Prediction of the pathogenicity of antithrombin sequence variations by in silico methods

Computational prediction tools have been developed to aid in the interpretation of novel sequence variations, but their utility within the diagnostic setting of antithrombin (AT) deficiency has not been evaluated to date. The aim of our study was to test the performance of different bioinformatic tools (Meta-SNP, MutPred, nsSNPAnalyzer, PANTHER, PhD-SNP, PMut, SIFT, SNAP, SNPs&Go, PolyPhen-2, PON-P2, and PredictSNP) in predicting the pathogenicity of AT sequence variations. We analysed all naturally occurring SERPINC1 missense mutations that have been previously characterised to be damaging with regard to the secretion or function of the AT molecule. Additionally, we analysed all reported non-synonymous exonic polymorphisms within SERPINC1 with a population allele frequency >1.0%. The in silico tools had accuracies of 62-96%, sensitivities of 59-98%, and specificities of 33-100% for the prediction of the pathogenicity of AT sequence variations; receiver operating characteristic analysis had area under the curves between 0.54-0.97. When mutations were grouped according to their effect on the phenotype of AT deficiency [type I or type II with a thrombin (IIRS) or heparin (IIHBS) binding defect or pleiotropic effects (IIPE)], we observed the lowest performance characteristics of the tools for mutations causing AT deficiency type IIHBS. Only three tools (MutPred, PhD-SNP, PolyPhen-2) detected mutants causing type IIHBS AT deficiency with high sensitivity (93%), the sensitivities of the other tools ranged between 36% and 79%. This study demonstrates that bioinformatic tools are useful for pathogenicity prediction for AT sequence variations, but they have substantially different performance characteristics, particularly for type IIHBS AT deficiency.