Introduction

Despite the recent progress in mapping and sequencing technologies, the identification of genes involved in human diseases remains a very demanding task. Indeed, genome-wide techniques such as linkage analysis, SNP profiling or even deep sequencing of genetically heterogeneous disorders may select hundreds of candidates, whose experimental verification is time- and resource-consuming.1 A deep knowledge of the modular organization of biological functions may significantly increase the efficiency of this identification process. Indeed, biological phenomena are emergent properties of complex interaction networks, composed of proteins, DNA, small molecules and different classes of RNA, capable of self-organizing in discrete functional modules.2, 3 The dissection of the molecular basis of many diseases has evidenced that in most cases abnormal phenotypes are caused by the derangement of an entire module, owing to single gene defects, to a combination of genetic abnormalities or to the interaction between gene variants and environmental factors.3, 4 Thus, when trying to identify the best candidates for a given phenotype or to establish the phenotypic significance of a particular gene variant, it would be very helpful to know whether the genes under study are involved in functional processes directly relevant to that phenotype. In theory, the use of functional gene annotations would represent the most straightforward support for this task. However, although this strategy has been used successfully in many cases,5, 6 it is clearly limited by the lack of complete information about the function of most human genes.

Coexpression relationships derived from microarray data represent an extremely rich and less biased source of information, potentially relevant for functional annotation and disease gene prediction. Indeed, it has been shown extensively that functionally interacting genes tend to show very similar expression profiles, as a result of common regulatory mechanisms.7, 8 Moreover, the probability for two genes to be functionally correlated is remarkably higher when they are strongly coexpressed in more than one species (conserved coexpression).9, 10 Accordingly, the systematic integration of phenotype information with conserved coexpression may allow the efficient prediction of the best positional candidates in wide genomic regions associated to genetic diseases.11, 12 Despite these important results, most coexpression studies reported so far have been performed on heterogeneous data sets covering multiple tissues and organs, thus partly overlooking the complex control of gene expression within specific tissues or organs that is essential for higher eukaryotes. This is a very important shortcoming, because the transcriptional units that compose the human genome show an extremely flexible organization, allowing complex qualitative and quantitative control of gene expression in the different cell types.13

In this report we show that the study of tissue-specific conserved coexpression allows an extended exploration of the transcriptional co-regulation of mammalian genes, with strong implications for their functional annotation. As an example, we identify a cluster of genes potentially implicated in the transcriptional programs of pluripotent stem cells. Moreover, we show that simultaneous use of multi-tissue and tissue-specific conserved-coexpression networks (CCNs), combined with phenome analysis, allows efficient candidate gene prediction. In particular, we analyze the potential of combining our predictive strategy with high-throughput mutational screenings using X-linked mental retardation (XLMR) as a case study. Finally, we provide a user-friendly web resource allowing both access to pre-computed predictions and the execution of custom analysis for functional annotation and disease gene identification.

Methods

Gene expression database annotation and normalization

The gene expression series that we used comprise 5188 human and 2310 mouse microarray experiments, performed on the Affymetrix platforms Human Genome U133 Plus 2.0 and Mouse Genome 430 2.0, respectively, obtained from the Gene Expression Omnibus (GEO). To allow efficient selection of tissue- and condition-specific subsets of the whole database, experiments were manually annotated according to the following scheme: First, the anatomical annotation of the samples was reported to the standard MeSH ontology, by associating every experiment to the most specific keywords that correctly described the sample. Second, we recorded for every sample whether the RNA was derived from whole tissues or from isolated cells. Third, we reported whether the sample corresponded to a normal, tumor-related or other diseased condition. Finally, we reported whether the sample was of adult or embryonic origin.

To avoid the spurious correlation links that may be introduced by the RMA procedure,14 we downloaded CEL files for each experiment and normalized them by using the MAS5 algorithm. This procedure was used as implemented in the software packages available from Bioconductor (http://www.bioconductor.org), using the default parameters.

Generation of CCNs

Generic and tissue-specific CCNs were generated using the procedures described in reference Ala et al.12 Briefly, after selecting corresponding subsets of the human and mouse gene expression datasets, we first generated single-species coexpression networks (SCNs) and then integrated them on the basis of human–mouse orthology. The SCNs were generated by calculating the Pearson correlation coefficients between all the probesets of the expression matrix. A directed edge was established from probeset p1 to probeset p2 if p2 fell within the top 1% of all the correlation values calculated with respect to p1. The directed networks where then converted into undirected SCNs by mapping the probesets to the corresponding Entrez Gene identifiers. Finally, CCNs were built from SCNs by mapping every Entrez Gene identifier to the corresponding homology cluster, defined as the union of the information contained in the Homologene (built 63) and the Ensembl homology (release 53) databases.

