Cancer drug sensitivity estimation using modular deep Graph Neural Networks.

Area of Science:

• Computational biology
• Machine learning in oncology
• Drug discovery and development

Background:

• Current drug sensitivity models struggle with generalizing to new drugs due to limitations in molecular representation (e.g., SMILES).
• Graph-attention networks offer high capacity but require extensive training data, which is often unavailable for drug sensitivity prediction.

Purpose of the Study:

• To develop a novel modular drug-sensitivity graph-attentional neural network architecture.
• To improve the prediction of drug-tumor interactions for precision oncology and drug discovery applications.

Main Methods:

• Developed a modular graph-attentional neural network for drug sensitivity prediction.
• Pre-trained model components (graph encoder, pooling layer) on related tasks with larger datasets.
• Utilized publicly available Genomics of Drug Sensitivity in Cancer (GDSC) data for experiments.

Main Results:

• The developed model outperforms existing reference models in predicting drug sensitivity.
• The model demonstrates superior ability in identifying specific drug-cell line interactions beyond general cytotoxicity and cell line survivability.
• Achieved better prediction accuracy for precision oncology use cases.

Conclusions:

• The modular graph-attentional neural network offers a promising approach for enhancing drug sensitivity prediction.
• This method advances precision oncology by enabling more accurate identification of targeted therapies.
• The model's architecture facilitates better generalization and prediction of specific drug-target interactions.