We focused on genes associated with hearing loss in children. The American Speech-Language-Hearing Association (Alsarraf et al., 1998; Dhooge, 2003) defines such hearing loss in children as either acquired or associated with otitis media or congenital deafness. We downloaded the relevant genes from DisGeNet (Piñero et al., 2017) (https://www.disgenet.org/, version 7.0, accessed in April 2021). In total, 175 genes were associated with otitis media, while 72 were associated with congenital deafness and 2 were associated with acquired hearing loss; thus, we did not study acquired hearing loss. The genes associated with congenital deafness and otitis media are listed in Supplementary Tables S1, S2, respectively. We used a network-based method to identify novel candidate genes associated with either pathogenesis.

PPIs are widely used to explore protein or gene-related problems. Several studies have reported that compared with non-interacting proteins, interacting proteins are more likely to have similar functions (Ng et al., 2010; Hu et al., 2011; Chen et al., 2016a; Cai et al., 2017; Zhao et al., 2019; Gao et al., 2021). Such interactions can be used to identify novel genes that are associated with known disease-related genes. We used the STRING database (https://www.string-db.org/, version 10.0) (Szklarczyk et al., 2015) to construct a PPI network; we then applied the powerful, network RWR algorithm (Kohler et al., 2008; Macropol et al., 2009) to discover novel candidate genes associated with otitis media or congenital deafness. Human PPI information collected in STRING is contained in “9606.protein.links.v10.txt.gz”. Each PPI features two proteins identified by their Ensembl IDs, as well as a confidence score indicating the PPI strength. Each score ranges from 1 to 999 and is derived by considering several types of PPIs. In fact, PPIs in STRING can not only indicate the interactions between proteins but also reflect functional associations of proteins. Thus, they can widely measure protein associations. We used the PPIs to build a network in which all 19,247 proteins served as nodes. Two nodes were considered adjacent if and only if they formed a PPI; thus, each edge was a PPI. We assigned a weight to each edge for indicating the strengths of the PPI, which was defined as the confidence score of the corresponding PPI. The network was termed N.

The RWR algorithm is powerful. It simulates a walker that commences at a node set and then randomly moves in the network. The start nodes are termed seed nodes. The walker delivers probabilities of seed nodes to all other nodes in the network. Given a network and k seed nodes, each seed node is assigned a probability of 1/k; the other nodes are assigned probabilities of zero. These probabilities form a vector termed P ₀. The vector is repeatedly updated as follows: P t + 1 = ( 1 − r ) A T P t + r P 0 , (1) where A is the column-wise, normalized adjacency matrix of the network and r is the restarting probability, which was set to 0.8 in this study. Updating stops when P t + 1 and P t are sufficiently close; closeness is given by ‖ P t + 1 − P t ‖ L 1 < 10 − 6 . P t + 1 is the required outcome of the algorithm. Based on this outcome, each node is assigned a probability transmitted from the seed nodes. A higher node probability is indicative of stronger associations with seed nodes.

We used the RWR program established by Li and Patra (Li and Patra, 2010). Genes associated with congenital deafness or otitis media were fed into the program, which ran on the PPI network N. Nodes with probabilities higher than 10⁻⁵ served as raw candidate genes for congenital deafness or otitis media.

Some raw candidate genes associated with congenital deafness or otitis media can be identified using a network-based method. However, several false-positives may be included in the results. To eliminate such genes and select only valid candidates, we used a screening procedure that featured three sequential tests.

The RWR algorithm was executed on the PPI network N to discover raw candidate genes. The structure of N may influence the outcome. Some nodes are readily assigned high probabilities because of their special locations in the network. However, they may have low or no associations with congenital deafness or otitis media. Thus, there is a need to test the statistical significance of the probability that each raw candidate gene is valid. Accordingly, we randomly generated 1,000 gene sets, each of which had the same number of genes associated with congenital deafness or otitis media. For each gene set, such genes were set as the seed nodes of the RWR algorithm. Thus, each candidate gene was assigned a probability in each random gene set. When all 1,000 sets had been tested, each candidate gene had been assigned 1,000 probabilities. By comparing the probability on actual seed nodes to the probabilities on randomly generated sets, the statistical significance of each probability was revealed. We used the Z-score to evaluate significance as follows: Z − s c o r e ( g ) = P ( g ) − P M ( g ) P S T D ( g ) , (2) where g is a raw candidate gene identified by the network-based method, P ( g ) is the probability on actual seed nodes, and P M ( g ) and P S T D ( g ) are the respective mean and standard deviation of the probabilities on randomly produced sets. We set the selection threshold for candidate genes to 1.96; this is a widely accepted threshold when statistical significance is essential.

