Single Nucleotide Polymorphisms Predict Symptom Severity of Autism Spectrum Disorder

Abstract

Autism is widely believed to be a heterogeneous disorder; diagnosis is currently based solely on clinical criteria, although genetic, as well as environmental, influences are thought to be prominent factors in the etiology of most forms of autism. Our goal is to determine whether a predictive model based on single-nucleotide polymorphisms (SNPs) can predict symptom severity of autism spectrum disorder (ASD). We divided 118 ASD children into a mild/moderate autism group (n = 65) and a severe autism group (n = 53), based on the Childhood Autism Rating Scale (CARS). For each child, we obtained 29 SNPs of 9 ASD-related genes. To generate predictive models, we employed three machine-learning techniques: decision stumps (DSs), alternating decision trees (ADTrees), and FlexTrees. DS and FlexTree generated modestly better classifiers, with accuracy = 67%, sensitivity = 0.88 and specificity = 0.42. The SNP rs878960 in GABRB3 was selected by all models, and was related associated with CARS assessment. Our results suggest that SNPs have the potential to offer accurate classification of ASD symptom severity.

Appendices

Appendices

Appendix 1: Histogram of CARS Scores

Please see Fig. 3.

Fig. 3

Histogram of CARS scores

Appendix 2: Tree Model Generated by ADTree

Please see Fig. 4.

Fig. 4

Tree model generated by ADTree. A negative number indicates that a subject was classified as belonging to the mild/moderate group, whereas a positive number indicates that a subject was classified as belonging to the severe group. The rs878960 SNP was selected as the priority root node

Appendix 3: Data-Mining Methods

In this section, we provide an overview of DS, ADTree, and FlexTree.

DS is a single-level decision-tree model with a categorical or numeric class label (Iba and Langley 1992). It tends to find the main predictor variable in one step. It is widely used when researchers seek the single most significant feature with respect to classification (Iba and Langley 1992).

An ADTree is a method based on combining weak hypotheses generated during boosting into a single interpretable representation (Freund and Mason 1999). An ADTree model is more compact than standard boosting-based decision-tree models, which generate more than one tree (Freund and Mason 1999). As a result, an ADTree model is relatively straightforward to interpret. The application of boosting procedures may improve classification performance for ADTrees. The structure of an ADTree has three characteristics: (1) the root node is a prediction node, and has a numeric score only, which is based on the total weights of the positive and negative instances that satisfy the conditions in the training data (Holmes et al. 2002); (2) the nodes in the next layer are decision nodes, and are essentially a collection of decision-tree stumps; (3) the subsequent layers alternate layers of prediction nodes and decision nodes. To classify a new instance with an ADTree model, all paths for which all decision nodes are true are followed, summing any prediction nodes that are traversed by these paths.

FlexTree, a general supervised-learning method, extends the binary tree-structured approach [Classification and Regression Trees, CART (Breiman et al. 1984)], although it differs greatly in its selection and combination of predictors (Huang et al. 2004). It is particularly applicable for assessing gene–gene and gene-environment interactions as they bear on complex diseases. FlexTree creates a simple rooted binary tree with each split defined by a linear combination of selected variables. The linear combination is determined by regression with optimal scoring; the variables are selected by a backward pruning procedure. Using a selected variable subset to define each split increases interpretability, improves predictive robustness, and prevents overfitting. FlexTree deals with additive and interactive effects simultaneously. Sampling units can be families or individuals, depending on the application. Generally, FlexTree demonstrated performance that is better than many alternatives to which it was compared, particularly when a small fraction of candidate genes are useful for classification (Huang et al. 2004).

