BioVNN

BioVNN predicts cancer dependencies by integrating biological pathway knowledge into a visible neural network architecture to provide interpretable, pathway-informed predictions.

Key Features:

• Pathway Knowledge Integration: BioVNN embeds detailed biological pathway information into its neural architecture and enables correlating predictions with neuron output states of relevant pathways.
• Interpretable Predictions: The model provides interpretability by correlating dependency predictions with neuron output states, allowing inference of pathway-level contributions to dependencies.
• Performance Efficiency: Despite using fewer parameters than conventional neural networks, BioVNN marginally outperforms traditional NNs and converges faster during training.
• Comparison with Randomized Pathways: In comparative analyses, BioVNN significantly outperforms a neural network based on randomized pathways.
• Feature Importance Analysis: Feature importance analysis recapitulates known reaction partners and proposes novel ones, expanding understanding of cancer dependency networks.

Scientific Applications:

• Oncology research: Identify potential drug targets by predicting cancer dependencies.
• Mechanistic exploration: Use interpretable neuron output states to explore mechanisms underlying cancer dependencies.
• Therapeutic development: Inform development of targeted therapies by linking dependencies to specific pathway components.
• Personalized medicine: Support tailoring predictions to specific cancers or individual patients.

Methodology:

BioVNN implements visible neural networks that incorporate pathway knowledge into the model architecture, and achieves interpretability by correlating dependency predictions with neuron output states.