Comparison of Various Variant Prediction Algorithms and Conservation Scores by Machine Learning Approaches

Machine learning methods are widely used for prediction of the pathogenicity and deleteriousness of the different types of single nucleotide variants (SNVs). This study encompasses a preliminary approach towards the application of different types of machine learning algorithms on the dbNSFPv4.7 (database of Nonsynonymous SNPs Functional Predictions) database to compare the various prediction algorithms and conservation scores to determine which one performs the best. This paves the way for selecting the best prediction algorithm or the combination of the better prediction algorithms in annotation of the variants and classifying them as pathogenic or likely pathogenic, benign or likely benign. The different learning approaches were compared based on the metrices like, accuracy, recall, precision and F1 score. Among the different variant prediction algorithms and conservation scores available in the dbNSFP database, BayesDel AF was the best perfoming followed by BayesDel noAF, MetaRNN, MCAP and MutPred. bStatistic, fitCons, fathmm XF, MutationTaster and phastCons17way primate performed the worst among 31 prediction algorithms and 9 conservation scores selected for this study. The best performing features or instances can be prioritized over others while selecting the functional prediction algorithms, which are used to determine the driver mutations or the pathogenic variants from omics datasets.