Using interpretable deep learning to model cancer dependencies.

Area of Science:

• Computational Biology
• Bioinformatics
• Systems Biology

Background:

• Cancer dependencies are crucial for identifying potential drug targets.
• Inter-individual and inter-cancer variability in dependencies necessitates personalized approaches.
• Visible neural networks (VNNs) offer robust performance and interpretability for biomedical applications.

Purpose of the Study:

• To develop a novel, interpretable neural network model for predicting cancer dependencies.
• To leverage biological pathway knowledge for enhanced prediction accuracy and mechanistic insights.
• To facilitate the discovery of novel drug targets and therapeutic strategies.

Main Methods:

• Design of the Biological Visible Neural Network (BioVNN) incorporating pathway knowledge.
• Comparative analysis against traditional neural networks (NNs) and NNs with randomized pathways.
• Evaluation of prediction accuracy, convergence speed, and parameter efficiency.
• Analysis of neuron output states for pathway-level mechanistic interpretability.

Main Results:

• BioVNN demonstrates marginally superior performance compared to traditional NNs with fewer parameters.
• BioVNN exhibits faster convergence rates.
• Predictions from BioVNN are interpretable through correlations with relevant pathway neuron states, suggesting dependency mechanisms.
• Feature importance analysis identifies known and proposes novel reaction partners.

Conclusions:

• BioVNN provides a robust and interpretable framework for understanding cancer dependencies.
• The model's interpretability aids in elucidating underlying dependency mechanisms.
• BioVNN holds promise for advancing the development of targeted cancer therapies.