Predicting candidate genes for human deafness disorders: a bioinformatics approach

Hearing loss, acquired or genetic, is a major worldwide public health concern. Numerous genes have been linked to hearing disorders [1]. These disorders may be syndromic or nonsyndromic; conductive, sensorineural, or mixed; and prelingual or postlingual [2]. The various genetic forms of hearing loss are distinguished based on otologic, audiologic and physical examination combined with linkage analysis. Some representative deafness genes that have been identified include the Alport syndrome (COL4A3, COL4A4 or COL4A5 genes), branchio-oto-renal syndrome (EYA1 gene), Mohr-Tranebjaerg syndrome (TIMM8A gene), Pendred syndrome (SLC26A4 gene), Jervell and Lange-Nielsen Syndrome (KVLQT1 and KCNE1 genes), Usher syndrome with its several types, Norrie disease (NDP gene), DFNB1 (GJB2 gene), DFN3 (POU3F4 gene), DFNB4 (SLC26A4 gene), DFNA6/14 (WFS1 gene), and several others [3, 4]. The mutational analysis of genes such as GJB2 (encoding the protein connexin 26) and GJB6 (encoding the protein connexin 30) [3, 5, 6] has aided diagnosis and geneticcounselling.

Syndromic hearing loss is associated with a variety of other clinical findings and is relatively less prevalent. In contrast, nonsyndromic hearing loss accounts for more than 70% of deafness cases and involves autosomal as well as X or Y -linked deafness phenotypes [7]. The molecular causes of nearly all nonsyndromic hearing loss are associated with inner ear structural damage, and changes in both the inner and the middle ear [8]. Mutations in genes such as the ACTG1, COCH, COL11A2, DFNA5, EYA4, GJB2, GJB6, KCNQ4, MYO6, MYO7A, TECTA, TMC1, and WFS1, as well as altered expression of genes such as GJB3 and MYO1A have been associated with the autosomal dominant types that are generally progressive and involve changes in inner ear [9–11]. The autosomal recessive phenotypes are associated with mutations in genes such as the CDH23, CLDN14, ESPN, GJB2, GJB6, MYO15A, MYO6, MYO7A, OTOF, PCDH15, SLC26A4, STRC, TECTA, TMC1, TMIE, TMPRSS3, and USH1C, as well as altered expression of GJB3 [8].

The map locations of a large number of nonsyndromic autosomal recessive deafness phenotypes are known, but the specific genes responsible for all these phenotypes have not been identified [4]. The cloning of genes involved in such phenotypes requires refinement of the suspected genomic interval to as short a region as possible by linkage analysis. However, it is not always possible to map a gene within an interval that is amenable for mutation analysis. The mutation analysis of all genes encoded by a large genomic interval is extremely labor-intensive. We describe here a bioinformatic approach that can reduce the candidate genes to a manageable number for mutation analysis. Initially, all the genes from a particular locus are cross-referenced to the databases of expressed mouse inner ear genes and the expressed human cochlear genes. The alternative procedure included a search for interacting proteins for the gene products mapping to the candidate region. As presented here, this approach has led to a set of specific candidate genes.

The locations of 23 autosomal dominant and 27 autosomal recessive nonsyndromic deafness phenotypes mapped to several chromosomes downloaded from hereditary hearing loss homepage are shown in Tables 1 and 2[4]. Additional loci for nonsyndromic conditions are mapped to chromosomes 1, 8, X and Y [4]. The hereditary hearing loss homepage is updated on a regular basis. The marker boundaries of these locations encompass between 1.4 and 18.6 million basepairs (Mbp) for various loci. To generate a set of candidate genes for the listed loci, a strategy schematically represented in Figure 1 was followed. The determination of coding sequences and/or genes in a genomic region was made by Unigene [12]. However, the genes encoded in a large genomic interval are too many to be characterized by mutational analysis in a gene-by-gene approach. Therefore, we used the human cochlea and mouse inner ear expression databases [13, 14] to eliminate from the candidate list certain genes that were not expressed in these organs. Such in silico expression analysis relies on the assumption that the expression databases are comprehensive. However, the characterization of all transcripts expressed in the ear is far from complete. We, therefore, introduced another step in our candidate gene strategy by taking advantage of the human protein reference database (HPRD) and generated a list of interacting genes for every gene mapping to candidate deafness loci [15]. The rationale for protein interaction is as follows. If a gene encoded in the candidate region interacts with a gene that is either involved in inner ear development/function, or a protein shows interaction with more than one candidate genes mapping to different loci, then such a gene is likely to be involved in the phenotype in question. The interaction pattern of the gene products from Usher syndrome is a good example to illustrate this point. The known gene products for several Usher syndrome loci are known to form interactions in vivo [16]. The mutation of each one of these genes affects protein interactions and influences Usher type 1 phenotype [17]. The five forms of Usher syndrome have defects in myosin VIIA, harmonin, cadherin 23, protocadherin 15, and a putative scaffolding protein sans. Harmonin binds sans, and it also binds myosin VIIA and protocadherin 15 [34]. The role of cadherins in mediating cell-cell interaction is well-characterized. Furthermore, interactions of harmonin (USH1C) with USH2A, USH2C and USH2B are mediated by PDZ domains [35, 36]. In retrospect, if the interacting protein strategy had been used to select candidate genes for Usher syndrome subtypes, it is likely that several genes could have been eliminated from consideration. Therefore, it is reasonable to assume that physical interactions will exist between proteins that are involved in inner ear developmental pathway or inner ear signal transduction pathways, and mutations in any one protein of the pathway is likely to give the same altered phenotype. If proteins of interacting networks can be identified or predicted, then such genes are natural candidates for a given phenotype. The above hypothesis is the underlying rationale for incorporating interacting proteins as a criterion for selecting candidate genes presented in this paper. Briefly, the strategy is as follows. First, assemble the genes encoded in all candidate intervals, list the proteins that interact with genes in the candidate region, and then search for candidates on different loci that interact with a common protein. Such a criterion will fulfil the rationale of putative involvement of proteins at two different loci involved in a common biological process, and by association the respective genes mapping to two different loci will be considered as candidates.

