In the total data set of 1,321 affected sibling pair families from the UK, the US and Scandinavia, data mining/decision tree analyses identified major T1D predictive signals (marginal markers; Table 1) corresponding to T1D linkage signals found by classic non-parametric linkage analysis 25. As the original T1DGC publication 25 included data on 254 additional affected sibling-pair families not part of the present analyses, direct comparison of results is not possible. However, substantial agreement existed between the analyses (Table 1). Ranking of markers is according to their T1D predictive signal determined by Pearson's χ² statistics and corresponding P value. As we evaluated only a limited number of the genotyped markers in the total data set, we endeavored to see if supplementary information could be extracted from more complete subsets of data (UK/US and Scandinavian). As seen in Table 1, the group of markers corresponding to the HLA region shows a much higher predictive signal (by several orders of magnitude) than the rest of the markers. D6S283 and D6S300 are markers for IDDM15 (6q21) 26, which in linkage studies generally require separate analysis to differentiate its effect from MHC 2526. Markers for the regions 2q31-q33, 16p12-q11.1, 11p15.5, 16q22-q24 and 10p14-q11 identified by linkage analysis 25 also showed high predictive signals in the current study, either in the total data set or in the data subsets (Table 1). In addition, a few new markers were found to show predictive signals (P < 0.05) when evaluated independently of chromosome 6 markers, for example, D17S798, D2S125, D9S175, D8S261 and D4S403. The D21S270 marker was identified in the Scandinavian subset and corresponds to a T1D linkage region on chromosome 21, which we have recently identified and fine mapped 2227. In the UK/US data set, the 2q31-q33 region (the CTLA4 region) seems of higher predictive value than in the total data set (Table 1). D4S403 corresponds to a region previously linked to T1D 2228 containing the WFS1 gene associated with Wolfram syndrome (MIM #222300), which involves T1D 28.

The importance of HLA is well established, and we are, by the methods used here, able to evaluate important markers in sibling pairs sharing just one HLA allele. The top scoring marginal marker for the HLA region was the tumor necrosis factor alpha (TNFA) micro satellite marker, located centrally in the HLA region. To determine candidates for the next level, we searched for interacting markers with the HLA region, in the subgroups of sibling pairs with TNFA IBD status = 1 (TNFA = 1) and TNFA IBD status = 2 (TNFA = 2), respectively. No interactions with TNFA = 0 could be generated due to the low number of affected sibling pairs in this group. Specific combinations of markers corresponding to statistically significant genetic interactions in the combined data set are shown in Table 2. The marker combination TNFA = 1 - D11S910 was shown to be of protective value, since sibling pairs sharing one TNFA allele, but two alleles of D11S910, were strongly protected against T1D (of 25 sibling pairs with this combination, one was concordant for T1D, 24 were non-T1D). The other combinations of markers detected implied increased susceptibility to T1D. None of the interacting markers from Table 2, except D4S403, correspond to previously known regions associated with T1D 29.

Genetic interaction analysis was performed for the marginal markers with the highest predictive signals, and was also performed independent of HLA (TNFA) IBD sharing status. When evaluating epistasis independent of HLA, we searched specifically for epistasis between the three highest ranking markers, D17S798, D2S152 and D2S125, after chromosome 6 markers were removed. In the combined data set, however, only combinations including the marker on chromosome 17 predicted genetic interaction (that is, D17S798 = 1 - D5S429 (P = 0.029) and D17S798 = 1 - D1S197 (P = 0.031), and between D17S798 = 2 - D2P25 (P = 0.041)). These combinations reached statistical significance, and demonstrated increased susceptibility to T1D (Table 2). Relationships could only be inferred for two markers at a time due to the high number of missing and non-informative values for many markers.

We searched for protein networks spanning the regions shown to interact genetically (P values < 0.05; Table 2). This was performed using a high-confidence human protein interaction network 4. Input proteins were proteins corresponding to a defined genetic region surrounding the interacting markers included in the different marker combinations. For all markers except TNFA, 5 Mb on each side of the marker in question was used as input. This region size was selected since linkage peaks (LOD - 1 intervals) from genome scans that use a similar number of markers often corresponded to regions of this size. For the HLA region, we have exclusively used the classic MHC region (4 Mb) for analysis, due to the well examined nature of this region with a high degree of linkage disequilibrium, as well as the large number of genes clustered in this specific region 30. The classic MHC region comprises the TNFA marker in a central position (positioned at bases 31,643,403-31,643,437 on the physical map of chromosome 6, corresponding to 46.7 cM).

