PhenoLinker: Phenotype-gene link prediction and explanation using heterogeneous graph neural networks

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Linking human phenotypes to genetic variants is crucial but challenging.
• Existing methods often lack comprehensive data integration or explainability.
• Understanding genotype-phenotype relationships is key to biomedical research.

Purpose of the Study:

• To introduce PhenoLinker, a novel graph-based system for scoring phenotype-gene relationships.
• To leverage heterogeneous information networks and graph convolutional neural networks for accurate predictions.
• To provide explainability for predicted associations using Integrated Gradients.

Main Methods:

• Developed a graph-based system, PhenoLinker.
• Utilized heterogeneous information networks integrating gene and phenotype attributes.
• Employed a convolutional neural network model for graphs.
• Incorporated Integrated Gradients for prediction explainability.

Main Results:

• PhenoLinker demonstrated superior performance compared to existing models.
• The system achieved high accuracy in both retrospective and temporal validation tasks.
• The explainability feature provides insights into the predicted phenotype-gene associations.

Conclusions:

• PhenoLinker offers a powerful and explainable approach to phenotype-gene association.
• The system can significantly aid in discovering novel genetic associations.
• It enhances the understanding of human genetic variation's consequences.