Predicting cancer drug response using parallel heterogeneous graph convolutional networks with neighborhood...

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in oncology

Background:

• Cancer heterogeneity necessitates personalized treatments for improved efficacy.
• Accurate anticancer drug response prediction is crucial for tailoring therapies.
• Existing graph neural network methods often overlook intrinsic differences between cell line and drug nodes.

Purpose of the Study:

• To develop an end-to-end graph neural network method for anticancer drug response prediction.
• To address limitations in existing models by incorporating neighborhood interactions.
• To improve the accuracy of predicting patient-specific responses to cancer drugs.

Main Methods:

• Proposed a Neighborhood Interaction (NI)-based heterogeneous graph convolution network (NIHGCN).
• Constructed a heterogeneous network integrating drugs, cell lines, and drug response data.
• Employed a parallel graph convolution and NI layer to capture both node-level and element-level interactions.

Main Results:

• NIHGCN achieved state-of-the-art performance on GDSC and CCLE datasets.
• Demonstrated superior accuracy in predicting responses for new cell lines, new drugs, and targeted therapies.
• Showcased excellent transferability from in vitro cell line data to in vivo PDX and TCGA datasets.

Conclusions:

• The proposed NIHGCN model effectively captures complex cell line-drug relationships for accurate response prediction.
• This approach holds significant potential for advancing personalized cancer therapy regimens.
• The model's robustness and transferability support its application in real-world clinical settings.