Gene-Specific Function Prediction for Non-Synonymous Mutations in Monogenic Diabetes Genes

The rapid progress of genomic technologies has been providing new opportunities to address the need of maturity-onset diabetes of the young (MODY) molecular diagnosis. However, whether a new mutation causes MODY can be questionable. A number of in silico methods have been developed to predict functional effects of rare human mutations. The purpose of this study is to compare the performance of different bioinformatics methods in the functional prediction of nonsynonymous mutations in each MODY gene, and provides reference matrices to assist the molecular diagnosis of MODY. Our study showed that the prediction scores by different methods of the diabetes mutations were highly correlated, but were more complimentary than replacement to each other. The available in silico methods for the prediction of diabetes mutations had varied performances across different genes. Applying gene-specific thresholds defined by this study may be able to increase the performance of in silico prediction of disease-causing mutations.

MODY caused by different gene mutations may have different severities of diabetes and different drug responses to diabetes medications [18]. For example, MODY2 (accounts for ,20% of all MODY cases [19]) caused by GCK gene mutations tends to have mild hyperglycaemia without obvious glycosuria. Patients with MODY2 are often asymptomatic [20] or only identified in women during pregnancy and diagnosed as gestational diabetes [21]. Most patients with MODY2 can have blood glucose satisfactorily controlled by diet therapy and don't need hypoglycemic medication [22,23]. In contrast, MODY3 caused by HNF1A mutations, the most common type of MODY that accounts for ,63% of all MODY cases [19], tends to have obvious glycosuria because of impaired glucose-stimulated insulin secretion [24], as well as decreased renal threshold for glucose [25]. MODY3 patients tend to have good response to sulphonylurea treatment and don't rely on insulin therapy [18,26]. Because of the implications of pharcogenetics and personalized medicine, molecular diagnosis of MODY has clinical importance for clinical decision and for genetic counseling [18,26]. However, because of unavailability and expense of MODY molecular diagnosis, it is not uncommon that MODY patients are classified as type 2 diabetes [27,28] and occasionally as type 1 [29].
The rapid progress of advanced genomic technologies has been providing new opportunities to address the need of MODY molecular diagnosis. The identification of mutations in MODY genes by sequencing technologies will enable the molecular diagnosis of MODY, whereas a new issue is emerging. Most mutations causing MODY are nonsynonymous single-nucleotide mutations causing the change of an amino acid residue (according to The Human Gene Mutation Database [30], http://www.hgmd. org/). High throughput sequencing technologies enable screening of a large number of patients and parallel sequencing of a large number of genes. If a known MODY gene mutation is identified in a patient suspected of MODY, the molecular diagnosis of MODY can be established. However, the increased throughput of sequencing technologies is likely to produce increased numbers of missense variants whose causative role in MODY can be questionable. Bioinformatics tools, e.g. SIFT (http://sift.jcvi.org/) RadialSVM score .0.5 RadialSVM score is an ensemble-based approach integrating multiple scoring systems (function prediction and conservation Score) by radial support vector machine (SVM) [3].
LR score .0.5 LR score is an ensemble-based approach integrating multiple scoring systems (function prediction and conservation Score) by logistic regression (LR) [3].
* Extensively comparisons of these methods genome-widely have been studied by Thusberg et al. [1,2] and Dong et al. [3]. Deleterious thresholds for PhyloP, GERP++ RS and SiPhy are defined according to the study by Dong et al. [3].  [31] and PolyPhen (http://genetics.bwh.harvard.edu/pph2/ index.shtml) [32], are often used to assess the pathogenicity of a nonsynonymous mutation [27]. Because the limitations of in silico methods, the functional prediction of a nonsynonymous mutation lacks a gold standard. To date, a number of bioinformatics methods besides SIFT and PolyPhen, based on different algorithms, have been developed [1][2][3]. The purpose of this study is to compare the performances of different bioinformatics methods in the functional prediction of nonsynonymous mutations in each MODY gene, and to provide reference matrices to assist the molecular diagnosis of MODY.

