Classification of genetic variants using machine learning

Recent advances in genomic sequencing technology have resulted in an abundance of genome sequence data. Despite the progress in interpreting those data, there remains a broad scope for their translation into clinical and societal benefits. Loss-of-function variations in the human genome can be causal in disease development. Precise identification of such variations and pathogenicity prediction may lead to better drug targeting, among other benefits. Machine learning comes across as a promising method for its proven predictive ability. We have curated a novel dataset for the classification of LOF variants using high-quality databases of genetic variation. We trained and validated seven different classification algorithms using the new dataset to classify the variants as Benign, Pathogenic and Likely pathogenic. We recorded the best overall performance using the XG-Boost algorithm with an F1-score of 0.88 on the test set. We observed fair performance on Pathogenic samples with high recall and moderate precision and subpar performance on Likely pathogenic class, albeit with moderate precision. Overall, the encouraging results make our final model a promising candidate for further real-world tests.