVM and FDA results. Given the limited sample size (n=159), a leave-one-out cross-validation approach was employed. This technique proceeds by removing a single sample from the dataset and replicating the multivariate technique on the abridged dataset. The sample removed was then classified based on the model generated from the remaining data. This process was repeated for every individual sample within the dataset Cross-validation can assess whether overfitting has occurred, controlling and optimizing the number of variables considered for this analysis to mitigate overfitting. In this work, the C-statistic of the receiver operator curve (ROC) following cross-validation was assessed for every possible model with 2–6 measurements.

Building upon the findings from Howsmon et al.^[15] this work sought to further explore and characterize trends amongst the top metabolite models. An extension of the original FDA analysis, all top performing 5-marker models as determined by their AUROC were noted, with the top 1000 models retained. The reason this number of markers per model was selected was that in the original work^[15] the performance of the 5-marker models were determined to be comparable to those with greater numbers of constituents with marginal to little changes in cross-validated performance. The top 1000 models for both FDA and SVM approaches were recorded, along with their specificity and sensitivity.

One of the additional objectives of this work was to characterize the degree to which behavioral symptom severity could be elucidated from metabolomic measurements. Kernel partial least squares regression (KPLS) was used to accomplish this task. ^[15] An extension of linear partial least squares regression, this technique proceeds by first performing a non-linear transformation f = ψ(x) on the predictor set x. Subsequently, the output y is regressed onto the high dimensional feature space f. For the behavioral data represented using the Vineland Adaptive Behavior Composite score, a Gaussian kernel was used for the feature space. An exhaustive search approach was applied to examine all combinations of 2–6 variable multivariate regression models and quantify those that attained the greatest R² value.

3. RESULTS

A total of 24 measurements related to the FOCM/TS pathway were examined. The results of the univariate analysis were largely in concordance with the original study.^[26] The measurement with the highest AUROC was the percentage of oxidized glutathione, which had an AUROC value of 0.93. Of the measurements examined that compared the ASD and TD cohorts, 22 were able to achieve a p-value of less than 0.05 and an FDR value of less than 0.10, when subjected to hypothesis testing. Additionally, 21 of 24 measurements attained an individual AUROC value of 0.70 which indicates fair to excellent performance in discriminating between two groups.^[29] No metabolite was found to be significantly correlated with ASD behavioral symptom severity, as measured by the Vineland Adaptive Behavior Composite score. An overview of the univariate findings for each metabolite measurement is given in Table 1. In total, 48 significant correlations between measurements were identified (Figure 2).

Table 1.

Univariate and Correlation Analysis Results Ordered by AUROC. Univariate analysis was performed by both determining optimal hypothesis testing to carry out between the ASD and TD groups as well as calculating the AUROC between them. Correlation analysis was performed between each marker and the Vineland Adaptive Behavior Composite.

	AUROC	p-Value	FDR	R
fGSH/GSSG	0.93	1.9E-20	0.00	0.09
% oxidized Glutathione	0.93	1.9E-20	0.00	−0.08
tGSH/GSSG	0.91	4.8E-19	0.00	0.08
8-OHG	0.85	1.5E-14	0.00	0.00
tGSH	0.85	4.1E-14	0.00	−0.02
Nitrotyrosine	0.84	6.7E-14	0.00	−0.12
fGSH	0.83	1.5E-12	0.00	0.05
GSSG	0.82	1.8E-12	0.00	−0.03
Chlorotyrosine	0.82	3.7E-12	0.00	0.00
Methionine	0.82	4.0E-12	0.00	0.10
SAM/SAH	0.80	1.1E-10	0.00	−0.05
fCystine/fCysteine	0.79	1.7E-10	0.00	0.14
Cysteine	0.79	5.4E-11	0.00	−0.01
% DNA methylation	0.77	2.6E-09	0.00	0.03
SAM	0.76	1.2E-08	0.00	0.05
Glu.-Cys.	0.75	3.0E-08	0.00	0.03
fCystine	0.72	1.8E-06	0.00	0.02
SAH	0.70	1.6E-05	0.00	0.02
Adenosine	0.64	2.8E-03	0.00	−0.01
fCysteine	0.63	4.1E-03	0.00	−0.17
Cys.-Gly.	0.62	1.9E-02	0.00	−0.10
Tyrosine	0.61	1.8E-02	0.00	0.00
Homocysteine	0.57	1.6E-01	0.88	−0.05
Tryptophane	0.56	1.6E-01	0.88	−0.04

Figure 2.

Frequency of appearance of each of the metabolites in the top 1000 five-metabolite Fisher discriminant analysis (FDA) and Support Vector Machine (SVM) models. The metabolites are ranked from highest to lowest area under the receiver operating characteristic curve (AUROC) as shown in Table 1. The metabolites which achieved maximal separation following cross-validation are shown indicated by letter (A) % Oxidized Glutathione, (B) 8-OHG, (C) Chlorotyrosine, (D) fCystine/fCysteine, and (E) % DNA methylation

FDA and SVM multivariate models were determined using all possible combinations of 2–6 measurements. Models were evaluated for their performance following leave-one-out cross validation and the optimal model for each number of measurements was determined. Among the optimized models, there was considerable commonality between their constituents. For the FDA models, percentage of DNA methylation and percentage of oxidized glutathione appeared in every optimized metabolite panel except when only considering 3 measurements (Table 2). Among the SVM metabolites, fCystine/fCysteine appears in every optimum model, and methionine appears in all but two (Table 3).

