Our goals are to exploit the functional coherence of genes involved in a given disease to identify genes that underlie diverse disorders, and to find previously unknown links between phenotypically dissimilar human diseases. We pursue these goals by first integrating genomic features from disparate data sources to establish quantitative functional links among human genes. Since each data source usually characterizes only one type of functional association between genes, and covers a relatively limited set of genes, functional associations from various sources need to be combined to attain maximal coverage and accuracy. We systematically assemble a set of 16 genomic features, which incorporates 32 sub-features. These genomic features include diverse functional genomics data in human, as well as functional associations mapped through orthology from five model organisms (yeast, worm, fly, mouse, and rat; Table 1).

We then use a naïve Bayes classifier to compute functional links between human genes by integrating these genomic features. Each functional link is weighted by a log likelihood ratio (LR) score, which reflects the probability of the linked gene pair sharing the same biological process after summing over evidence from all available data sources (see Materials and methods). Such extensive data integration outperforms individual data sources in terms of inferring functional linkages (Figure 2), demonstrating the importance of data integration.

After data integration, we choose a permissive linkage weight cutoff (LR score is higher than 1; Equation 1 in the Materials and methods) such that two genes are linked if the overall evidence supports the functional linkage. This threshold is very intuitive, as it retains edges with more evidence for functional association than against it, and removes edges with more evidence against functional association than for it. In addition, the same cutoff has also been successfully used by Lee et al. 41 to predict perturbation phenotypes of genes based on an integrated FLN in worm. The resulting genome-scale FLN network consists of 21,657 genes (covering approximately 85% of RefSeq annotated human genes 42) and 22,388,609 weighted links (Additional data file 1). Despite its high coverage, our network retains its accuracy because each link is weighted and the linkage weight is proportional to the linkage precision (Figure 2). The average number of linked neighbours per gene is around 2,000. Such high linkage density together with linkage weighting allows a quantification of functional associations between thousands of genes. Such high coverage is critical to the successful utilization of the FLN for both disease gene prioritization and mapping the disease-disease associations at the molecular level.

Given a network of functionally linked genes, our first goal is to use the information in this network to identify genes most likely to be associated with a particular disease. The motivation for using the FLN to identify potentially disease-related-genes is the hypothesis that genes whose dysfunction contributes to a disease phenotype tend to be functionally related 12345678910111213141516. Our approach exploits this concept by using genes known to be associated with a particular disease as network 'seeds', and identifying those genes whose connectivity with the seeds indicates a strong functional relation. In particular, for a given disease, each gene in the network is prioritized according to the sum of the weights of its network links to the known disease (seed) genes (Equation 4; see Materials and methods) 4143. This prioritization rule is referred to as neighbourhood weighting.

To test this approach, we first extract from the OMIM database 40 1,025 known disease genes, and assemble them into 110 seed gene sets, representing 110 disorders covering a wide spectrum of human disease phenotypes (Additional data file 2). Each seed set contains at least 5 genes, and the average seed count is 11 (with some seed genes associated with more than one disease). Next we use the FLN to identify new candidate disease genes for each of the 110 diseases, based on the neighbourhood weighting rule 414344. As a result, on average, nearly half of the genome is prioritized for each disease. In Additional data file 3, we list the top 100 ranked new candidate disease genes for each disease. To help investigators estimate the prediction precision at a particular rank cutoff, for each disease we provide a plot of precision estimate versus different rank cutoffs (see Materials and methods; Additional data file 4). We assess the performance of our FLN-based disease gene prediction method using leave-one-out cross validation, with so-called disease-centric 4143 and gene-centric approaches 1923 (see Materials and methods).

Next, we investigate the difference in prioritization performance between monogenic diseases and complex diseases. It is potentially important to distinguish these two classes of diseases, as complex diseases tend to be caused by dysfunctions in multiple biological processes, and this lack of functional coherence may reduce the utility of the FLN in predicting novel disease genes. Kohler et al. 19 published a gene-disease association benchmark dataset that explicitly separates monogenic diseases (83 diseases), polygenic diseases (12 diseases) and cancers (12 diseases). We adopt their categorization so that we can evaluate the three disease groups separately. In particular, using the gene-centric evaluation (see Materials and methods), we evaluate how well each known disease gene is ranked relative to a background gene set in a leave-one-out cross-validation (see Materials and methods). The background gene set is composed of the 100 nearest genes flanking the test disease gene physically on the chromosome. As expected, the results are best for monogeneic diseases (Figure S4 in Additional data file 5). The lower performance for complex diseases and cancers is not surprising, since disease gene prediction methods are based on the assumption of functional coherence among genes contributing to the same disease, while as mentioned above, complex diseases tend to perturb multiple biological processes, making the contributing genes less functionally coherent. Despite the lower performance for complex disorders, the results are still far better than random control; for example, 65% of tested disease genes ranked in the top 20 among the background gene set composed of 100 genes.

