Explainable drug sensitivity prediction through cancer pathway enrichment

Area of Science:

• Computational biology
• Pharmacogenomics
• Cancer research

Background:

• Precision anticancer therapeutics require accurate drug sensitivity prediction.
• Generalizable and explainable models are crucial for clinical translation but often secondary to performance.
• Existing models may lack interpretability and broad applicability.

Purpose of the Study:

• To propose PathDSP, a novel pathway-based computational model for predicting drug sensitivity.
• To integrate chemical structure, pathway enrichment, and multi-omics data (gene expression, mutation, copy number variation) for enhanced prediction.
• To ensure model generalizability, explainability, and utility in personalized medicine and drug development.

Main Methods:

• Developed PathDSP, a deep neural network model integrating chemical structure and cancer signaling pathway enrichment.
• Utilized multi-omics data (gene expression, mutation, copy number variation) from the Genomics of Drug Sensitivity in Cancer dataset.
• Validated model performance and generalizability on the Cancer Cell Line Encyclopedia dataset.

Main Results:

• PathDSP outperformed state-of-the-art deep learning models in drug sensitivity prediction.
• The model demonstrated strong generalizability on an independent dataset.
• Explainable results were achieved, supported by case studies aligning with existing knowledge.
• Good performance was observed for predicting responses to unseen drugs and cell lines.

Conclusions:

• PathDSP offers a powerful, explainable, and generalizable approach to predict anticancer drug sensitivity.
• The model has potential utility for guiding individualized cancer treatment and accelerating drug development.
• Integrating pathway information enhances the predictive power and interpretability of computational drug sensitivity models.