Graph2GO

Graph2GO predicts protein functions by learning multi-modal graph-based representations that integrate heterogeneous interaction networks and intrinsic protein features to assign Gene Ontology (GO) terms.

Key Features:

• Integration of Heterogeneous Data: Combines sequence similarity networks and protein-protein interaction networks with intrinsic protein features including amino acid sequences, subcellular locations, and protein domains.
• Advanced Representation Learning: Employs attributed network representation learning to model both interaction networks and protein features within a unified graph representation.
• Predictive Performance: Outperforms baseline methods such as BLAST and approaches like Mashup and deepNF in benchmark evaluations and has been tested across multiple species.
• Extensibility: Architecture supports incorporation of additional protein features to improve predictive accuracy.

Scientific Applications:

• Hypothesis Generation: Generates testable hypotheses to guide large-scale biological experiments for protein function characterization.
• Large-scale Functional Annotation: Enables computational annotation of protein functions within the Gene Ontology framework to prioritize targets for experimental follow-up.
• Bioinformatics and Systems Biology: Supports analyses that require accurate protein function predictions across species and network contexts.

Methodology:

Uses multi-modal graph-based and attributed network representation learning to integrate interaction networks and intrinsic protein features, constructs latent representations, and uses those representations as feature inputs for machine learning models to predict Gene Ontology terms.