KPNN

KPNN encodes prior biological network structures into neural network architectures to enable interpretable prediction and quantification of node importance from data such as single-cell RNA-sequencing.

Key Features:

• Knowledge-Based Network Structure: Uses a predefined network structure with nodes labeled as proteins or genes and edges reflecting prior regulatory interactions such as signaling pathways.
• Interpretability of Node Weights: Extracts node weights or importance scores after training to quantify the contribution of individual genes or proteins to predictive tasks.
• Stabilization and Quantification: Applies techniques to stabilize node weights in the presence of redundancy to improve quantitative interpretability.
• Control for Uneven Connectivity: Controls for biases introduced by uneven network connectivity to ensure accurate assessment of node importance.

Scientific Applications:

• Validation on Simulated Data: Validated on simulated datasets with known ground truth to assess interpretability and performance.
• Single-Cell RNA-Seq (cancer): Applied to single-cell RNA-sequencing data from cancer cells for interpretable prediction and mechanism discovery.
• Single-Cell RNA-Seq (immune): Applied to single-cell RNA-sequencing data from immune cells for interpretable prediction and mechanism discovery.
• General Network-Driven Analyses: Applicable to other research areas where prior domain knowledge can be represented as networks.

Methodology:

Encode prior knowledge as a predefined network of genes/proteins and signaling interactions, train neural network architectures constrained by that structure, extract node weights/importance scores post-training, apply stabilization for redundant nodes, and control for uneven connectivity when quantifying importance.