Measure of phenotype similarity

To measure the pairwise similarity of disorders (described by independent OMIM phenotype entries) we downloaded the OMIM database on 17 June 2009 and processed it using MimMiner, essentially as described by van Driel et al.15

The MimMiner scores are normalized and range from 0 (unrelated) to 1 (highly related or identical). Since van Driel et al showed that similar phenotypes can be identified with reasonable accuracy considering a minimum score of 0.4, we used the same threshold for our work. Hence, when mentioning ‘similar phenotypes’, if not otherwise stated, we refer to pairs of OMIM phenotype entries that have a MimMiner similarity score of at least 0.4 in our updated database.

Generation of OMIM PCNs

Genes (nodes) in a given CCN were mapped to the disorders (OMIM phenotype IDs) in which they are known to be involved using Entrez (mim2gene, downloaded on 16 June 2009). Each edge in the CCN involving two genes both associated to OMIM phenotypes was translated into a corresponding OMIM–OMIM edge. For genes involved in multiple phenotypes, several corresponding OMIM–OMIM edges could be obtained. For the final phenotype coexpression network (PCN), however, only OMIM–OMIM edges involving two similar phenotypes (MimMiner score ≥0.4; see above) were retained, whereas OMIM–OMIM edges that connect unrelated phenotypes were dropped. Therefore, PCNs can be considered to capture the valuable knowledge contained in CCNs for the purpose of disease gene prediction by means of identification of candidate genes that show a conserved coexpression with genes known to cause similar disorders.

Disease gene prediction

We applied our previously described method for disease gene prediction12 to one multi-tissue and 13 tissue-specific CCNs. Here, we briefly summarize the method.

We define gene clusters in the network, consisting of a given gene plus all its nearest neighbors (ie, all genes that show conserved coexpression with the given gene). Hence each network composed of N genes will yield N gene clusters (one for each gene).

For these clusters we establish associations to OMIM phenotype entries in the following manner: each cluster containing at least two genes known to be involved in a given phenotype p or similar phenotypes (see above) is consider as ‘potentially disease-relevant’ with respect to p.

To identify promising candidate disease genes for phenotypes with a currently unknown molecular basis (OMIM category ‘%’, as of June 2009) but an already mapped gene map locus (eg, by linkage analysis), we determine which (if any) of the candidate genes fall within the respective disease-relevant gene clusters taken from our networks. In other terms, we select the ‘best’ candidates as those that show a conserved coexpression, within a given context depending on the CCN (eg, tissue-specific), with other genes involved in the given or similar phenotypes.

FDR for the disease gene prediction

In order to estimate the quality of our predictions, we performed 20 randomizations of each CCN (keeping the network structure intact and instead randomizing only gene labels) and applied the same candidate selection method as described above. Please note that the false discovery rate (FDR) is computed for each CCN separately, hence there is some limited amount of multiple testing not yet considered by the computed FDRs. Therefore, with the exception of the evaluation of XLMR candidates, we present only predictions that satisfy a stringent FDR of 1%, instead of the 10% threshold used in our previous study.12

The leave-one-out test procedure

To verify the effectiveness of the method in identifying disease genes, we performed a leave-one-out cross-validation on all genes known to be involved in the phenotypes described in OMIM (Entrez mim2gene, downloaded on 16 June 2009), with the exception of those genes that could not be mapped to their precise genome location using Ensembl 55 (http://www.ensembl.org). For each evaluated gene–phenotype pair we constructed an artificial locus, centered around the disease gene, containing 149 genes (Ensembl 55), corresponding to the average number of genes present in the orphan loci associated to phenotypes with a currently unknown molecular basis (OMIM; 17 June 2009). Additionally, we removed all associations regarding possible known disease genes from the given phenotype. We then applied the prediction procedure on the same CCNs and considered as positive predictions those that satisfied a CCN-specific P-value threshold corresponding to the 1% FDR of the true predictions. We defined the precision as the fraction of cases in which the positive predictions contained the true disease gene.

