INSnet

INSnet detects insertions in genomic data using deep learning to identify structural variations from long-read alignments and to determine insertion sites and lengths.

Key Features:

• Input data and features: Extracts five key features at each locus through alignments between long reads and the reference genome.
• Genome partitioning: Divides the reference genome into continuous sub-regions for localized analysis.
• Feature extraction network: Uses a depthwise separable convolutional network to extract spatial and channel information from alignment-derived features.
• Convolutional Block Attention Module (CBAM): Applies CBAM to emphasize relevant spatial and channel-wise alignment features within sub-regions.
• Efficient Channel Attention (ECA): Employs ECA to capture efficient channel attention and complement CBAM for feature selection.
• Sequential modeling (GRU): Integrates a Gated Recurrent Unit network to model relationships between adjacent sub-regions and extract SV signatures.
• Detection outputs: Predicts whether each sub-region contains an insertion and determines the insertion site and length.
• Performance: Reported experimental results show improved F1 score relative to existing methods on real datasets.

Scientific Applications:

• Genetic disease research: Enables identification of insertions that contribute to genetic disorders.
• Personalized medicine: Supports detection of patient-specific insertions relevant for therapeutic decision-making.
• Genomic data analysis: Provides insertion calls and precise coordinates for downstream structural variation studies.

Methodology:

The method divides the reference genome into continuous sub-regions, extracts five features per locus from long-read alignments, applies a depthwise separable convolutional network with CBAM and ECA attention modules for feature extraction, uses a GRU to model adjacent sub-region relationships, and predicts insertion presence, site, and length.