Three Steps Novel Machine Learning Method Classifies Uncertain MEFV Gene Variants

Abstract

Introduction: The International Study Group for Systemic Autoinflammatory Diseases (INSAID) consensus criteria revealed that the clinical outcomes of more than half of the MEFV gene variants are uncertain. In this study, we estabilished a novel approach for more accurate classification of MEFV gene variants by using the optimal number of amino acid prediction scores and machine-learning algorithms. Our goal was to determine a more accurate classification of MEFV variants while also reducing the uncertainties. Material-Methods: We extracted variants of the MEFV gene from the infevers database ,and point mutations were included, others excluded from the study. We then determined the optimal number of in silico instruments for our model. On the training dataset, we implemented seven machine learning algorithms on MEFV gene variants with known clinical effects. We evaluated the effectiveness of our model in three steps: First, we performed machine-learning algorithms on the training dataset and implemented those with a prediction accuracy of greater than 90 percent. Second, we compared our prediction results to existing algorithms and studies. Third, we evaluated our outcomes functional and clinical level. Results We included 266 of 381 MEFV gene variants and four computational tools in a study. Our algorithm classified Likely pathogenic (LP) variants with an accuracy of 96.6% while classifying 97.6% of Likely Benign (LB) variants. Among the machine learning methods used to classify MEFV variants, our classification method yielded the most accurate results on training datasets. Most of the predictors classified LB variants with higher accuracy than 90% however, LP classification showed a wide range of variety in accuracy scores between 2% − 62.5%. Disease-causing MEFV variants are frequently located in domains. Functional and clinical level evaluation compatible with our classification results. Discussion The comparison indicated that LP variant prediction is the biggest problem in variant classification, and our method might be a candidate for solving this problem with the 96.67% accuracy. Considering that 60% of the clinical effects of MEFV gene variants are unresolved, evaluating our methods in conjunction with the clinical manifestations of patients significantly simplifies the interpretation of unknown variants