We were able to identify 22 protein sub-networks that connect proteins from the different regions corresponding to the significant two-marker predicted genetic interactions. The union of these sub-networks resulted in 13 putative functional modules (Figure 2).

The significance of each putative functional module was assessed by comparison to search results for randomly selected genetic regions. This assessment was made for both the results of marker-region pairs (2-interval) and for the resulting merged modules containing genes from two or more intervals (k-interval). Four 2-interval modules that included TNFA-region genes, two of which were found to be significant, were merged into a single highly significant 5-interval module (Figure 3, module A). This concordance strongly suggests that the four TNFA-region genes TUBB, RPS18, ITPR3 and BAT1 may be important in explaining the mechanism of the four genetic interactions. From the interacting chromosomal regions, the WDR1, LMO7, HNRPLL and RPS15A genes are potential T1D candidate genes. These genes are involved in transcriptional regulation, DNA binding, RNA binding, ion channel activity, ATP synthesis, actin binding and natural killer cell mediated cytotoxicity and cell proliferation. Candidate genes from the four significant functional modules (Figure 3) are listed in Table 3. Other networks with TNFA include genes involved in signal transduction, regulation of transcription, protein biosynthesis and folding, histone activity, ubiquitin-protein ligase activity, as well as response to oxidative stress (Table 3), also of potential relevance in T1D pathogenesis.

A region on chromosome 17 also conferred a high predictive value for T1D and was found to have genetic interactions with three other marker regions. Searches conducted for genes from the three marker pairs (D17S798-D2P25, D17S798-D5S429 and D17S798-D1S197) resulted in six putative functional modules after the initial results were combined (Figure 2). Several of the proteins in these networks are involved in signal transduction, anti-apoptosis, RNA binding regulation of transcription, kinase activity, oxidoreductase activity, DNA and ATP binding as well as oxygen transporter activity (Table 3), making them potentially important in T1D pathogenesis. One of these modules (Figure 3, module D) was found to be significant (P < 0.05) and contained protein interactions between members of three genetic interaction marker pairs. GO terms for molecular function and biological process for all candidate genes in significant functional modules are listed in Table 3. These findings shed light on the pathways the candidate genes in these two regions are likely to be involved in, and may help in understanding the possible effect in T1D suggested by this interaction.

The data set was generated by T1DGC as part of the combined analysis of the existing T1D genome scans 222325. In this process all genotyping data were intra-familially recoded, when possible, to show IBD status rather than exact allele calls. The Scandinavian data set comprised 392 families (411 affected sibling pairs) that were genotyped for 335 microsatellite markers. The combined UK/US data set included 763 families (910 affected sibling pairs) and genotyping of 1,283 markers. In order to analyze markers only genotyped in all data sets the number of markers was reduced to 298. Thus, the total data set used in the analysis comprised 1,321 affected sibling pairs with genotyping data on 298 markers.

As the total data set included only a few unaffected sibling pairs and the analytical methods applied in the present study take advantage of information from non-diseased subjects 18, we simulated data for non-affected sibling pairs 1415161741. A data matrix for unaffected sibling pairs was generated from the data matrix representing the affected sibling pairs. For each marker the number of missing values from the affected was maintained for unaffected sibling pairs. The rest of the matrix for unaffected sibling pairs was completed with values reflecting normal IBD 0, 1 and 2 frequencies, that is, 0.25, 0.5 and 0.25. No correction was made in the simulation for the actual frequency of homozygous parents. The number of unaffected sibling pairs (simulated) was two times the number of affected sibling pairs. The final data matrix then contained 1,311 affected sibling pairs and 2,622 non-affected sibling pairs.

Identification of marginal markers and evaluation of interaction between markers were done as detailed previously 18, with minor modifications. Briefly, data mining algorithms and decision trees were used to predict the most informative markers. We have used the concept of marginal markers and the interactive tree model in SAS Enterprise Miner (SAS Institute Inc., Cary, NC, USA) to calculate all marginal markers using Pearson's χ² statistics and corresponding P value. The tree algorithm determines marginal markers as the roots (the highest level of the trees), as described previously 18. The list of marginal markers identified by this method is produced by Pearson's χ² statistics and corresponding P value.

When searching for interactions between a marginal marker and markers on different chromosomes, we also used Pearson's χ² statistics. Data sets were created including sibling pairs with TNFA IBD status = 1 (TNFA = 1), TNFA IBD status = 2 (TNFA = 2), D17S798 IBD status = 1 (D17S798 = 1) and D17S798 IBD status = 2 (D17S798 = 2) to search for interactions between these, the highest ranked, marginal markers and other markers. Pearson's χ² statistics was then used to search for association between T1D and a marker in these individual data sets. Searching for interactions between markers on the same chromosome was not performed, because the random methods used here do not allow for linkage disequilibrium of adjacent markers on a chromosome.

