Machine-learning algorithm for identifying and predicting amyotrophic lateral sclerosis causal mutations

Abstract

We propose a machine learning (ML) method to classify ALS-causative and non-ALS-causative variants based on 24 variables in five different datasets. The proposed ML method classifies the five datasets with very high accuracy. In particular, it predicts the ALS variants with 100 percent accuracy, while its accuracy for the non-ALS variants is up to 99.31 percent. The trained classifier also identifies the nine most influencial mutation assessors that help distinguishing the two classes from each other. They are FATHMM_score, PROVEAN_score, Vest3_score, CADD_phred, DANN_score, meta-SVM_score, phyloP7way_vertebrate, metaLR, and REVEL. Thus, they may be used in future studies in order to reduce the time and cost of collecting data and carrying out experimental tests, as well as in studies with more focus on the recognized assessors.