We started with the MBP gene (which encodes myelin basic protein, an integral element of the myelin sheath surrounding the neuronal extensions). This gene is critical in triggering the immune reaction in the demyelination process for experimental allergic encephalomyelitis (EAE), a rodent model of MS. Importantly, MBP has been implicated both in linkage and in association studies conducted in MS pedigrees of Scandinavian origin (see Haines and coworkers 4 for references).

We applied the second genome-wide LA search method (Figure 2c) to the NCI_cDNA database. By treating MBP as the query gene X, we evaluate the LA score for every pair of genes (Y,Z). Because there are 9,076 genes in this database, about 49 million LA scores are computed and compared with each other. The output of a short list of 25 gene pairs with the best LA scores each from the positive and the negative ends is given in Additional data file 1 (Table S1). The statistical significance of the results of this gene search procedure is discussed in Additional data file 2 (Supplementary Text 3). We find that the gene A2M (encoding α₂-macroglobulin, a cytokine transporter and protease inhibitor) appears many times. We further find an interesting biologic functional association between A2M and MBP from some literature about the pathogenesis of MS. Following demyelination in human MS and rodent EAE, immunogenic MBP peptides are released into cerebrospinal fluid and serum (see Oksenberg and coworkers 2 for references) and A2M represents the major MBP-binding protein in human plasma 17. A significant increase in α₂-macroglobulin is found in plasma of MS patients 18. Analogously, in rodent EAE, infusion of α₂-macroglobulin significantly reduces disease symptoms 19.

Among the genes to which A2M is paired, three are found to have functional association with immunologic neurodegenerative diseases. LYST (lysosomal trafficking regulator, also known as CHS1) is the causal gene for Chediak-Higashi syndrome, an inherited immunodeficiency disease, and CHM (Rab escort protein 1) is responsible for an inherited human retinal blindness known as choroideremia. (For details, see Online Mendelian Inheritance in Man of the National Center for Biotechnology Information 20.) TRIB2 (tribbles homolog 2) was identified as an autoantigen in autoimmune uveitis, a term encompassing a group of ocular inflammatory disorders with unknown causes 21. Additionally, MPDZ (multiple PDZ domain protein) encodes a tight junction protein that is detected in noncompact regions of myelin, and it is thought to be required to maintain the cytoarchitecture of myelinating Schwann cells 22. The biologic connection for other genes, many still of unknown function, is not clear. We compute the correlation between these genes and find that most of them have significant correlations. (See in Additional data file 2 [Supplementary Text 4] for more discussion.)

Four major loci linked to MS have been identified in Finnish families: HLA on 6p21, MBP on 18q, and loci on 17q22-24 and 5p14-p12 23. These loci have also been implicated in other MS study samples from more heterogeneous populations 2425. The large locus on 17q was further refined to a 3 megabase (Mb) region in the Finnish MS families 23. However, little information is available in the literature concerning how various loci are related to each other biologically. Most recently, association of specific PRKCA alleles at 17q24 with MS both in Finnish and Canadian MS study samples has been reported 7. Involvement of PRKCA in MS was also validated by an association reported in a UK population 26. PRKCA encodes a regulator of immune response, making it a highly suitable candidate gene for MS. A potential functional link between the MBP and PRKCA genes was identified by Feng and coworkers 27, who showed that a golli product of the myelin basic protein gene (MBP) can serve as a negative regulator of signaling pathways in T lymphocytes, particularly the protein kinase C pathway.

To study the co-expression pattern between MBP and PRKCA, we took them as genes X and Y to explore the GNF_2002 database using our system. The gene with the greatest LA score was the gene SLC1A3 (glial high affinity glutamate transporter, member 3; see Additional data file 1 [Table S2]). Interestingly, SLC1A3 is located on 5p13.2 (36.6 to 36.7 Mb), within the previously identified MS locus on 5p 28, which is syntenic to the EAE2 locus in mouse.

We wished to test whether there is any genetic relevance of SLC1A3 to MS. We selected five single nucleotide polymorphisms (SNPs) flanking the SLC1A3 gene (Table 1) to be genotyped in our primary study set, consisting of 61 MS families from the high-risk region of Finland. The most 5' SNP, namely rs2562582, located within 2 kilobases from the initiation of the SLC1A3 transcript, exhibited initial evidence for association with MS (P = 0.005) in the transmission disequilibrium test (TDT) analysis, suggesting a possible functional role for this variant in the transcriptional regulation of this gene. Moreover, as shown in Table 1, stratification of the Finnish MS families according to HLA genotype (using the SNP rs2239802, which exhibited strongest evidence for association in the Finnish families in the report by Riise Stensland and coworkers 28), strengthened the association between the SLC1A3 SNP and MS (P = 0.0002, TDT). Thus, based on LA, and supported by association analyses in an MS study sample, the presence of SLC1A3 serves as a potential candidate to connect all four major MS loci identified in Finnish families, elucidating a potential functional relationship between genetically identified genes and loci. We consider further evidence in the following discussion.

