KidneyNetwork: Using Kidney-derived Gene Expression Data To Predict And Prioritize Novel Genes Involved in Kidney Disease

DISCUSSION 

We present KidneyNetwork, a publicly available co-expression network with optimized expression and phenotype annotation data for application to kidney diseases. A signifificant proportion of patients with a suspected genetic kidney disease remain without a genetic diagnosis, as lists of disease genes for many conditions are incomplete. Identifying which genes are involved in kidney disease is essential for improving the diagnostic yield in kidney disease patients and for studying disease pathogenesis to approach treatment avenues. Establishing novel disease genes requires careful biological validation. Implicating genes worthy of such investigations is critical. Application of KidneyNetwork in conjunction with WES or GWAS data by nephrologists, clinical geneticists, or researchers will help each of these groups to participate in gene implication. KidneyNetwork combines a co-expression network based on a kidney sample dataset with the previously published multi-tissue dataset used to build GeneNetwork. Combining the datasets into KidneyNetwork improved phenotype predictions related to kidney disease when compared to networks based on the two datasets separately. As proof of principle, we show that the candidate gene list for the combined phenotype of kidney and liver cysts generated by KidneyNetwork prioritized a manageable list of candidate genes from a long list of genes containing rare variants in our patient with this phenotype.

CLICK HERE TO GET CISTANCHE FOR KIDNEY FUNCTION

Our implication and exploration of ALG6 as a potential candidate gene for kidney and liver cysts result in a plausible candidate gene, supported by the co-occurrence of the ALG6 loss of function variants and polycystic liver and kidney disease in several patients with supportive familial segregation of affected patients in two families, and by the statistically signifificant enrichment of the truncating ALG6 c.257 + 5 G > A variant in a phenotypically defined cohort of unsolved ADPKD/PCLD cases. Biological validation will be necessary to finally determine if ALG6 is a disease gene for autosomal dominant polycystic kidney and liver phenotypes.

The biological plausibility is suggested by known functional similarities and tight transcriptional coregulation of ALG6 to established disease genes as highlighted by KidneyNetwork. ALG6, similar to the established polycystic kidney and liver disease gene ALG8, is a member of the α3-glucosyltransferase family [22]. In addition to ALG8 [20], ALG9 heterozygous variants have recently also been implicated in the etiology of kidney and liver cyst phenotypes [16]. These three genes each play an essential role in the biosynthetic pathway for lipid-linked oligosaccharides prior to their transfer onto asparagine (N) residues of nascent proteins as so-called N-glycans in the endoplasmic reticulum [23]. Interestingly, while kidney or liver cysts have been described, among multi-organ involvement in fetuses or children with ALG9-CDG or infrequently in ALG8-CDG, cysts have not been described for ALG6-CDG [19]. Parents of CDG patients have not yet been studied for cysts. Given the mild phenotype, cysts are likely to go unnoticed in many cases, especially in early parenthood, which is when children are most often diagnosed with CDG.

Fig. 4 Imaging from patients. An abdominal CT illustrating polycystic kidneys and liver in SAMPLE6. Some cysts are highlighted by red arrows, with the largest hepatic cyst measuring 7.7 cm (red asterisk). B abdominal MRI of an affected child of SAMPLE6 shows multiple cysts in the left kidney (several highlighted with red arrows), some hypo-intense on T2, and a few cysts in the right kidney. C abdominal MRI of YU378 showed extensive polycystic liver disease and two kidney cysts. D abdominal MRI of YU481 shows multiple liver cysts. The left kidney has a 9 cm cyst and a few small cysts, and the right kidney has no cysts. E abdominal MRI illustrating polycystic liver disease in LE1. Hepatic cysts are highlighted by red arrows, with the largest cyst located in liver segment IV (red asterisk), necessitating surgical intervention for progressive cholestasis. Of note, both kidneys presented with normal morphology in the absence of any cystic lesions.

The phenotype in the genetically unsolved polycystic kidney and liver patients we identified to carry ALG6 variants is relatively mild, in many cases liver predominant and asymptomatic, consistent with the phenotype described for patients carrying a heterozygous ALG8 or ALG9 pathogenic variant. The potentially pathogenic variants we identified are also found in individuals in the gnomAD database [17]. One explanation for this observation could be incomplete penetrance of the disease. The fact that some of the individuals in our cohort reported no known affected family members, could be an indication of incomplete penetrance, although segregation is lacking in many families. However, we did not identify unaffected individuals carrying the variant. An alternative explanation could be that the observed phenotype is relatively mild and subclinical. For example, the kidney and liver cysts observed in SAMPLE6 were discovered as incidental finding. If no abdominal imaging is done in individuals carrying these variants, the cysts can go unnoticed. Also for ALG8 and ALG9 Besse et al. contemplate on the relatively mild phenotype and propose this can likely be determined by two factors [16, 20]. First, it is expected that a somatic second hit is needed to get a cystic phenotype. The relative infrequency of these somatic second hit mutations that inactivate the normal copy of ALG8/ALG9 and the incomplete effect this has on Polycystin-1 is expected to cause a relatively mild phenotype.

