SIMLINK Enables Accurate Variant Pathogenicity Prediction through Modeling the Gene-Variant-Feature Association...

Area of Science:

• Genetics and Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate prediction of variant pathogenicity is essential for clinical genetics.
• Current methods fail to explicitly model gene-variant-feature associations and the linear component of variant-pathogenicity relationships.
• These limitations hinder the performance of existing pathogenicity prediction tools.

Purpose of the Study:

• To introduce SIMLINK, a novel approach for simultaneous modeling of linear and nonlinear components in variant pathogenicity prediction.
• To leverage knowledge graphs for explicit modeling of gene-variant-feature associations.
• To improve the accuracy and reliability of predicting variant pathogenicity.

Main Methods:

• Constructed a variant-centered knowledge graph with over 8 million triplets to model gene-variant-feature associations.
• Employed a combination of linear models and graph neural networks to learn both linear and nonlinear components.
• Trained and evaluated SIMLINK on ClinVar variants and independent test sets.

Main Results:

• SIMLINK demonstrated superior prediction performance for both missense and synonymous variants compared to state-of-the-art methods like CADD and AlphaMissense.
• The approach effectively distinguished high- and low-confidence variants in Autism Spectrum Disorder.
• Genes with top-ranked variants identified by SIMLINK were found to be highly pathogenic.

Conclusions:

• SIMLINK effectively addresses limitations in current variant pathogenicity prediction by explicitly modeling complex associations and disentangling linear components.
• The method offers improved accuracy and provides valuable insights for genetic variant interpretation.
• The developed knowledge graph and modeling approach represent a significant advancement in clinical genetics.