Variant predictions in congenital adrenal hyperplasia caused by mutations in CYP21A2

CYP21A2 deficiency represents 95% of congenital adrenal hyperplasia (CAH) cases, a group of genetic disorders that affect steroid biosynthesis. The genetic and functional analysis provide critical tools to elucidate complex CAH cases. One of the most accessible tools to infer the pathogenicity of new variants is in silico prediction. Here, we analyzed the performance of in silico prediction tools to categorize missense single nucleotide variants (SNVs) of CYP21A2. SNVs of CYP21A2 characterized in vitro by functional assays were selected to assess the performance of online single and meta predictors. SNVs were tested separately or in combination with the related phenotype (severe or mild CAH form). In total, 103 SNVs of CYP21A2 (90 pathogenic and 13 neutral) were used to test the performance of 13 single-predictors and four meta-predictors. All SNVs associated with the severe phenotypes were well categorized by all tools, with an accuracy of between 0.69 (PredictSNP2) and 0.97 (CADD), and Matthews’ correlation coefficient (MCC) between 0.49 (PoredicSNP2) and 0.90 (CADD). However, SNVs related to the mild phenotype had more variation, with the accuracy between 0.47 (S3Ds&GO and MAPP) and 0.88 (CADD), and MCC between 0.18 (MAPP) and 0.71 (CADD). From our analysis, we identified four predictors of CYP21A2 variant pathogenicity with good performance, CADD, ConSurf, DANN, and PolyPhen2. These results can be used for future analysis to infer the impact of uncharacterized SNVs in CYP21A2.

1 Introduction

One of the most common autosomal recessive genetic disorders is the impairment of the steroid 21-hydroxylase (CYP21A2). The CYP21A2 deficiency represents about 95% of cases in congenital adrenal hyperplasia (CAH), a group of enzymatic disorders that affect cortisol biosynthesis. The CYP21A2 enzyme is a member of the cytochrome P450 superfamily (CYPs) and catalyzes the conversion of 17α-hydroxyprogesterone (17OHP) into 11-deoxycortisol and progesterone into 11-deoxycorticosterone. Other enzymes subsequently convert these steroids into cortisol and aldosterone (Miller and Auchus, 2011).

Clinically, the CYP21A2 deficiency in humans has a wide spectrum of phenotypes, from severe to mild or asymptomatic (New et al., 2013; Witchel, 2017). The classic severe CAH has salt-wasting (SW) and simple-virilizing (SV) forms. The classical SW form has no enzyme activity and is related to severe virilization and electrolyte imbalance. In contrast, the classical SV form has enough residual enzyme activity to prevent adrenal crisis (Witchel, 2017). Mild CAH is the non-classical (NC) form of CAH and has CYP21A2 activity associated with hyperandrogenism and mild late-onset CAH (New et al., 2013). Furthermore, there is a relatively good genotype-phenotype correlation for CYP21A2 deficiency, which allows the categorization of variants according to the residual enzyme activity (obtained from in vitro studies) and their expected phenotype (New et al., 2013). The classical CAH (CL) has less than 10% of wild-type (WT) enzyme activity in 95% of the cases, while the NC form has an activity of between 10 and 78% of the WT in 90% of the cases, as reported by Simonetti et al., (Simonetti et al., 2018).

The CYP21A2 gene is a tandemly arranged module (RCCX: RP-C4-CYP21-TNX) and shows 96–98% of sequence identity with its pseudogene, CYP21A1P (Rodrigues et al., 1987). These features make the CYP21A2 gene analysis a complex endeavor, with many different types of mutations—from single nucleotide variants (SNVs) to genetic rearrangements—and further complicated by the fact that most carriers have compound heterozygous mutations (New et al., 2013). However, only ten mutations described in the general population are sampled by CYP21A2 deficiency screening programs. Whole gene sequence analysis by Sanger sequencing is an alternative method in exceptional cases due to the cost and time-consuming nature of such studies (Stenson et al., 2017; Baumgartner-Parzer et al., 2020).

So far, with the whole CYP21A2 gene sequenced, genetic studies have reported more than 1,300 variants in the CYP21A2 gene. Out of the 230 variants reported as affecting human health, 153 are missense variants (Simonetti et al., 2018). The advancement of next-generation sequencing (NGS) to analyze a large number of genes has facilitated the detection of rare single nucleotide variants (SNVs). A few years ago, this technology was not applied to screen the CYP21A2 gene defects due to its high sequence identity with CYP21A2P, which hampers the proper analysis of this genomic region (Rodrigues et al., 1987). However, recently, some groups have found alternative ways to perform NGS for the CYP21A2 gene through a combination with other methods, such as multiplex ligation-dependent probe amplification (Gangodkar et al., 2021; Lee et al., 2021). These genetic analysis strategies of the CYP21A2 gene with NGS technology represent a promising tool for the future, opening the window to identify new variants while improving the diagnosis of CYP21A2 deficiency and establishing a more reliable estimate of mutation frequencies.

The gold standard for the characterization of new CYP21A2 variants is the in vitro functional assay. However, this approach takes too much time, and it is not a viable option for the analysis of all the new variants detected by sequencing studies. One of the most accessible tools to predict the pathogenicity of variants is in-silico analysis, which usually has free access, a friendly interface, and provides quick results. Many online predictors are available that have different features and approaches, from single characteristic analysis to meta-predictors with different compositions and algorithms. Some studies have shown the general performance of these tools against a whole database with few predictors (Hicks et al., 2011; Tang et al., 2020). However, studies with variants on protein-specific analysis showed that general analysis results cannot be extrapolated for all proteins as each protein has unique characteristics, which is a key limitation of predictor programs (Pshennikova et al., 2019; Hart et al., 2020; Montenegro et al., 2021). Therefore, it is essential to be careful when choosing the prediction tools and to consider their variable accuracies for each gene (Tang et al., 2020).

