MISTIC: A prediction tool to reveal disease-relevant deleterious missense variants

Abstract

The diffusion of next-generation sequencing technologies has revolutionized research and diagnosis in the field of rare Mendelian disorders, notably via whole-exome sequencing (WES). However, one of the main issues hampering achievement of a diagnosis via WES analyses is the extended list of variants of unknown significance (VUS), mostly composed of missense variants. Hence, improved solutions are needed to address the challenges of identifying potentially deleterious variants and ranking them in a prioritized short list. We present MISTIC (MISsense deleTeriousness predICtor), a new prediction tool based on an original combination of two complementary machine learning algorithms using a soft voting system that integrates 113 missense features, ranging from multi-ethnic minor allele frequencies and evolutionary conservation, to physiochemical and biochemical properties of amino acids. Our approach also uses training sets with a wide spectrum of variant profiles, including both high-confidence positive (deleterious) and negative (benign) variants. Compared to recent state-of-the-art prediction tools in various benchmark tests and independent evaluation scenarios, MISTIC exhibits the best and most consistent performance, notably with the highest AUC value (> 0.95). Importantly, MISTIC maintains its high performance in the specific case of discriminating deleterious variants from benign variants that are rare or population-specific. In a clinical context, MISTIC drastically reduces the list of VUS (<30%) and significantly improves the ranking of "causative" deleterious variants. Pre-computed MISTIC scores for all possible human missense variants are available at http://lbgi.fr/mistic.

Interpreting the pathogenicity of Joubert syndrome missense variants in Caenorhabditis elegans

“…VUS alleles, many of which are missense, are not clinically useful and require further characterisation to be classified as pathogenic or benign. Although in silico prediction tools, such as Sorting Intolerant From Tolerant (SIFT) ( Sim et al, 2012 ), MISsense deleTeriousness predICtor (MISTIC) ( Chennen et al, 2020 ), Rare Exome Variant Ensemble Learner (REVEL) ( Ioannidis et al, 2016 ), Combined Annotation-Dependent Depletion (CADD) ( Rentzsch et al, 2019 ) or (Polymorphism Phenotyping (Poly-Phen) ( Adzhubei et al, 2010 ), can predict the pathogenicity of missense variants, these methods are not always reliable and can give contradictory predictions. For example, SIFT, MISTIC and REVEL correctly predict P74S and G155S to be deleterious, whereas CADD predicts that they are likely to be benign.…”

A comparison on predicting functional impact of genomic variants

“…Thus, we constructed two independent test datasets based on the ClinVar and VariBench databases, which are widely used ( 14 , 16–18 ), reliable in quality and easily accessible. On these two datasets, we performed a comprehensive comparison of 14 functional impact prediction methods including CADD ( 19 , 20 ), DANN ( 21 ), FATHMM-MKL ( 22 ), FunSeq2 ( 23 ), PredictSNP2 ( 24 ), SIFT ( 25 ), PROVEAN ( 26 ), MetaLR ( 14 ), MetaSVM ( 14 ), MutationAssessor ( 27 ), PrimateAI ( 28 ), M-CAP ( 29 ), REVEL ( 30 ) and MISTIC ( 17 ). Based on the performance evaluation of these two datasets for 14 prediction methods, CADD and REVEL, obtained the best performance, respectively.…”

MvPPT: a highly efficient and sensitive pathogenicity prediction tool for missense variants

“…Nevertheless, the effects of most missense variants on proteins are unclear, as experimental validation of large numbers of variants is limited by efficiency and cost. To address these limitations, many computational tools have been developed to predict the potential impact of variants (5)(6)(7)(8)(9)(10)(11)(12)(13)(14)(15)(16)(17)(18)(19)(20)(21). Early prediction models compute deleterious scores based on single property of variants, such as evolutionary conservation (8,10,15) and protein structure/function (16,17), and recent ensemble methods achieve higher classification accuracy by integrating information from individual predictors (5)(6)(7)9,(11)(12)(13)20).…”

“…To address these limitations, many computational tools have been developed to predict the potential impact of variants (5)(6)(7)(8)(9)(10)(11)(12)(13)(14)(15)(16)(17)(18)(19)(20)(21). Early prediction models compute deleterious scores based on single property of variants, such as evolutionary conservation (8,10,15) and protein structure/function (16,17), and recent ensemble methods achieve higher classification accuracy by integrating information from individual predictors (5)(6)(7)9,(11)(12)(13)20). Although these existing tools have made significant contributions to the prediction of the hazard of genetic variants, the sensitivity of prediction still needs to be improved when assessing the pathogenicity of massive variants in clinical scenarios.…”