Independent Component Analysis of SNPs Reflects Polygenic Risk Scores for Schizophrenia

Abstract

Schizophrenia is a psychiatric disorder with high heritability. Recent genome-wide association studies have provided a list of risk loci reliably derived from unprecedentedly large samples. However, further delineation of the diagnosis-associated susceptibility variants is needed to better characterize the genetic architecture given the disease's complex nature. In this sense, a data-driven approach might hold promise for identifying functionally related clusters of genetic variants that might not be captured by hypothesis-based models. In the current study, independent component analysis (ICA) was applied to the Psychiatric Genomics Consortium's schizophrenia-related single nucleotide polymorphisms (SNPs) in 104 schizophrenia patients and 142 healthy controls of European Ancestry. We found that, for 13 out of 16 extracted independent components, the associated loadings correlated highly (r>0.5) with the polygenic risk scores for SZ of the corresponding SNPs. These correlations were likely not inflated by the linkage disequilibrium structure (permutation p<0.001). In brief, we demonstrate an example of ICA analysis on SNP data yielding functionally meaningful clusters, which motivates further application of data-driven approaches as a complimentary tool for hypothesis-based methods to enrich our knowledge on the genetic basis of complex disorders.

Introduction

Genome-wide association studies (GWAS) provide evidence for a polygenic model of schizophrenia (SZ) where 23% of the variance in liability can be explained by 915,354 common single nucleotide polymorphisms (SNPs) covered in the Psychiatric Genomic Consortium (PGC) schizophrenia study (Lee et al., 2012; Purcell et al., 2009; Ripke et al., 2011). A more recent roadmap study by PGC identified 128 independent genetic associations with SZ from ∼35,500 cases (Ripke et al., 2014). However, these diagnosis-associated susceptibility SNPs are likely highly heterogeneous in terms of their pathways to phenome, considering that SZ is a complex disorder. Echoing this speculation, one recent work investigated genetic influences on schizophrenia and subcortical brain volumes and found no notable overlap at overall common variants level (Franke et al., 2016). Hence, further delineation is required to characterize the genetic architecture and better understand the pathophysiology of SZ. In this sense, a data-driven approach to dissect large sets of SNPs may hold promise for identifying functionally related variants without a prior knowledge of their biological functions.

Independent component analysis (ICA) is a type of blind source separation approach which finds extensive applications in the neuroimaging fields (Calhoun et al., 2009). However, its application to genotype data has not been as popular (Pearlson et al., 2015), presumably due to the difficulty of interpretation, given the limited knowledge on gene-gene interactions and intergenic SNPs. In this work, we leverage the information derived from a large schizophrenia GWAS, and show, as an example, that ICA analysis yields valid clustering where the identified covariation patterns highly correlate with the polygenic risk scores for SZ (PRS-SZ) of those clusters of SNPs.

Materials and Methods

A total of 246 subjects of European Ancestry, including 142 healthy controls and 104 individuals with SZ, were aggregated from two cohorts that were not used in the PGC study, i.e., the Mind Clinical Imaging Consortium (MCIC) study and a Center of Biomedical Research Excellence (COBRE) study. The institutional review board at each site approved the study and all participants provided written informed consents. Table S1 lists the demographic information. Recruitment and clinical screening information were described in the previous study (Chen et al., 2013). The MCIC and COBRE data were genotyped using the Illumina chips and went through the standard imputation, quality control and population structure correction procedures as detailed in the supplemental information. Discrete numbers were assigned to the categorical genotypes: 0 for no minor allele, 1 for one minor allele, and 2 for two minor alleles. Finally, 5,468 SZ-related SNPs were mapped from the 108 PGC reported SZ-related regions (Ripke et al., 2014) and entered into the ICA analysis.