Data source
The diabetes mutation data analyzed in this study were acquired from the Human Gene Mutation Database (HGMD) 2013.4 release (http://www.hgmd.org/) [30]. As the purpose of this study is to assess the prediction performances of different in silico methods for nonsynonymous single-nucleotide mutations, insertion/deletion mutations (InDels) are not included in this study. Altogether, 1,130 nonsynonymous single-nucleotide mutations from 24 genes have been reported causing MODY or neonatal diabetes. Among these genes, 7 genes harbor more than 30 single-nucleotide mutations within each gene with the total of 1,091 diabetes mutations ( Control nonsynonymous single-nucleotide mutations in the diabetes genes were acquired from the NHLBI GO Exome Sequencing Project (ESP) [33,34], the ARIC samples [35] in the CHARGE Exome Sequencing Project [36], and the 1000 Genome Project [37], excluding mutations recorded in the HGMD database.

Functional prediction of nonsynonymous singlenucleotide mutations
Eleven methods, including PhyloP [38], GERP++ RS [39], SiPhy [40], SIFT [31], PolyPhen-2 [32], the likelihood ratio test (LRT) [41], MutationTaster [42], Mutation Assessor [43], FATHMM [44], RadialSVM score [3], and logistic regression (LR) score [3], were covered in the dbNSFP database [45,46] and compared in this study (Table 2). Among the 1,091 mutations involved in this analysis, 104 mutations from the genes GCK, HNF1A, HNF1B, HNF4A, and INS, are nonsense mutations, i.e. producing a pre-termination codon; two other mutations from the gene GCK replace a termination codon with an amino acid codon. For these mutations, the methods, PolyPhen-2 HDIV, PolyPhen-2 HVAR, MutationAssessor, FATHMM, Radial SVM score, LR score are not applicable to nonsense mutations or mutations assumed with highly damaging potential. Other methods, except MutationTaster, tend to have higher error rates (false negative rates, FNR), compared to the prediction of amino acid substitution mutations, i.e. SIFT FNR = 72%, GERP++ RS FNR = 40%, PhyloP FNR = 35%, SiPhy FNR = 26%, LRT FNR = 21%, and MutationTaster FNR = 3%, for the prediction of nonsense mutations. Compared with amino acid substitutions, the assessment of the functional effect of nonsense mutations tends to be less of an issue. The final analysis of this study involved 985 nonsynonymous single-nucleotide mutations. The quantitative performances of these methods were compared by the Spearman's rank correlation test [47] and the ANOVA test using the IBM SPSS Statistics 19 software ((IBM SPSS Inc., Chicago, IL, USA). To re-define gene-specific thresholds of deleterious mutations, the receiver operator characteristic (ROC) analysis was calculated by the sensitivity and specificity values of screening series of cutoffs of each method for each gene. A redefined threshold was identified using the maximum Matthews correlation coefficient (MCC) [48].