Results

Generation of CCNs and coexpression clusters

To evaluate the potential of tissue-specific conserved coexpression, we studied a large microarray data set downloaded from the Gene Expression Omnibus (GEO),16 covering many tissues, cell types and experimental conditions in both human (5188 experiments) and mouse (2310 experiments). Based on the description given in the GEO, the samples were manually annotated to allow the selection of condition-specific subsets of the desired anatomical depth (see section Methods), which we used to generate the corresponding SCNs or human–mouse CCNs, by the procedures described previously by Ala et al.12 Briefly, CCNs are obtained by intersecting a human SCN with a mouse SCN, which reflects the same selection of tissues and/or conditions, keeping only the coexpression links that are observed in both species. We generated multi-tissue CCNs, covering, for example, all samples from normal (non-tumoral) tissues, and tissue-specific CCNs from normal samples from heart, kidney, breast, etc. In multi-tissue coexpression networks, it has been shown that phylogenetic conservation is a strong filter for functionally relevant links,10, 12 because it is highly unlikely that the artifactual correlations generated in one species by possible sample outliers or by other factors are reproduced in a corresponding data set of a second species. We verified that this principle holds true also in tissue-specific CCNs (see Supplementary Table S2 and Supplementary Figure S1 for details).

From each of the CCNs we then extracted a set of ‘coexpression clusters’, consisting of a gene together with its nearest neighbors, that is, together with those genes that show a conserved coexpression with it.12

The complementarity and biological relevance of multi-tissue and tissue-specific conserved coexpression

As expected, a direct pairwise comparison of all the different networks showed in every case a highly significant overlap, consistent with the existence of transcriptional modules composed of genes highly coexpressed in most mammalian cell types. For instance, the least significant overlap was found between embryonic tissues and adipose tissue CCNs, where only 0.6% of the edges present in either network were found in both: such a small overlap is however much larger than expected by chance (about 100 standard deviations above the mean overlap obtained from 100 network randomizations).

However, the overlap between the different tissue-specific CCNs or between the tissue-specific CCNs and multi-tissue CCNs was always much lower than the intersection between different multi-tissue CCNs (Figure 1a). Altogether, these data indicate that tissue-specific CCNs capture a large number of conserved-coexpression relationships that cannot be detected using the promiscuous approach and vice versa.

Figure 1
figure 1

(a) A heatmap representing the fraction of common edges between different CCNs. The two random cases are representative examples of networks obtained from 10% of the experiments that compose the indicated data set, chosen at random. The scale bar represents the percentage of common links (intersection/union). (b) The number of GO keywords significantly enriched in the indicated number of CCNs.

Full size image

To address the functional relevance of these correlations, we evaluated their capability to provide functionally characterized coexpression clusters, in comparison with those obtained from multi-tissue CCNs. Therefore, for every CCN, we analyzed the overrepresentation of GO keyword in all coexpression clusters extracted from the CCN. Both tissue-specific CCNs and multi-tissue CCNs showed a strongly increased number of functionally enriched clusters (Supplementary Figure S2) when compared with 100 randomized versions of the same networks. Moreover, the distribution of the enriched GO terms showed that the different networks are strongly complementary, because a high percentage of them were specifically identified in one network and the majority was found in only a few CCNs (Figure 1b). Most importantly, we found that evolutionary conservation as a filter tends to preserve those coexpression edges that are of higher functional significance (see Supplementary information for details). Finally, tissue-specific CCNs showed also a high prevalence of edges between proteins that are known to physically interact from the HPRD.17 (Supplementary Figure S2)

As a case study that illustrates the added value of tissue-specific CCNs, we report the example of mammalian pluripotency genes. The elucidation of the molecular circuitry underlying the establishment and the maintenance of pluripotency is a crucial issue in biology,18 especially after the discovery that the expression of a small set of genes can reprogram differentiated cells to a pluripotent state.19 Proteomic screenings have so far been quite successful in mapping the interactome of some crucial pluripotency factors,20, 21 but cannot identify functional correlations that do not require stable protein–protein interactions. We thus explored the possibility that our approach may identify pluripotency factors. In particular, we asked which genes are connected to the crucial pluripotency gene POU5F1 in the above-mentioned multi-tissue and tissue-specific CCNs. The network obtained from normal adult tissues contained no edges for this gene and very few edges were found in most tissue-specific networks (Supplementary Table S3). By contrast, POU5F1 had a very high connectivity in the network obtained from embryonic tissues and was even more connected in a CCN obtained specifically from human and mouse stem cell data sets (Supplementary Table S3). Strikingly, the neighborhood of POU5F1 in the latter networks contained most of the other genes that have been used successfully to reprogram differentiated cells to induced pluripotent stem (iPS) cells, such as NANOG, LIN28, SOX2 and MYCN, and many other genes that have been functionally associated with the establishment and maintenance of the pluripotent state (Figure 2, Supplementary Table S3 and Supplementary Figure S3). This result indicates that in the stem cell CCN, most of the key pluripotency factors form a very tight cluster. Considering that the identification of this cluster by our method was completely independent from previous knowledge, it is very likely that it may comprise also a significant portion of the unknown core pluripotency machinery, including both other master control genes and at least some of the most direct targets of the core transcription factors.

