DrugSAGE: an aggregation-based method for drug response imputation

Accurate prediction of drug response is critical for optimizing the clinical application of cancer therapies. Nevertheless, owing to genetic and genomic heterogeneity and the complex tumor microenvironment, cancer samples may exhibit heterogeneous responses to the same therapeutic agents, even when harboring identical driver mutations or genes. In this study, we developed DrugSAGE, a novel framework that leverages Graph Neural Networks (GNNs) to model cellular heterogeneity and accurately predict drug response from transcriptomic data. DrugSAGE enhances prediction by aggregating features from a sample and its most similar counterparts. Critically, we introduce a customized linear layer incorporating gene-pathway annotations to provide biological interpretability, facilitating the identification of key pathways driving the prediction. We benchmarked DrugSAGE using TCGA bulk data, six public bulk datasets, and four scRNA-seq datasets. Predictions from DrugSAGE showed significant associations between drugs and their known target genes (e.g., HER2 + inhibitors, MET inhibitors, BRAF inhibitors) or significant differences between patient groups stratified by their treatments (e.g., Erlotinib, PLX4720, 5-Fluorouracil). Furthermore, we show DrugSAGE can effectively predict drug response at the single-cell level. Additionally, by analyzing pathway-based embeddings, we identified pathways crucial to the models, including those involved in cancer, MAPK signaling, endocytosis, and JAK-STAT signaling, among others. Compared to previous methods, DrugSAGE demonstrated superior or comparable performance, offering a novel approach for predicting drug responses.