The second test directly evaluated the associations between candidate genes and congenital deafness or otitis media. For each candidate gene, such associations can be measured by associations between that gene and other genes associated with either condition. Proteins that interact in STRING always exhibit strong associations that can be quantified using confidence scores. For proteins p and q, the confidence score is denoted as Q ( p , q ) . For each candidate gene g, we computed the maximum association score (MAS) as follows: M A S ( g ) = M a x { Q ( g , g ′ ) : g ′  is a gene associated with congenital deafness or otitis media } (3)

Genes with high MAS values are strongly associated with at least one gene linked to congenital deafness or otitis media. Thus, such genes may also be highly related to either condition. We set the threshold for selection of essential candidate genes to 900; this is the cutoff of the highest STRING confidence score.

The last test further filtered candidate genes according to the similarities between their functional terms and the functional terms of genes associated with congenital deafness or otitis media. If the functional terms of a candidate gene are similar to the functional terms of a gene that is validly associated with either condition, that gene may also be linked to one of the conditions. We first used enrichment theory (Carmona-Saez et al., 2007; Huang et al., 2011; Huang et al., 2012; Chen et al., 2016b; Chen et al., 2019) to evaluate the associations between genes and functional terms (GO terms and KEGG pathway terms). Given one gene and one functional term, the gene set containing that gene and genes with which it interacted (in the PPI network of STRING) was constructed; another gene set containing genes annotated by the functional term was built. The associations between the gene and the functional term were calculated as the −log₁₀ of the hypergeometric test p-value of the gene sets constructed above. For any gene g, its associations with all functional terms were computed and collected in a vector denoted V ( g ) . The similarity of two genes g and g ' (based on their functional terms) can be evaluated by comparing their vectors as follows: Λ ( g , g ′ ) = V ( g ) ⋅ V ( g ′ ) ‖ V ( g ) ‖ ⋅ ‖ V ( g ′ ) ‖ (4)

In a manner similar to MAS calculation, for each candidate gene g, the maximum function score (MFS) was computed as follows: M F S ( g ) = M a x { Λ ( g , g ′ ) : g ′  is a gene associated with congenital deafness or otitis media } (5)

Essential genes can be selected by choosing an appropriate MAS threshold.

To explore biological functions associated with identified genes, we applied gene ontology (GO) enrichment analyses using R package topGO (https://bioconductor.org/packages/release/bioc/html/topGO.html, v.2.42.0). The threshold of p-value was set to 0.001 for selecting enriched GO terms in three subclasses: biological processes (BP), cellular components (CC) and molecular functions (MF).

Genes associated with congenital deafness were fed into the RWR algorithm, which ran on PPI network N. Each node in the network was assigned a probability. The selection threshold for raw candidate genes was set to 10⁻⁵; this yielded 5,426 genes (Supplementary Table S3). We then engaged in screening (i.e., filtration) to identify essential genes. First, we used the permutation test to evaluate the statistical significance of probability that each raw candidate gene was essential; the Z-scores for all genes are listed in Supplementary Table S3. In total, 367 candidate genes were assigned Z-scores greater than 1.96. These were fed into the association test, which assigned an MAS to each gene (Supplementary Table S3). At a threshold of 900, 117 genes were selected; these were finally evaluated using the function test. The MFS values are listed in Supplementary Table S3. At an MFS threshold of 0.9, 18 genes were chosen. These “putative genes” were considered to be closely associated with congenital deafness; they are listed in Supplementary Table S4.

For the obtained putative genes, their associations with validated genes were investigated. We extracted all PPIs between putative and validated genes. The confidence scores of these PPIs are illustrated in a heat map, as shown in Figure 2. It can be observed that each putative gene had some interacting genes with confidence scores no less than 900, suggesting strong associations with validated genes. This can be further inferred that putative genes had special relationships with congenital deafness.

