Machine learning-based prediction of human structural variation and characterization of associated sequence...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Structural variants (SVs) are crucial for genetic diversity, evolution, and human diseases.
• Quantifying the influence of local sequence context on SV formation is challenging.

Purpose of the Study:

• To develop machine learning models for predicting SV occurrence in the human genome.
• To identify genomic determinants that contribute to SV formation.

Main Methods:

• Developed a sequence-only convolutional neural network (CNN) model.
• Utilized a random forest approach integrating genomic annotations.
• Employed model interpretability techniques to identify key genomic contributors.

Main Results:

• Both models achieved high predictive performance (>90% AUROC), improved by ensemble methods.
• Identified sequence motifs (microhomology, non-canonical DNA structures) and SV hotspots as key determinants.
• Different SV classes (transposable elements, inversions) show distinct sequence signatures.
• Predicted SV probability correlates with allele frequency and gene functional constraint.

Conclusions:

• Local sequence context accurately predicts SV-prone genomic regions.
• Machine learning models provide a framework for quantifying SV susceptibility.
• Findings support the utility of these models for variant effect prediction in personalized genomics.