ALG6 has previously been suggested to be involved in one individual with ADPKD [24]. However, that patient, who carried two missense variants with inconclusive predictions that have not been functionally assessed, had a very severe phenotype that did not match the expected phenotype for ALG6.

Strengths and limitations

Building gene co-expression networks requires a large number of RNA-sequencing samples [6] derived from various cell types and developmental stages in order to achieve accurate function predictions. This sample diversity, combined with high numbers of samples are not often available for specifific tissues. To overcome this issue, one earlier approach used hierarchical similarities between tissue types [25]. However, this solution requires a priori gene selection due to its computational burden. In contrast, our method can be used to make unbiased genome-wide predictions. Moreover, the hierarchical approach would have to be repeated for each new tissue of interest, whereas the multi-tissue dataset can be re-used to build a different tissue-specific network using our method. Another approach used differential expression between different tissue types [26]. Here, the top 10% of most differentially expressed genes were correlated with kidney-related GWAS loci. Using differential expression allows predictions to be made regardless of previous knowledge on gene-phenotype interactions. However, this also requires applying a differential expression cut-off. In contrast, our approach makes use of underlying biological structures in RNA-sequencing data to obtain a prediction score for every gene. While combining differential expression with GWAS summary statistics allows for unbiased gene predictions, the reliability of experimentally validated HPO annotations is higher than that of GWAS results. Integrating the HPO database thus results in more reliable predictions. Moreover, we make simultaneous predictions for all HPO terms, whereas the GWAS-based approach needs to be repeated for each GWAS of interest.

Combining kidney-specific RNA-sequencing samples with the multi-tissue dataset allowed us to overcome both the issue of sample size and the challenges in observing tissue-specific differential expression when using only tissue-specific expression datasets. In addition, during the development of KidneyNetwork, we did not have to limit the number of genes that the network is built upon. Furthermore, KidneyNetwork users can get predictions for all possible genes in an unbiased approach, and gene prioritizations for a combination of HPO terms can be obtained.

A downside of using bulk RNA-sequencing data is that we have limited power to make inferences for lowly expressed genes, which is particularly important for genes that are specifific to rare cell types. As more cell-type specifific and single-cell RNA-sequencing data become available in the future, creating co-expression networks based on different kidney cell types might solve this for genes that are expressed more abundantly within specifific cell types. Another limitation of using only RNA-sequencing data is that other biological processes potentially involved in disease development, for example, post-translational modifications and protein-protein interactions, are currently not considered by our prediction model.

Apart from identifying new plausible candidate genes, KidneyNetwork can also be well used to prioritize known kidney disease genes. This can be particularly useful after an initial negative diagnostic result after exome-based gene panel analysis is performed, which might not include analysis of all known kidney disease genes.

Currently, KidneyNetwork is optimized for intrinsic kidney disease. However, kidney disease can also be present because of a pathogenetic process in other systems, such as the immune system. While we can also make inferences on gene prioritization for non-kidney phenotypes, these predictions can improve by building networks specifific for different tissues in the future.

We realize that based on the present literature alone, ALG6 would be a candidate gene for the cyst phenotype in SAMPLE6. To prove the involvement of ALG6 in this phenotype, functional follow-up is required. However, this also proves the strength of our method; out of 322 genes with potentially deleterious variants this plausible candidate gene was prioritized to the top 3, making going into exome-wide sequencing data ‒ for more patients, with various phenotypes ‒ time-efficient and worthwhile.

Improved gene function predictions 

We show that our improved method for assigning gene functions and kidney-related HPO terms to genes outperforms our previously published model. Our leave-one-out cross-validation approach ensures that predictions are not overfitted, that the reported AUC values are not inflated and that our method is robust. Furthermore, before predicting gene‒phenotypes associations, we excluded gene-disease associations from the HPO database that had little experimental evidence, because prediction accuracy is dependent on the accuracy of annotated gene-phenotype associations. Prediction accuracy is based on true positive and true negative gene predictions, which means that more accurate mapping of known genes to phenotypes results in better predictions. Gene-phenotype association accuracy will improve once more genes are annotated and validated for each phenotype. Therefore, we expect an improvement in network prediction accuracy as gene-phenotype association knowledge increases and is added to the HPO database.