Appendix 4: Rationale for SNP Selection

GABA is the major inhibitory neurotransmitter in the adult brain, although it mediates excitatory transmission during development. As a result, many GABA_A receptors encoding genes were involved in our study. Previous autism pathophysiology studies reported that: (1) The numbers of GABA_A receptors were significantly decreased in brains of children with autism (Blatt et al. 2001); (2) Plasma GABA, and its essential precursor glutamate, were elevated in children with autism (Moreno-Fuenmayor et al. 1996; Dhossche et al. 2002; Aldred et al. 2003); (3) Benzodiazepines, which are effective in treating the seizures, anxiety, and social phobia that occur in the setting of autism, bind to, and act on, GABA_A receptors (Olsen and Macdonald 2002); (4) GABA-ergic transmission has important trophic actions during development. Based on these data, the GABA_A receptor subunit genes, particularly those in 15q11-q13 (Cook et al. 1998; Buxbaum et al. 2002; McCauley et al. 2004; Ashley-Koch et al. 2006), represent excellent candidates, allelic variants of which could confer genetic susceptibility for development of autism (McCauley 2005).

TDO2 (Nabi et al. 2004), SLC25A12 (Ramoz et al. 2004; Segurado et al. 2005), and BDNF (Cheng et al. 2009) were also found to be associated with ASD, so we included these genes in our studies.

Appendix 5: Genotyping

The first step in genotyping was PCR; primers were designed using Primer Premier 5.0 software, based on published DNA sequences. The primers were synthesized and HPLC purified by the TaKaRa Company (P.R. China). All reverse primers were modified with an acrylamide group at the 5’-terminal, in order to covalently bond to the polyacrylamide gel. After several cycles of PCR amplification, we used ethanol to precipitate PCR products.

Step 2 was immobilization of PCR products. We dissolved acrylamide-modified PCR products, and spotted them on 3-methacryloxypropyltrimethoxy silane-modified glass slides, using a microarrayer (Captial Biochip Corporation, P.R. China). Each slide was placed into a humid, 1,000 Pascal (Pa) pressure-sealed chamber full of tetramethylethylenediamine, to induce copolymerization between acrylamide groups and acryl groups. We then used electrophoresis to obtain single stranded DNA (ssDNA) for hybridization.

Step 3 was hybridization. We designed a pair of probes for every SNP locus, such that the probes could be matched with the polymorphic portion of the targets, and labeled with Cy3 or Cy5. For every SNP genotyped, we mixed the labeled probes in equimolar amounts, and suspended them in unihybridization solution (3:1 dilution) to obtain a final concentration of 2 μM. We achieved hybridization in a humid chamber at 37°C for 2–4 h.

The fourth and fifth steps were post-hybridization and scanning, respectively. We rinsed the slide in water and air dried it, after which we completed electrophoresis at 2 V/cm for 8 min in 1X TBE buffer at 4º C. We scanned the hybridization slides at 70% laser power and 65% photomultiplier tubes gain with a confocal scanner (Luxscan-10 K/A, CapitalBio Company, P.R. China) that had been fitted with filters for Cy3 and Cy5. We used QuantArray software (Packard BioChip Technologies, Billerica, MA) to analyze these images.

Appendix 6: rs878960 Genotype and Allele Distributions Between Our Study and Other Asian Cohorts

The genotypes and the allele distributions for rs878960 polymorphisms in our study (n = 118) and for the HapMap Han Chinese group (n = 43, total 45 subjects but 2 missing values) are presented in Table 5. There is no significant difference in genotype or allele distribution between these two groups. Similarly, there is no significant difference in genotype or allele distributions between our Chinese subjects and Japanese subjects (HapMap, n = 86).

Table 5 Genotypes and allele distributions for rs878960 polymorphisms in our study and for other Asian groups

Appendix 7: Haplotype Analysis for Each Gene

Table 6 shows the results of haplotype analysis using methods described in (Li et al. 2009).

Table 6 Haplotype analysis for each gene

We found that haplotypes of GABRA4, but not of GABRB3, are significantly associated with ASD symptom severity. The reason for the lack of association with GABRB3 is that we tested 6 SNPs in GABRB3, and some of these SNPs may contribute noise to the analysis. In particular, when we remove one SNP—rs1432007—and repeat the analysis (see Table 7), we obtain χ² = 19.5, p value = 0.007. That is, haplotypes of GABRB3 are significantly associated with symptom severities of ASD when we remove SNP rs1432007 from the analysis.

Table 7 Haplotype analysis of GABRB3 after removing rs1432007