SZGR: a comprehensive schizophrenia gene resource

Association data

For association, we first collected the data from the recently established SchizophreniaGene database (http://www.schizophreniaforum.org/). The downloaded genotyping data was processed by a data cleaning and risk-allele evaluation pipeline developed in our recent combined odds ratio (OR) method.¹⁶ We selected the genes that had significant P value using our combined OR method or at least one positive association result in publication. Currently, there were 281 genes in this category, all of which have been genotyped with positive association signal in at least one study.

Linkage data

We selected linkage bins identified by the two genome scan meta-analyses (GSMA). The first GSMA was applied to data from 20 schizophrenia genome-wide linkage scans⁹ and identified 12 bins whose PAvgRnk and Pord are both <0.05. We obtained 2158 genes from these bins based on their genomic locations and gene annotations in NCBI (ftp://ftp.ncbi.nlm.nih.gov/gene/GeneRIF/) and defined them as “GSMA_I” in our database. The second GSMA was applied to 32 schizophrenia genome scans.¹⁰ We obtained genes from 10 bins whose PSR are <0.05 for all the samples and 6 bins only for European-ancestry samples. These two lists were defined as “GSMA_IIA” (2295 genes) and “GSMA_IIE” (1474 genes). The P value of each linkage bin was assigned to the genes within the bin.

Expression data

We downloaded gene expression data from the Stanley Medical Research Institute (SMRI, https://www.stanleygenomics.org/). The data is based on meta-analysis of 12 individual gene expression datasets from 988 microarrays for schizophrenia and bipolar disorder.¹¹ We extracted 726 genes that were differentially expressed between schizophrenia post-mortem and control samples (P < 0.05) and considered them schizophrenia candidate genes.

Literature search

Co-occurrence of a gene and a schizophrenia-related keyword in an abstract may indicate that the gene is likely associated with schizophrenia. We performed a high throughput literature search based on this assumption using six schizophrenia-related keywords: “schizophrenia”, “schizophrenias”, “schizophrenic”, “schizophrenics”, “schizotypy” and “schizotypal”. We used the Linkout e-retrieval utility in NCBI Entrez Programming Utilities (http://eutils.ncbi.nlm.nih.gov/corehtml/query/static/eutils_help.html) to perform the literature search followed by manual checks of errors.

Gene Ontology

In psychiatric genetics as well as other complex disease studies, investigators have often selected functional candidates falling under the heading referred to as “the usual suspects” such as the genes suggested by neurotransmitter psychopharmacology. Searching genes by appropriate GO terms is useful and efficient for this purpose. A list of neuro-developmental terms was compiled based on expert recommendations (Supplementary Table S1). The related GO terms and their corresponding genes were identified based on keyword matching. We restricted those GO terms whose level were higher than 3 in GO tree because lower level GO terms tend to be non-specific (e.g. molecular function).

Protein-protein interaction network

Gene networks play important role in causing complex diseases. We first constructed the human interactome by integrating the experimentally verified PPI pairs from six databases: Human Protein Reference Database (HPRD),¹⁷ BIND,¹⁸ IntAct,¹⁹ MINT,²⁰ Reactome²¹ and DIP.²² Then, we collected a small set of genes with “best” evidence so far to serve as “seeds” in network analysis. In this study, we selected 38 genes that had significant meta-analysis results⁵ or had been reviewed with extensive evidence.²³ We named them core genes; the details are shown in SZGR. Among these 38 genes, 32 appeared in the human interactome. In network topology, proteins in the shortest path tend to have same or similar biological process.^{24, 25} For any pair of core genes, we identified its shortest path in the human interactome. Then, the genes in the shortest path were selected and scored (see next subsection). A total of 1035 genes were selected based on this network feature.

miRNA and _targets

Schizophrenia related miRNAs were collected from two independent studies which identified 18 miRNAs differently expressed in brain cortex of schizophrenia patients versus control samples.^{26, 27} We also collected and curated 87 non-schizophrenia specific brain expressed miRNAs from miRNA microarray expression studies and miRNA regulation surveys.^{26, 28, 29} Finally, we collected the miRNAs expressed in non-brain tissues from two large-scale miRNA expression atlas studies.^{30, 31} After removing schizophrenia or brain specific miRNAs, the remaining miRNAs were considered non-brain expressed.

