Gene Specific Pathogenicity Predictor for Chromatin-Remodeling BAF Complex-Associated Neurodevelopmental Disorders.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Whole genome sequencing identifies numerous variants of uncertain significance (VUS), creating a diagnostic challenge.
• Genetic counselors face a bottleneck in identifying causative variants for patient conditions.
• Existing machine learning (ML) tools require gene-specific calibration for accurate pathogenicity prediction.

Purpose of the Study:

• To develop a workflow for creating accurate, gene-specific ensemble-learning ML models.
• To leverage variant location, evolutionary conservation, and algorithm outputs for pathogenicity prediction.
• To improve diagnostic accuracy for rare neurodevelopmental diseases linked to BAF complex proteins.

Main Methods:

• Screened 15 ML algorithms using variants in SMARCA2 and SMARCA4 associated with neurodevelopmental diseases.
• Tuned a random forest learner, achieving 0.93 accuracy on holdout data.
• Developed a final predictor for BAF complex proteins, reaching 0.91 accuracy.

Main Results:

• A gene-specific ML model achieved 0.93 accuracy for SMARCA2/SMARCA4 variants.
• Generalizing the predictor to other BAF complex proteins significantly reduced performance.
• A final predictor for BAF complex proteins demonstrated 0.91 accuracy and superior AUROC compared to other predictors.

Conclusions:

• Gene-specific calibration is crucial for accurate ML-based pathogenicity prediction.
• The developed workflow offers a rapid, cost-effective method for enhancing ML tools for genetic counselors.
• This approach aids in diagnosing rare diseases by efficiently identifying causative genetic variants.