Here we have done a meta-analysis of the performance of online predictor tools to classify missense SNVs of CYP21A2. Missense SNV is the most common group of variants in the human genome, one at every kilobase. In the CYP21A2 gene, this type of SNV represents about 60% of the CYP21A2 variants in The Human Gene Mutation Database (HGMD, RRID:SCR_001888) (Stenson et al., 2017) and 65% of those affecting human health (Simonetti et al., 2018). Additionally, missense SNV is one of the hardest variant types to interpret (Khan and Vihinen, 2007; Simonetti et al., 2018; Pignatelli et al., 2019). In total, we analyzed 17 predictors with multiple algorithms, approaches, and datasets. Thirteen of these were based on single features: CADD (RRID:SCR_018393) (van der Velde et al., 2017), ConSurf (RRID:SCR_002320) (Ashkenazy et al., 2016), DANN (Quang et al., 2015), FATHMM (Shihab et al., 2013), MAPP (RRID:SCR_014375) (Stone and Sidow, 2005), MutPred2 (RRID:SCR_010778) (Pejaver et al., 2020), PANTHER-PSEP (RRID: SCR_005145) (Tang and ThomasPANTHER-PSEP, 2016), PhD-SNP^g (RRID: SCR_010782) (Capriotti and FariselliPhD-SNPg, 2017), PolyPhen-2 (RRID:SCR_013189) (Adzhubei et al., 2013), PROVEN (RRID: SCR_002182) (Choi et al., 2012), SIFT (RRID:SCR_012813) (Vaser et al., 2016), SNAP2 (RRID:SCR_002127) (Hecht et al., 2015), and SNPs&GO (RRID:SCR_005788) (Calabrese et al., 2009). Four meta-predictors: PredictSNP (Bendl et al., 2014), PredictSNP2 (Bendl et al., 2016), Meta-SNP (Capriotti et al., 2013), and S3Ds&GO (Capriotti and Altman, 2011)) (Figure 1). We excluded nonsense and frameshift variants from our analysis since they have specific settings in some predictors which are not applied for single amino acid substitution and a high agreement ratio between tools.

FIGURE 1

FIGURE 1. Composition of the four meta-predictors studied. The PredictSNP algorithm comprises the outputs of six single-predictors, the PredictSNP2 of six, the Meta-SNP of 4, and the S3Ds&GO of three predictors.

2 Results

2.1 Data of the selected SNVs

From variants in the CYP21A2 gene reported in the literature and databases, we selected missense SNVs with clinical significance, using the criteria described in Section 4.1. We obtained 96 valid SNVs out of 299 missense variants in the list (Simonetti et al., 2018), 85 out of 614 missense variants in dbSNP, 66 out of 459 missense variants in Ensembl, 45 out of 71 missense variants in GeneCards, 47 out of 83 missense variants in ClinVar, 26 in OMIM, and 81 found in the UniProt database.

By removing SNVs that were duplicated and with no functional characterization, we obtained 103 SNVs, 51 classified as classical, 39 as non-classical, and 13 as neutral. The SNVs selected with the respective enzyme activity are described in Supplementary Table S1. All studies presented the CYP21A2 activity measured by the hydroxylation of 17OHP, while 86 also measured progesterone hydroxylation. Mutations of the CL group have a mean enzyme activity for the 17OHP hydroxylation of 1.5 ± 2 (SD)% and the progesterone hydroxylation of 1.3 ± 1.6%. While mutations of the NC group have 17OHP hydroxylation activity of 42.9 ± 23% and progesterone hydroxylation activity of 37.4 ± 21%. Finally, mutations in the neutral group have 17OHP hydroxylation activity of 100.17 ± 11% and progesterone hydroxylation activity of 94.1 ± 8.25%.

2.2 General analysis of mutation groups

We obtained 22 SNVs in the CYP21A2 gene with the correct prediction for all tested predictors, although the exact number was incorrectly predicted for at least half of them (Figure 2). There was no SNV, as wrongly predicted by all predictors. We compared the hit and miss by the 17 predictors for all SNVs affecting CYP21A2 activity, the CAH group, and all the non-pathogenic SNVs of the neutral group. We showed that 24% (22 of 90) SNVs from the CAH group obtained the correct score by all predictors, while 23% (21 of 90) were wrongly predicted by at least nine tools. The neutral group got two of its 13 SNVs (15.4%) rightly predicted by all and one by nine predictors. Moreover, we divided the SNV of the CAH group into the CL and NC groups. We got 37% (out of 51 SNVs) from the CL and 2.6% (of 39 SNVs) from the NC group of SNVs correctly predicted by all tools. While 5.9% and 46% of CL and NC groups, respectively, were wrongly predicted by nine tools.

FIGURE 2

FIGURE 2. Frequency of hit and miss obtained for each SNV by mutation group. Each SNV (vertical list) was analyzed by seventeen predictors (horizontal measurement) performed with the default setting for missense mutation. Please, refer to Supplementary Table S2–S4 for details. TP, true positive; TN, true negative; FN, false negative; FP, false positive.

