Abstract Drug response prediction (DRP), accounting for the diverse biological characteristics of cancer types that affect sensitivity or resistance to treatment, is crucial for anticancer drug selection and discovery. Although numerous deep learning models for DRP have been developed, it has not been investigated whether these models can maintain reliable predictive power when applied to omics datasets not used for training, which is crucial for real-world applications. Moreover, they have not been rigorously examined through systematic benchmark tests to investigate the relationships between model architecture and performance. We present a new model, GCNPath, that exploits both graph convolution network (GCN) architectures and pathway-based feature reduction of gene expression data, which have previously been shown to improve DRP model performance. In comprehensive benchmark tests utilizing multiple cell omics platforms, GCNPath shows robust and competitive performances compared with state-of-the-art models including prediction of unseen drugs and the ability to overcome batch effects across various RNA datasets. This study demonstrates the validity of the pathway-level GCN model in DRP and suggests directions for developing DRP models with improved adaptability to diverse and heterogeneous datasets and enhanced usability for practical applications.
Yoon et al. (Wed,) studied this question.