GraphLoc

GraphLoc predicts protein subcellular localization from immunohistochemistry (IHC) images to identify protein distribution patterns and location biomarkers in cancer tissues.

Key Features:

• Graph Neural Network Architecture: Employs a multi-label multi-instance model based on deep graph convolutional networks (GCNs) to model complex protein distribution patterns across cell types and states.
• Multi-Instance Learning: Constructs a graph representation from multiple IHC images per protein to learn comprehensive protein-level representations via graph convolutions.
• Dynamic Threshold Method: Uses a dynamic threshold approach to generate multi-label subcellular location predictions.

Scientific Applications:

• Protein Localization Prediction: Analyzes IHC images to infer tissue-level and subcellular protein distributions.
• Biomarker Identification: Identifies candidate location biomarkers and potential members of protein networks for experimental prioritization in cancer research.
• Model Interpretability: Provides interpretable predictions to support biological interpretation of localization patterns.

Methodology:

Builds a graph from multiple IHC images associated with each protein, applies deep graph convolutions to learn protein-level representations, and applies a dynamic threshold to produce multi-label subcellular location predictions.