Some critical amino acid positions have two or three pathogenic replacements (e.g., p.P31Q/L, p.V282G/L, and p.R357W/Q/P) (Figure 2). Two tools of prediction were able to identify the critical amino acid position for all of those double/triple mutants: CADD and ConSurf (Supplementary Table S2–S4). As those variants have the functional data available, CADD was able to use that information together with gene annotation, epigenetic and evolutionary data, outputting the right prediction. On the other hand, ConSurf obtained the same result using structural, phylogenetic, and evolutionary data.

2.3 Performance of predictors to identify SNVs detrimental to CYP21A2 activity

We analyzed the performance of 17 predictors to identify the 90 SNVs that affect CYP21A2 function against the 13 SNVs with a neutral effect (Table 1). All the predictors tested obtained a good PPV rate (>0.90). However, only CADD (0.73) and DANN (0.56) showed negative predictive values (NPVs) higher than 0.5. The PANTHER-PSEP found no result for the NPV, as it could not identify benign variants.

TABLE 1

TABLE 1. Performance of 17 programs to predict the effect SNVs in the CYP21A2. We performed the analysis with 103 functionally characterized variants, 90 damaging the protein functionality, and 13 neutral—Color scores from blue (good result) to yellow (not good).

We obtained both sensitivity and specificity higher than 0.8 for three predictors, CADD (sensitivity, e, = 0.96 and specificity, sp, = 0.85), ConSurf (se = 0.88 and sp = 0.90), and PolyPhen-2 (se = 0.87 and sp = 0.85). Moreover, five predictors obtained accuracy between excellent and good: CADD (0.94), PANTHER-PSEP (0.91), DANN (0.89), ConSurf (0.88), and PolyPhen-2 (0.86) (Table 1). The Matthews’ correlation coefficient (MCC) test showed positive values for that of almost all predictors (except by PANTHER-PSEP), being five of them with an MCC >0.5 (Table 1). The greatest performance, a value closer to +1, was obtained by CADD (0.75), followed by ConSurf (0.58), PolyPhen-2 (0.57), DANN (0.56), and PROVEN (0.51). Figure 3 shows a Venn diagram of the four predictors with better accuracy and MCC values.

FIGURE 3

FIGURE 3. Overlaps of the hit for four predictors with good-excellent performance for SNVs on CYP21A2. Total, indicates all 103 SNVs tested. (Image generated with https://bioinformatics.psb.ugent.be/webtools/Venn/).

2.4 Performance of predictors to identify SNVs affecting the specific CAH groups

We analyzed the performance of the selected predictors to identify 51 SNVs of the CL group and 39 SNVs of the NC group against 13 SNVs with a neutral effect (Table 2). Seventeen predictors obtained an excellent positive predictive value (PPV) rate (>0.90) for the SNV CL group and 15 for the SNV NC group. Four tools obtained excellent-good (>0.8) negative predictive values (NPV) values for the CL group: CADD (1.0), DANN (0.9), PolyPhen-2 (0.85), and ConSurf (0.82). However, for the NC group, only CADD (0.73), DANN (0.6), and PolyPhen-2 (0.52) showed NPV >0.5. Taking the sensitivity and specificity balance, we obtained 12 tools with excellent-good values (>0.8) for the CL group. However, for the NC group, only CADD (se = 0.9 and sp = 0.85) obtained both sensitivity and specificity with excellent-good values. The accuracy was excellent for seven predictors in the CL group: CADD (0.97), ConSurf (0.95), PolyPhen-2 and PROVEN (0.94), DANN and MutPred2 (0.92), and Meta-SNP (0.91); and good for four tools in the NC group: CADD (0.88), DANN and PANTHER-PSEP (0.81), and ConSurf (0.8).

TABLE 2

TABLE 2. Performance of 17 programs for the specific CYP21A2 groups. We predict the effect of SNVs in the CYP21A2 by dividing them by the two levels of protein damage: severe (classical mutation, CL group) and mild (non-classical mutation, NC group). We performed the analysis with 103 SNVs of known effect, 51 being CL, 39 NC, and 13 neutral. Color score from blue (good result) to yellow (not good).

Finally, the MCC test with positive values was obtained by almost all predictors (except for PANTHER-PSEP). The MCC was higher than 0.5 for 16 predictors in the CL group and five in the NC group (Table 2). For the CL and NC groups, the same predictor, CADD, got the MCC value closer to +1, with MCC = 0.9 for the CL group and MCC = 0.71 for the NC group.

3 Discussion

Genetics analysis is an essential approach for elucidating complex CYP21A2 deficiency cases, mainly to confirm asymptomatic carriers and unfollow false-positive cases (Pignatelli et al., 2019; Baumgartner-Parzer et al., 2020). Therefore, fast and accessible tools to infer variants’ pathogenicity are essential to quickly deduce the harm of unknown variants. In silico prediction is one of the most accessible tools to infer the pathogenicity of SNVs. Here, for the first time, we analyzed the performance of in silico prediction tools to discriminate between pathogenic and neutral variants of CYP21A2. We focus on the performance of 13 single predictors and four meta predictors chosen accordingly to the popularity and performance of free-access programs. Although some programs are based on the same data set, each one has particular features (e.g., evolutionary, epigenetic, functional, gene annotation, protein structure), algorithms (e.g., machine learning, matrix of effect, alignment), and databanks (e.g., ClinVar, SwissVar, UniProt, UniRef90, HumVar). Furthermore, they also differ in the input data (protein change, nucleotide change, chromosomal position, and accession number of the gene). The 17 programs chosen in the present study are composed of distinct combinations of features, databanks, and algorithms (Supplementary Table S5). All of them were able to identify pathogenic variants. Nonetheless, only PANTHER-PSEP could not distinguish neutral variants, which is unacceptable for testing variants of the CYP21A2. Moreover, all tools showed better performance with variants of the CL group than the NC group, as expected, since the CL group gathers the most harmful variants.

