ECOLE: Learning to call copy number variants on whole exome sequencing data

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Copy number variants (CNVs) are significant contributors to the development of genetic disorders.
• Accurate detection of CNVs from whole exome sequencing (WES) data is crucial for clinical applications.
• Existing algorithms for CNV detection in WES data often exhibit low precision and recall on gold-standard datasets.

Purpose of the Study:

• To introduce ECOLE, a novel deep learning-based caller for somatic and germline CNVs in WES data.
• To enhance the accuracy and reliability of CNV detection in clinical settings.
• To develop a method capable of detecting CNVs without requiring control samples in specific cancer types.

Main Methods:

• Developed ECOLE using a transformer architecture variant for per-exon CNV calling.
• Utilized high-confidence calls from matched whole genome sequencing (WGS) data for initial model training.
• Employed transfer learning with expert-curated calls and tumor samples for fine-tuning.

Main Results:

• ECOLE achieved unprecedented performance on expert-labeled data, with 68.7% precision and 49.6% recall.
• Demonstrated significant improvements over existing methods, with 18.7% higher precision and 30.8% higher recall.
• Successfully detected RT-qPCR-validated variations in bladder cancer samples using fine-tuned tumor samples without a control.

Conclusions:

• ECOLE represents a significant advancement in CNV detection from WES data, offering high precision and recall.
• The developed fine-tuning strategy enables robust CNV detection in cancer samples, even without control data.
• ECOLE holds promise for improved clinical diagnostics of genetic disorders and cancer.