We used Ensembl (version 39) as the annotation reference database. Homology between human and mouse genes was derived via BioMart. The total number of genes under study comprises 15,277 Ensembl mouse genes representing the union of the homologue genes from all data sources. An overview about the T2DM specific datasets is given in Table 1.

Several of the available resources are based on microarrays. Each individual microarray study was normalized using the GC RMA method of the R/BioConductor package 33 34 36 . Recently, it has been reported that the remapping of the oligoprobes, as an alternative to the annotation given by the chip provider, enhances data analysis to a significant extent 37 . Using this mapping we indeed observe differences in annotation resulting from recent changes in genome annotation. Annotation comparisons showed improvement of reproducibility and specificity (data not shown). Re-annotation affects the main fraction of genes, for example in the remapped Mouse Genome U74A version 2 platform the top ten genes of our candidate list are represented by oligoprobe sets of 8 to 15 probes (11 with annotation of the chip provider) with an average of 13 probes (Serpina1a 11 probes, Ybx1 8, Pdk4 16, Cstb 14, Adipoq 15, Agt 14, Lgals1 11, Serpine2 15, Mt2 10, Gpi1 16 probes).

Numerical scores were computed for all genes in each individual study, the scores were summarized and the summarized scores were compared against a random sample at the 99.9 percentile as described in Materials & Methods. This procedure determines a cut-off score value of 3.05 and identifies a set of 213 genes with a score exceeding this cut-off.

Randomly, we would expect 15 out of the 15,277 genes to exceed the threshold. Cutting at the 99 percentile results in 943 genes (expecting 153 by chance), cutting at the 98, 97, 96 and 95 percentiles would result in 1352, 1587, 1792 and 1972 selected genes (305, 458, 611 and 764 randomly expected genes). Thus, the ratio of detected vs. expected significant genes increases with percentile of the random sample from 2.6 to 14.2, indicating a necessary precondition for the validity of our selection procedure [see Figure 1 in Additional file 1]. Since we have analyzed data from multiple tissues in human and mouse, it is likely that for some cases an individual experiment is dominating the score, for example, if the gene is active only in a single tissue. In order to identify those genes we have computed an entropy-based numerical criterion (see Materials & Methods). Entropy is high if many experiments contribute equally to the overall score, it is low if a single (or few) experiment accounts for a large fraction of the score. For example, the gene Serpina1b has the top score (7.9, rank 1/15,277) in our study. This is due to a very high fold-change in a single experiment; consequently, entropy is low (1.17, rank 4,590/15,277). In contrast, other genes show more consistent alterations across many different studies, for example Pdk4 (6.7, rank 3/15,277) indicated by higher entropy (3.0, 167/15,277). Differential expression of Pdk4, a major regulator of glucose metabolism, has been found in fat, pancreatic islets and skeletal muscle but not in liver.

The thirty genes with highest scores are listed in Table 2 [The entire candidate list is given in Additional file 2].

Adiponectin (Adipoq) has a score of 6.1 (rank 5/15,277) and the high entropy indicates a consistent behavior across data sets. Adipoq is a hormone from adipocytes that modulates insulin sensitivity and thus regulates glucose and lipid metabolism and energy homeostasis. Expression of Adipoq is reduced in obesity, certain genotypes are associated with increased risk of T2DM in humans 38 . The protein is secreted from fat tissue and has insulin-sensitising and anti-inflammatory properties. Additionally, we find strong changes in the expression in muscle both in human and mice. Adipoq is an oxidative regulator. The systemic oxidative stress causes the metabolism to share the burden from fat to muscle 39 . Adipoq is also responsible for the crosstalk between the three KEGG pathways 'PPAR signaling', 'Adipocytokine signaling' and 'type II diabetes mellitus'. It has been tested for transcriptional regulation but no binding to the TFs under study could be detected. A negative regulation has been described for Tnf that has not been tested in the underlying studies.

Pdk4 phosphorylates and inhibits pyruvate dehydrogenase complex thereby contributing to the regulation of glucose metabolism. Expression of this gene is regulated by glucocorticoids, retinoic acid and insulin. This is in accordance with a consistent differential expression in fat, muscle and pancreatic islets resulting in high entropy. On the other hand, possible regulation of the gene by Hnf4a and Usf1 is reported in liver.

