Improving drug response prediction based on two-space graph convolution.

Area of Science:

• Computational biology
• Genomics
• Pharmacogenomics

Background:

• Cancer patients with identical cancer types exhibit varying genomic profiles, leading to differential drug sensitivities.
• Accurate prediction of patient drug responses is vital for guiding cancer treatment strategies and enhancing patient outcomes.
• Current computational methods using graph convolutional networks often overlook similarities within homogeneous nodes (e.g., cell lines or drugs).

Purpose of the Study:

• To develop an advanced computational model for predicting anticancer drug response.
• To address the limitation of existing methods by incorporating similarity information among homogeneous nodes.

Main Methods:

• Proposed Two-Space Graph Convolutional Neural Network (TSGCNN) algorithm.
• Constructed separate cell line and drug feature spaces for graph convolution operations to diffuse similarity information.
• Generated a heterogeneous network integrating cell line-drug relationships for further graph convolution.
• Integrated self-features, feature space representations, and heterogeneous space representations for final feature extraction.
• Utilized a linear correlation coefficient decoder to reconstruct the cell line-drug correlation matrix for prediction.

Main Results:

• TSGCNN demonstrated superior performance in predicting drug response compared to eight other state-of-the-art methods.
• The model was validated on the Cancer Drug Sensitivity Data (GDSC) and Cancer Cell Line Encyclopedia (CCLE) databases.
• The approach effectively captures both homogeneous node similarities and heterogeneous network interactions.

Conclusions:

• TSGCNN offers a significant advancement in computational drug response prediction for cancer.
• The model's ability to integrate diverse feature spaces enhances prediction accuracy.
• This approach holds promise for guiding personalized cancer therapy decisions.