Variant Combination Pathogenicity Predictor (VarCoPP) 2.0

Variant Combination Pathogenicity Predictor (VarCoPP) 2.0 predicts the pathogenicity of variant combinations in gene pairs (digenic/bilocus) to identify potentially disease-associated bilocus configurations from single-individual SNVs and small insertions/deletions.

Key Features:

• Predictive scope: Predicts pathogenicity of digenic/bilocus variant combinations derived from SNVs and small insertions/deletions in a single individual.
• Machine learning training data: Trained using variant data from OLIDA (Oligogenic Diseases Database) and variants from the 1000 Genomes Project.
• Balanced Random Forest model: Implements a Balanced Random Forest of 400 decision trees with reported sensitivity of 95% in cross-validation and 98% in testing and a 5% false positive rate.
• Runtime improvement: Achieves approximately a 150-fold reduction in running time compared to the predecessor version.
• Confidence labels: Outputs 95% and 99% confidence labels indicating the probability that a bilocus combination is pathogenic.
• Interpretable predictions: Provides explanations that highlight the biological information contributing to each pathogenicity prediction.
• Input filtering recommendation: Applied to variant sets after initial filtering, typically restricting analysis to up to 150 genes.
• Training enrichment: Positive training set is enriched with variant combinations confidently associated with pathogenicity based on OLIDA confidence scores.
• Feature integration: Integrates multiple variant-level and gene-pair features, combining biological information at different levels for prediction.

Scientific Applications:

• Oligogenic and multilocus disease analysis: Supports identification and study of oligogenic/multilocus disease mechanisms by predicting pathogenic bilocus variant combinations.
• Medical genetics and rare disease research: Aids prioritization of candidate variant combinations to reduce missing diagnoses and enable analysis of larger oligogenic datasets.

Methodology:

VarCoPP 2.0 employs a Balanced Random Forest model of 400 decision trees trained on variants from the 1000 Genomes Project with a positive set enriched from OLIDA using OLIDA confidence scores and integrates variant-level and gene-pair features.