VEF

VEF applies supervised decision-tree ensemble learning to filter genetic variants in VCF files from single non-cancerous whole-genome sequencing (WGS) Human samples.

Key Features:

• Supervised Learning Approach: Uses gold-standard known true variants to train decision-tree ensemble models, framing variant filtering as a supervised learning problem rather than relying on VQSR or Hard Filtering (HF).
• Robust Performance: Demonstrated superior performance to VQSR and HF on WGS Human datasets with available gold standards.
• Generalization Capabilities: Maintains high filtering accuracy in the presence of missing feature values, coverage differences between training and testing datasets, and when used with sequencing pipelines other than GATK.
• Efficiency: Requires a single training run and substantially reduces SNP filtering time in a WGS Human sample (reported reduction from ~50 minutes with VQSR to ~4 minutes).

Scientific Applications:

• WGS variant filtering: Filters variants in WGS Human VCF files to remove incorrectly called variants using models trained on gold-standard datasets.
• Single-sample non-cancer genomics: Applied to single non-cancerous samples to improve the precision of downstream genetic and genomic analyses.

Methodology:

Trains decision-tree ensemble models on variant call data containing known true variants and applies those models to filter variants in VCF files; implemented as a bash script that calls Python scripts vef_clf.py (training) and vef_apl.py (application) and produces two .clf model files and two .vef.vcf filtered VCF files.