Our databank for performance tests comprises all missense variants of CYP21A2 that are functionally characterized. With this strategy, we could get a more realistic result on the prediction evaluation. However, the number of variants was imbalanced between the two categories, with 90 pathogenic and 13 neutral. Therefore, the primary statistics data considered for the performance evaluation were the accuracy and MCC, which consider all values—true positive (TP), true negative (TN), false positive (FP), and false negative (FN) (Vihinen, 2012; Chicco et al., 2021). As sensitivity and specificity are calculated with half of the information, they cannot represent all the performance by themselves, so we considered the sensitivity-specificity balance. Additionally, we calculated PPV and NPV but, as both are more sensitive to data disbalance, they were not considered for the program performance (Vihinen, 2012).

The main feature assessed by most single predictors tested is the evolutionary data since residue conservation over time can indicate critical residues for the protein function. Four of the tools tested use only this feature for the prediction calculation, FATHMM (Shihab et al., 2013), PhD-SNPg (Capriotti and FariselliPhD-SNPg, 2017), PROVEAN (Choi et al., 2012), and SIFT (Vaser et al., 2016). A similar performance was obtained between these four tools, with a fair accuracy ranging from 0.73 (FATHMM) to 0.79 (SIFT), and MCC from 0.42 (PhD-SNPg) to 0.51 (PROVEAN). Additionally, SIFT had the most excellent sensitivity-specificity balance between programs, similar to the performance shown previously (Vaser et al., 2016). In another study, Montenegro et al. (Simonetti et al., 2018), with HSD17B3, NR5A1, AR, and LHCGR genes, SIFT and PROVEAN also had the same performance, with an accuracy of 0.74–0.75, and MCC of 0.5. In yet another study (Pshennikova et al., 2019), SIFT and PROVEAN showed the best results among nine programs tested for GJB2, GJB6, and GJB3 genes, with an accuracy of 0.89, while FATHMM produced a large number of erroneous predictions with an accuracy of 0.33. FATHMM also had poor performance in another study (Montenegro et al., 2021), with an accuracy of 0.56 and a MCC of 0.04.

Changes in the secondary and tertiary structures by missense mutations are likely to affect protein activity (Ashkenazy et al., 2016). Therefore, it is no surprise that the second most evaluated feature is structural information, being present in ConSurf (Ashkenazy et al., 2016), MutPred2 (Pejaver et al., 2020), PolyPhen2 (Hicks et al., 2011), and SNAP2 (Hecht et al., 2015). In addition, ConSurf includes phylogenetics relationships, MutPred2 functional proprieties, and SNAP2 uses a matrix of effect probabilities with a neural network method (Hecht et al., 2015; Ashkenazy et al., 2016; Pejaver et al., 2020). While PolyPhen2 has two trained datasets as options, HumVar and HumDir. The first trained dataset is suggested for diagnostics of Mendelian diseases, which requires variants with a drastic difference effect (Adzhubei et al., 2013). Moreover, PolyPhen2 has a low dependency on the sequence alignment employed (Hicks et al., 2011). PolyPhen2 showed good prediction performance (ac = 0.88, MCC = 0.64), with the same sensitivity but a better sensitivity-specificity balance than reported in (Hicks et al., 2011). ConSurf obtained a similar performance with the setting used in the test (ac = 0.88, MCC = 0.58). However, we cannot compare the ConSurf performance with other studies since there is no default setting for the alignment sequence, database, and algorithms available, as mentioned by ConSurf’s developers (Ashkenazy et al., 2016). MutPred2 and SNAP2, both had sensitivity-specificity imbalances of 1.4 and 1.5-folds, respectively, and almost the same fair performance. However, these values were relatively better than reported for MutPred2 by (Pejaver et al., 2020) with the ClinVar and UniProt databases and SNAP2 by (Hecht et al., 2015) with a databank with more than 9,500 variants from human genes.

For meta predictors, which work with many databases and combine outputs from other predictors to generate their own, we expected to obtain one of the best performances. However, counterintuitively, our study showed an intermediate performance compared with the single predictors tested, and the number of tools combined was not related to the prediction improvement. Meta-SNP and PredictSNP performance were better than PredictSNP2 and S3Ds&GO. Furthermore, compared with the developer tests, we obtained for Meta-SNP (Capriotti et al., 2013) and PredictSNP (Bendl et al., 2014) similar accuracy, while the performance for PredictSNP2 (Bendl et al., 2016) and S3Ds&GO (Capriotti and Altman, 2011) was lower. Meta-SNP and PredictSNP share three single predictors, PhD-SNP, SNAP, and SIFT.

