SUN-035 Performance of Mutation Pathogenicity Prediction Tools on Missense Variants Associated with 46XY Disorders of Sex Development

INTRODUCTION: Single nucleotide variants (SNV) are the most common type of genetic variation among humans. The high-throughput sequencing methods increase the number of identified variants in the human genome but functional studies for disease-associated variants is laborious and time consuming. Different computational methods have been developed to predict SNP pathogenicity and phenotypic effect. They are based on different variant parameters, such as sequence homology, proteins structure and evolutionary conservation. OBJECTIVE/METHODS: To evaluate the performance of the 10 widely used pathogenicity prediction tools available in the internet in relation with functional studies: Fathmn, Mutation Acessor, Phanter and Sift sites (based on evolutionary information) and Mutation Taster, Polyphen-2, Align GVGD, CAAD, Provean and SNP-and-Go sites (based in a combination of protein structural, functional parameters and evolutionary information). We analyzed 40 described pathogenic mutation in 4 different genes associated with disorder of sex development (DSD): 17β-hydroxysteroid dehydrogenase (HSD17B3), Steroidogenic factor 1 (SF-1/NR5A1), Androgen receptor (AR) and Luteinizing hormone/chorionic gonadotropin receptor (LHCGR). All mutations are already published and functional studies showed loss of function proteins. To evaluate the false discovery rate of each tool, we analyzed 36 frequent (MAF >0.01) benign SNVs described in the same 4 DSD. The quality of the predictions was analyzed by five parameters: accuracy, precision, sensitivity, specificity and Matthews correlation coefficient (MCC). RESULTS: No method was in full accordance with the functional study. The best accuracy was observed in the Polyphen-2 and SNP-and-Go programs (0.81 and 0.82); the best precision was observed in Mutation Accessor and SNP-and-Go (0.96 and 0.84). The best specificity was observed for Mutation Accessor and SNP-and-GO (0.96 and 0.83). Five programs (Phanter, Sift, Mutation Taster, Polyphen-2 and CAAD) showed sensitivity > 0.80, but only SNP-and-GO program had specificity of 0.83. Performance ranged from poor (Align GVGD - MCC 0.13) to reasonably good (SNP-and-GO - MCC 0.63). CONCLUSION: Computational algorithms are important tools for SNV analysis but their correlation with functional studies is extremely variable. The overall best performing methods among the 10 prediction tools was SNP-and-GO, with accuracy reaching 0.82.