Genetic variant effect prediction by supervised nonnegative matrix tri-factorization.

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Machine Learning in Genetics

Background:

• Distinguishing deleterious from neutral non-synonymous single nucleotide variants (nsSNVs) identified via whole exome sequencing (WES) presents a significant challenge in genetic research.
• Existing machine learning methods for variant consequence prediction often rely on analyzing protein sequences, structures, or integrating gene-level and phenotype data, facing challenges with feature heterogeneity and integration.
• Effective integration of diverse data sources is crucial for developing robust predictive models for nsSNV effects.

Purpose of the Study:

• To propose a novel supervised nonnegative matrix tri-factorization (sNMTF) algorithm for integrating variant prediction scores with gene-level data and disease networks.
• To construct an enhanced feature space using sNMTF for improved classifier training in predicting nsSNV consequences.
• To evaluate the performance of the proposed sNMTF method against existing nsSNV effect prediction approaches.

Main Methods:

• Development and implementation of a supervised nonnegative matrix tri-factorization (sNMTF) algorithm.
• Integration of existing variant prediction scores, gene-level data, and disease networks into a unified feature space using sNMTF.
• Training and assessment of a classifier using the newly constructed feature space on two benchmark datasets of nsSNPs (non-synonymous single nucleotide polymorphisms).

Main Results:

• The proposed sNMTF method demonstrated superior performance compared to existing nsSNV effect prediction approaches on two benchmark datasets.
• The model achieved an average prediction accuracy increase of 15% over other ensemble scores.
• Excluding any integrated data source led to a significant decrease in deleterious variant prediction accuracy, highlighting the importance of data integration.

Conclusions:

• The novel sNMTF algorithm provides a powerful and accurate method for predicting the consequences of nsSNVs by effectively integrating heterogeneous data sources.
• The proposed model offers a significant improvement in prediction accuracy over current methods, aiding in the interpretation of genetic variants.
• The developed framework is extensible, allowing for the incorporation of additional heterogeneous data sources for future enhancements in variant effect prediction.