Construction of LDGIs

The data analysis workflow is shown in Figure 1. We first extracted human haplotypes from the HapMap Phase II and III data [18], which were generated using the PHASE software [19]. Only the SNPs that are located within the promoter or exonic regions of protein-coding genes (with reference to the Ensembl annotations [20]) were considered. Note that the promoter regions encompass 2 kb sequences upstream of the transcriptional start sites, and exonic regions include coding regions (CDSs) and untranslated regions (UTRs). In view of population stratification, we clustered the individuals examined in the HapMap Phase II and III projects into subpopulations using the PLINK package (version 1.07) [21] (Table 1). Here we consider only the subpopulations that contain at least 20 individuals. For each subpopulation, we calculated LD scores (i.e., r² and D′[14]) for all combinations of SNP pairs. Two SNPs were considered to be a long-range LD-linked SNP pair (designated as an “LRLD-SNP pair”) if they satisfied both of the following criteria: (1) to avoid the inclusion of accidentally linked SNPs, an LRLD-SNP pair had to be subject to a strong LD (r² ≥ 0.8); (2) to minimize the probability of hitchhiking or lack of recombination, the two SNPs had to be located in different chromosomes or be separated by at least one recombination hotspot retrieved from the International HapMap Project. The latter criterion may considerably decrease the probability that the identified LRLD-SNP pairs belong to the same “LD blocks” (or “haplotype blocks”, which represent regions where recombination events occur rarely, and consequently LD is maintained) even if they are located in the same chromosomes. Accordingly, we identified 801,340 LRLD-SNP pairs, which contained 94,876 SNPs (Table 1). Genes connected by these LRLD-SNP pairs were considered human LD-based gene interactions (LDGIs). The LDGIdb is composed of a collective total of about 646,203 gene linkages, which contain 21,240 genes (Table 1). Since population stratification was also considered, the LDGIdb also provides potential population-specific gene interactions, which may be useful for investigations of population-specific traits/diseases.

Calculation of r² and D′ values

Data retrieval

HapMap Phase II (release 22) and III (release 2) haplotype data and the corresponding recombination hotspot information were retrieved from the International HapMap Project [22]. The human protein-coding genes were downloaded from the Ensembl genome browser (release 53). The human PPI data (designated as “collected PPIs” in the LDGIdb) were collected from seven experiment-supported PPI databases: HPRD [23], DIP [24], MINT [25], IntAct [26], REACTOME [27], BioGRID [28], and MIPS [29]. The extracted PPI collection included a total of 76,955 interactions. The CRG (Centre for Genomic Regulation) human interactomes (designated as “CRG PPIs” in the LDGIdb) were downloaded from Bossi and Lehners’ study [30], which comprised 80,922 interactions. Human gene co-expression data were downloaded from the TMM database [4], which contained 203,043 high-confidence co-expression links that were observed in at least three microarray data sets. The biological interactions inferred from the above databases (i.e., collected PPIs, CRG PPIs, and co-expression links) were integrated into the LDGIdb for comparison with LDGIs. If an LDGI was not found in any of the databases, it was considered to be a potentially uncharacterized gene interaction. The GWAS [16] data were downloaded on August 23rd, 2011 [31]. For LRLD-linked genes, more detailed information was provided including protein domain descriptions (according to Interpro [32], SMART, and PFAM), KEGG pathways [33], and disease association information (OMIM, HIV interaction, and the Genetic Association Database [34]), which were all downloaded from the DAVID knowledgebase [35].

Web interface

Users can search for LRLD-SNP pairs and LDGIs (which are linked by LRLD-SNP pairs) by setting three adjustable parameters: HapMap data source (Phase II or III), P value for PLINK population clustering (P < 0.01 or P < 0.001), and r ² value for linkage disequilibrium (≥0.8, ≥0.9, or 1) (Figure 2A). Note that we only considered population clusters containing at least 20 individuals (Table 1). Also note that LDLR-SNP pairs with r ² = 1 are subject to a “complete” LD. The LDGIdb supports four types of queries. Users can search for LRLD-SNP pairs/LDGIs by specifying the types of genomic location of LRLD-linked SNPs, SNP ID, gene accession number(s), or genomic coordinates (Figure 2B). GWAS-related LRLD-SNP pairs are also provided (Figure 2C). As shown in Figure 2D, the LRLD-SNP pairs/LDGIs are categorized, according to the types of genomic location of the linked SNPs, into promoter-promoter, promoter-UTR, promoter-CDS, CDS-CDS, CDS-UTR and UTR-UTR interactions. The CDS-related LDGIs are further categorized according to whether the LD-linked SNPs are nonsynonymous or synonymous (Figure 2D). Therefore, the user can choose LRLD-SNP pairs that occur in different genomic regions and that (in the case of coding SNPs) represent changes at the RNA or protein levels (the user can choose more than one type of interaction). The user can further select one or more population of interest to retrieve population-specific LDGIs. The results are downloadable (Figure 2E). For simplicity, the web interface displays only the first 10 records of each query (Figure 2F). The user can find detailed information of allele combinations of LRLD-linked SNPs and genomic regions where the linked SNPs are located in the results (Figure 2G). For the identified LDGIs, the interface also provides human PPI data collected from eight experiment-supported databases (i.e., collected PPIs and CRG PPIs) and high-confidence co-expression interactions for comparison. More detailed information of LDGI genes is also provided, including protein domain annotations, biological pathways, and disease associations.

