synVep

synVep predicts the biological impact of synonymous single nucleotide variants (sSNVs) in the human genome to prioritize variants that may alter gene expression, splicing, or contribute to disease.

Key Features:

• Machine Learning Framework: Uses extreme gradient boosting models to classify sSNVs as effect or no-effect.
• Training Data Innovation: Trains on observed variants from gnomAD combined with synthetically generated possible human-genome variants to broaden variant representation.
• Positive-Unlabeled Learning: Applies positive-unlabeled learning to filter likely unobservable variants from the training set.
• Performance Metrics: Achieves approximately 90% precision/recall on previously unseen variant sets.
• Conservation Correlation: Does not explicitly use sequence conservation signals but its predictions correlate with evolutionary distances between orthologs in cross-species analyses.
• Pathogenicity and Splice-Site Differentiation: Distinguishes pathogenic versus benign sSNVs and identifies splice-site disrupting variants (SDVs) versus non-SDVs.

Scientific Applications:

• sSNV Annotation and Prioritization: Enhances annotation by prioritizing sSNVs most likely to have biological effects.
• Disease Variant Prioritization: Assists in identifying candidate disease-associated synonymous variants.
• Splicing Analysis: Supports detection and characterization of splice-site disrupting variants (SDVs).
• Evolutionary and Cross-Species Variation Studies: Enables correlation of predicted effects with evolutionary distances among orthologs.

Methodology:

Training combined observed variants from gnomAD with synthetically generated possible human-genome variants, used positive-unlabeled learning to remove likely unobservable variants, trained extreme gradient boosting models, and evaluated model performance (≈90% precision/recall) with analyses including correlation of predictions with evolutionary distances and classification of SDVs.