Disease genes have previously been prioritized using network-based strategies that used only protein-protein interactions (PPI) 202123. Here we have integrated multiple data sources with the expectation that such integration will improve performance. To assess whether this is in fact the case, we compare disease-prioritization performances between our FLN and a PPI network that combines human PPI links from seven major curated PPI databases 45464748495051, along with high-throughput PPI data from yeast two-hybrid and mass spectrometry 525354, and interactions mapped from PPI of other model organisms 55. To avoid bias, interactions from different sources in the PPI network are weighted using the same procedure as FLN construction (Equation S1 in Additional data file 5). In the end, a total of 105,361 interactions among 11,886 genes are included in the PPI network.

As expected, data integration does improve performance (Figure 5). Using the gene-centric evaluation with the artificial chromosome region background, 62% of disease genes rank in the top 10 among 100 using the integrated FLN (excluding text miming), in contrast to 40% in the PPI network. Similar results were also found using the disease-centric assessment (Figure S5a in Additional data file 5; see Materials and methods for the description of disease-centric assessment). Further support for using an FLN-based approach is the increased gene coverage. In the PPI network only 40% of disease genes are connected to seed genes, and thus only 40% can be prioritized. In contrast, in the integrated network more than 92% of disease genes are linked to seeds and can, therefore, be prioritized. Finally, the benefit of data integration is also evident when we evaluate the prioritization performance of the FLN at different linkage weight cutoffs. After the application of the permissive linkage weight cutoff (LR > 1; Equation 1), we explore other higher cutoffs but find no improvement in the prioritization performance (Figure S5b, c in Additional data file 5). This further demonstrates that functional links are assigned proper weights after data integration, and that the neighbourhood weighting decision rule (Equation 4) allows links with lower weights to contribute to performance.

The performance evaluations described above are based on leave-one-out cross-validation. Here we evaluate the predictive performance for unknown disease genes by simulating the search for new disease genes. We first manually check the date of the landmark reference for each gene-disease association recorded in the OMIM database. Next, disease genes with references published after January 2007 are set aside for testing, while all other disease genes with reference dates before 2007 are used for seed genes. For the purpose of this evaluation we included text mining from the STRING database, which was curated before January 2007 56.

Within the FLN, there are a total of 61 disease genes associated with 31 diseases that are published after January 2007. These are used for evaluating the FLN-based predictions. Among them, 45 disease genes associated with 24 diseases are also present in the representative PPI network. These genes are used for evaluating PPI-based predictions. These recently identified disease genes and their landmark references are listed in Additional data file 6.

Again, the FLN shows improvement over the PPI network (Figure 6). For instance, using the gene-centric evaluation with the artificial chromosome region background, 45% of disease genes are ranked in the top 5 in the FLN, in contrast to fewer than 25% in the PPI network. It is noteworthy that there is a drop in performance for both PPI and FLN relative to the cross-validation analysis presented above. In particular, the fold enrichment drops from 16 to 8.2 for PPI and from 35.5 to 16.5 for FLN. This indicates that cross-validation tends to overestimate performance, and it is important to consider this when interpreting cross-validation results.

Obesity is a polygenic disorder involving genes from various processes, such as nutrient catabolism and appetite control 57. Our FLN includes 24 obesity-associated genes in the OMIM database, and 334 additional obesity-associated genes collected from the literature by Hancock et al. 58. We will subsequently refer to this set of 334 genes as 'ObesHancock' genes. There is no overlap between the 24 OMIM obesity genes and the 334 ObesHancock genes. Here, we rank obesity-related genes using the 24 OMIM genes as seeds, then evaluate the utility of our ranking using the non-overlapping set of ObesHancock genes. Since ObesHancock genes are collected from the literature, we exclude the text mining data source from FLN construction.

