Drug sensitivity prediction across cancer types using graph isomorphism networks and biological pathway features: A dual-branch deep learning approach.
Summary
This study introduces an innovative dual-branch deep learning approach for predicting drug sensitivity (IC50) across various cancer types. The model integrates Graph Isomorphism Networks (GIN) to represent drug chemical structures with a Multilayer Perceptron (MLP) that incorporates 50-dimensional ssGSEA pathway activities derived from gene expression. Trained on the GDSC2 dataset, this model achieved superior performance compared to existing methods, with an R2 of 0.8553 and a Pearson Correlation Coefficient of 0.9249. The inclusion of biological pathway features was demonstrated to be crucial for enhancing prediction accuracy.
Analysis
CLINIQUE: This computational research holds significant translational impact by improving drug sensitivity prediction. Enhanced IC50 prediction could potentially guide therapeutic decisions in precision medicine, enabling the selection of more effective treatments for cancer patients. By reducing the costs and time associated with experimental pharmacogenomic screenings, this approach could accelerate drug discovery and the identification of predictive biomarkers. The time to clinical impact could be 3-5 years, following experimental validation and integration into clinical trials. BIOINFO: The GIN+Pathway MLP model represents an advancement in drug sensitivity prediction by innovatively integrating 3D drug structures via GIN and the biological context of cell lines through ssGSEA pathway activities. The approach is validated on the extensive GDSC2 dataset, demonstrating robust performance with an R2 of 0.8553 and a PCC of 0.9249, outperforming benchmark models like GraphDRP and DeepCDR. A variant ablation study confirms the critical importance of biological features, showing that pathway integration improves R2 by over 0.15. This model could be clinically deployed to prioritize drug-patient combinations for Molecular Tumor Boards (MTBs) or for clinical trial design, by identifying promising drug candidates based on tumor molecular profiles.