PhenoSV: interpretable phenotype-aware model for the prioritization of genes affected by structural variants

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Structural variants (SVs) are a key source of genetic variation linked to diverse phenotypes and diseases.
• Long-read sequencing identifies numerous SVs, but interpreting their functional impact, especially noncoding SVs, remains a challenge.
• Current methods for disease-related SV identification are limited, often focusing only on deletions/duplications and struggling to prioritize specific genes.

Purpose of the Study:

• To develop PhenoSV, a phenotype-aware machine learning model for comprehensive interpretation of all major structural variant types.
• To enable prioritization of genes affected by SVs, including those in noncoding regions.
• To improve the identification of disease-associated SVs by integrating phenotype information.

Main Methods:

• PhenoSV segments and annotates SVs using diverse genomic features.
• A transformer-based architecture with a multiple-instance learning framework predicts SV impacts.
• Gene-phenotype associations are utilized to prioritize phenotype-related SVs.

Main Results:

• PhenoSV demonstrates superior performance compared to existing methods across extensive human SV datasets encompassing all SV types.
• The model effectively interprets the functional consequences of various SVs, including noncoding variants.
• Applications in disease datasets show PhenoSV's capability to pinpoint disease-related genes affected by SVs.

Conclusions:

• PhenoSV offers a powerful, comprehensive approach to interpreting structural variants and their impact on human health.
• The tool advances the field by enabling precise identification of disease-causing genes from complex SVs.
• PhenoSV is accessible via a web server and command-line tool, facilitating broader research application.