CoGO

CoGO predicts disease similarity by learning joint embeddings of genes and Gene Ontology (GO) terms using graph deep learning and cross-view contrastive learning to compare diseases via gene-based representations.

Key Features:

• Integration of multiview data sources: Combines gene interaction networks and Gene Ontology (GO) domain knowledge as complementary biological views.
• Graph deep learning models: Encodes genes and GO terms using graph deep learning on their respective graph structures.
• Nonlinear projection into a common embedding space: Projects encoded gene and GO features into a unified embedding space using nonlinear projections.
• Cross-view contrastive loss: Applies cross-view contrastive loss to maximize agreement between corresponding gene–GO associations and improve representation quality.
• Disease similarity inference: Infers disease similarity by computing cosine similarity between disease representation vectors derived from related gene embeddings.

Scientific Applications:

• Disease similarity prediction and benchmarking: Predicts disease similarities and outperforms baseline methods, reporting improvements in AUROC and AUPRC, including a 19.57% increase in AUPRC to 0.7733.
• Genetic basis of disease analysis: Supports identification and comparative analysis of diseases based on shared genetic underpinnings.
• Case studies of similar disease pairs: Enables detailed case studies on top similar disease pairs with findings that can be corroborated against existing literature.

Methodology:

Encode gene and GO features using graph deep learning models tailored to each data structure, project encoded features into a unified embedding space via nonlinear projections, apply cross-view contrastive loss to align gene and GO views, and compute cosine similarity of disease representation vectors for disease similarity inference.