Pathway-based outlier method reveals heterogeneous genomic structure of autism in blood transcriptome

Abstract

Background: Decades of research strongly suggest that the genetic etiology of autism spectrum disorders (ASDs) is heterogeneous. However, most published studies focus on group differences between cases and controls. In contrast, we hypothesized that the heterogeneity of the disorder could be characterized by identifying pathways for which individuals are outliers rather than pathways representative of shared group differences of the ASD diagnosis.

Methods: Two previously published blood gene expression data sets--the Translational Genetics Research Institute (TGen) dataset (70 cases and 60 unrelated controls) and the Simons Simplex Consortium (Simons) dataset (221 probands and 191 unaffected family members)--were analyzed. All individuals of each dataset were projected to biological pathways, and each sample's Mahalanobis distance from a pooled centroid was calculated to compare the number of case and control outliers for each pathway.

Results: Analysis of a set of blood gene expression profiles from 70 ASD and 60 unrelated controls revealed three pathways whose outliers were significantly overrepresented in the ASD cases: neuron development including axonogenesis and neurite development (29% of ASD, 3% of control), nitric oxide signaling (29%, 3%), and skeletal development (27%, 3%). Overall, 50% of cases and 8% of controls were outliers in one of these three pathways, which could not be identified using group comparison or gene-level outlier methods. In an independently collected data set consisting of 221 ASD and 191 unaffected family members, outliers in the neurogenesis pathway were heavily biased towards cases (20.8% of ASD, 12.0% of control). Interestingly, neurogenesis outliers were more common among unaffected family members (Simons) than unrelated controls (TGen), but the statistical significance of this effect was marginal (Chi squared P < 0.09).

Conclusions: Unlike group difference approaches, our analysis identified the samples within the case and control groups that manifested each expression signal, and showed that outlier groups were distinct for each implicated pathway. Moreover, our results suggest that by seeking heterogeneity, pathway-based outlier analysis can reveal expression signals that are not apparent when considering only shared group differences.

Figure 1

Outlier-enriched pathways in the TGen data set. A) Heatmap marking outliers for each outlier-enriched pathway: neuron development, nitric oxide (NO) signaling, and skeletal development (Figure 4, step 4). Samples sorted by outlier status followed by diagnosis. B) Venn diagram showing the overlap of outlier samples for the three pathways in Figure 1A. The numbers of case and control outliers, respectively, are shown in parentheses. C) A comparison of Mahalanobis distance values (Figure 4, step 2) in neuron development and NO signaling. Marginal density plots show the distributions of case (red) and control (blue) samples.

Figure 2

TGen outlier-group differentially expressed genes (Figure 4, step 5). Differential expression was calculated using t-tests on log (base 2) expression values at FDR < 5%. A) Overlap of differentially expressed genes for the three outlier groups. B) Comparison of the log (base 2) expression levels of the two most differentially expressed genes in the neuron development pathway, SPON2 and FEZ1. Marginal density plots show the distributions of case (red) and control (blue) samples.

Figure 3

Neurogenesis outliers in the Simons data set. The y-axis is in units of Mahalanobis distance from the pooled centroid of cases and controls, and the x-axis is samples. One control sample was misclassified as an outlier due to the choice of ${\chi }_{0.975}^2$ as the hard threshold.

Figure 4

Analysis flow chart. In brief, the steps of the method were: 1) project samples onto the principal components (PC) of pathway subspaces, 2) calculate the 1-dimensional Mahalanobis distance distribution of samples in the PC space of each pathway, 3) classify samples as “outliers” or “non-outliers” in each pathway based on the 97.5^th percentile of the chi-squared distribution, 4) employ Fisher’s exact test to identify pathways that are specifically enriched for case or control outliers, using the False Discovery Rate (FDR) q-value of Storey and Tibshirani to correct for multiple hypothesis testing [88], 5) for each subgroup of samples corresponding to the outliers in a candidate pathway, perform standard differential expression analysis to determine the genes responsible for the grouping.