SV-MeCa: an XGBoost-based meta-caller approach for structural variant calling from short-read data

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate calling of structural variants (SVs) from whole genome short-read data is challenging due to limitations in existing tools.
• Meta-caller approaches, combining multiple SV callers, are widely used to improve accuracy and robustness.
• Current meta-callers often rely on the number of supporting tools rather than variant-specific quality metrics.

Purpose of the Study:

• To introduce SV-MeCa, the first structural variant meta-caller that incorporates variant-specific quality metrics.
• To develop a scoring system for ranking consensus SV calls based on their likelihood of being true positives.
• To enhance the accuracy and robustness of structural variant detection from whole genome sequencing data.

Main Methods:

• SV-MeCa utilizes seven standalone SV callers and merges results using SURVIVOR.
• Caller-specific quality metrics from individual VCF files are extracted for consensus SV calls.
• XGBoost decision tree classifiers, trained on benchmark data, predict the probability of consensus SV calls being true positives.

Main Results:

• SV-MeCa demonstrated superior performance compared to four other meta-caller approaches based on F-scores for deletions (0.58) and insertions (0.42).
• While ConsensuSV showed higher precision, SV-MeCa achieved competitive precision (0.64 for deletions, 0.53 for insertions).
• SV-MeCa exhibited strong recall, outperformed only by Meta-SV for deletions (0.55 vs 0.53).

Conclusions:

• SV-MeCa outperforms existing SV meta-caller approaches by leveraging variant-specific quality measures.
• The XGBoost prediction probabilities provide a flexible scoring mechanism, allowing users to adjust sensitivity and precision.
• SV-MeCa is publicly available, offering an improved tool for structural variant detection.