GCNPath: introspecting drug response prediction with pathway-guided graph convolution networks.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Drug response prediction (DRP) is vital for selecting effective anticancer treatments, but current deep learning models struggle with diverse cancer data.
• The reliability of these models on external datasets and the impact of model architecture on performance remain underexplored.

Purpose of the Study:

• To develop and validate a novel deep learning model, GCNPath, for robust drug response prediction across heterogeneous omics datasets.
• To systematically benchmark GCNPath against state-of-the-art models, assessing its performance on unseen drugs and its ability to mitigate batch effects.

Main Methods:

• GCNPath integrates graph convolution network (GCN) architectures with pathway-based feature reduction of gene expression data.
• Comprehensive benchmark tests were conducted using multiple cell omics platforms and various RNA datasets.

Main Results:

• GCNPath demonstrated robust and competitive performance compared to existing state-of-the-art models.
• The model successfully predicted responses to unseen drugs and effectively overcame batch effects across different RNA datasets.

Conclusions:

• The pathway-level GCN model (GCNPath) is validated as a powerful tool for drug response prediction.
• Findings suggest future DRP models should prioritize adaptability to diverse datasets and practical usability.