We find that the ObesHanecok set is overrepresented in the top scoring FLN genes, with 22 of them occurring in the top 100 (P < 1.0 × 10^-13; see Additional data file 5 for P-value calculation). The list of the 22 ObesHancock genes and their supporting evidence are provided in Additional data file 7. Detailed analysis of the subset of the top 100 ranked obesity genes that does not overlap with the ObesHancock set reveals additional genes with potential roles in obesity. For instance, NR1H3, ranked 24th, is predominantly expressed in adipose tissue and plays an important role in cholesterol, lipid, and carbohydrate metabolism 5960616263. Recently, Dahlman et al. 64 also found that one NR1H3 single nucleotide polymorphism (SNP), rs2279238, is associated with the obesity phenotype. Similarly, NMUR2, ranked 35th, is exclusively expressed in the central nervous system as a receptor for neuromedin U, a neuropeptide regulating feeding behavior and body weight 65. Additionally, Schmolz et al. 66 found that a NMUR2 variant potentially related to obesity in a mouse model.

As described in the Introduction, human diseases tend to form an interrelated landscape. We hypothesize that the basis for these relationships stems from multiple diseases resulting from dysfunctions in the same genes, and more broadly, multiple diseases resulting from dysfunction of the same or related biological processes 1234. Associations between diseases potentially stemming from common causal genes were previously reported by Goh et al. 4. Here we focus on quantifying associations between diseases based on perturbation in common biological processes by developing the concept of 'mutual predictability' (see Materials and methods). The mutual predictability between two diseases measures the extent to which genes known to be associated with either member of a disease pair can be used to identify genes known to be associated with the other member (see Materials and methods). We hypothesize that disease pairs with high mutual predictability will be closely related to each other, as a high mutual predictability should be indicative of high connectivity in the FLN between the two gene sets associated with two diseases, and hence should quantify the functional relatedness between diseases.

We validate our mutual-predictability-based disease-disease associations at the molecular and gene network level, using disease-disease associations based on the classification in Goh et al. 4, where the diseases in OMIM were manually partitioned into 22 classes based on physiological system-level phenotypic observations. After calculating the mutual predictability (Equation 7) between every possible disease pair (all mutual predictability scores are provided in Additional data file 8), we threshold pair selection with increasing score cutoffs. At each cutoff, we examine the fraction of disease pairs belonging to the same disease class (excluding 'unclassified class' and 'multiple class'; the former has insufficient information for disease class assignment, and the latter lacks physiological system specificity). As seen in Figure 7, the fraction of pairs placed in the categories defined by Goh et al. increases rapidly with increasing score cutoff. This demonstrates that FLN-based mutual predictability can capture disease-disease association in a quantitative way.

To visualize disease-disease association estimated by mutual predictability, we create a network of disease associations, in which the nodes represent individual diseases and the weighted edges represent mutual predictability. Figure 8a shows a high confidence subset of the disease network obtained by selecting the top 100 pairs (out of 5,995, that is, the top 1.7%; Figure S9 of Additional data file 5), corresponding to a mutual predictability cutoff of approximately 0.85. These 100 pairs cover a total of 66 diseases. At this cutoff, the disease pairs are four times more likely to share the same disease class than expected at random (Figure 7). Moreover, 97 of the 100 disease pairs are supported by various types of evidence, such as the classification scheme of Goh et al. (within the same disease class) or other literature evidence (Additional data file 9). These results suggest that disease pairs with high mutual predictability tend to be related.

We compare our method with another available disease-disease association identification method proposed by Goh et al., which identifies the associations between two diseases by counting their overlapping disease genes 4. However, Goh et al.'s method could only identify the associations between diseases with known overlapping disease genes. In contrast, our method is able to identify additional associations for those diseases sharing no known disease genes but having dense functional links between their corresponding disease gene sets by taking advantage of the FLN.

Among the 97 potentially related disease pairs with literature support, 48 pairs share known disease genes and are identified by both methods (pairs connected by blue links in Figure 8a). However, the remaining 49 pairs share no known disease genes, and their associations are identified solely based upon functional links among associated genes (pairs connected by red links in Figure 8a). An example of a non-trivial disease-disease linkage in the latter group is the association between Alzheimer's disease, a neurological disorder, and hypercholesterolemia, a metabolic disorder. Since the two diseases share no disease genes in OMIM, their predicted association is based entirely on the strong and dense functional links between the corresponding disease gene sets (Figure 8b). Importantly, associations between diseases identified using the FLN provide immediate insight into the molecular mechanisms underlying different diseases, and thus generate novel hypotheses for therapeutic strategies. For instance, based on the association of hypercholesterolemia and Alzheimer's diseases, we propose that high cholesterol may play an important role in the development of Alzheimer's disease and that modulation of cholesterol levels might help to reduce or delay the risk of Alzheimer's disease, which is indeed supported by recent literature 676869. Besides Alzheimer's disease and hypercholesterolemia, there are diverse disease-disease associations that are identified only by the FLN but not by the disease gene sharing method. These include night blindness/Leber's congenital amaurosis, which are both ophthalmological; pseudohypoaldosteronism/Bartter syndrome - both involved in ion transport deficiency); and holoprosencephaly/Waardenburg syndrome - both involved in developmental deficiencies (Additional data file 9). We provide a more quantitative comparison between our mutual predictability method and disease gene sharing method in Additional data file 5.