Figure 2
figure 2

The neighborhood of the POU5F1 gene in the CCN obtained from stem cell microarray experiments. The size of the nodes and their distance from the center are a function of their connectivity. The genes shown in orange have been used successfully to reprogram differentiated cells to iPS cells. The genes shown in yellow have been linked experimentally to the pluripotent state or are considered as pluripotency markers.

Full size image

In summary, these results indicate that the tissue-specific approach significantly extends the potential of human–mouse conserved coexpression to provide new functional hypotheses about mammalian genes. The observed complementarity to the multi-tissue approach suggests that it is best to use both approaches jointly, because both coexpression across multiple tissues and coexpression within specific tissues can yield important clues to gene function.

Tissue-specific conserved coexpression and phenotype information

In our previous work we have shown that the analysis of conserved coexpression in multi-tissue microarray data sets can be a powerful tool to evaluate candidate genes for human genetic diseases.12 Therefore, we asked whether tissue-specific CCNs as well can provide important clues for disease gene prediction, and whether the disease-related information they contain overlaps with or rather complements the information obtained from multi-tissue CCNs.

For this purpose we derived OMIM PCNs from 13 different tissue-specific CCNs and from one multi-tissue CCN, obtained from normal tissue data. In these PCNs, the nodes represent OMIM phenotypes and two nodes are joined by an edge if the phenotypes are similar,15 and there is at least one edge in the CCN between genes associated to the two phenotypes. In other words, PCN edges indicate that, in the underlying CCN, genes involved in one phenotype are directly linked to genes involved in a similar phenotype and therefore that the CCN edges contain useful information to evaluate candidate genes for related disease phenotypes with so far unknown molecular basis.

Interestingly, 505 (34.2%) of 1477 OMIM phenotype IDs and 3348 (73.4%) of 4562 OMIM–OMIM edges were present in only one PCN. Although the multi-tissue PCN provides more unique OMIM IDs and edges than any of the tissue-specific PCNs, 68.1% of all unique phenotypes and 61.8% of all unique edges can be found exclusively in one of the tissue-specific networks (Figure 3a). These data suggest that tissue-specific CCNs can provide valuable information for disease gene prediction that is not contained in multi-tissue CCNs.

Figure 3
figure 3

(a) The number of OMIM–OMIM links unique to each of the PCNs. (b) Charcot–Marie–Tooth disease type-4D (CMT4D; OMIM 601455) and its first- and second-level neighbors in the CNS and NT PCNs. Gray nodes and black edges were found in both PCNs. Red and green nodes/edges were specifically found in the CNS or in the NT PCNs, respectively. Legend: CMT, Charcot–Marie–Tooth disease; MMZ, myopathy, myofibrillar, ZASP-related; NEM, nemaline myopathy; HMN, neuronopathy, distal hereditary, motor; SNCV, slowed nerve conduction velocity; CHN, neuropathy, congenital, hypo-myelinating; HSAN, neuropathy, hereditary, sensory and autonomic; RLHAD, Roussy–Levy hereditary areflexic dystasia; HNPP, neuropathy, hereditary, with liability to pressure palsies; HNDS, hypertrophic neuropathy of Dejerine–Sottas.

Full size image

A good example of this complementarity is given by the small sub-network of OMIM phenotypes directly linked to Charcot–Marie–Tooth disease type-4D (CMT4D; OMIM 601455) and their respective neighbors in both the central nervous system (CNS) and the normal tissues (NT) PCNs (Figure 3b). Whereas many of the phenotypes and of the relationships between them can be found in both PCNs, a significant part can be found in only one of the two PCNs.

Identification of high-confidence disease gene candidates by multi-tissue and tissue-specific conserved coexpression

As we have shown in our previous work, multi-tissue conserved coexpression can be efficiently combined with phenotype correlation data to provide high-confidence candidates within genetic disease loci (see reference Ala et al,12 and section Methods). The results described above indicated not only that high-confidence candidates could be obtained from tissue-specific expression data, but also that such results would likely be complementary to those obtained from multi-tissue CCNs.

