DeepSVFilter

DeepSVFilter filters false positive structural variant (SV) calls from short read whole genome sequencing using deep learning to improve SV call specificity.

Key Features:

• Deep Learning Framework: Encodes structural variant signals from read alignments into image formats and classifies them with convolutional neural networks (CNNs).
• Transfer Learning with Pre-trained CNNs: Employs transfer learning with pre-trained CNN models trained on well-characterized samples containing high-confidence structural variants to improve generalization to new datasets.
• Versatility in Application: Operates as a standalone SV filter or integrates with existing structural variant detection tools to enhance specificity of SV calls.
• Performance Validation: Validated on two well-characterized samples, demonstrating reduction of false positive SV calls when coupled with existing SV detection methods.
• Implementation: Implemented in Python.

Scientific Applications:

• Genetic disease studies: Enhances accuracy of SV calls used to identify disease-associated structural variants in research cohorts.
• Cancer genomics: Reduces false positive SVs in tumor genome analyses to improve detection of somatic structural alterations.
• Population genetics: Improves specificity of SV catalogs derived from short read whole genome sequencing for population-scale studies.
• Clinical genomics: Supports interpretation of SVs in clinical sequencing by lowering false positive rates.

Methodology:

Encodes SV signals from read alignments into image formats and classifies them with convolutional neural networks (CNNs), using transfer learning with pre-trained CNN models trained on well-characterized samples containing high-confidence structural variants.