XGraphCDS: An explainable deep learning model for predicting drug sensitivity from gene pathways and chemical...

Area of Science:

• Genomics
• Pharmacology
• Artificial Intelligence

Background:

• Cancer exhibits significant heterogeneity, complicating drug development and personalized treatment.
• Predicting cancer drug sensitivity from genomics data remains a challenge despite data availability.

Purpose of the Study:

• To develop an explainable graph neural network framework (XGraphCDS) for predicting cancer drug sensitivity.
• To integrate cancer gene expression and drug chemical structures for enhanced predictive accuracy.

Main Methods:

• Developed XGraphCDS, a framework using comparative learning on a unified heterogeneous network.
• Integrated molecular graphs for drugs and gene enrichment scores for cell lines.
• Employed transfer learning to build an in vivo prediction model from in vitro data.

Main Results:

• XGraphCDS outperformed state-of-the-art methods with R² = 0.863 and AUC = 0.858.
• The in vivo prediction model achieved good predictive power (AUC = 0.808).
• The framework provides interpretable insights into drug resistance mechanisms.

Conclusions:

• XGraphCDS demonstrates significant potential for developing targeted anti-cancer drugs.
• The framework supports personalized dosing strategies within precision medicine.
• Explainable AI in drug sensitivity prediction offers valuable mechanistic insights.