Table 2.

Fitting and cross-validation results for the best combinations of two, three, four, five, and six metabolites used in FDA. The cross-validated TPR and TNR are shown for classification thresholds associated with different values of β

Number of Metabolites	Metabolite Combination	Fitted AUROC	Cross-Validated Results
			β	TPR	TNR
2	% DNA methylation	0.96	0.01	1.00	0.40
			0.05	0.95	0.83
	% oxidized Glutathione		0.10	0.90	0.89
			0.20	0.80	0.04
3	tGSH/GSSG	0.98	0.01	1.00	0.81
	Chlorotyrosine		0.05	0.96	0.92
	Nitrotyrosine		0.10	0.90	0.95
			0.20	0.80	0.99
4	% DNA methylation	0.99	0.01	1.00	0.83
	8-OHG		0.05	0.95	0.95
	Glu.-Cys.		0.10	0.90	0.96
	% oxidized Glutathione		0.20	0.80	1.00
5	% DNA methylation	1.00	0.01	0.99	0.85
	8-OHG		0.05	0.95	0.96
	Glu.-Cys.		0.10	0.90	1.00
	fCystine/fCysteine		0.20	0.80	1.00
	% oxidized Glutathione
6	% DNA methylation	1.00	0.01	0.99	0.89
	8-OHG		0.05	0.95	0.96
	Glu.-Cys.		0.10	0.90	1.00
	fCystine		0.20	0.80	1.00
	fCysteine
	% oxidized Glutathione

Table 3.

Cross-validated Classification Loss (CVCL), TPR and TNR for best combinations of two, three, four, five, and six metabolites used in SVM

Number of Metabolites	Metabolite Combination	CW CL	Cross-Validated Results
			TPR	TNR
2	tGSH/GSSG	0.08	0.91	0.91
	fCystine/fCysteine
3	tGSH	0.04	0.96	0.96
	fGSH/GSSG
	fCystine/fCysteine
4	Methion	0.04	0.96	0.96
	fGSH/GSSG
	Chlorotyrosine
	fCystine/fCysteine
5	Methion.	0.03	0.98	0.97
	Glu.-Cys.
	Chlorotyrosine
	fCystine/fCysteine
	% oxidized Glutathione
6	Methion	0.03	0.98	0.97
	Glu.-Cys.
	fGSH
	fCystine/fCysteine
	% oxidized Glutathione

The majority of top performing 5-measurement models featured many of the same metabolites as those observed among the optimal panels for FDA and SVM. Specifically, % oxidized Glutathione, which was a component of both the optimum 5-metabolite SVM and FDA models was also featured in more than 40% of all top 1000 models for both groups (Figure 2). While for most metabolites there was a degree of commonality between both the FDA and SVM techniques in terms of model composition, two metabolites that were notably different in their prominence were Chlorotyrosine and fCystine/fCysteine. Chlorotyrosine was observed in 50.2 % of FDA models but only 29.9% of SVM ones. In contrast fCystine/fCysteine was observed in 67.8 % of SVM but only 27.0% of SVM models.

Metabolites in the FOCM/TS pathway were examined using KPLS to distinguish their capability to predict adaptive behavior. In order to account for non-linear relationships in the underlying data measurements, a Gaussian kernel was utilized for the analysis. Utilizing all possible combinations of 2–6 metabolite measurements, the highest cross-validation R² value attained was 0.45 with a p-value less than 0.01. ^[15] This model consisted of 5-metabolites and used GSSG, tGSH/GSSG, Nitrotyrosine, Tyrosine, and fCysteine as inputs.

4. DISCUSSION

Multivariate statistical approaches have significant promise for the diagnosis and characterization of ASD. In both the original work^[15] and this reassessment, the use of metabolites related specifically to the FOCM/TS pathway have demonstrated potential towards this goal. Individually, 22 of 24 measured quantities were deemed to significantly differ between the two groups. The measurement that attained the highest AUROC (0.93) and lowest p-value (<0.01) was fGSH/GSSG. No individual metabolite was found to be strongly or even moderately correlated with the Vineland Adaptive Behavior Composite score. The metabolite that was observed to have the strongest correlation with behavioral symptoms was free cysteine in plasma (fCysteine), with a Pearson correlation coefficient of −0.17. However, the p-value attained for the significance of this correlation was greater than 0.05.

