A knowledge graph representation learning approach to predict novel kinase-substrate interactions

Area of Science:

• Biochemistry
• Bioinformatics
• Systems Biology

Background:

• The human proteome involves complex kinase-substrate interactions.
• Many kinases remain understudied, limiting therapeutic target identification.
• Predicting novel kinase interactions is crucial for understanding cellular functions.

Purpose of the Study:

• To develop a novel knowledge graph representation learning approach.
• To predict novel interaction partners for understudied kinases.
• To gain insights into the biology of understudied kinases.

Main Methods:

• Constructed a phosphoproteomic knowledge graph integrating iPTMnet, protein ontology, gene ontology, and BioKG.
• Employed directed random walks and modified SkipGram/CBOW models for representation learning.
• Utilized a supervised classification model to predict novel kinase-substrate interactions.

Main Results:

• Successfully predicted novel interactions for understudied kinases.
• Demonstrated the utility of the phosphoproteomic knowledge graph for interaction prediction.
• Provided a post-predictive analysis and ablation study for biological insights.

Conclusions:

• The developed approach effectively predicts novel kinase-substrate interactions.
• Knowledge graph representation learning offers a powerful tool for exploring understudied kinases.
• This work facilitates the identification of new therapeutic targets within the human kinome.