Machine Learning to Detect Conflicting Clinical Classifications of Genetic Variants

Abstract

Identifying conflicting variants in the clinical classification of pathogenicity is important, as diagnoses directly affect the treatment plans for patients. Machine learning models can effectively categorize and analyze multidimensional data, and the incorporation of feature selection algorithms into these models allows us to identify relationships between clinical features that can contribute to conflicting clinical classifications of pathogenicity. We use the ClinVar dataset to address this need, which serves as a public archive for annotations of human genetic variants. In ClinVar, variants are manually categorized into different classes: benign, likely benign, uncertain significance, probably pathogenic, and pathogenic by researchers. However, there are inconsistencies in the annotations across clinical laboratories that can create confusion when assessing the impact of a variant on a patient's condition. This project proposes the development of a machine learning model trained on the ClinVar dataset to address this issue as well as feature selection to identify the properties of the variants that are most predictive of conflicting pathogenicity labels. The model leverages variant annotations such as genetic features, clinical data, and other critical information to identify patterns and relationships that harmonize conflicting classifications. We trained a random forest model and studied the importance of the input features using both tree-based and lasso feature selection.  The five most significant features based on the tree-based feature selection, which inherently handles the nonlinear relationship between features, are (1) the score of the deleteriousness of variants, (2) allele frequencies emitted by ExAC, (3) Phred Scaled Score, (4) LoFtool’s gene intolerance score, and (5) allele frequencies emitted by GO-ESP. The utilization of machine learning for identifying conflicting clinical classifications of genetic variants helps ensure precise and consistent interpretation of variants. This, in turn, plays a crucial role in improving clinical genomics, making diagnoses more accurate, and enabling personalized treatment options.