The potential miRNA _target sites, family annotations, and sequence conservation information were extracted from the files downloaded from _targetScan (version 4.2, April 2008, http://www._targetscan.org/vert_42/). Then, the miRNA information was matched to schizophrenia candidate genes and made available in SZGR.

Summary

We collected experimental data from association, linkage and expression studies for schizophrenia, performed high throughput literature search and GO term analysis using the keywords or terms related to schizophrenia or neurodevelopment, and extracted schizophrenia candidate genes based on network features. Overall, the three datasets (“Association”, “Linkage”, and “Expression”) represent experimental data while the other three datasets (“Literature”, “GO_Annotation”, and “Gene_Network”) represent schizophrenia candidate genes with weak evidence. We also collected schizophrenia-specific or brain-specific miRNAs and their _target sites in the candidate genes. In total, there were 7855 non-redundant genes whose symbols could be found in the EntrezGene database (Table 1). The overlap between datasets is shown in Table 2.

Data browse

In the main page (Figure 2A), the user may browse data by: (1) clicking one of the eight data categories (“Association”, “Linkage”, “Expression”, “Literature”, “GO_Annotation”, “Gene_Network”, “KEGG_Pathway”, and “miRNA__target”); (2) selecting one chromosome; (3) selecting one of the “Datasets” on the function bar on the top of the web page; or (4) clicking one of the four lists of prioritized candidate genes generated in our recent studies.

All datasets are relationally linked. Once the user clicks a gene ID, a detailed gene page is shown. It includes the following information.

(1) A summary for the gene, including gene symbol, synonyms, description, type, map location, and external links to other public databases (Figure 3A).

(2) Data sources of the gene. This summarizes the evidence in six major data categories including a category-specific score and web link to the related databases when available (e.g. SchizophreniaGene database or PubMed) (Figure 3A).

(3) Gene expression profile in 79 human tissues extracted from the Gene Atlas (version 2).³² Gene expression in a tissue was measured using the arithmetic mean of the average difference (AD) values of their corresponding probe sets. The ADs for a gene are dynamically plotted in a single graph (Figure 3B).

(4) Gene Ontology annotations. The neuro-related GO terms are highlighted (Figure 3C).

(5) Protein-protein interactions between the protein encoded by the gene and other proteins in the human interactome. It presents local PPI environment by listing its direct interactors (distance 1) and distance-2 interactors (proteins that directly interact with the distance-1 proteins) (Figure 3D).

(6) KEGG pathways in which the gene involves (Figure 3E).

(7) miRNA _target sites. It includes the information such as miRNA families, prediction (e.g., the start and end positions of each predicted _target site) and an external link to the MiRBase database (http://microrna.sanger.ac.uk/http://microrna.sanger.ac.uk/) (Figure 3F).

Data search

SZGR provides multiple search options in a user-friendly environment. Besides a quick search function on the top right of the web page, the user may search by gene id, symbol, synonym, chromosomal region, specific data source, or by a user-defined combined setting (Figure 2B). It also provides an option to search genes involved in a pathway by pathway ID or name.

Data integration

Data integration includes two functions: union and intersection. Union is to combine the genes appearing in both gene sets, while intersection is to find genes common in both gene sets (Figure 2C). This simple tool may help the user quickly identify genes with specific evidence.

Gene ranking algorithm

In SZGR, we extended and implemented a multi-dimensional evidence-based gene ranking algorithm developed in our recent study.¹⁵ In this algorithm, in each data category, each gene is assigned a category-specific score. There are four data categories (association, linkage, expression, and literature) in Sun et al.;¹⁵ here we extended to six data categories (GO and network). For most data categories, gene score is calculated by –log₁₀ P when P values are available. For literature search, we assigned score for a gene based on the number of keywords being hit in the search. Similarly, we assigned a score for a gene based on the number of neuro-related GO terms annotated to the gene.