Hydroxysteroid (11-beta) dehydrogenase 1 (Hsd11b1, score 5.5, rank 13/15,277) is a critical enzyme for cortisol metabolism. Hsd11b1 is increased in obese subjects and transgenic mice over-expressing Hsd11b1 develop visceral obesity 40 . Inhibition of Hsd11b1 decreases blood glucose in hyperglycaemic mice. Selective antagonists are currently developed and tested as anti-obesity and anti-diabetes drugs.

Scd1 (5.6, 11/15,277) is the rate-limiting enzyme in monounsaturated fatty acid synthesis. It has been shown to exert a critical role in hepatic lipogenesis and lipid oxidation. Scd1 knock-out mice are lean due to increased energy expenditure, show increased insulin sensitivity and are resistant to diet-induced obesity and liver steatosis.

Nicotinamide nucleotide transhydrogenase (Nnt, 5.5, 12/15,277) is a mitochondrial enzyme involved in proton transport into the mitochondrial matrix. Nnt was identified as a novel candidate gene in a quantitative trait locus for glucose intolerance 41 . Nnt has been recently shown to regulate insulin secretion in pancreatic beta cells. Nnt deficiency results in defective insulin secretion and inappropriate glucose homeostasis 42 . It has been proposed that Nnt detoxifies reactive oxygen species 43 implicating a possible role of Nnt in regulating ATP production in mitochondria and function of the ATP sensitive K+ channel Kir6.2 (Kcnj11) in insulin producing beta cells.

For eighteen genes only limited functional information is available as a basis for assessing a possible relationship to T2DM: Ccrn4l, Serpina12, Htatip2, Mest, Pgcp, Tmsb4x, Angptl4, Mrpl33, Ndfip1, Yipf5, Tmem30a, Asnsd1, Oact5, Larp5, Thrsp, 1810015C04Rik, 2310003F16Rik, and 2610002J02Rik. High genetic variation is known for Pgcp in mouse. Serpina12, a target of Hnf4a, is massively changed in liver and 1810015C04Rik in pancreatic islets. Using the entropy criterion we observe medium to high entropy in these genes, like in Ndfip1 (entropy 2.9), what points to the fact that high scoring of these genes was not due to single outlier experiments but that these genes are truly affected by the disease and, thus, exhibit a high potential for further functional experiments.

In order to allow users to screen the disease potential of any given gene of interest we developed T2DM-GeneMiner, a web interface summarizing the results of our work (Figure 1, 35 ). The user interface is shown for the well-known Adipoq and the resulting bar plots for two other genes, Pdk4 and Cfd, with lower content of available information. The resource is searchable by gene or protein IDs (for example Ensembl ID or gene symbol). The score distribution is shown as a bar plot and, where available, functional information is displayed. The two rightmost bars show the entropy, indicating uniform or specific score distribution, and the score. The red line at the score bar indicates the cut-off.

From fourteen genes in the OMIM description of T2DM (Diabetes mellitus, noninsulin dependent, #125853, 44 ) five genes have a significant score in our study: Retn, Gpd2, Vegfa, Irs2 and Tcf2 (see Table 3). Retn represents an adipocytokine which has been implied to play roles in obesity, diabetes, and insulin resistance 45 46 . Interestingly, Retn is only deregulated in one of two studies involving adipose tissue. In contrast, differential expression for Irs2, Vegfa and Tcf2 was observed in pancreatic islets whereas Gpd2 did not show tissue-specific expression. Several previous studies have already published T2DM candidate lists allowing us to assess common content. The overlap to the list of the Diabetes Genome Anatomy Project 23 , being also the source of some of the transcriptome data sets used for this meta-analysis 12 13 14 , results in a P-value of 9.9E-03. Using the same resource, with a less conservative selection of data sets, Liu et al. identified 82 genes related to insulin signaling with an overlap of seven genes to our candidate list containing several strongly connected proteins (see below) 24 . More selective is a review of sequencing candidates leading to a P-value of 5.28E-13 47 . In Tiffin et al. 99 candidates were published as partial overlap of several electronic candidate prediction methods 22 . This results in a P-value of 1.9E-05 comparing it with our list (Figure 2 shows a Venn diagram of the absolute gene numbers). In summary, the T2DM candidate gene list includes a small amount of candidate genes from previous studies and, further, leads to an additional set of 191 genes not identified in the other studies. Subtracting those genes for which we have disease information from the incorporated reviews our approach identifies 128 novel T2DM candidate genes.