The performance of the best individual metabolites and measurements were eclipsed by the performance of multivariate approaches to both ASD prediction and severity correlation. The optimal 5-marker FDA model consisting of the percentage of DNA methylation, 8-OHG, Glu.-Cys., fCystine/fCysteine and the percentage of oxidized Glutathione was able to attain an AUROC of 0.99. Multivariate classification using FDA resulted in 96.1% of all typically developing participants being correctly identified as such while still correctly identifying 97.6% of the ASD cohort. SVM performed similarly with a 97.5% accuracy after cross-validation. Using KPLS it was possible to attain a cross-validated R² value of 0.45 for the behavioral score using the five metabolites GSSG, tGSH/GSSG, Nitrotyrosine, Tyrosine, and fCysteine.

DNA methylation percentage was the most prevalent measurement featured in the top-1000 FDA 5-metabolite models, with 71% of classifiers including it as a component. Furthermore, this marker was not found to be significantly correlated with any other measurement observed in this study. DNA methylation is defined as a biological process by which gene transcription is repressed via a methyl group binding to a DNA segment. As an epigenetic mechanism, the relationship between DNA methylation, the environment, and ASD has been explored in the literature.^[30,31] DNA methylation has been speculated to reveal insight into the conditions underlying etiology.^[32,33] DNA methylation patterns observed in brain samples derived from individuals with ASD have shown notable differences that impact pathways associated with immune response, synaptic pruning and microglial cell specification.^[34] Certain other studies have also observed that DNA methylation patterns observed in blood are statistically distinct between ASD and TD cohorts.^[35,36] However, this is far from a universal indicator of ASD, with some studies showing mixed results for DNA methylation as being a reliable predictor for ASD. Specifically, no individual CpG site was found to be statistically significantly distinct in ASD in a comprehensive meta-analysis consisting of 796 cases and 858 controls. ^[37]

The ratio of free plasma cystine to cysteine (fCystine/fCysteine) was found to be the most prominent constituent marker among top SVM models. In total, 67.8% of the 1000 top 5-marker SVM models utilized this ratio. This metabolite ratio retained some prominence among the FDA models as well with almost 30% featuring it as a constituent including the top 5-marker FDA panel. The cystine/cysteine cycle plays a major role in maintenance of redox homeostasis.^[38] Free plasma cystine and cysteine have been investigated for their relationship in ASD. In a comprehensive meta-analysis on studies examining cysteine in blood, it was noted that children with ASD had an average of 14% reduction in the concentration of this metabolite.^[39]

The dataset was collected from children with a range of symptom severity. ASD diagnoses were confirmed by developmental pediatricians using criteria laid out by the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV 299.0), the Autism Diagnostic Observation Schedule (ADOS), and/or the Childhood Autism Rating Scales. An average the Vineland Adaptive Behavior Composite score of 69 was observed for the children with ASD, which represents a standard deviation of two from the general mean of 100. The average sensitivity among the top FDA models was between 95–96% and the children observed to be frequently misclassified tended to be among the ones within the top quartile of the Vineland Adaptive Behavior Composite scores.

Many other metabolites related to the FOCM/TS pathways have been shown to be observed in different concentrations between ASD and TD cohorts in the literature. ^{[25, 40–42]} Comprehensively leveraging multivariate techniques has proven to be useful for gaining insights from these observed differences. Multivariate models using these metabolites were found to differentiate mothers of children with ASD from those with only typically developing children.^[43] Subsequent studies using multivariate models with metabolites that could be tied to the FOCM/TS pathways have also shown strong capability for ASD diagnosis prediction.^[44] While these results are promising, they should be replicated in future studies and assessed for greater robustness. Nonetheless, these analyses suggest combinations of metabolites related to the FOCM/TS pathways can have considerable potential as ASD biomarkers.

While there is currently no clinically utilized approach to ASD diagnosis that takes advantage of patients’ biochemical profiles, both this work and the literature highlight promising avenues by which clinicians can potentially leverage such toolsets to assess and diagnose children more quantitatively. This in turn can lead to earlier access to treatment and intervention options. Furthermore, by comprehensively examining both diagnostic biomarker panels and the molecular pathways that they are associated with, it may be possible to not only diagnose but better understand ASD etiology on an individual level. As an extremely heterogeneous condition, the potential to better categorize children diagnosed with ASD can allow for a more personalized approach towards treatment and symptom management.

5. CONCLUSIONS

The computational analysis outlined in this paper underscores how traditional population-level statistics can be enhanced using multivariate techniques for medical diagnostic purposes. Using techniques such as FDA, SVM and KPLS it was possible to identify patterns and relationships that would otherwise not have been uncovered. Using these techniques, the metabolites of the FOCM/TS pathways were shown to readily distinguish between ASD and TD cohorts, while also providing insight into behavioral/symptom severity. Although not currently clinically utilized, these metabolites demonstrate significant potential as candidates for biochemical biomarkers of ASD. Nonetheless, future work should continue to corroborate these findings and assess their robustness in other sub-populations involving children with ASD.

Figure 1:

Correlation network for significant measurements. The strength of correlation is visualized by 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 0.40, FDR less than 0.10 and p-value less than 0.05. See Table S1 for a complete list of relationships and Pearson correlation coefficient values.