Lightweight Graph Network for Efficient and Interpretable Cancer Drug Response Prediction
Abstract
Accurate prediction of cancer drug response is essential for advancing precision oncology; however, existing deep learning approaches, particularly graph neural networks (GNNs), often rely on computationally intensive architectures and lack interpretability, limiting their applicability in real-world clinical settings. This paper proposes a lightweight and interpretable graph neural network (LIGNN) framework designed to efficiently predict drug response by integrating multi-omics data. The proposed model constructs a heterogeneous graph comprising genes, drugs, and cell lines, capturing complex biological interactions through protein-protein and drug-target networks. To reduce computational overhead, shallow graph architecture with sparsity-aware propagation and node pruning is employed, significantly decreasing model complexity while preserving predictive performance. Furthermore, an embedded attention-driven interpretability module identifies biologically relevant features and sub-graphs, enabling transparent decision-making. Experimental evaluation on benchmark datasets such as GDSC and CCLE demonstrates that LIGNN achieves competitive prediction accuracy of 97.7% and 97.2%, and AUC of 0.96 and 0.95 while reducing model size and inference time compared to state-of-the-art models.