A detailed description of the method is provided in the Supplemental information. Briefly, ICA decomposes the input data X into a linear combination of sources (or components) such that X=AS. S and A denote the component and loading matrix, respectively. Each row of S represents a component while each column of A represents a loading vector associated with the component. For interpretation, the component is generally normalized (z-score) and the top contributing SNPs are selected with |z-score|>2. While ICA decomposition essentially identifies a set of variables covarying with an embedded pattern as reflected in the loading vector, this covariation pattern is largely contributed to by the top SNPs. Thus the top SNPs of each component can be considered as a cluster where the cluster members covary with the loading vector. In the current work, given the 246 samples and 5,468 SZ-related SNPs, we extracted 16 independent components (ICs) based on the criterion of minimal description length (Rissanen, 1978) using Infomax ICA (Amari, 1998; Bell and Sejnowski, 1995). Top SNPs were then selected with |z-score|>2. For each component, we computed the loadings accounted by the top contributing SNPs (denoted as A_top). Then correlations between A_top and A were computed to assess the contributions of top SNPs to the loading patterns accounted by all the input SNPs. The PRS-SZ was also computed for the top SNPs of each component. It was a linear weighted sum of the top SNPs' genotype data where weights were natural logarithms of the odds ratios of the same SNPs released by the PGC GWAS of SZ (De Jager et al., 2009; Ripke et al., 2014). Correlations were then computed between ICA loadings accounted by the top SNPs and their PRS-SZs. Furthermore, we investigated with a permutation test whether the observed A_top-PRS_top correlations might be inflated by the linkage disequilibrium (LD) structure.

Results

Table 1 lists the number of top SNPs, A-A_top correlation, as well as A_top-PRS_top correlation for each IC. The top SNPs of each component are summarized in Table S2. Table S3 details a representative component (IC1) with SNP positions and hosting genes listed. As shown in Table 1, all 16 components showed very high correlations (r>0.9) between the loadings accounted by all SNPs (A) and those accounted by the top SNPs (A_top), confirming that the ICA components reflected the covariation patterns among the top SNPs. For 13 out of the 16 components, A_top showed high correlations (|r|>0.5, p<5.76×10^-17) with the corresponding PRS-SZs, indicating that the identified covariation patterns among top SNPs reflected their cumulative risk for SZ identified previously in much large samples. For most of the components, the top SNPs were distributed across multiple LD blocks, as shown in Figure S1, suggesting that ICA captures covariation including but not limited to LD structure. In addition, a significant p-value<0.001 was observed from the permutation test, confirming that the observed A_top-PRS_top correlations were not likely fully attributable to the LD structure. A speculation is that, a cluster of top SNPs covarying with the identified pattern might be reflecting their participation in the same genetic network where functional adaptation and selective pressure may play a role in maintaining the covariation (Jones et al., 2014; Rohlfs et al., 2010). A DAVID functional annotation analysis (Huang et al., 2009a, b) on IC1 revealed a number of enriched pathways, including synapse, mitotic cell cycle, chromatin modification, as summarized in Table S4. The annotation results suggest some possibility of interactions.

Table 1. Correlations between the top SNPs' ICA loadings (A_top) and polygenic risk scores for schizophrenia (PRS_top).

IC index	Number of Top SNPs	A-Atop correlation (r-value)	Atop-PRStop correlation (r-value)
1	273	0.9179	-0.5784
2	231	0.9245	-0.7578
3	143	0.9311	0.7733
4	116	0.9139	-0.6993
5	184	0.9206	0.1900
6	163	0.9227	-0.7092
7	146	0.9307	-0.5833
8	199	0.9228	-0.4623
9	260	0.9281	0.4999
10	130	0.9255	-0.8827
11	131	0.9307	0.6596
12	278	0.9367	-0.5136
13	122	0.9448	0.9700
14	148	0.9377	-0.8777
15	166	0.9490	-0.9280
16	80	0.9797	0.9934

Discussion

In summary, we demonstrate that ICA decomposition of SZ-related SNPs yielded meaningful clusters. The ICA components were dominated by top SNPs whose covariation patterns (reflected in loadings) correlated highly with their polygenic risk scores for SZ. And the observed correlations between ICA loadings and PRS-SZs were significantly higher than those observed from randomly weighted genotype data with the same data structure, justifying the ICA decomposition. Given that the PRS-SZs were calculated using odds ratios derived from a GWAS including ∼35,500 cases, it is noteworthy that ICA is able to identify patterns highly correlated with the PRS-SZs in just 246 subjects. Meanwhile, although this observation increases the confidence that the identified components are likely functionally meaningful, how these SNPs are functionally interactive awaits further investigation. Overall, this type of blind source separation analysis holds promise for identifying subsets of functionally related genetic variants that may not be recognized by hypothesis-based models, which might help characterize the genetic architecture of the susceptibility SNPs identified from GWAS and contribute to delineating the pathophysiological pathways underlying the complex disorder.