Since our disease-disease association identified at the molecular level correlates with disease-disease associations based on phenotypic level classification (Figure 7), it is not surprising that some diseases in the same disease class are found to be connected in our disease network. For example, prostate cancer and ovarian cancer both belong to the cancer class, and are connected in our network. Potentially of more interest is the observation that among the 97 potentially associated disease pairs identified by high mutual predictability and supported by the literature, 54 disease pairs belong to different disease classes, as defined by Goh et al. 4. This finding suggests that diseases with different phenotypes can share common etiology at the molecular level 12. Examples of such disease pairs include Alzheimer's disease (neurological class) and hypercholesterolemia (metabolic class) described above, central hypoventilation syndrome (respiratory class) and Hirschsprung disease (gastrointestinal class), both of which are potentially caused by a defective developmental process 2, and ceroid-lipofuscinosis (neurological class) and glaucoma (ophthamological class), both of which are involved in abnormalities in lysosomal function 7071 (Additional data file 9). These results suggest that standard disease classifications may be much less informative than currently thought.

In addition to validating individual links, we visualize the topology of our disease-disease association network based on the network connectivity. We find that some network modules (such as those defined by rectangles in Figure 8a) can be readily identified through visual inspection. Specifically, diseases within a module tend to be more connected than diseases between different network modules. These disease modules tend to be enriched with homogeneous diseases belonging to the same or related classes, or sharing similar molecular mechanisms (Figure 8a, homogeneous modules are colored correspondingly; cluster details are listed in Additional data file 10 and Figure S9 in Additional data file 5). Examples are the ophthalmological disease clique of night blindness/Leber's congenital amaurosis/retinitis pigmentosa/cone or cone-rod dystrophy, and the mitochondrion deficiency clique of mitochondrial complex I/II/III deficiency/combined oxidative phosphorylation deficiency/Leigh syndrome/mitochondrial DNA depletion syndrome. These results confirm that the FLN-based mutual predictability measure can be used to identify biologically meaningful disease-disease associations.

We have demonstrated that our FLN-based disease gene prioritization approach successfully predicts new candidate disease genes for diverse diseases. The method compares favorably to previous methods, providing 50% more coverage than 19 with accuracy comparable to 192472. (Detailed descriptions of the performance comparisons are provided in Additional data file 5.)

A potential strategy for further improving on our predictive performance is to consider not only functional linkages between neighbors as was done here, but to also incorporate global network topology into disease gene identification 19. It should be noted, however, that the use of a global propagation scheme should be done with extra care. A global propagation scheme may more comprehensively exploit the information contained in the FLN by considering non-adjacent nodes, but at the same time it can be more susceptible to noise. Such issues are exemplified in recent work by Wu et al. 23, where they compared a direct interaction based local rule and a shortest path rule, based on their abilities to prioritize disease genes. Surprisingly, the shortest path rule, which uses global network topological information, performs worse than the direct interaction rule 23. A likely reason is that although, in principle, rules utilizing global network topological information are preferred, they are more sensitive to the false positive and false negative links in the network. In practice, however, a properly implemented local rule can perform well, when combined with a comprehensive and weighted FLN.

Among the data sources used for constructing the integrated FLN, literature-based text mining, GO annotation (GO annotations in the molecular function and cellular component categories are included in the integration), and curated PPI might be biased towards well characterized genes (Table 1). Similarly, text mining and curated PPI data mapped from mouse and rat (organisms close to human in the evolutionary tree) via orthology may also have the same bias. Consequently, predictions based on these sources tend to be biased towards well characterized genes. On the other hand, the remaining data sources are relatively unbiased, and include sequence data, high-throughput experiments, and functional associations mapped from yeast, worm, and fly (organisms distant to human in the evolutionary tree). Predictions from these latter sources can reveal new disease genes that are not well characterized. To test the extent to which our predictions depend on the relatively biased data sources, we remove them and use the remaining unbiased sources for FLN construction and disease gene prioritization. The prediction performance remains far better than random control (Figure S6 in Additional data file 5). Using the gene-centric assessment, over 40% of disease genes are still ranked within the top 10 positions in the artificial chromosome region background composed of 100 genes. The disease-centric assessment also shows similar results. Therefore, our predictions do not solely rely on the potentially biased text mining, molecular function and cellular component annotations of GO, and curated PPI data sources; other relatively unbiased sources also make significant contributions to the disease gene prediction and biological insights that go beyond well characterized genes.

