GRGNN

GRGNN reconstructs gene regulatory networks from gene expression data by applying graph neural networks to formulate GRN inference as a link-prediction problem.

Key Features:

• End-to-End Approach: Performs end-to-end reconstruction of gene regulatory networks from gene expression data using graph neural networks.
• Supervised and Semi-Supervised Frameworks: Uses both supervised and semi-supervised training frameworks to improve accuracy and generalization.
• Semi-Supervised Scheme: Implements a semi-supervised scheme that leverages labeled and unlabeled data to enhance classifier quality.
• Graph Classification / Link Prediction: Formulates inference as a graph classification problem by classifying subgraphs centered on a transcription factor and a target gene as linked or unlinked.
• Node Feature Integration: Integrates node features derived from explicit gene expression data and graph embeddings.
• Noisy Starting Graph: Initializes with a noisy starting graph built from Pearson’s correlation coefficient and mutual information.
• Ensemble Refinement: Refines the initial noisy graph through an ensemble technique to guide the inference process.
• Topological Neighbor Propagation: Incorporates topological neighbor propagation to propagate information across network neighborhoods.
• Benchmark Performance: Demonstrated superior performance compared to established methods on the DREAM5 GRN inference benchmarks.
• Biological Interpretability: Enables interpretation of inferred regulatory relationships following the biological network motif hypothesis.

Scientific Applications:

• Gene Regulatory Network Reconstruction: Reconstructs complex gene regulatory networks from gene expression datasets.
• Regulatory Mechanism Analysis: Investigates genetic interactions and regulatory mechanisms through inferred links and motifs.
• Biological Interpretation: Facilitates interpretation of network structure under the network motif hypothesis.
• Translational Research: Supports applications in disease modeling, drug discovery, and personalized medicine.
• Method Benchmarking: Serves for benchmarking and method comparison on DREAM5 GRN inference data.

Methodology:

Applies graph neural networks and formulates inference as subgraph-based graph classification/link prediction using node features from gene expression and graph embeddings, initializes with a noisy graph from Pearson’s correlation coefficient and mutual information which is refined via an ensemble technique, and employs supervised and semi-supervised training with topological neighbor propagation.