We next took MBP and SLC1A3 as the query genes to conduct a genome-wide LA search in all four gene expression databases. Figure 3 and Table 2 highlight a set of genes whose biologic functions are most relevant to our MS study according to the literature. The detail LA outputs are given in Additional data file 1 (Tables S3 to S6). All LA plots are easy to generate online using our website. The one for the triplet including MBP, PRKCA, and SLC1A3 is shown in Figure 4.

For GNF_2002 data, the gene with greatest LA score is GRM3 (glutamate receptor, metabotropic 3), followed by several genes involved in nervous diseases and neural development/functioning: GFAP, CDR1, ROM1, CACNA1A, and GRIA3. We also find IL7R, IGHG3, IGLJ3, and HLA-A among the highest scoring genes in this query. The identification of the IL7R by the LA analysis is particularly interesting because this gene was found to be associated with MS in the recent large international Whole Genome Association study 8.

The locus of HLA on 6p21 is the only consensus MS locus replicated by genetic studies across different populations. Importantly, in the recent fine mapping effort with 1,068 SNPs covering the HLA locus and providing the SNP density of 1 SNP per 2 kilobases in the study sample of 4,200 individuals from Finnish and Canadian MS families 29, susceptibility to MS proved to be determined by HLA-DRB1 alleles and their interactions. Therefore, it is especially interesting that for the GNF_2004 data, eight of the 25 genes with the best LA scores are from the HLA locus: HLA-A (twice), HLA-B (twice), HLA-C (twice), and HLA-G (twice). Other HLA genes with very high LA scores include HLA-E, HLA-F, HLA-DRA, and HLA-DPB1. We also find B2M (which encodes β₂-microglobulin, the light chain of MHC class I antigen) and a MS susceptibility gene, namely CD45 (a T-cell receptor for galectin-1).

The LA lists from NCI_cDNA data and NCI_Affy data also yielded several highly relevant candidate genes for MS, such as MAG, IRF1, APOE, EIV2A, and PDGFA. Also of interest are SIAT8A, SIAT1, SOX4, SOX9, and EPHA2. The protein encoded by MAG (myelin-associated glycoprotein) is involved in the process of myelination 30 and binds to sialic acid. SIAT1 and SIAT8A are both sialyltransferases. SOX4 and SOX9 are involved in central nervous system development 3132. SOX4 is required for the development of lymphocytes and thymocytes 33.

Results from the two NCI datasets also contain genes from the 6p21.3 locus: TAP2, TRIM10, and HLA-DQB1. A further investigation into the expressional association of the HLA family with SLC1A3 using the LA method finds two highly significant genes, namely GMFB and PDGFRA (see Additional data File 1 [Tables S7 and S8]). Also, these genes are be biologically relevant. GMFB (which encodes glia maturation factor beta) is reported to increase in astrocytes around the lesioned area after cortical cryogenic brain injury 34. PDGFRA (the gene encoding platelet-derived growth factor receptor-α) is a well known marker for remyelination. The PDGFA supply may control oligodendrocyte progenitor cell numbers in the adult central nervous system as well as during development 35. Interestingly, CTNND2 (catenin delta 2; neural plakophilin-related arm-repeat protein) is the fifth most correlated gene for SLC1A3 in the GNF_2004 data (see Additional data file 1 [Table S9]). It is also highly correlated with MBP (see Additional data file 1 [Table S10]).

Although all of the putative genes identified using LA method must be confirmed with genetic association studies in multiple populations and eventually in targeted functional studies, the putative genes identified here are highly relevant to MS. Our transcript regulatory findings portray a coherent web of molecular evidence, which supports the glutamate-induced excitotoxicity hypothesis of MS. SLC1A3 is highly expressed in various brain regions including cerebellum, frontal cortex, basal ganglia, and hippocampus. It encodes a sodium-dependent glutamate/aspartate transporter 1 (GLAST). Glutamate and aspartate are excitatory neurotransmitters that have been implicated in a number of pathologies the nervous system. Glutamate concentration in cerebrospinal fluid rises in acute MS patients 37, whereas the glutamate antagonist amantadine reduces MS relapse rate 38. In EAE, the levels of GLAST and GLT-1 (SLC1A2) have been reported to be downregulated in spinal cord at the peak of disease symptoms, and no recovery was observed after remission 39. We consider it encouraging that several lines of evidence, including both genetic association and gene expression association, are consistent with the glutamate-induced excitotoxicity hypothesis, which states that glutamate-induced excitotoxicity results in demyelination and axonal damage in MS 40.

The recent international MS Whole Genome Association scan 8 provided additional evidence supporting an association between MS susceptibility and SLC1A3. A major component of the study is the use of Affymetrix 500K to screen common genetic variants of 931 family trios. Using the online supplementary information provided by the International MS Genetics Consortium 8 we found two SNPs, namely rs4869676 (chromosome 5: 36641766) and rs4869675 (chromosome 5: 36636676), with TDT P values of 0.0221 and 0.00399, respectively, which are in the upstream regulatory region of the SLC1A3 gene. In fact, within the 1 Mb region of rs486975 there are a total 206 SNPs in the Affymetrix 500K chip. No other SNPs have P values less than that of rs486975. The next most significant SNPs in this region are rs1343692(chromosome 5: 35860930) and rs6897932 (chromosome 5: 35910332; the identified MS susceptibility SNP in the IL7R axon). The MS marker we identified, rs2562582 (chromosome 5: 36641117), less than 5 kilobases away from rs4869675, was not used in the Affymetrix chip.