A variety of genetic studies have been performed in the last decades. At least nineteen genetically engineered mouse models with T2DM phenotype have been studied in detail 2 3 . Of those, five genes show a significant score in our meta-analysis: Slc2a4, Irs2, Ptpn1, Slc2a2 and Irs1. Consistent with previous reports, the insulin-regulated glucose transporter GLUT4 (Slc2a4) is down-regulated in the insulin resistant state in adipose tissue but not in skeletal muscle. Likewise, down-regulation of Irs2 and the glucose transporter GLUT2 (Slc2a2) in pancreatic islets confirms previous reports and reflects deterioration of beta cell function in the course of insulin resistance and diabetes.

On the other side Slc2a2 is also changed in liver. Ptpn1 is expressed in all tissues showing only small fold-changes. Several genes from OMIM or KO-mice do not change at all on the expression level. This indicates that only the complete loss of the associated protein alters the system whereas the gene's expression is not altered in T2DM. For KO-mice we also see a strong tendency to genes only expressed in mice.

Recently, a total of nine candidate genes for T2DM have been identified and replicated in humans through multiple genome-wide association studies of common variants by using high-density SNP mapping approaches: Cdkal (score 0),Cdkn2a (score 0)/Cdkn2b (1.925, 1165/15277), Fto (1.798, 1364/15277), Hhex (1.213, 2456/15277), Igf2bp2 (0.855, 3555/15277), Kcnj11 (2, 1056/15277), Pparg (2.528, 500/15277), Slc30a8 (0.076, 8056/15277), and Tcf7l2 (2.192, 811/15277) 5 6 7 8 9 10 11 . Interestingly, none of these genes shows a high score in our meta-analysis, although Pparg and Tcf7l2 are significant on the less restrictive 0.01 level. On the other hand, from the data we could infer that Fto and Hhex act in pancreatic islets indicated by the T2DM-GeneMiner result for these genes. Cdkal1 and Cdkn2a are not expressed in the transcriptional studies. These genes show very low expression levels or might be active in tissues not included in our study. Since our meta-analysis approach takes into account several data sets from DNA microarrays, our candidate genes have a bias towards transcripts whose expression is changed in the context of T2DM. Moreover, the gene variants from association studies may not result in altered gene expression and, for most SNPs found in association studies, there is a lack of functional information since the variation mostly occurs in non-coding regions of the genes. In order to correlate the T2DM genes with genetic variation we plotted the number of known SNPs for the genes [see Figure 2 in Additional file 1]. No general tendency to highly variable genes is observable. Two genes of the candidate list show high variation, Pgcp (9,098 SNPs) and Sorbs1 (4,130). Particularly interesting is Pgcp, because it has not been related to T2DM before and its functional role is also undetermined.

A further issue of our study was the chromosomal localization of the T2DM genes. The marker genes are scattered over the entire mouse genome [see Figure 3 in Additional file 1]. Using the hypergeometric distribution on local sliding windows across the chromosome we could identify significantly enriched chromosomal regions. However, none of these regions convinced since they are sparsely occupied. For example on chromosome 2 with 2,249 genes and 15 T2DM candidates, a window of 20 genes containing two T2DM candidates leads to a P-value of 0.007. None of the windows of 20 genes contained more than three candidates. Rather conversely, we observe that T2DM affects a wide range of physiological phenomena spanning loci in the entire genome.

Enrichment analyses based on the hypergeometric distribution were carried out in order to assess whether our T2DM candidate list is over-represented with respect to a certain functional category (Table 4). Categories on the physiological level comprise three major pathway resources (KEGG, Reactome, BioCyc) 28 29 30 and the GO tree 27 . Altogether, we have analyzed 4,555 gene sets, whereof 314 (6.9%) are significant with a P-value below 0.05 [see Additional file 3].

As greater parts of the metabolism are affected by T2DM, multiple pathways have a significant enrichment P-value. For example, in KEGG 45 out of 182 pathways have a P-value lower than 0.05. Table 5 shows the results of pathways with a P-value lower than 1.0E-04. Results for different pathways are not independent. For example, the 136 genes annotated with 'Insulin signaling pathway' and the 46 genes annotated with 'type II diabetes mellitus' share 32 genes. The first four pathways help to validate our significant gene set. 'PPAR signaling', 'Adipocytokine signaling' and 'Insulin signaling pathway' are well related to T2DM.