Therefore, for the CYP21A2 variants tested, CADD presented the best performance to categorize the pathogenicity of the missense variants, with an overall accuracy of 0.94 (CL 0.97; NC 0.88) and MCC of 0.75 (CL 0.9, NC 0.71). The specificity (0.85) and sensitivity (0.9) also reached a good balance. Interestingly, the accuracy and specificity obtained for CYP21A2 were even higher than reported by the software developers (van der Velde et al., 2017) with the ClinVar database, which were 0.85 and 0.57, respectively. ConSurf, DANN, and PolyPhen-2 showed similar performance, giving the second-best results according to the accuracy (CAH group 0.86–0.89; CL 0.92–0.95; NC 0.77–0.81) and MCC (CAH group 0.56–0.58; CL 0.75–0.83; NC 0.51–0.56) values. The sensitivity and specificity for ConSurf and PolyPhen2 were well balanced, while DANN had 1.3-fold less specificity than sensitivity. The original article of DANN (Quang et al., 2015) presents only the area under the curve (AUC) ROC, which was 0.95 using the ClinVar database for the performance test. PolyPhen2 showed a better sensitivity-specificity balance for the CYP21A2 variants than the values presented by (Hicks et al., 2011), testing the tool with gene-specific mutations (BRCA1, MSH2, MLH1, and TP52). We obtained a sensitivity similar to a previous analysis (Hicks et al., 2011), but the specificity was lower, at 0.85 and 0.60, respectively.

Considering the individual errors of the four predictors with the greatest performance for the CYP21A2 variants analyzed in this study, CADD had six errors, ConSurf 12, DANN 11, and PolyPhen2 14. However, computing their predictions together, we would have four false results from 103 missense SNVs in the CYP21A2, one neutral (p.L13M), and three pathogenic from the NC phenotype group (p.P106L, p.R225W, and p.M474I). The variant p.P106L was correctly categorized by SNAP2 and PANTHER-PSEP. In turn, PROVEAN (Choi et al., 2012) could type the other three variants correctly, even with a lower sensitivity value than the other four tools, mainly for the pathogenic SNVs of the NC group (0.49). Nonetheless, we obtained an intermediated performance with PROVEAN (ac = 0.77; MCC = 0.51), which could be because it uses the neighborhood sequences as input, which can be a trick for enzymes since they have some residues with high conservation, making critical connections between variable residues. In comparison using the developer’s test (Choi et al., 2012) with the UniProt database (se = 0.78; sp = 0.79), we had imbalanced sensitivity-specificity performance, with similar overall sensitivity (0.73) and higher specificity (1.0).

One of the limitations associated with this analysis is the unbalanced number of variants in each group analyzed due to the limited number of variants characterized. In order to diminish this issue, only statistical approaches considering the whole data were used for the performance comparison, as recommended in the literature (Vihinen, 2012; Chicco et al., 2021). Another limitation was the selection of the tools available online since each predictor has a specific input such as chromosomal version, the number of variants per input, and data input type. For that reason, we established parameters to choose a limited number of predictors based on the features and literature citations.

4 Materials and methods

4.1 SNVs selection and categorization

To select CYP21A2 missense SNVs with clinical significance, we used the list of variants reported to affect human health, as reviewed by Simonetti et al. (Simonetti et al., 2018) Complementarily, we searched for SNVs reported by dbSNP, Ensembl, and GeneCards, applying the following filters when present: “missense,” “clinical significance,” “pathogenetic,” or “benign,” “without conflicting interpretation”, and “human or homo sapiens”. We excluded nonsense and frameshift mutations. In addition, we performed a cross-check of the databases with original articles or reviews to remove variants without enzyme activity data available.

To standardize the effect of the SNVs selected, we categorized them into three groups according to the CYP21A2 activity observed by Simonetti et al. (Simonetti et al., 2018) for at least one of the steroid substrates: i) CL group has SNVs with the CYP21A2 activity level <10% relative to WT; ii) NC group has SNVs with the activity level between >10% and <78% relative to WT; and iii) the neutral group has SNV with the enzyme activity >78% relative to WT. The CAH group is composed of all SNVs from the CL and NC. The mean and standard deviation (SD) of enzyme activity were calculated for each steroid and mutation group.

4.2 Selection of predictor tools

To choose predictors with different features, we reviewed the literature for software applied to in silico analysis of SNPs or SNVs. Predictors used in more than two studies by different research groups or significant performance in a large study were selected. In addition, we filtered for tools with free access and online availability, thus not requiring local powerful computational resources. The characteristics of each predictor chosen are shown in Supplementary Table S5–S6.

4.3 Data treatment

The default setting for missense mutation was used on all predictors. However, when there was no set instruction for that purpose highlighted for the program, we followed the developers’ recommendation from the tutorial or original paper. In addition, three scores were extracted indirectly from meta-predictors: MAPP (v.28.6.2005) and SIFT (v.4.0.4) scores were obtained from PredictSNP, and DANN (v.1.2) score from PredictSNP2. For statistical purposes, we standardized two variables for the outputs of all the predictors: “damage” for SNVs with the potential to affect CYP21A2 and “neutral” for SNVs with no or very low potential to affect the enzyme. The following outputs were standardized as “damage”: score >0.5 to Meta-SNP, SNP&GO, S3D&GO, MutPred2, FATHMM-MKL (weighted) and PhD-SNP^g; “deleterious” message to PredictSNP, PredictSNP2, MAPP, and SIFT; score >0.9 to DANN; score > -2.5 to PROVEN; score >10 to CADD (GRCh38-v1.5–6); score >0.45 PolyPhen-2 (HumVar); score <0 to ConSurf; score >0 to SNAP2; and score >450 millions of years to PANTHER. Otherwise, we classified the outputs as “neutral".

