A convolutional attention model classifies copy number variants from whole exome sequencing

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Copy number variants (CNVs) are crucial biomarkers for genetic diseases and cancer.
• Existing whole-exome sequencing (WES) CNV callers often lack positional context and perform inconsistently across platforms.

Purpose of the Study:

• To develop a novel deep learning model for accurate and robust CNV detection.
• To improve CNV calling by integrating genomic coordinates and chromosome identity.

Main Methods:

• A dual-input convolutional neural network (CNN) with attention mechanism was developed.
• The model ingests normalized read depth, genomic coordinates, and chromosome identity.
• Pretraining on ECOLE-labeled 1000 Genomes data and fine-tuning on expert-annotated samples were performed.

Main Results:

• The model achieved a macro F1 score of 0.83 and macro PR-AUC of 0.93 on a test set.
• Cross-platform evaluations demonstrated high overall F1 scores up to 0.96 on various sequencers.
• The model showed strong sensitivity and precision, with performance comparable to established WES CNV callers.

Conclusions:

• The developed CNN with attention offers sensitive and robust CNV detection across diverse sequencing technologies.
• This method shows potential for clinical applications such as genetic disease triage and large-scale cancer screening.
• The model's architecture effectively handles noisy signals and small CNV events.