Predicting deleterious missense genetic variants via integrative supervised nonnegative matrix tri-factorization.

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Missense variants are key genetic variations that can disrupt protein function and lead to human diseases.
• Existing machine learning methods for variant classification utilize diverse features but struggle to integrate heterogeneous data effectively.
• A reliable algorithm is needed to merge complex network interactions for accurate variant effect prediction.

Purpose of the Study:

• To develop a novel, accurate algorithm for predicting the deleteriousness of missense variants.
• To effectively integrate heterogeneous data sources, including protein-protein interaction networks and gene disease associations.
• To address limitations of existing methods, such as type 2 circularity error.

Main Methods:

• Proposed a new method based on non-negative matrix tri-factorization clustering.
• Developed two versions: a two-source algorithm (prediction methods + PPI network) and a three-source algorithm (adding gene disease associations).
• Validated the algorithms using four benchmark datasets for internal and external assessment.

Main Results:

• The proposed method, particularly the three-source algorithm, significantly outperformed most state-of-the-art variant prediction tools across all datasets.
• The method successfully avoided type 2 circularity error, a common issue in ensemble-based predictors.
• Demonstrated superior performance for variants in genes with limited prior pathogenicity information.

Conclusions:

• The non-negative matrix tri-factorization approach provides a robust framework for integrating diverse data in variant effect prediction.
• The developed algorithms offer improved accuracy and reliability in identifying deleterious missense variants.
• This work advances the field of genetic variant interpretation and its application to human disease.