Locus Name	Location
DFNA7	1q21–q23
DFNA16	2q24
DFNA18	3q22
DFNA21	6p21
DFNA23	14q21–q22
DFNA24	4q
DFNA25	12q21–24
DFNA27	4q12
DFNA30	15q25–26
DFNA31	6p21.3
DFNA32	11p15
DFNA34	1q44
DFNA37	1p21
DFNA41	12q24–qter
DFNA42	5q31.1–32
DFNA43	2p12
DFNA44	3q28–29
DFNA47	9p21–22
DFNA49	1q21–q23
DFNA50	7q32
DFNA53	14q11–q12
DFNA54	5q31

Locus Name	Location
DFNB5	14q12
DFNB13	7q34–36
DFNB14	7q31
DFNB15	3q21–q25, 19p13
DFNB17	7q31
DFNB19	18p11
DFNB20	11q25–qter
DFNB24	11q23
DFNB25	4p15.3–q12
DFNB26	4q31
DFNB27	2q23–q31
DFNB28	22q13
DFNB32	1p13.3–22.1
DFNB33	9q34.3
DFNB35	14q24.1–24.3
DFNB38	6q26–q27
DFNB39	7q11.22–q21.12
DFNB40	22q
DFNB42	3q13.31–q22.3
DFNB44	7p14.1–q11.22
DFNB46	18p11.32–p11.31
DFNB48	15q23–q25.1
DFNB49	5q12.3–q14.1.
DFNB55	4q12–q13.2
DFNB58	2q14.1–q21.2
DFNB60	5q22–q31

In principle, the interactions-based strategy can be targeted to identify candidates for deafness if a database for interacting proteins involved in inner ear development and function is available. For example, oncomodulin and prestin are expressed in outer hair cells [20]. The protein interaction approach could link the possible candidate genes to specific cochlear cells by identifying known interactants. If the interactors happen to map to a region harboring a deafness gene, such interactors are obvious candidates for mutational analysis. However, such an approach will require identification of interacting proteins. The primary limitation of the in silico approach described here is inadequate description of interacting protein networks.

The strong candidate list includes genes such as various cadherins, collagens, some cytoskeletal components and a number of growth factors and inner ear specific transcripts. For example, HAT (Human airway trypsin-like protease) from the DFNA27 locus is known to enhance cell growth and IL-8 production. It has been implicated in induction of PAR-2 (protease activated receptor)-mediated IL-8 release in psoriasis vulgaris [21]. Because HAT is expressed in the ear, and protease activated receptor (PAR-2) has the ability to activate G-proteins followed by an increase in calcium ion concentration, we consider HAT as a candidate. KDR(kinase insert domain receptor), a vascular endothelial growth factor(VEGF) receptor-type 2, from the same locus shows age-dependent expression in the inner ear [22]. Our analyses indicated that only a fraction (200/2400) of genes mapping to various genomic intervals was expressed in the inner ear. We attribute these observations to depth of inner ear libraries. It is likely that the genes being scored in these libraries have multiplicity for certain transcripts and absence of other transcripts. For example, out of 153 genes at the DFNA7 locus, only 18 genes are present in the cochlear library. We cannot reasonably rule out the expression of the remaining 135 genes in the inner ear. Therefore, the approach presented here will be more comprehensive if we do not include ear expression in this scheme. Consequently, in a second attempt to mine the protein-interaction data obtained from the HRPD, we analyzed all genes encoded in the candidate intervals for their interactors. The interaction data were considerably exhaustive and resulted in many more possible candidates with their expression not reported in the ear expression library. A summary of gene numbers at different loci before and after interacting proteins analysis using the ear-expression scheme is presented in Table 6. The mouse syntenic genes are also indicated in these results. The number of unfiltered candidate genes for each locus obtained by interacting proteins analysis is shown in Table 7. To elucidate the relevance of genes not found in the ear-expression library as possible candidates, we performed a literature search cross-referencing the identified gene with any reported hearing-associated condition in humans or other model animals. Some of these genes were linked to ear-development or hearing impairment as a secondary or unrelated symptom of other conditions. For example, Neurod1 gene mapping to DFNB27 locus was not reported in any of the inner ear libraries. However, it appears to participate in the development of the auditory system as NeuroD1 null mice exhibited severe reduction of sensory neurons in the cochlear-vestibular ganglion [23]. E2F3, a transcription factor of the E2F family mapping to the DFNA21 locus, may be indirectly implicated by its ability to regulate cell proliferation possibly during the developmental stages [24]. Other candidate genes from the unfiltered candidate analysis for the various loci are listed in Table 8. Thus the unfiltered strategy adds 51 candidates for 25 loci and expands the candidate list to 92 genes for further mutation analysis.

We have used an in silico strategy to assemble a list of candidate genes that are positionally linked to and could be causing specific nonsyndromic hereditary hearing loss conditions. As presented here, a list of 2378 genes mapping to various genomic intervals have been narrowed down to 92 genes as candidates. These candidates may be analyzed for mutations in various deafness phenotypes in parallel with attempts to further narrow down the suspected region by genetic linkage analysis. It warrants mention that the potential of the approach presented here will be better harnessed as more information becomes available about inner ear transcripts and protein interaction networks.