CNVoyant a machine learning framework for accurate and explainable copy number variant classification.

Area of Science:

• Genomic Medicine
• Computational Biology
• Rare Genetic Diseases

Background:

• Classifying copy number variants (CNVs) is crucial for diagnosing rare genetic diseases (RGDs).
• Existing methods struggle to accurately differentiate benign, uncertain, and pathogenic CNVs.
• This limitation hinders precise diagnosis and treatment in genomic medicine.

Purpose of the Study:

• To introduce CNVoyant, a machine learning framework for multi-class classification of CNV clinical significance.
• To improve the accuracy and interpretability of CNV classification compared to existing approaches.
• To provide a reliable tool for genomic researchers and clinicians interpreting CNVs.

Main Methods:

• Developed CNVoyant, a multi-class machine learning model trained on 52,176 ClinVar entries.
• Incorporated diverse genomic features: position, gene annotations, dosage sensitivity, and conservation scores.
• Validated performance using 21,574 CNVs from the DECIPHER database and real-world RGD cases.

Main Results:

• CNVoyant demonstrated improved precision-recall and ROC AUC for binary pathogenic classification.
• The framework successfully performed multi-class classification (benign, uncertain, pathogenic) with SHAP explainability.
• CNVoyant accurately classified all diagnostic CNVs in real-world RGD cases as pathogenic with high confidence.

Conclusions:

• CNVoyant offers superior accuracy in classifying the clinical significance of CNVs.
• The tool enhances the interpretation of CNVs in the context of rare genetic diseases.
• CNVoyant has the potential to significantly aid clinical geneticists and genomic researchers.