VEF: a variant filtering tool based on ensemble methods.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Genomic analysis pipelines often produce incorrectly called variants.
• Current filtering tools like Variant Quality Score Recalibration (VQSR) and Hard Filtering (HF) are user-dependent and can fail.
• There is a need for more robust and automated variant filtering methods.

Purpose of the Study:

• To introduce VEF, a novel variant filtering tool based on decision tree ensemble methods.
• To address the limitations of existing variant filtering approaches (VQSR and HF).
• To treat variant filtering as a supervised learning problem using gold-standard data.

Main Methods:

• Developed VEF using decision tree ensemble methods for variant filtering.
• Trained VEF on variant call data with known true variants (gold standard).
• Applied VEF to filter variants in VCF files, assuming similar feature characteristics between training and testing data.

Main Results:

• VEF consistently outperformed VQSR and HF on whole genome sequencing (WGS) Human datasets.
• VEF demonstrated good generalization capabilities with missing features, varying coverage, and different sequencing pipelines.
• VEF significantly reduced processing time compared to VQSR (approximately 4 minutes vs. 50 minutes for WGS human SNP filtering).

Conclusions:

• VEF is a more accurate and efficient variant filtering tool than VQSR and HF.
• VEF offers a robust solution for variant filtering, overcoming key limitations of current methods.
• The supervised learning approach provides a significant advantage in variant filtering accuracy and computational efficiency.