MTGCN

MTGCN employs a multi-task graph convolutional network to identify cancer driver genes by integrating gene features with protein-protein interaction (PPI) networks and producing probabilistic gene-level driver scores.

Key Features:

• Graph Convolutional Network (GCN): Utilizes GCNs to embed and process gene features within a protein-protein interaction (PPI) network for graph-based learning and feature integration.
• Multi-Task Learning Framework: Simultaneously addresses node prediction and link prediction to enable joint learning of driver gene labels and PPI relationships.
• Node Prediction: Identifies cancer driver genes as a node-level prediction task within the PPI graph.
• Link Prediction: Predicts interactions within the PPI network as an auxiliary task to support feature learning.
• Bayesian Task Weight Learner: Dynamically balances the relative importance of node and link prediction tasks using a Bayesian weighting mechanism during training.
• Probabilistic Output: Assigns a probability score to each gene indicating its likelihood of being a cancer driver gene.
• Feature Augmentation via PPI Embedding: Augments gene features using their representations derived from the PPI network.
• Feature Propagation and Aggregation: Propagates and aggregates augmented features across GCN layers to capture multi-hop network information.

Scientific Applications:

• Pan-cancer driver prediction: Applied to predict cancer driver genes across pan-cancer datasets.
• Single cancer-type driver prediction: Applied to identify cancer driver genes within specific single cancer types.

Methodology:

Gene features are augmented using representations in PPI networks and embedded with graph convolutional networks; a multi-task learning process jointly performs node prediction (driver identification) and link prediction with feature propagation and aggregation across GCN layers, and a Bayesian task weight learner dynamically balances the tasks, producing probabilistic gene-level scores.