Since we used several pathway resources in parallel, we can compare the findings for consistency, assuming the resources are independent. For example, we found enrichment of the KEGG pathway 'Fatty acid metabolism' what is complemented by the BioCyc pathways 'fatty acid elongation – saturated', 'fatty acid elongation – unsaturated' and by the GO categories 'positive regulation of fatty acid biosynthesis', 'positive regulation of fatty acid metabolism', 'fatty acid binding' and 'fatty acid oxidation'. The KEGG pathway 'Complement and coagulation cascades' is complemented by the Reactome pathways 'Initial triggering of complement', 'Complement cascade' and the GO categories 'defense response' and 'complement activation, alternative pathway'.

For 116 T2DM gene candidates there is information on the associated biochemical pathways according to the KEGG database. Whereas most genes (106) are associated with a single or a few (up to five) pathways, some genes exhibit a higher interconnection such as Mapk1 (22 pathways), Pik3r1 (19), Aldh9a1 (15), Mapk9 (11), Sh3glb1 (9), Pla2q12a (9), Pkm2, Nfkbia, Dhrs7, Actb (all 6). The importance of Mapk1, Pik3r1, Rasa1 and Socs2 is also supported by Liu et al. as members of an insulin signaling subnet derived from protein-protein-interactions 24 .

In order to identify crosstalk between pathways we scored the pathways according to their counts and their overlap in the T2DM candidate list (Figure 3). An important module with a large number of genes represented in the T2DM list and a high overlap is visible with the pathways "Insulin signaling", "Type II diabetes mellitus", "PPAR signaling", "Adipocytokine signaling", and "Fatty acid metabolism" pointing to the interplay between obesity and insulin resistance. Another path of signaling action is activation of the RAS/RAF/MEK MAPK cascade resulting in cell growth and gene expression alterations expressed by the crosstalk between "Type II diabetes mellitus" and "MAPK signaling" pathways.

Protein-protein interactions have been taken from the IntAct database denoting the number of interactions and interactors registered for the T2DM candidate genes. The ratio of interactors to interactions indicates whether the protein participates in big complexes or binds with single proteins. Figure 4 shows the number of interactions and the score for the genes under study. There is no trend for preferential selection of highly interacting genes in our T2DM candidate list. The high-scored genes comprehend a few genes with many interactions like Mapk1, Pik3r1 and Rela in mouse with more than 15 interactions. The large number of interactions of Mapk1 and Pik3r1 is consistent with their participation in many of the signaling pathways (Figure 3). Actb, Cltb, Hspa5 and Grn have more than 600 interactors, indicating big polymers. In human Tsc22d1, Tnfrsf1b, Ndrq1 and Nme1 have most interactions. Lmna is the only gene with more than 300 interactors.

Mapping the interactions on Ensembl genes and coerce the human net and mouse net we derive a graph with 5,179 nodes and 216,446 edges (data not shown). If we consider the edges between significant genes and their non-significant nearest neighbors we still have 1,471 nodes and 11,378 edges. This shows that the disease genes strongly interact with main physiological triggers and deregulate essential parts of the metabolic network. Reducing the interactions on the 213 T2DM genes we end up with 45 nodes and 167 edges [see Figure 4 in Additional file 1].

In order to study the information content of the set of selected disease genes on gene regulation, we have analyzed a) the TFs present in our significant set and b) known target sets of TFs for enrichment. Analysis is often hampered because TFs are known to be expressed at a very low level and fold changes are commonly low. Moreover, many TFs are regulated by phosphorylation (e.g. Foxa's) and/or ligand binding (e.g. Ppar's). As a result, important core regulators including Onecut1 (score 1.2, rank 2461/15,277), Hnf4a (2.36, 642/15,277), Tcf1 (3, 243/15,277), and Foxa2 (1.4, 2055/15,277) are not in our candidate list.