4.4 Analytical parameters

We analyzed the performance of each predictor in two ways. First, to assess the performance to discriminate the effects of CYP21A2 SNVs, we compared SNVs of the CAH group with the neutral group. Second, to get the number of hits and misses per group, we analyzed CL and NC groups separated from the neural group. We used Microsoft Excel for the data organization and, together with IBM SPSS Statistics software v.2.1, we performed the statistical analysis.

4.5 Statistical methods

We considered the TP result for correct “damage” prediction, TN for correct “neutral” prediction, FP for incorrect “neutral” prediction, and FN for incorrect “damage” prediction. We calculated the PPV to access the ratio of TP results for all positive results (Eq. 1), and the NPV to the ratio of TN for all negative results (Eq. 2).

$\mathit{P}\mathit{P}\mathit{V}=\frac{TP}{TP+FP}(1)$

$\mathit{N}\mathit{P}\mathit{V}=\frac{TN}{TN+FN}(2)$

The proportion of correct SNVs identified as harmful was assessed with the sensitivity (Se) equation (Eq. 3), while the correct neutral identification was assessed with the specificity (Sp) (Eq. 4). Besides that, we obtained the accuracy (Ac) by the ratio of true results (TP and TN) (Eq. 5). The accuracy was classified as excellent (0.9 < Ac < 1.0), good (0.8 < Ac < 0.9), fair (0.7 < Ac < 0.8), and not good (0.6 < Ac < 0.7).

$\mathit{S}\mathit{e}=\frac{TP}{TP+FN}(3)$

$\mathit{S}\mathit{p}=\frac{TN}{TN+FP}(4)$

$\mathit{A}\mathit{c}=\frac{TP+TN}{TP+FP+TN+FN}(5)$

Finally, we applied the MCC to measure the two-class quality (harmful and neutral). This method is suitable for imbalanced data and has been used to evaluate in silico prediction approaches. MCC score ranges from 1 (perfect prediction) to -1 (total disagreement between the results predicted and observed), with 0 being no better than random prediction (Eq. 6) (Chicco et al., 2021).

$\mathrm{M}\mathrm{C}\mathrm{C}=\frac{TPXTN-FPXFN}{\sqrt{(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P})(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N})(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P})\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}\right)}}(6)$

5 Conclusion

The cost-effective and easy method of SNV analysis for CYP21A2 has important value as a screening tool, especially with a large number of genetic variants being available in massive genome projects. In the present study, we compared the abilities of 17 online and free tools for predicting the pathogenicity of SNVs on the CYP21A2 gene, a highly complex gene. Based on a curated databank composed of SNVs on the CYP21A2 enzyme functionally characterized, we reported the four highest-performing predictor programs for characterizing the pathogenicity of the CL group—CADD, ConSurf, DANN, and PolyPhen2. One of them, CADD, also showed the best performance for identifying mild mutations from the NC group, followed by ConSurf and DANN. Therefore, according to those results, CADD, ConSurf, DANN, and PolyPhen2 are high-recommended to be run for the first screening of each uncharacterized CYP21A2 SNV. Moreover, there is a great probability of a missense variant of the CYP21A2 being “pathogenic” when at least two of those four tools obtain that result. These results may be applicable in the future analysis of new missense variants of CYP21A2 and emphasize the relevance of using multiple predictors together.

Data availability statement

The original contributions presented in the study are included in the article/Supplementary Material; further inquiries can be directed to the corresponding authors.

Author contributions

Conceptualization, MP, RL-B, AZ, and AP; methodology, MP, AP, and RL-B; formal analysis, MP; investigation, MP; resources, AP and AZ; writing—original draft preparation, MP; writing—review and editing, RL-B, AZ, MR, and AP; visualization, MP; supervision, AP, AZ, MR, and RL-B; funding acquisition, MP, AP, and AZ. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by a Swiss Government Excellence Scholarship (ESKAS), grant number 2020.0209, and a Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) Scholarship. Open access funding was provided by the University of Bern and swissuniversities.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors, and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fphar.2022.931089/full#supplementary-material

References

Adzhubei, I., Jordan, D. M., and Sunyaev, S. R. (2013). Predicting functional effect of human missense mutations using PolyPhen‐2. Curr. Protoc. Hum. Genet. 76, Unit7. doi:10.1002/0471142905.hg0720s76

PubMed Abstract | CrossRef Full Text | Google Scholar

Ashkenazy, H., Abadi, S., Martz, E., Chay, O., Mayrose, I., Pupko, T., et al. (2016). ConSurf 2016: An improved methodology to estimate and visualize evolutionary conservation in macromolecules. Nucleic Acids Res. 44, W344–W350. –W350. doi:10.1093/nar/gkw408

PubMed Abstract | CrossRef Full Text | Google Scholar

Baumgartner-Parzer, S., Witsch-Baumgartner, M., and Hoeppner, W. (2020). EMQN best practice guidelines for molecular genetic testing and reporting of 21-hydroxylase deficiency. Eur. J. Hum. Genet. 28, 1341–1367. doi:10.1038/s41431-020-0653-5

PubMed Abstract | CrossRef Full Text | Google Scholar

Bendl, J., Musil, M., Štourač, J., Zendulka, J., Damborský, J., and Brezovský, J. (2016). PredictSNP2: A unified platform for accurately evaluating SNP effects by exploiting the different characteristics of variants in distinct genomic regions. PLoS Comput. Biol. 12, e1004962. doi:10.1371/journal.pcbi.1004962

