Ideafix

Ideafix refines variant calls in formalin-fixed and paraffin-embedded (FFPE) DNA sequencing data to reduce FFPE-associated sequencing artefacts and improve accuracy of variant calling for cancer genomics.

Key Features:

• Target data: Specifically addresses variant calling in formalin-fixed and paraffin-embedded (FFPE) DNA sequencing datasets.
• Training dataset: Built from a dataset of over 1.6 million variants derived from 27 paired FFPE and fresh-frozen breast cancer samples.
• Machine learning evaluation: Assessed five machine learning algorithms and identified XGBoost (extreme gradient boosting) and random forest as top performers.
• Cross-validation performance: Achieved area under the ROC curve (AUC) values greater than 0.86 in leave-one-sample-out cross-validation.
• Independent validation: Validated on two independent datasets with AUC values up to 0.96, exceeding previously published tools (maximum AUC 0.92).
• Discriminating features: Leverages read pair orientation bias, genomic context, and variant allele frequency among its variant features.

Scientific Applications:

• Clinical variant interpretation: Refinement of variants in FFPE-derived sequencing data to support more accurate genomic interpretation for cancer treatment decisions.
• Molecular testing with FFPE biopsies: Improvement of variant call reliability in routine molecular testing workflows that use FFPE tissue.
• Precision oncology research: Enabling more reliable use of FFPE samples in research applications requiring accurate somatic variant calls.

Methodology:

Constructed a comprehensive set of variant features from paired FFPE and fresh-frozen samples, evaluated five machine learning algorithms including XGBoost and random forest using leave-one-sample-out cross-validation, and validated performance on two independent datasets while leveraging features such as read pair orientation bias, genomic context, and variant allele frequency.