Applications of KidneyNetwork 

We have developed https://kidney.genenetwork.nl/ through which we provide gene-HPO term prediction. Using the same prediction algorithm that we used to assign genes to HPOterms, we also predicted which genes are likely to be involved in GO, KEGG, and Reactome pathways. Here we also provide an online version of GADO that can be used to prioritize relevant genes for patients with suspected rare kidney disease. It is possible to specify the phenotype of a patient using HPO terms and provide a list of genes harboring potential disease-causing variants. These genes will then be ranked using KidneyNetwork, thereby allowing the identifification of genes that are more likely to be involved in the patient’s disease. Since it is not necessary to upload personal genetic information, this method respects patient privacy. We advise using the KidneyNetwork scores in conjunction with WES or GWAS data to increase the prediction accuracy.

Future directions

Application of KidneyNetwork to unsolved cases from diagnostics, large research cohorts and, for instance, GWAS datasets will result in more insight into kidney physiology and pathophysiology. To further improve the accuracy of kidney phenotype prediction, we plan to build cell-type specifific networks by incorporating single-cell RNA-sequencing data, which we expect will yield more detailed and accurate gene-phenotype predictions. Conclusion We present KidneyNetwork, a kidney-specific co-expression network that accurately predicts which genes have kidney-specific functions. The method we developed to combine multi-tissue data with tissue-specific data can easily be extended to other tissues, allowing improved predictions for other tissue-specific diseases. Using KidneyNetwork, we highlight ALG6 as a candidate gene for kidney and/or liver cysts. KidneyNetwork provides a useful tool to help with the interpretation of genetic variants. It can therefore be of great value in translational nephrogenesis and ultimately improve the diagnostic yield in kidney disease patients. DATA AVAILABILITY The publicly available datasets analyzed during the current study are available in the European Nucleotide Archive (ENA) repository (https://www.ebi.ac.uk/ena/browser/ home). The GTEx-derived datasets are available from the database of Genotypes and Phenotypes (dbGaP), but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from dbGaP under accession number phs000424.v8.p2. The patient-derived WES datasets analyzed during the current study are not publicly available for privacy reasons. The results are available on kidney.genenetwork.nl.

REFERENCES 

1. Groopman EE, Marasa M, Cameron-Christie S, Petrovski S, Aggarwal VS, MiloRasouly H, et al. Diagnostic utility of exome sequencing for Kidney disease. N. Engl J Med. 2019;380:142–51. 

2. Snoek R, van Jaarsveld RH, Nguyen TQ, Peters EDJ, Elferink MG, Ernst RF, et al. The genetics-first approach improves the diagnostics of ESKD patients younger than 50 years. Nephrol Dial Transplant. 2020;37:349–57.

3. Connaughton DM, Kennedy C, Shril S, Mann N, Murray SL, Williams PA, et al. Monogenic causes of chronic kidney disease in adults. Kidney Int. 2019;95:914–28. 4. Cooper GM, Shendure J. Needles in stacks of needles: finding disease-causal variants in a wealth of genomic data. Nat Rev Genet. 2011;12:628–40. 5. van Dam S, Võsa U, van der Graaf A, Franke L, de Magalhães JP. Gene co-expression analysis for functional classification and gene–disease predictions. Brief Bioinform. 2017;19:575–92. 6. Deelen P, van Dam S, Herkert JC, Karjalainen JM, Brugge H, Abbott KM, et al. Improving the diagnostic yield of exome-sequencing by predicting gene–phenotype associations using large-scale gene expression analysis. Nat Commun. 2019;10:2837. 7. Köhler S, Carmody L, Vasilevsky N, Jacobsen JOB, Danis D, Gourdine JP, et al. Expansion of the Human Phenotype Ontology (HPO) knowledge base and resources. Nucleic Acids Res. 2019;47:D1018–27. 8. Online Mendelian Inheritance in Man, OMIM®. McKusick-Nathans Institute of Genetic Medicine, Johns Hopkins University (Baltimore, MD) [Internet]. Available from: https://omim.org/ 9. Orphanet: an online database of rare diseases and orphan drugs. Copyright, INSERM 1997. [Internet]. Available from: http://www.orpha.net 10. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics. 1988;44:837–45.

Supportive Service:

Email:wallence.suen@wecistanche.com 

Whatsapp/Tel:+86 15292862950

Shop:

https://www.xjcistanche.com/cistanche-shop