Discussion and future development

Here we propose a new resource for studies of potential human gene interactions (i.e., LDGIs) based on haplotype data. In LDGIs, the linked genes are located in different chromosomes or LD blocks but are connected by one or more exonic/promoter SNP pairs that are subject to strong linkage disequilibrium (r ² ≥ 0.8, ≥ 0.9, or 1). We suggest that this LRLD approach and the LDGIdb can be potentially applied to the following areas. First, LDGIs may represent potential uncharacterized gene interactions, in which the functional associations between the LDGI genes may not be explicitly indicated in other biological networks. Second, although we constructed the LDGIdb using SNP data in this study, the LRLD approach can actually be expanded to include other types of genomic variants such as copy number variation and insertion/deletion. Third, given enough haplotype information, population-specific LDGIs/LRLD-SNP pairs may be identified for more focused studies, particularly in the field of pharmacogenetics. Fourth, the correlation between the LDGIs/LRLD-SNP pairs and disease-associated SNPs such as those identified in GWAS studies can be explored. For example, SNP rs393152, which is associated with Parkinson’s disease [36], forms an LRLD-SNP pair with rs12185268. Interestingly, rs12185268 was demonstrated to be connected to the same disease [37] two years after the publication (i.e., Ref #36) of the association of rs393152 with the disease. Another example is the LRLD-SNP pair: rs9858542–rs3197999. The two SNPs in this pair were shown to be related, respectively, to the Crohn’s disease [38–41] and the ulcerative colitis [42, 43]. These examples show that two SNPs that are associated with the same (or related) human diseases/traits can be identified by our approach. Moreover, there are also cases in which GWAS SNPs and their LDGI partners are associated with the same (or related) human diseases. For example, the GWAS SNP rs5215 in KCNJ11 is known to be associated with Type II diabetes [44, 45]. This SNP forms an LRLD-SNP pair with rs757110, which is located within the CDS of ABCC8. Mutations and deficiencies in the protein encoded by ABCC8 have been suggested to be associated with hyperinsulinemic hypoglycemia of infancy and non-insulin-dependent diabetes mellitus type II [46, 47]. The above examples suggest that the LRLD-SNP linkages may reflect biological interactions in the human molecular system and have the potential to detect previously uncharacterized gene interactions. As disease-association data accumulate, the LDGIdb may become an increasingly powerful tool by which to identify potentially uncharacterized disease-associated gene interactions, contributing to network-based studies of human diseases. Notably, however, since the majority of HapMap SNPs are relatively common variants, the linkages of rare alleles may not be represented in LDGIdb.

This study actually examined whether observed non-physical SNP linkages occur simply by chance or whether they are biologically meaningful. The above examples suggest that the inferred LDGIs may be functionally relevant. One interesting question is what are the molecular mechanisms underlying the inferred gene interactions. For the CDS-CDS LDGIs that involve only nonsynonymous changes, the functional association is speculated to result from direct or indirect protein-level interactions. Of course, the LDGIs may also represent adventitious linkages or false positives that result from unknown population substructures. Meanwhile, the biological meanings of the LDGIs that involve UTR SNPs (i.e., CDS-UTR and UTR-UTR linkages) or synonymous SNPs (i.e., nonsynonymous-synonymous and synonymous-synonymous linkages) may be more subtle. These potential interactions may be associated with translational regulation. Specifically, 5′UTRs may contain multiple sequence features that are involved in translational regulation, including upstream open reading frames, secondary structures, internal ribosome entry sites, and iron regulatory protein binding sites [48]. The disruption of these functional elements may cause changes in the efficiency of protein translation. On the other hand, 3′UTRs are known to be the major binding target of microRNAs, which can also suppress protein expression [49]. In addition, 3′UTRs may harbor protein-interacting secondary structures or the signals of nonsense-mediated decay or polyadenylation [48], both of which can affect the efficiency of protein translation. Meanwhile, synonymous coding SNPs are known to affect mRNA stability and splicing, leading to changes in the corresponding protein products [50]. Since both the UTR and synonymous SNPs may affect protein abundance, dosage imbalance and unidentified, indirect protein interactions may be possible explanations for the observed linkages.