VPatho

VPatho predicts the pathogenicity and functional impact of genetic variants, with a focus on distinguishing gain-of-function (GOF) and loss-of-function (LOF) mutations to support interpretation of disease-associated variants.

Key Features:

• Two-Stage Deep Learning Approach: Uses a two-stage framework where the first stage employs a random under-sampling multi-scale residual neural network (RUS-Wg-MSResNet) with a novel weighted-loss function to predict pathogenicity using gnomAD_NV + GOF/LOF data.
• Comprehensive Feature Extraction: Extracts 138 variant-level, 262 protein-level, and 103 genome-level features for model training.
• Second-Stage Functional Impact Prediction: Applies an XGBoost-based outlier detection model (XGBOD) to predict functional impact and refine first-stage predictions.
• Training Dataset: Trained on a curated set of 9,619 pathogenic GOF/LOF variants and 138,026 neutral variants.
• Benchmarking and Validation: Demonstrated superior performance versus state-of-the-art predictors in benchmarking, including assessment on CAGI6 and independent tests.
• Identification of Misclassified Variants: Identified 31 non-LOF variants that were previously labeled as LOF or uncertain in the gnomAD database.

Scientific Applications:

• Genetic Research: Facilitates identification of disease-causing mutations by predicting variant pathogenicity and functional consequences.
• Clinical Diagnostics: Refines variant classification in genetic testing through GOF/LOF and pathogenicity predictions.
• Drug Development: Informs target discovery and validation by characterizing variant effects on protein function.

Methodology:

First-stage training used a random under-sampling multi-scale residual neural network (RUS-Wg-MSResNet) with a weighted-loss on gnomAD_NV + GOF/LOF data, followed by a second-stage XGBoost-based outlier detection (XGBOD); training used 138 variant-level, 262 protein-level, and 103 genome-level features from a curated set of 9,619 pathogenic GOF/LOF and 138,026 neutral variants.