pathDNN

pathDNN predicts drug sensitivity in cancer cell lines by integrating gene-level inputs with a pathway-layer representation to link drug targets to pathway activities and drug response.

Key Features:

• Integration of Biological Pathways: pathDNN incorporates a pathway-node layer that connects with input gene nodes, representing groups of molecules regulating cellular functions such as proliferation and apoptosis to embed biological knowledge into the model.
• Enhanced Predictive Performance: pathDNN demonstrates superior performance compared to a canonical deep neural network and seven classical regression models on multiple independent drug sensitivity datasets by leveraging pathway information.
• Pharmacological Interpretability: During forward propagation with drug target inputs, pathDNN reveals significant decreases in activity of disease-related pathway nodes consistent with pathway inhibition associated with cancer progression.

Scientific Applications:

• Drug sensitivity prediction: Predicts compound sensitivity across cancer cell lines to help prioritize potential therapeutic candidates.
• Mechanistic interpretation: Links predicted drug responses to specific biological pathways to provide insights into mechanisms of drug action and pathway inhibition in cancer.

Methodology:

The approach reshapes a canonical deep neural network by integrating biological pathways as an intermediary layer between gene input nodes and output predictions, with forward propagation propagating drug-target effects to pathway node activities.