PubMed Abstract | CrossRef Full Text | Google Scholar

Bendl, J., Stourac, J., Salanda, O., Pavelka, A., Wieben, E. D., Zendulka, J., et al. (2014). PredictSNP: Robust and accurate consensus classifier for prediction of disease-related mutations. PLoS Comput. Biol. 10, e1003440. doi:10.1371/journal.pcbi.1003440

PubMed Abstract | CrossRef Full Text | Google Scholar

Calabrese, R., Capriotti, E., Fariselli, P., Martelli, P. L., and Casadio, R. (2009). Functional annotations improve the predictive score of human disease-related mutations in proteins. Hum. Mutat. 30, 1237–1244. doi:10.1002/humu.21047

PubMed Abstract | CrossRef Full Text | Google Scholar

Capriotti, E., Altman, R. B., and Bromberg, Y. (2013). Collective judgment predicts disease-associated single nucleotide variants. BMC genomics 14, S2. Suppl 3. doi:10.1186/1471-2164-14-s3-s2

PubMed Abstract | CrossRef Full Text | Google Scholar

Capriotti, E., and Altman, R. B. (2011). Improving the prediction of disease-related variants using protein three-dimensional structure. BMC Bioinforma. 12, S3. doi:10.1186/1471-2105-12-S4-S3

PubMed Abstract | CrossRef Full Text | Google Scholar

Capriotti, E., and FariselliPhD-Snpg, P. (2017). PhD-SNPg: A webserver and lightweight tool for scoring single nucleotide variants. Nucleic Acids Res. 45, W247–W252. doi:10.1093/nar/gkx369

PubMed Abstract | CrossRef Full Text | Google Scholar

Chicco, D., Tötsch, N., and Jurman, G. (2021). The Matthews correlation coefficient (mcc) is more reliable than balanced accuracy, bookmaker informedness, and markedness in two-class confusion matrix evaluation. BioData Min. 14, 13–22. doi:10.1186/s13040-021-00244-z

PubMed Abstract | CrossRef Full Text | Google Scholar

Choi, Y., Sims, G. E., Murphy, S., Miller, J. R., and Chan, A. P. (2012). Predicting the functional effect of amino acid substitutions and indels. PLoS ONE 7, e46688. doi:10.1371/journal.pone.0046688

PubMed Abstract | CrossRef Full Text | Google Scholar

Gangodkar, P., Khadilkar, V., Raghupathy, P., Kumar, R., Dayal, A. A., Dayal, D., et al. (2021). Clinical application of a novel next generation sequencing assay for CYP21A2 gene in 310 cases of 21- hydroxylase congenital adrenal hyperplasia from India. Endocrine 71, 189–198. doi:10.1007/s12020-020-02494-z

PubMed Abstract | CrossRef Full Text | Google Scholar

Hart, S. N., Polley, E. C., Shimelis, H., Yadav, S., and Couch, F. J. (2020). Prediction of the functional impact of missense variants in BRCA1 and BRCA2 with BRCA-ML. npj Breast Cancer 6, 13. doi:10.1038/s41523-020-0159-x

PubMed Abstract | CrossRef Full Text | Google Scholar

Hecht, M., Bromberg, Y., and Rost, B. (2015). Better prediction of functional effects for sequence variants. BMC Genomics 16, S1. doi:10.1186/1471-2164-16-S8-S1

PubMed Abstract | CrossRef Full Text | Google Scholar

Hicks, S., Wheeler, D. A., Plon, S. E., and Kimmel, M. (2011). Prediction of missense mutation functionality depends on both the algorithm and sequence alignment employed. Hum. Mutat. 32, 661–668. doi:10.1002/humu.21490

PubMed Abstract | CrossRef Full Text | Google Scholar

Khan, S., and Vihinen, M. (2007). Spectrum of disease-causing mutations in protein secondary structures. BMC Struct. Biol. 7, 56. doi:10.1186/1472-6807-7-56

PubMed Abstract | CrossRef Full Text | Google Scholar

Lee, C., Yen, H.-Y., Zhong, A. W., and Gao, H. (2021). Resolving misalignment interference for NGS-based clinical diagnostics. Hum. Genet. 140, 477–492. doi:10.1007/s00439-020-02216-5

PubMed Abstract | CrossRef Full Text | Google Scholar

Miller, W. L., and Auchus, R. J. (2011). The molecular biology, biochemistry, and physiology of human steroidogenesis and its disorders. Endocr. Rev. 32, 81–151. doi:10.1210/er.2010-0013

PubMed Abstract | CrossRef Full Text | Google Scholar

Montenegro, L. R., Lerário, A. M., Nishi, M. Y., Jorge, A. A. L., and Mendonca, B. B. (2021). Performance of mutation pathogenicity prediction tools on missense variants associated with 46, xy differences of sex development. Clinics 76, e2052–e2055. doi:10.6061/clinics/2021/e2052

PubMed Abstract | CrossRef Full Text | Google Scholar

New, M. I., Abraham, M., Gonzalez, B., Dumic, M., Razzaghy-Azar, M., Chitayat, D., et al. (2013). Genotype-phenotype correlation in 1, 507 families with congenital adrenal hyperplasia owing to 21-hydroxylase deficiency. Proc. Natl. Acad. Sci. U. S. A. 110, 2611–2616. doi:10.1073/pnas.1300057110