Our current framework assumes that known disease genes ('seed genes') contribute equally to the disease, but in reality some seed genes may contribute more than others, especially for complex diseases. An example is homozygosity in the APOE4 allele, which confers a 40% chance of late-onset Alzheimer's disease 73. Incorporation of this notion of 'seed strength' into our framework would be desirable. In addition, our framework requires the knowledge of a number of known disease genes as seed genes in order to predict new candidate genes. A further improvement might be to borrow seed genes from closely related diseases when the disease under study does not have enough seeds, adjusting the strength in accordance with how closely related the two diseases are.

Additionally, FLN has been made available in VisANT 74, a web-based platform for the integrative visualization and analysis of biological networks and pathways 757677787980. Users can interactively query the FLN for genes of interest by choosing FLN as the current database in the VisANT toolbar; compare the FLN with other interaction data by updating gene interactions stored in the Predictome database; and filter the FLN with preferred weights and visualize the weight using edge color, edge thickness or both. In addition, we are developing a series of functions to integrate the methods of FLN-based prediction of new disease genes and disease-disease associations into VisANT, including making predictions for the new seed sets provided by users. Each disease will be represented as a metanode 78 - a special type of node that contains associated sub-nodes (disease genes), which allows direct integration of disease information and the FLN. These new functions will extend the applications developed here to other areas, such as the prediction of new proteins for protein complexes, cross-talk between pathways and so on.

Here, we carry out a proof-of-principle study with our mutual predictability measure to demonstrate that the functional links provided by the FLN can be used to quantify disease-disease associations at the molecular level as illustrated in the Results section. Further studies are needed to improve our measures as well as to explore related measures of disease-disease associations based on the FLN. Our approach for assessing disease-disease associations requires the knowledge of a few known disease genes for each disease, and cannot be applied to diseases with no known disease genes. In these cases, phenotype-based methods such as 3 are good alternatives if there is enough disease phenotype information available, bearing in mind that it is possible for two diseases to share related molecular pathology mechanisms but be less similar at the phenotypic level 12. In general, it is likely that the identification of disease-disease associations can be further improved by combing our FLN-based approach with phenotype-based approaches.

All methods are applied to a common set of 25,564 protein encoding genes (downloaded from RefSeq in November 2007 42) with the Entrez gene ID as the unique identifier 81. For data sources using different gene or protein identifiers, the original identifiers are all mapped to Entrez gene IDs using the cross-reference files from Entrez Gene 81, the International Protein Index (IPI) database 82, or Biomart 83.

For a given disease, the known disease associated genes are used as seeds. New candidate genes can then be ranked based on their association with these seed genes in the FLN.

Given two diseases, I and II, each of which is associated with a set of known disease genes, we define a mutual predictability score that reflects the degree of association between the two diseases (Figure S7 in Additional data file 5). The mutual predictability score accounts for both the number of genes in common between the two disease gene sets, and the functional links between the two sets in the FLN. In essence, the mutual predictability score evaluates the degree to which genes associated with disease I can be used as seed genes to predict genes associated with disease II (predictability _I-II), and vice versa (predictability _II-I). High mutual predictability implies close association between the two disease gene sets in the FLN, suggesting related molecular mechanisms.

To quantify predictability _I-II, we first use genes associated with disease I as seeds, and rank all other genes based on the S_i score (Equation 4). In addition, disease II genes that overlap with the seed genes from disease I are given a S_i score of infinity and ranked on the top. Next, we plot a ROC curve of sensitivity versus 1 - specificity on how well the disease II genes are ranked in the sorted gene list. Sensitivity and (1 - specificity) are defined in Equations 5 and 6, where TP is the number of disease II genes above a particular S_i cutoff, TN is the number of genes associated with neither disease below the cutoff, FP is the number of genes associated with neither disease above the cutoff, and FN is the number of disease II genes below the cutoff. Finally, we calculate the AUC of the ROC curve (AUC_I-II) as a measure for predictability _I-II.

Similarly, we can calculate AUC_II-I as a measure of predictability_II-I. The mutual predictability between disease I and II is defined as the geometric mean of AUC_I-II and AUC_II-I:

Since AUC ranges from 0 to 1, the mutual predictability also ranges from 0 to 1.