A Knowledge-Guided Graph Learning Approach Bridging Phenotype- and Target-Based Drug Discovery.

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Integrating phenotype-based drug discovery (PDD) and target-based drug discovery (TDD) is crucial but challenging due to biomedical data complexity.
• Existing methods struggle with data heterogeneity, noise, and bias, hindering effective therapeutic molecule discovery.

Purpose of the Study:

• To develop a robust graph representation learning approach for integrating multimodal biomedical data.
• To enhance the efficiency and accuracy of drug discovery processes, including screening and target identification.

Main Methods:

• Developed Knowledge-Guided Drug Relational Predictor (KGDRP), a graph representation learning model.
• Integrated multimodal data (network, gene expression, chemical structures) into a heterogeneous graph (HG) structure.
• Utilized a heterogeneous graph neural network (HGNN)-based architecture incorporating a biomedical HG (BioHG).

Main Results:

• KGDRP achieved a 12% improvement in real-world drug screening scenarios compared to previous methods.
• Biology-informed representations from KGDRP enhanced drug target prioritization by 26%.
• Demonstrated high success rates in identifying potential drugs for COVID-19 using zero-shot evaluation and elucidated drug mechanisms through cell-target-drug interaction analysis.

Conclusions:

• KGDRP offers a robust infrastructure for seamless multimodal data and biomedical network integration.
• The approach effectively accelerates phenotype-based drug discovery (PDD) and guides therapeutic target discovery.
• KGDRP expedites the overall discovery of novel therapeutic molecules.