Cue: a deep-learning framework for structural variant discovery and genotyping.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Structural variants (SVs) significantly contribute to human genetic diversity and disease.
• Current SV detection methods struggle with the complexity and diversity of SVs, limiting precision medicine advancements.

Purpose of the Study:

• To introduce Cue, an extensible deep-learning framework for accurate SV calling and genotyping.
• To overcome limitations of traditional SV callers by learning complex SV patterns directly from sequencing data.

Main Methods:

• Cue converts sequence alignments into images encoding SV-specific signals.
• A stacked hourglass convolutional neural network processes these images to predict SV type, genotype, and location.
• The framework is designed for extensibility to various sequencing platforms.

Main Results:

• Cue demonstrates superior performance over state-of-the-art methods for detecting multiple SV classes.
• The framework shows strong performance on both synthetic and real short-read sequencing data.
• Cue is adaptable to different sequencing technologies with competitive results.

Conclusions:

• Cue offers a powerful, data-driven approach to SV detection, surpassing existing heuristic methods.
• This deep-learning framework enhances the discovery of genetic variants crucial for precision medicine.
• The adaptability of Cue promises broader applications across diverse genomic research and clinical settings.