We used the method described above to identify putative otitis media-associated genes. The RWR algorithm with genes associated with otitis media as seed nodes was performed on the PPI network N. The probabilities of all nodes were obtained. We selected nodes with probabilities over 10⁻⁵; this yielded 5,631 genes (Supplementary Table S5). These genes were filtered as described above. The Z-scores, MAS values, and MFS values are listed in Supplementary Table S5. Use of thresholds of 1.96 for the Z-score, 900 for the MAS, and 0.96 for the MFS yielded 87 “putative genes” (Supplementary Table S6).

Likewise, the PPIs between putative and validated genes were investigated. A heat map was plotted to indicate the strength of these PPIs, as shown in Figure 3. Also, each putative genes had one or more interacting genes with highest confidence (confidence score ≥900). It is suggested that these putative genes may have special associations with otitis media.

We identified 18 putative genes, of which 5 were chosen for detailed analysis (Table 2). The first is PRKACB (ENSP00000359719), which encodes a catalytic subunit of cAMP-dependent protein kinase. The enzyme is expressed in hearing-associated organs in utero. In 2017, researchers from Southeast University showed that mouse PRKACB regulated the development of Lgr5+ hair cells (inner ear progenitor cells) (Cheng et al., 2017). Therefore, PRKACB is functionally associated with cochlear development; the cochlea is a sensorineural hearing organ. Cochlear impairment and abnormalities are reportedly associated with congenital hearing loss in children (O’malley et al., 1995; Korver et al., 2017; Van Wieringen et al., 2019). It is thus reasonable to expect that a regulator of cochlear development, such as PRKACB, would be associated with congenial pediatric deafness. We identified another putative gene with a similar biological function. PRKACG (ENSP00000366488) encodes another protein of the same complex. In 2016, researchers from the University of Bristol confirmed that the gain-of-function variant DIAPH1 caused macrothrombocytopenia and hearing loss (Stritt et al., 2016). PRKACG acts downstream of DIAPH1, thus participating in DIAPH1-related biological effects. PRKACG may also be functionally connected to pediatric hearing loss.

The next putative gene is PAX2 (ENSP00000396259), which is regarded as a key transcription factor that regulates the development of multiple systems, including the central nervous system (Ziman et al., 2001) and the eyes (Adam et al., 1993). In 2006, researchers from the McLaughlin Research Institute for Biomedical Sciences reported that PAX2 interacted with EYA1 to regulate the development of sensory regions in the inner ear (Zou et al., 2006). Developmental abnormalities of these regions are directly associated with congenital hearing loss (Kimura et al., 2018), implying that PAX2 is a relevant putative gene involved in congenital pediatric deafness.

PRKX (ENSP00000262848) is also associated with congenital pediatric hearing loss. A 2019 review concerning chromosomal aberrations associated with endocrine abnormalities in children confirmed that PRKX regulated the development of hearing (Haltrich, 2019). PRKX is located on the X chromosome; it is functionally connected to X-linked congenital hearing loss (Song et al., 2012).

The next putative gene is MATK (ENSP00000378485); this regulates signal transduction in hematopoietic cells (Grgurevich et al., 1997; Lee et al., 2006). In 2017, a clinical case report in JAMA Otolaryngology—Head and Neck Surgery stated that MATK was associated with unilateral hearing loss and otorrhea (Costello et al., 2017). Acute megakaryoblastic leukemia has been functionally connected to unilateral, congenital hearing loss; the pathogenetic backgrounds are related (Costello et al., 2017). MATK encodes megakaryocyte-associated tyrosine kinase, which is structurally similar to C-terminal Src kinase; notably, megakaryocyte-associated tyrosine kinase is associated with acute megakaryoblastic leukemia (Jhun et al., 1995). Therefore, MATK might be involved in the development of ear tumors that cause adaptive hearing loss.

We identified 87 genes putatively associated with otitis media (Supplementary Table S6); we subjected 5 of these genes to detailed analysis (Table 3 ). The first such gene is RAC3 (ENSP00000304283). Although there is insufficient direct evidence that RAC3 is involved in otitis media, a clinical genomic database (ClinVar Miner) (Henrie et al., 2018) indicates that the Talkowski Laboratory of Massachusetts General Hospital has demonstrated associations of RAC3 variants with otitis media. The next gene is HCK (ENSP00000365012); this member of the Src tyrosine kinase family regulates the innate immune response (Ernst et al., 2002). HCK was previously reported to be specifically associated with chronic otitis media and its major chronic complications in children with hearing loss (Suri et al., 2016), validating our findings. The next putative gene is ITK (ENSP00000398655), which encodes an IL2-and T cell-associated kinase. In 2008, the gene was reported to potentially mediate the inflammation of otitis media (Juhn et al., 2008). Furthermore, a report concerning early diagnosis of PI3Kδ syndrome in a 2-year-old girl revealed an association between ITK deficiency and recurrent otitis media (Saettini et al., 2017).