Although the results reported here should be considered preliminary, we propose that the genes and networks identified should be targets for additional analyses of MS in different study populations.

One unique feature of our approach to finding candidate genes is the use of public domain gene expression databases, of which the original experiments were not designed to study our disease of interest. For example, the two NCI-60 cell line (a panel of 60 diverse human cancer cell lines) gene expression data have primarily been used to aid anticancer drug screening, not for the study of MS. With our promising initial findings, we expect our functional genomics approach to be applicable in the initial identification of involved molecular pathways in the pathogenesis of other complex diseases. Investigators may apply our LA method or bring in other computational methods to data mine the numerous free public gene expression databases, thus reducing the time and expense associated with disease gene identification.

Four large-scale gene expression databases are employed in this study, with various numbers of conditions and genes. The first two databases give expression profiles for the 60 representative cell lines from seven cancer types that have been used in NCI's anticancer drug screen. The NCI_cDNA database uses the cDNA microarray reported by investigators from P Brown's laboratory at Stanford University 10, whereas the NCI_Affy uses Affymetrix oligonucleotide high-density HU6800 arrays 11. The two other databases 1213 are samples from diverse array of human tissues. GNF_2002 has a probe set for a total of 12,533 genes/clones and 101 chips (using Affymetrix U95A arrays) and GNF_2004 has a probe set for 33,689 genes and 158 chips (using Affymetrix HG-U133A and GNF1H; data downloaded from the Gene Expression Atlas 41). The corresponding numbers for NCI_Affy are 5,611 and 60 (data downloaded from the supplementary data file of Staunton and coworkers 42), and for NCI_cDNA they are 9,703 and 60 (data downloaded from the NCI60 Cancer Microarray Project) 43.

To compute the liquid association score LA(X, Y|Z) for a triplet of genes, normal score transformation is first applied for each gene. After transformation, LA(X, Y|Z) is given by the average of triple product between X, Y and Z: LA(X, Y|Z) = (x₁y₁z₁ + ... x_my_mz_m)/m. For a given pair (X,Y), the test of significance of an LA score is conducted by permutation, as previously described 1416 and the P value is reported in each of our LA output tables. In addition, to help with the interpretation of the effect size of the LA score, two algorithms were used to find the correlation change between the state of high expression of the LA scouting gene and the state of low expression (see Additional data file 2 [Supplementary Text 5]). The LA website 36 was created to facilitate the online computation of LA. High score output genes are returned to user's browser for immediate connection to Entrez Gene. The website also generates LA graphs, performs standard correlation analysis, and provides summary information regarding gene location, functional annotation, and so on.

The study set used for the association analysis contained 28 multiplex MS families with multiple affected individuals, and 41 nuclear MS families (MS patient and his/her parents and, in case of a missing parent, healthy siblings were included). Twenty-two of the 28 multiplex families and all trio families originated from Southern Ostrobothnia region of Finland, which has an especially high incidence and prevalence of MS. All families were Finnish and of Caucasian descent, and they have been described in more detail by Saarela and coworkers 23. Diagnosis of MS in affected individuals strictly followed Poser's diagnostic criteria 44. All individuals gave informed consent and the study was approved by the Ethics Committee for Ophthalmology, Otorhinolaryngology, Neurology, and Neurosurgery in the Hospital District of Helsinki and Uusimaa. (decision 46/2002, DNRO 192/E9/02).

To control for sample mix-ups, all samples were genotyped for determining the sex and four microsatellite markers using the ABI 3730 (Applied Biosystems, Foster City, CA, USA). The data were compared with the known sex of the samples and checked for Mendelian errors. No Mendelian discrepancies were observed in this study set. To select the initial set of SNPs used for the association analysis, we set up the following criteria: each of the markers should be highly polymorphic in Centre d'Etude du Polymorphisme Humain (CEPH) reference families genotyped in the HapMap project and should belong to unique solid line linkage disequilibrium haplotype blocks as defined by Haploview's version 3.11 45. We found five highly polymorphic SNPs within and in the proximity of SLC1A3 gene that belonged to separate haplotype blocks, according to the HapMap data. The SNPs were genotyped using multiplexed allele-specific primer extension on microarrays 46. Primers for multiplex polymerase chain reactions were designed using in-house scripts written for the Primer3 program 47, and an in-house built software package SNPSnapper, version 1.38beta, was utilized to call genotypes automatically.

Allele and genotype frequencies were determined from the data, and deviation from the Hardy-Weinberg equilibrium was tested using Pearson's χ² test. We used the ANALYZE package to conduct TDT analyses to test for association between MS and SLC1A3 gene 48.