Results and Discussion
In our analysis, the prediction scores by different methods of the diabetes mutations are highly correlated ( Table 3). The highest correlations are seen between RadialSVM score and LR score (r = 0.957), PolyPhen-2 HDIV and PolyPhen-2 HVAR (r = 0.89),  and phyloP and GERP++ RS (r = 0.871), while the other correlations have r,0.80. Therefore, in spite of the high statistical significance of correlations between different methods, different methods may not be able to replace each other except for the above three pairs. Especially, the FATHMM method has no obvious correlation with PhyloP, GERP++ RS, and LRT, while the correlation with MutationTaster is less significant. On the other hand, we observed significantly varied performances in detecting deleterious mutations by different methods (Table 4) Our investigation further disclosed significant differences of the quantitative performances of different methods, except SIFT, across different genes (Table 5). Varied performances across genes highlight another aspect of limitation of these in silico methods. The distribution of the prediction scores presented in Table 5 may be able to serve as a matrix to assist the assessment of functional effects of new mutations in these diabetes genes.
The varied performances of these methods in different genes and the different scores of each method for different genes suggest that using gene-specific thresholds for deleterious mutations may improve the prediction performance of these in silico methods. We screened each gene and identified the gene-specific threshold with maximum MCC. Nonsynonymous single-nucleotide mutations in the diabetes genes from the NHLBI GO Exome Sequencing Project (ESP) [33,34], the ARIC samples [35] in the CHARGE Exome Sequencing Project [36], and the 1000 Genome Project [37], were used as controls without including mutations recorded in the HGMD database. Shown by our analysis (Table S1), we have been able to improve the prediction performance of each method in most cases, with the FATHMM method as an exception because of its nil/low FNRs in those diabetes genes. For example, the FNR of GERP++ RS for HNF4A mutations and the FNR of LRT for HNF1B mutations were decreased without any obvious change of their FPRs. On the other hand, redefined thresholds are able to decrease the FPRs of LRT for INS mutations, MutationTaster for ABCC8 mutations, LR score for INS mutations, LRT for ABCC8 mutations, MutationTaster for INS mutations, and MutationTaster for HNF1B mutations, without obviously increasing the FNRs. The general performances of different methods were summarized in Table 6. From low to high MCCs, the methods were sorted from left to right and from top to bottom. The average difference of MCCs and P value of each two methods was shown.
The varied performance of different methods in different genes is related to specific molecular mechanisms of diabetes mutations. For the 41 INS mutations involved in this study, 34 mutations cause neonatal diabetes. These mutations exert diabetic effects by causing misfolding of the insulin protein, rather than inactivating the gene [49,50]. The dominantly inherited mode of the disease is from dominant negative mechanism, instead of haploinsufficiency. The misfolded insulin protein interferes cellular processes, leading to severe endoplasmic reticulum stress and potentially b cell death by apoptosis [50]. In contrast, a heterozygous individual with one copy of inactivating INS mutation may still have a sufficient response to metabolic regulation, thus without neonatal diabetes. For the prediction of neonatal diabetes mutations in the INS gene, a protein structure-based prediction method may thus have better performance than others. In this study, we see that PolyPhen-2 with structure-based predictive features has better performance than the more sequence-based SIFT method (Table S1). Unlike other monogenic diabetes genes, the neonatal diabetes mutations in ABCC8 and KCNJ11 are gain-of-function mutations [51]. Sequence-based method like SIFT has also lower performance for these mutations than PolyPhen-2.
We acknowledge the current publication bias of diabetes mutations (i.e. the bias towards identifying and reporting diabetes-causing mutations in the general human population). The diabetes mutations have been identified by studies involving much larger number of human individuals, while the genome sequencing projects involved limited number of human subjects. For a disease-causing mutation, no matter its low frequency, as long as the mutation is identified, it will be included. For example, in the case of GCK and HNF1A genes, the numbers of reported diabetes mutations are much larger than control mutations (479 vs. 22, 324 vs. 78, respectively). We also want to emphasize the application of gene-specific mutations as functionally neutral controls. Our analysis showed that different methods using redefined thresholds by genome-wide control mutations, instead of gene-specific controls, tend to have poor performances (data available upon request). To acquire a satisfactory MCC tends to need a large number of both diabetes mutations and functional neutral mutations. The gene-specific prediction model proposed by our study will have further improved performance with the availability of sequencing data of a larger number of human individuals.
In conclusion, the available in silico methods for the prediction of diabetes mutations have varied performances across different genes. In spite of the high statistical significance of correlations between different methods, different methods may not be able to replace each other. Because of varied performances across genes, applying gene-specific thresholds when possible (i.e. for genes with a number of disease mutations identified and the ROC analysis feasible) may be able to increase the performance of prediction. For genes without sufficient numbers of mutations for the ROC analysis, a consensus threshold should be used [52]. Nevertheless, the limitations of the above methods warrant that new methods are being developed continuously. For example, Johansen et al. recently developed a sequence conservation-based artificial neural network predictor called NetDiseaseSNP [53]. Capriotti et al. developed a Meta-SNP algorithm for the detection of diseaseassociated nsSNVs, which integrates four different methods: PANTHER, PhD-SNP, SIFT and SNAP. They showed these methods are orthogonal with different biologically relevant relationships, and the integration of different methods achieved higher accuracy [54].

Supporting Information
Table S1 Method comparisons for gene-specific function prediction for non-synonymous mutations causing diabetes. (XLS)