Pathway-Guided Deep Neural Network toward Interpretable and Predictive Modeling of Drug Sensitivity

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Drug Discovery

Background:

• Drug research and development requires cost-effective and risk-reducing in silico methods for predicting drug sensitivity in cancer cells.
• Machine learning models are increasingly used for drug sensitivity prediction, but often lack interpretability or have limited performance.
• Biological pathways are crucial for cellular functions, and their dysregulation contributes to diseases like cancer.

Purpose of the Study:

• To develop an interpretable and predictive in silico model for cancer drug sensitivity prediction.
• To integrate biological pathway knowledge with deep neural network (DNN) architectures.
• To enhance the understanding of drug action mechanisms in cancer cells.

Main Methods:

• A novel pathway-guided deep neural network (DNN) model was designed by incorporating pathway nodes and connections into the DNN structure.
• The model was trained and evaluated on multiple independent cancer drug sensitivity datasets.
• Performance was compared against canonical DNN models and eight classical regression models.

Main Results:

• The pathway-guided DNN model significantly outperformed canonical DNN and other classical regression models in predicting drug sensitivity.
• The model demonstrated improved interpretability, showing decreased activity in disease-related pathway nodes upon drug target input.
• Empirical experiments confirmed the model's pharmacological interpretability and predictive accuracy.

Conclusions:

• The pathway-guided DNN model offers a more interpretable and predictive approach for cancer drug sensitivity prediction.
• This method enhances understanding of drug mechanisms by linking drug targets to pathway activity.
• The developed tool provides a valuable resource for drug research and development, with code and data publicly available.