DrugSAGE: a transcriptome aggregation approach using cell lines for drug response imputation

Accurate drug response prediction is essential for optimizing cancer therapy, yet genomic heterogeneity drives variable responses even among tumors with identical driver mutations. We developed DrugSAGE, a Graph Neural Network framework that predicts drug response from transcriptomic data by aggregating features from each sample and its most similar counterparts. A customized linear layer incorporating gene-pathway annotations provides biological interpretability. Benchmarking across independent bulk and single-cell datasets showed significant associations with known drug targets and treatment-stratified patient groups. DrugSAGE effectively predicts single-cell drug responses and identifies key genes and pathways, offering a novel, interpretable approach with superior or comparable performance.