For gene network, the genes in the shortest path to a pair of core genes were assigned scores to measure their closeness to the phenotype (i.e. schizophrenia). We modified Wu et al.³³ method to calculate the closeness of a gene in the shortest path to schizophrenia (i.e. core genes). The closeness of a gene g in the shortest path to a schizophrenia core gene is calculated by Gaussian kernel $e^{-L_{gg\text{'}}^2}$ where g’ is the core gene and L is the distance between genes g and g’ in the shortest path. Therefore, the final score of gene g is the sum of its closeness to all core genes:

where C is the set of core genes.

Next, we searched an optimal weight matrix that weighs the score in each data category differently. The search of the optimal weight matrix was described in Sun et al.,¹⁵ which is based on two steps evaluated by the core genes and independent GWAS P values. The final score of a gene is calculated by

where i is the data category index and w_i is the corresponding weight in the weight matrix.

Gene ranking tool

SZGR provides online tool for gene ranking. Currently, it has two options (Figure 4A). The first one is based on the optimal weight scheme that was recommended in our recent multi-dimensional evidence-based candidate gene prioritization method.¹⁵ The score for each data category is initially calculated and then a combined score by weighing the category-specific scores is calculated. Genes are then ranked by their combined scores. The second option is custom weight scheme. The user may choose any weight for each data category based on his/her prior knowledge or special interest.

To help the user evaluate and fine-tune the weight scheme, SZGR provides two graphical presentations of the ranking results. The first one is to show the rank positions of the core genes among all the ranked genes (Figure 4B). Since there has been no gold positive genes for schizophrenia yet (genes that have been confirmed to cause schizophrenia), we use core genes for this purpose, but they can be replaced by user-defined genes. Assuming that the core genes may have better evidence than other candidate genes, an efficient weight matrix is expected to rank the core genes, or their majority, on top of all candidate genes. The second one is a comparative distribution of the scores of the core genes and all genes. Genes are separated into different bins by their scores (e.g., 1–2, 2–3). An ideal distribution is that most of the core genes are ranked on the top while only few are ranked in the middle or at the bottom, as illustrated in Figure 4C.

Prioritized candidate genes

There are multiple approaches to select and prioritize candidate genes for complex diseases. A straightforward approach is to evaluate and weigh genetic significance information in multiple studies in one data category. We demonstrated this by ranking more than 500 genes in more than 2000 association studies and prioritizing 75 candidate genes using combined OR method.¹⁶ This gene list, named “75 genes by COR”, is available in SZGR.

We also demonstrated a multi-dimensional evidence-based gene prioritization approach for schizophrenia genetic data.¹⁵ By using the optimal weight matrix and evidence in four categories of data (association, linkage, expression, and literature), we prioritized 160 genes using the first version of GSMA results⁹ (named “160 by Lewis et al.”) and 173 genes using the just released second version GSMA results¹⁰ (named “173 by Ng et al.”). Both gene lists are available in SZGR.

Gene network analysis

The prioritized candidate genes can be further applied to follow up bioinformatics analysis and experimental verification. Here we demonstrated it by one example. For the 160 candidate genes, we extracted their network from the human interactome using the Steiner minimal tree algorithm.³⁴ The network had 233 nodes including 135 prioritized genes (named SZGenes) and 98 new genes (named non-SZGenes). We examined association signal of these non-SZGenes in an independent genotyping project (unpublished data). In this project, a total of 3660 SNPs from 191 schizophrenia genes were genotyped in our Irish Study of High Density Schizophrenia Families (ISHDSF) sample. Among the 98 non-SZGenes, 6 genes were included for genotyping, three had P values <0.05. In one gene, five SNPs had P < 0.05 and the smallest P value was 0.00091.