GraphDRP

GraphDRP predicts drug response by applying graph convolutional networks to molecular graphs and integrating cell-line genomic aberration vectors to model drug–cell line interactions.

Key Features:

• Graph-Based Drug Representation: Drugs are represented as molecular graphs that capture atoms and bonds.
• Integration with Genomic Data: Cell lines are encoded as binary vectors representing genomic aberrations.
• Feature Learning through Convolution Layers: Convolution layers learn representative features from both drug molecular graphs and genomic aberration vectors, which are combined per drug–cell line pair.
• Predictive Modeling with Neural Networks: A fully-connected neural network predicts response values for each drug–cell line pair using the combined feature representations.
• Variants of Graph Convolutional Networks: Four different variants of graph convolutional networks are implemented to optimize learning and predictive performance.
• Interpretability via Saliency Maps: Saliency maps are generated to identify and interpret contributions of specific genomic aberrations to predicted drug responses.

Scientific Applications:

• Drug response prediction: Predicts response values for drug–cell line pairs to study drug sensitivity.
• Interpretation of genomic contributions: Uses saliency maps to identify and interpret which genomic aberrations contribute to predicted responses.
• Benchmarking: Demonstrates superior performance compared to tCNNS across reported experiments.

Methodology:

Drugs are represented as molecular graphs and cell lines as binary genomic-aberration vectors; convolution layers and four variants of graph convolutional networks learn features from these inputs, combined features are fed to a fully-connected neural network to predict responses, and saliency maps are produced for interpretation.