DeepCCDS: Interpretable Deep Learning Framework for Predicting Cancer Cell Drug Sensitivity through Characterizing...

Area of Science:

• Computational biology
• Genomics
• Pharmacology

Background:

• Accurate characterization of cellular states is crucial for precision oncology and predicting drug sensitivity.
• Current deep learning methods for cellular state characterization are limited as they focus on isolated genetic markers, neglecting complex regulatory networks.

Purpose of the Study:

• To develop a deep learning framework, DeepCCDS (Deep learning framework for Cancer Cell Drug Sensitivity prediction through Characterizing Cancer Driver Signals), to improve cancer cell drug sensitivity prediction.
• To enhance the characterization of cellular states by incorporating prior knowledge of cancer driver signals.

Main Methods:

• Developed DeepCCDS, a deep learning framework integrating a prior knowledge network with a self-supervised neural network.
• Utilized cancer driver signals to represent key mechanisms influencing cancer cell development and drug response.
• Applied DeepCCDS to multiple datasets and The Cancer Genome Atlas (TCGA) solid tumor samples.

Main Results:

• DeepCCDS demonstrated superior performance in predicting drug sensitivity compared to existing state-of-the-art methods.
• The framework exhibited powerful feature representation capabilities and enhanced interpretability due to integrated prior knowledge.
• Identified embedding features with potential applications in drug screening for new indications.

Conclusions:

• DeepCCDS offers an improved approach to cancer cell drug sensitivity prediction by characterizing driver signals.
• The framework's interpretability and feature representation capabilities can aid in understanding drug response mechanisms.
• Integrating DeepCCDS into clinical decision-making may enhance personalized treatment strategies for cancer patients.