PubMed Abstract | CrossRef Full Text | Google Scholar

Pejaver, V., Urresti, J., Lugo-Martinez, J., Pagel, K. A., Lin, G. N., Nam, H. J., et al. (2020). Inferring the molecular and phenotypic impact of amino acid variants with MutPred2. Nat. Commun. 11, 5918. doi:10.1038/s41467-020-19669-x

PubMed Abstract | CrossRef Full Text | Google Scholar

Pignatelli, D., Carvalho, B. L., Palmeiro, A., Barros, A., Guerreiro, S. G., and Maçut, D. (2019). The complexities in genotyping of congenital adrenal hyperplasia: 21-Hydroxylase deficiency. Front. Endocrinol. 10, 432. doi:10.3389/fendo.2019.00432

CrossRef Full Text | Google Scholar

Pshennikova, V. G., Barashkov, N. A., Romanov, G. P., Teryutin, F. M., Solov’ev, A. v., Gotovtsev, N. N., et al. (2019). Comparison of predictive in silico tools on missense variants in GJB2, GJB6, and GJB3 genes associated with autosomal recessive deafness 1A (DFNB1A). ScientificWorldJournal. 2019, 5198931. doi:10.1155/2019/5198931

PubMed Abstract | CrossRef Full Text | Google Scholar

Quang, D., Chen, Y., and Xie, X. D. A. N. N. (2015). Dann: A deep learning approach for annotating the pathogenicity of genetic variants. Bioinformatics 31, 761–763. doi:10.1093/bioinformatics/btu703

PubMed Abstract | CrossRef Full Text | Google Scholar

Rodrigues, N. R., Dunham, I., Yu, Y., Carroll’, M. C., Porter2, R. R., and Campbell, R. D. (1987). Molecular characterization of the HLA-linked steroid 21-hydroxylase B gene from an individual with congenital adrenal hyperplasia. EMBO J. 6, 1653. doi:10.1002/j.1460-2075.1987.tb02414.x

PubMed Abstract | CrossRef Full Text | Google Scholar

Shihab, H. A., Gough, J., Cooper, D. N., Stenson, P. D., Barker, G. L. A., Edwards, K. J., et al. (2013). Predicting the functional, molecular, and phenotypic consequences of amino acid substitutions using hidden markov models. Hum. Mutat. 34, 57–65. doi:10.1002/humu.22225

PubMed Abstract | CrossRef Full Text | Google Scholar

Simonetti, L., Bruque, C. D., Fernández, C. S., Benavides-Mori, B., Delea, M., Kolomenski, J. E., et al. (2018). CYP21A2 mutation update: Comprehensive analysis of databases and published genetic variants. Hum. Mutat. 39, 5–22. doi:10.1002/humu.23351

PubMed Abstract | CrossRef Full Text | Google Scholar

Stenson, P. D., Mort, M., Ball, E. v., Evans, K., Hayden, M., Heywood, S., et al. (2017). The human gene mutation database: Towards a comprehensive repository of inherited mutation data for medical research, genetic diagnosis and next-generation sequencing studies. Hum. Genet. 136, 665–677. doi:10.1007/s00439-017-1779-6

PubMed Abstract | CrossRef Full Text | Google Scholar

Stone, E. A., and Sidow, A. (2005). Physicochemical constraint violation by missense substitutions mediates impairment of protein function and disease severity. Genome Res. 15, 978–986. doi:10.1101/gr.3804205

PubMed Abstract | CrossRef Full Text | Google Scholar

Tang, B., Li, B., Gao, L.-D., He, N., Liu, X.-R., Long, Y.-S., et al. (2020). Optimization of in silico tools for predicting genetic variants: Individualizing for genes with molecular sub-regional stratification. Brief. Bioinform. 21, 1776–1786. doi:10.1093/bib/bbz115

PubMed Abstract | CrossRef Full Text | Google Scholar

Tang, H., and ThomasPANTHER-Psep, P. D. (2016). PANTHER-PSEP: Predicting disease-causing genetic variants using position-specific evolutionary preservation. Bioinformatics 32, 2230–2232. doi:10.1093/bioinformatics/btw222

PubMed Abstract | CrossRef Full Text | Google Scholar

van der Velde, K. J., de Boer, E. N., van Diemen, C. C., Sikkema-Raddatz, B., Abbott, K. M., Knopperts, A., et al. (2017). Gavin: Gene-aware variant INterpretation for medical sequencing. Genome Biol. 18, 6. doi:10.1186/s13059-016-1141-7

PubMed Abstract | CrossRef Full Text | Google Scholar

Vaser, R., Adusumalli, S., Leng, S. N., Sikic, M., and Ng, P. C. (2016). SIFT missense predictions for genomes. Nat. Protoc. 11, 1–9. doi:10.1038/nprot.2015.123

PubMed Abstract | CrossRef Full Text | Google Scholar

Vihinen, M. (2012). How to evaluate performance of prediction methods? Measures and their interpretation in variation effect analysis. BMC genomics 13 (4), S2. doi:10.1186/1471-2164-13-S4-S2

PubMed Abstract | CrossRef Full Text | Google Scholar

Witchel, S. F. (2017). Congenital adrenal hyperplasia. J. Pediatr. Adolesc. Gynecol. 30, 520–534. doi:10.1016/j.jpag.2017.04.001

PubMed Abstract | CrossRef Full Text | Google Scholar