Challenging popular tools for the annotation of genetic variations with a real case, pathogenic mutations of lysosomal alpha-galactosidase

Bioinformatics

In some cases manifestations of FD occur at an early age with general, neurological, cardiovascular and renal signs, in other cases in adulthood and with a limited subset of symptoms. For this reason a qualitative phenotypic classification of mutations based on the symptoms observed in the patients, has been attempted and classic or severe ones have been distinguished from mild, late onset or variant forms [42].

Fabry-database.org [43, 44] provides a list of mutations and their qualitative phenotypic classification. Since FD is X linked and the association between genotype and phenotype is clearer in males [51], only the 175 hemizygous cases have been gathered from Fabry-database.org and form the first dataset analysed in this paper. The variants were annotated with wANNOVAR [3] and the output is provided in Additional file 1 with the original qualitative phenotypic description in the last column. In the first place it can be noticed that only 51 cases are also present in ClinVar, which is a public archive of reports of the relationships among human variations and phenotypes [52].

To test whether it is possible to broadly distinguish FD mutations collected from Fabry-database.org by the qualitative predictions provided by wANNOVAR annotation, the observed phenotypes were reduced to two classes, a severe group POS of 152 cases, which clusters mutations originally defined as “severe” or “classic”, and a mild group NEG of 23 cases, which clusters those mutations originally defined as “mild”, “late onset”, “variant” or “atypical variant” in fabry-database.org. For the predicted phenotypes, if the tool provides binary classification, like in the case of SIFT [4], the more deleterious one, D in the case of SIFT, is considered as predicted POS, the other one, T in the case of SIFT is considered as predicted NEG. If the tool provides multiple classes, as in the case of PolyPhen-2 [12], the most deleterious one, D in the case of PolyPhen-2, is considered as predicted POS, the other ones, P and B in the case of PolyPhen-2, is considered as predicted NEG. The results are summarized in Table 1. Since the two classes have different sizes, Matthews correlation coefficient should be preferred for the evaluation of predictors [53].

Table 1

Accuracy Indexes

B

SIFT

0.749

0.549

0.241

0.092

B

LRT

0.794

0.576

0.280

0.162

B

MutationAssessor

0.191

0.460

0.247

−0.106

B

FATHMM

0.846

0.500

0.000

0.000

B

PROVEAN

0.737

0.557

0.258

0.103

Meta

MetaSVM

0.846

0.500

0.000

0.000

Meta

MetaLR

0.846

0.500

0.000

0.000

Meta

M-CAP

0.846

0.500

0.000

0.000

ML

Polyphen2_HDIV

0.771

0.592

0.310

0.175

ML

Polyphen2_HVAR

0.691

0.621

0.341

0.188

ML

MutationTaster

0.194

0.433

0.230

−0.156

ML

FATHMM-MKL

0.829

0.505

0.063

0.022

For most tools the values are quite low and in some cases no discrimination is possible.

A different way of ordering by severity, relies on the residual activity of AGAL mutants measured in vitro in HEK293 or COS cell transiently transfected with expression plasmids. Values for 280 mutations have been collected gathering results of several laboratories [33, 45, 5462]. They form the second dataset analyzed in this paper. wANNOVAR annotation for these mutants can be found in Additional file 2 with the relative residual activity in the last column.

The intersection between the two datasets is formed by 67 mutations of the severe group POS and 12 of the mild group NEG, for which relative residual activity is available. The median residual activity of severe mutations POS is 0.1 (Fig. 1). This finding suggests that severe cases have null, or very close to null activity, when tested in transfected cells. The box plot in Fig. 1 shows 20% outliers with high residual activity in POS population that might represent an overestimation in the original literature.

Fig. 1

Distribution of residual activities for phenotypically annotated GLA mutations. The boxplot shows the distribution of residual activity in the subpopulations of mutations with severe and mild effects. The red bars represent outliers

Contrary to what occurs in the first dataset of mutations whose phenotypic annotation is derived from clinical literature (Additional file 1), the second dataset, whose annotation is based on residual activity (Additional file 2), is balanced with half of the mutations with values above 0.

The box plot in Fig. 2 shows the distribution of rank scores for mutations showing 0 residual activity. Rank scores were created by wANNOVAR to make the functional prediction scores and conservation scores more comparable to each other and monotonic (a higher score indicating “more likely to be damaging”) [63]. As can be observed FATHMM [7], metaSVM [10], metaLR [10], M-CAP [11]correctly assign high scores to very severe cases. On the other side, the histograms in Fig. 3 show the rank scores assigned by the predictors to 6 non pathological mutation whose residual activity is comparable or higher than that of wild type. The same predictors, FATHMM, metaSVM [10], metaLR [10], M-CAP [11], give a constantly high score and tend to over-estimate the damage caused by a mutation. In Table 2 the correlation between the rank scores of the predicting tools and the residual activity of all the mutations in the second dataset (Additional file 2), is shown. Results obtained by some predictors used by wANNOVAR, for example VEST3 (Pearson correlation coefficient 0.71; p p 

Fig. 2

Distribution of rank scores for mutations with null residual activity. The boxplot show the distribution of the rank scores for all the predictors used by wANNOVAR. The red bars represent outliers. Predictor category label is B for “biologically based prediction method”, ML for “Machine Learning based prediction method”, Meta for “Meta prediction method” and Cons for “Conservation scoring tool”

Fig. 3

Rank scores for mutations with residual activity equal or greater than wild type alpha-galactosidase. The histograms show the rank scores of the six mutations whose residual activity is greater or equal than the wild type alpha-galactosidase, for each of the wANNOVAR predictors. Mutations are color coded, and are detailed inset

Table 2

Correlations

B

SIFT

− 0.493

7.87E-19

B

LRT

−0.486

2.76E-18

B

MutationAssessor

−0.573

5.22E-26

B

FATHMM

−0.054

1.85E-01

B

PROVEAN

−0.546

1.86E-23

Meta

VEST3

−0.699

1.08E-42

Meta

MetaSVM

0.285

1.00E + 00

Meta

MetaLR

−0.482

5.77E-18

Meta

M-CAP

−0.255

8.09E-06

ML

POLYPHEN2 HDIV

−0.672

1.67E-38

ML

POLYPHEN2 HVAR

−0.648

4.53E-35

ML

MutationTaster

−0.499

2.42E-19

ML

CADD

−0.595

1.78E-28

ML

DANN

−0.388

8.51E-12

ML

FATHMM-MKL

−0.434

1.35E-14

ML

GenoCanyon

−0.282

7.95E-07

n

GERP++

−0.405

9.34E-13

Cons

phyloP7way vertebrate

−0.441

4.79E-15

Cons

phyloP20way mammalian

−0.214

1.54E-04

Cons

phastCons7way vertebrate

−0.486

2.55E-18

Cons

phastCons 20 way mammalian

−0.256

7.35E-06

Cons

SiPhy 29way logOdds

−0.389

7.65E-12

Articles You May Like

Former Astronaut On Watching For Life-Destroying Objects From Space
Hurricane Michael Was A Category 5, NOAA Finds – The First Since Andrew In 1992
BRB-seq: The quick and cheaper future of RNA sequencing
Defying the laws of physics? Engineers demonstrate bubbles of sand
Why language technology can’t handle Game of Thrones (yet)

Leave a Reply

Your email address will not be published. Required fields are marked *