Thus, using the 14 CCNs from Figure 3a, we applied our procedure to 1074 mapped ‘orphan loci’ associated to 1028 OMIM phenotype entries with unknown molecular basis. At a 1% FDR, estimated separately for each CCN, we identified an average of 7.1 candidates for 238 of these ‘orphan loci’, thus obtaining a total of 1692 high-confidence candidates (see the detailed table at http://87.253.99.109/ts-coexp/disease_gene_precomp_table.php).

This is a significantly reduced number, considering that the orphan loci contain on average 149 candidate genes. A leave-one-out test performed on artificial loci of the same size, centered on the known disease genes (see section Methods), showed a precision of 44%. Therefore, we expect that in approximately half of the cases the positional candidates selected by our procedure will contain the actual disease-causing gene.

Evaluation of XLMR candidates obtained by X-chromosome exome sequencing

Beside the proof of concept already provided by the leave-one-out test, we decided to specifically evaluate the usefulness of our method in the prioritization of the mutational candidates obtained by exome sequencing projects in genetically heterogeneous disorders. In particular, we concentrated on XLMR as a case study, because Tarpey et al22 have recently re-sequenced a large fraction of X-chromosome coding exons of patients or carriers from 208 different XLMR families. Interestingly enough, although this survey identified three novel genes (SYP, ZNF711 and CASK), as well as several other genes already implicated in XLMR, it also underscored the limits of such kind of mutational screenings.22 For example, truncating mutations were identified in 30 genes, but after extensive validation only a minority of them could be proven to be actually disease-causing. Moreover, despite the identification of a very large number of missense mutations on conserved residues and of several potential splice site mutations, the statistical power of the study allowed to define a reasonable genetic explanation of the phenotypes only for 25% of the cases.22

Therefore, to evaluate conserved coexpression as an additional criterion for scoring mutated genes, we asked how many known XLMR genes could be predicted as strong candidates using our approach. Importantly, to perform this test we considered as candidates all the genes that have been sequenced in the mentioned study and removed all known XLMR genes from the list of the reference genes, hence pretending XLMR to be a disease phenotype without any known molecular basis. As the number of reference genes (disease genes involved in similar phenotypes) was still large at a MimMiner phenotype similarity threshold score (see section Methods) of 0.4 (401 genes), we decided to use a more stringent MimMiner threshold of 0.5 that yielded a total of 64 reference genes. Afterwards, we used the previously mentioned CCNs to predict the candidates with a 10% FDR cut-off. We obtained a total of 222 predictions, corresponding to 102 candidate genes (Supplementary Table S4). Although this number may appear very large, it must be considered that 90 genes on the X-chromosome are known to be associated to some form of intellectual disability and that a similar number probably remains to be identified.23, 24 Among our candidates, only 25 were retrieved from the normal tissue CCN, while the majority was derived from tissue-specific networks, with 40% coming from the central nervous system or from brain networks.

Our list of candidates was strongly enriched for known XLMR genes, as 32% of them were included (P=2.5E−06, Supplementary Table S4). Moreover, the global prevalence of validated XLMR genes was 11.8% among the 288 genes in which the authors found missense substitutions of conserved residues, but almost three times as much among the 57 mutated genes that were also found as candidates by our approach (Supplementary Tables S4).

Even more strikingly, the comparison of the 30 genes, in which the authors found truncating mutations, with our candidates showed only seven common genes (AP1S2, BRWD3, CUL4B, SLC9A6, UPF3B, USP9X and ZNF711), six of which have been validated by the authors in the same study (P=2.2E−08). The only non-validated gene of this kind is USP9X, whose role has remained uncertain because only a single truncating mutation was found in patients. Thus, our data further underscore USP9X as a strong XLMR candidate.

Considering the efficiency of our approach in ‘re-discovering’ XLMR genes even when none of them was considered as a reference gene, we speculate that it may be even more effective in identifying new candidates when the validated genes are correctly considered as reference. We therefore provide in Table 1 a list of the candidates with an FDR of 10% or lower, obtained using only the known XLMR genes as reference. Also in this case, USP9X was highlighted as a promising candidate. The fact that it was predicted with two completely distinct sets of reference genes further supports the significance of its XLMR candidacy.