Collecting TFs from Odom et al. 17 18 , TransFac 31 and the GO category GO:0003700 in mouse and human with evidence codes IC, IMP, TAS or IDA we identify 490 TFs. Thereof 12 TFs received a high score in our T2DM set: Srebf1, Tcf2, Rela, Ybx1, Cebpb, Nr1d2, Klf10, Nfil3, Ccrn4l, Atf3, Nme1 and Drap1. Srebf1 and Ybx1 are expressed only in mouse but in every tissue. Cebp's and Srebp's are important regulators of lipid metabolism and adipogenesis and were found differentially expressed in the course of insulin resistance and T2DM. Consistent changes could be identified in the tissues under study (fat: all but Nfil3; liver: Srebf1, Ccrn4l, Ybx1, Bhlhb2, Klf10, Nme1; muscle: Atf3, Klf10, Nme1, Nfil3; pancreatic islets: Ccrn4l, Atf3, Ybx1, Bhlhb2, Klf10, Nme1, Nfil3). Pparg is expressed solely in fat where its expression is altered. In total, target sets of 187 TFs have been investigated as gene sets for enrichment analysis. Table 6 shows the TFs from Odom et al. 17 with significant P-value. For example, Cebpa is highly significant. It is expressed in adipose tissue and modulates the expression of leptin. Cebpa shows some correlation with the level of hyperglycemia in 16 . Alteration is also observable in liver.

A gene regulatory network comprising the regulatory interactions of the significant genes and the significant and enriched TFs is shown in Figure 5. Obvious are the five hubs, the core regulatory circuit derived from 17 . Well-regulated candidates can be identified like Acly and Fabp4. Target and regulator at the same time is Ipf1.

Data sets were selected from heterogeneous sources that target different levels of cellular information. For each gene and each source we computed a numerical value that expresses its likelihood for being T2DM relevant. Data categories are either binary or quantitative.

Binary data was introduced by incorporating medical reviews, phenotype information (for example from knock-out genes), results from proteome analysis 1 2 3 19 44 48 49 as well as published candidate gene lists from previous studies or models 22 23 47 55 . Binary information was assigned according to the fact whether the gene had been identified in the study or not.

Quantitative data was incorporated by evaluating data from differential gene expression and time series microarray studies 12 13 14 15 16 . In order to extract comparable information across the different studies we used data from the same technological platform (Affymetrix GeneChip studies). Furthermore, in order to conduct standardized data normalization (see below) only studies were taken into account that published and provided the raw data (CEL file level) [see Additional file 5].

Genes do not act as individual units, they collaborate in overlapping pathways, the deregulation of which is a hallmark for the disease under study. In order to integrate pathway information and to derive cellular network information on the selected genes, we added functional annotation from pathway databases such as KEGG, Reactome, BioCyc 28 29 30 , GO 27 , protein-protein interaction databases such as IntAct 56 and databases on transcription factors (TFs) such as TRANSFAC 31 .

Genetic variation of a gene was described with the number of associated SNPs. The number of SNPs in the coding and surrounding region of the gene is noted for mouse and human 57 . A particular biomedical interest is on genes that can be used for drug development. This characteristic has been previously assigned to the gene's ability to provide binding sites for biochemical well-characterized (i.e. druggable) compounds 58 59 . The selected candidates were evaluated with respect to this information. All collected information on the identified 213 T2DM candidate genes and further description of the data sets is given [see Additional file 2].

A central pre-requisite of any meta-analysis approach is the consolidation of the different ID types, for example coming from different organisms and from different versions of chips. We have used the Ensembl database 57 as the backbone annotation for all studies. IDs are mapped on their mouse Ensembl gene ID (version 41). Mapping and merging of the information has been done within R and the BioConductor package collection 33 34 . To ease the access for researchers we have added the more informative MGI marker symbols and HUGO ID's together with ENTREZ gene numbers and RefSeq IDs. In total, information on 15,277 Ensembl annotated genes has been mapped.

Affymetrix gene chip annotations were adapted from latest genome annotations 37 in version 8. Affymetrix data has been normalized with GC RMA using the R/BioConductor software platform 36 . For transcriptome studies that are targeting differential expression three bits of information are stored – the fold-change indicating the alteration of the gene when comparing the diabetic state with the normal state, the standard error of the fold-change computed from the replicated experiments in that study and the expression P-value (presence-call) that indicates whether or not the gene is expressed in the target samples under study. In time series studies we store the correlation between phenotypic characteristics, for example blood glucose, and the gene expression levels with the coefficient of variation and the expression P-value.