FGR (ENSP00000363115; also known as SRC2), another member of the Src tyrosine kinase family, is also associated with otitis media. This gene has roles in immune responses against pathogens in multiple organs, including ears (Kim et al., 2008). Additionally, the gene has been widely reported to participate in Epstein–Barr virus-associated malignancies (Klein et al., 1988). In 2017, Epstein–Barr virus infection was confirmed as a major etiological and pathological factor for secretory otitis media in children (Vogelnik and Matos, 2017), validating the link between FGR and otitis media.

CDC42 (ENSP00000314458) is an immune system-associated gene; we found that it was closely associated with otitis media. In 2021, CDC42 deficiency was shown to be associated with recurrent pneumonia, otitis media, and bacteremia (Kashani et al., 2021). CDC42 interacts with another effector gene, RAC1 (Hoppe and Swanson, 2004); a homolog of RAC1 (i.e., RAC3, discussed above) was shown to be associated with hearing loss, confirming that CDC42 is linked to otitis media.

In summary, several putative genes are associated with the two types of pediatric deafness. Their identification may provide insights concerning the pathogeneses involved.

For GO functional enrichment analyses on putative genes yielded by our computational method, multiple significant GO terms were identified. The detailed analyses on the top three enriched GO terms ranking by p-values for congenital deafness and otitis media were presented below.

For congenital deafness, the first enriched GO term is cytoskeletal anchor activity (GO:0008093). According to recent publications, mutations in cytoskeletal encoding proteins have been shown to be associated with congenital deafness (Riazuddin et al., 2006), reflecting the potential associations between congenital deafness and cytoskeletal anchor activity. The second enriched term is spectrin binding (GO:0030507). In 2017, a recessive mutation on spectrin associated gene has been shown to be associated with congenital central deafness (Knierim et al., 2017), validating this result. Furthermore, costamere (GO:0043034) is the third enriched GO term (in CC) associated with congenital deafness. According to recent next-generation sequencing analyses (Schraders et al., 2011), costameres has been shown to be associated with progressive hearing impairment.

More GO terms were enriched by putative genes associated with otitis media, including non−membrane spanning protein tyrosine kinase activity (GO:0004715) and interleukin mediated signaling pathway (GO:0038100, GO0035723). Non-membrane spanning protein tyrosine kinase has been shown to be associated with specific inflammatory effects and pathogen infections (Gu et al., 2009; Rocha-Sanchez et al., 2013). Considering that otitis media is associated with infection and inflammatory effects around middle ears, it is reasonable for otitis media associated genes to enrich in inflammatory effects. As for interleukin mediated signaling pathways, middle ear inflammation has been shown to be associated with interleukin related signaling pathways, validating this result (Kerschner et al., 2006; Shi et al., 2014).

By comparing the putative genes associated with congenital deafness and otitis media, only one shared gene CDH1 (ENSG00000039068) was identified. The pathogenesis of congenital deafness and otitis media are totally different according to recent studies. Congenital deafness means the hearing loss is present at birth linking the pathogenesis to genetic factors or stimulations during pregnancy. However, as for otitis media, generally, otitis media is caused by infections and happens after birth. CDH1 has been widely reported to be associated with hearing loss (Friedman and Avraham, 2009; Kanavy et al., 2019). Specifically, CDH1 has been reported to be associated with congenital deafness due to the pathogenic alteration of inner ear but not middle ear (Friedman and Avraham, 2009), which has totally different pathogenic regions comparing with otitis media. As for otitis media, CDH1 has been shown to participate in the pathogenesis of otitis media via regulation on the inflammatory proliferative responses against infections (Kurabi et al., 2013).Therefore, although both subtypes of hearing loss have been shown to be associated with gene CDH1, the contribution and regulatory role of CDH1 on them are totally different, reflecting the complex regulatory mechanisms for childhood hearing loss.

In summary, CHD1 is associated with two types of pediatric deafness. Its dentification may provide insights concerning the pathogeneses involved.