Table 1 List of the candidate genes for XLMR generated from CCNs with an FDR of 0.1 or less, considering as reference genes all known XLMR genes
Full size table

One of the most appealing features of our system is that, besides selecting promising candidate genes for a particular disease, it may give good indications about the underlying molecular mechanisms, through functional analysis of the candidates’ neighboring genes in the coexpression networks. Interestingly, analysis of the genes that are connected with USP9X showed a strong prevalence of enzymes involved in the ubiquitin cycle (Figure 4). This result is fully consistent with USP9X being a deubiquitinating enzyme,25 and offers testable predictions about the specific proteins that may functionally cooperate with it and/or about the possible targets of its activity.

Figure 4
figure 4

A representation of a subset of the USP9X gene’s neighborhood in the three CCNs from which it was predicted as a candidate XLMR gene (normal tissues, CNS and skeletal muscle) by means of related phenotypes. It includes all the nodes that were connected to USP9X in at least one network. The thick edges represent links found in all three networks. The genes shown in orange are functionally involved in the ubiquitin cycle, whereas those with a cyan border are involved in diseases related to XLMR (MimMiner score ≥0.4).

Full size image

Discussion

Although deep-sequencing techniques are setting new benchmarks for the analysis of gene expression, microarray data deposited in public databases still represent an incredibly rich source of information. Therefore, it is quite surprising to observe that, with notable exceptions,26, 27, 28, 29, 30 coexpression relationships based on DNA microarrays are not used extensively to support the identification of new disease-causing genes. The main reason for this discrepancy may well reside in the low specificity of the functional predictions obtained by coexpression analysis. In other words, it is well known that if two genes are involved in similar functions and disease phenotypes, they tend to have strongly correlated expression profiles, but it is known as well that two or more genes may show strongly correlated expression profiles even though they are not functionally related. The studies conducted by different groups, including our study,10, 11, 12 have shown that very simple filters, such as phylogenetic conservation, can strongly increase the predictive power of coexpression analysis. Here we have shown that another simple filter, that is, the selection of tissue-specific subsets of a large data repository, can significantly extend the potential usefulness of the coexpression-based predictions. It is important to underscore that, if used in isolation the tissue-specific approach would be even noisier than a multi-tissue approach, independently of the size of the starting microarray data set (Supplementary Table S1). However, when combined with conservation, the tissue-specific approach can provide a high percentage of accurate predictions that could not be obtained using heterogeneous data sets. In agreement with our previous studies, we showed that this information can be used effectively to identify high-confidence disease gene candidates. Moreover, and perhaps most importantly, we provide proof of concept that the combination of conserved coexpression and phenomic networks could be successfully used to support the identification of disease genes by massive sequencing strategies, which in the case of genetically heterogeneous disorders can identify a much higher number of mutated candidates than suspected previously.22

As a side-product of our analysis, we obtained a list of high-confidence candidates for XLMR. The relevance of this list is supported by two recent observations. Indeed, while the manuscript was in preparation, we realized that RAB39B had just been discovered as an actual XLMR gene.31 Moreover, during the revision process, the GSPT2 gene had been proposed independently as a strong XLMR candidate by a copy-number variation study.32

We think that the case of USP9X illustrates particularly well the potential usefulness of our approach for selecting the most promising candidates and making hypothesis about their functional interactions with the validated genes. The mutational screening by Tarpey et al22 found a single truncating mutation in this gene, which so far has not been validated by clear additional mutational evidence (Joseph Gecz, personal communication). USP9X is a de-ubiquitinating enzyme, encoded by a strongly conserved gene that escapes X inactivation.25, 33 Moreover, the Drosophila ortholog faf is known for its role in synaptic development and the expression pattern of the mouse ortholog is suggestive of a synaptic function.34, 35

Thus, the finding that the gene is consistently coexpressed with many other genes involved in intellectual disability (namely UBE3A, UBR1, OPA1 and CRBN) and with genes involved in the ubiquitin cycle is very suggestive. In particular, as UBE3A and UBR1, two of the most strongly connected genes, show both features, we propose the working hypothesis that the three proteins they encode may directly cooperate in regulating the ubiquitination/de-ubiquitination cycle of common substrates involved in synaptic function, some of which could be possibly included in our list of potential partners.

In conclusion, we propose multi-tissue and tissue-specific CCNs as highly valuable resources for predicting and exploring the functional properties of genes in relation to human diseases. To support the scientific community in this task, this paper is accompanied by the TS-CoExp browser (http://www.cbu.mbcunito.it/ts-coexp), a user-friendly web interface described in more detail in the Supplementary information.