In order to score the different categories of information, i.e. binary counts and quantitative gene expression values, for each category we summarized the scores of the individual experiments. For binary information the counts were grouped in sub-categories, for example knock-out mice described in two reviews only get a single count.

For quantitative information, the score of the ith gene in the jth study, s_ij, was computed as follows:

s i j = { | log ⁡ 2 ( r i j ) | ( 1 − e i j r i j ) ( 1 − p i j ) , p i j ≤ 0.1  and  e i j / r i j ≤ 1 0 , else . MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=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@7881@

Here, r_ij is the fold change, p_ij is the average detection P-value and e_ij is the standard error of the ratio derived from the experimental replicates of the study. Thus, the fold change is weighted with its reproducibility across the experimental replicates and with the likelihood of the gene being expressed in the study's target samples. A similar formula applies for correlation studies:

s i j = { | c i j | ( 1 − v i j ) ( 1 − p i j ) , p i j ≤ 0.1  and  v i j ≤ 1 0 , else . MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=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@67F4@

Here, c_ij is the correlation to a certain phenotypic parameter, v_ij the coefficient of variation of the gene's signal across experimental replicas. The formula is applied on the data of Nadler et al. 16 . Mice from three different strains (B6, BTBR and F2 intercrosses) are separated in five classes with increasing hyperglycemia. The Kendall rank correlation between the classes and the gene expression was calculated.

The total score of the gene was computed as the sum across all individual study scores.

In order to assess the significance of the overall gene scores we generated random gene scores. For this bootstrap 26 we draw a random score from each study. The sum of the drawn study scores determines the score for a virtual gene. The distributions of the original scores (black line) and the random scores (blue line) are shown in Figure 6. Using the random distribution as background sample we assigned those genes as "significant" that are above the 99.9 percentile of the background distribution.

For each gene, entropy of the score distribution was computed in order to quantify the relative influence of a certain study (for example a particular tissue) on the overall score. Let s_ij be the score of the ith gene in the jth study, then E_i is a measure for the uniformity of the score distribution over the individual experiments:

E i = − ∑ j s i j ∑ k s i k log ⁡ 2 ( s i j ∑ k s i k ) . MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemyrau0aaSbaaSqaaiabdMgaPbqabaGccqGH9aqpcqGHsisldaaeqbqcfayaamaalaaabaGaem4Cam3aaSbaaeaacqWGPbqAcqWGQbGAaeqaaaqaamaaqafabaGaem4Cam3aaSbaaeaacqWGPbqAcqWGRbWAaeqaaaqaaiabdUgaRbqabiabggHiLdaaaaWcbaGaemOAaOgabeqdcqGHris5aOGagiiBaWMaei4Ba8Maei4zaC2aaSbaaSqaaiabikdaYaqabaGcdaqadaqcfayaamaalaaabaGaem4Cam3aaSbaaeaacqWGPbqAcqWGQbGAaeqaaaqaamaaqafabaGaem4Cam3aaSbaaeaacqWGPbqAcqWGRbWAaeqaaaqaaiabdUgaRbqabiabggHiLdaaaaGccaGLOaGaayzkaaGaeiOla4caaa@5535@

Entropy is low if a single study has a major contribution on the overall score. On the other hand, entropy is high if most of the studies account equally for the score. A plot of the entropy versus the score is given in Figure 5 in Additional file 1.

Disease related networks were investigated with four different types of network information – biological pathways 28 29 30 , protein-protein interaction networks 56 , gene regulatory networks 17 18 31 and functional annotation using GO annotations 27 (see Table 4). These networks define – by annotation – groups of associated genes. The hypergeometric distribution compares the overlap between our superset and the gene group to the overlap of a random selection of two gene sets with the same size 52 . Thus we were able to assign each annotation item (pathway, GO term etc.) a P-value that reflects enriched occurrence of candidate genes. In case of GO terms we include only genes with evidence codes IC, IMP, TAS or IDA to rely on the same confidence level as in the above mentioned resources. We correct P-values for multiple testing using Q-values following Storey for the control of the false discovery rate 60 61 .

The same method we use in the leave-one-out cross-validation. Our qualitative studies are the benchmark for our scoring approach. The scoring, including a notional candidate set, is calculated without the respective qualitative study. The hypergeometric distribution of the qualitative study gene set and the notional candidate set assigns a P-value. This P-value reflects the success of the score to identify the genes from the qualitative study.