DeepDRK: a deep learning framework for drug repurposing through kernel-based multi-omics integration.

Area of Science:

• Computational biology
• Pharmacogenomics
• Machine learning in oncology

Background:

• Precision cancer medicine relies on integrating multi-omics data from patient-derived cell lines.
• Effective methods for multimodal and multisource data integration are crucial for clinical translation but remain a bottleneck.

Purpose of the Study:

• To develop DeepDRK, a machine learning framework for deciphering anticancer drug response using kernel-based data integration.
• To improve the prediction accuracy and robustness of drug response in cancer cell lines and patients.

Main Methods:

• Trained deep neural networks on over 20,000 pan-cancer cell line-anticancer drug pairs.
• Integrated multisource and multi-omics data (genomics, transcriptomics, epigenomics, chemical properties, drug-target interactions) using kernel-based similarity matrices.
• Validated the model on benchmark and newly established patient-derived cancer cell line datasets.

Main Results:

• DeepDRK surpassed previous approaches in accuracy and robustness on benchmark datasets.
• Achieved high performance on patient-derived cell lines with AUC of 0.84 and AUPRC of 0.77.
• Predicted clinical patient response, correlating well with outcomes and identifying drug repurposing candidates.

Conclusions:

• DeepDRK offers a computational method for predicting cancer drug response by integrating pharmacogenomic data.
• Facilitates prioritization of drug repurposing candidates for precision cancer treatment.
• Provides an open-source framework for advancing computational oncology research.