A human protein interaction network was generated 4. Briefly, protein interaction data were obtained from the databases BIND 4243, MINT 44, IntAct 45, KEGG annotated protein-protein interactions (PPrel), KEGG Enzymes involved in neighboring steps (ECrel) 46 and Reactome proteins involved in the same complex, indirect complex or same or indirect reaction 47. All human data were pooled, and to increase information interolog data (protein interactions among orthologous protein pairs in different organisms) from 17 eukaryotic organisms were also included to obtain protein-protein interaction networks 4. We devised and thoroughly tested a global confidence score for all interactions in the network. This confidence score is constructed to take into account factors like topology of the interaction network surrounding the interaction, number of publications the interaction had been detected in, that interactions are more reliable, if they have been reproduced in more than one independent interaction experiment, and, furthermore, the experimental set-up (large- or small-scale study). Interactions from large-scale experiments generally contain more false positives than interactions from small-scale experiments 48. Furthermore, the reliability of this score was confirmed by fitting a calibration curve of the score against overlap with a high-confidence set of about 35,000 human interactions, demonstrating that the score was a reliable measure of interaction confidence 4. Networks were constructed from proteins in defined intervals (corresponding to the respective rules) and their first order interaction partners using interolog data in a manner similar to that described by Lehner and Fraser 49. Proteins known to interact in other species were mapped to their human orthologs using the Inparanoid database 5051. In the resulting networks, each node represents all proteins encoded by a single human gene and their orthologs in other species. An edge between two nodes indicates one or more interactions between any of the proteins represented by the node. The protein interaction confidence score was implemented to use only interaction data above the interaction threshold separating 'high' from 'low' confidence interaction data. This threshold was found by using a genetic algorithm on the interaction network to obtain the optimal threshold for signal to noise ratio 4.

To further reduce noise in the networks we also devised a network score, implemented to retrieve sub-networks enriched in proteins from the selected regions that interact directly or through significant linker proteins (that is, proteins that connect proteins from the selected regions, but are not in any of the selected regions themselves). The network score reflects the amount of interaction partners allowed for each linker protein for it to be included in relation to the number of interaction partners from the selected regions. The score is calculated for every protein and is the result of 'number of interactions with input proteins' divided by 'total interactions' for each protein, making networks consisting of proteins with many interactions less important and reducing noise from highly interacting proteins from unselected regions in the genome. A very stringent threshold-score of 0.5 was used.

Positional genes and their corresponding proteins were obtained from the University of California Santa Cruz (UCSC) genome browser using 'Genes and Gene Prediction Tracks' 52 and 'Ensembl Genes' from the table browser 53. For two marker rules, proteins encoded by genes from 5 Mb on each site of the respective markers were used as input proteins. For the TNFA marker, proteins encoded by genes from an interval corresponding to the classic MHC region (position 29.26-33.90 Mb on chromosome 6) 29 were used.

For each protein belonging to an interval of interest, a query was made in the constructed human interaction network. Only interactions above the high-confidence threshold were maintained. Cytoscape version 2.3.1 was used to visualize the resulting networks 54. Genes were classified according to GO terms 1.

In an effort to determine the significance of the putative functional modules, we empirically estimated the probability of observing as many or more marker interval genes (n_i and n_j for interval i and j) in modules of size N or smaller in our protein interaction network G, that is:

P(x_i ≥ n_i, x_j ≥ n_j, X ≤ N|G).

This probability was estimated for each module with n_i > 0 and n_j > 0 found for queries based on genes from one of the nine 2-interval genetic interactions. Estimates were derived from the size and number of modules discovered from 100,000 random queries. Random queries were constructed from genes selected from random interval pairs with the same number of genes as in the two genetically interacting marker intervals. Random intervals were defined from consecutive genes on a chromosome. As each query generates a varying number of modules (connected components), the probability estimates were calculated from the frequency of queries that result in one or more connected components containing x_i ≥ n_i and x_j ≥ n_j genes from random interval i and j, respectively, with total number of X ≤ N genes.

After merging all connected components from the nine different 2-interval queries, six modules were found to contain genes from three or more intervals. For each of these k-interval modules, a new set of 100,000 random queries were performed with k random intervals of the same sizes as the actual intervals. Probabilities were estimated in the same way as before but now based on gene counts from k intervals where k ranged from 3 to 5.