ERBK

ERBK integrates structured axioms and unstructured textual definitions to learn representations of biological entities for encoding biological knowledge and predicting relations such as protein–protein interactions and gene–disease associations.

Key Features:

• Representation learning: Learns vector representations of bio-entities by combining information from structured axioms and natural language definitions.
• Dual encoding: Applies knowledge graph embedding methods to structured axioms and deep convolutional neural models to textual definitions.
• Knowledge graph embeddings: Encodes machine-readable ontology axioms using knowledge graph embedding algorithms.
• Textual encoding: Encodes natural language definitions using deep convolutional neural networks.
• Predictive performance: Demonstrated improved prediction of protein-protein interactions and gene-disease associations relative to existing methods.
• Zero-shot robustness: Maintains performance under zero-shot scenarios where entities lack labeled training examples.
• Generality: Produces representations intended to generalize across multiple types of bio-relations beyond the initially tested tasks.

Scientific Applications:

• Protein-protein interaction prediction: Uses learned representations to predict protein–protein interactions.
• Gene-disease association prediction: Uses representations to predict gene–disease associations.
• Zero-shot inference: Enables inference for entities or relations without labeled training data (zero-shot).
• Broad bio-relation discovery: Applicable to discovery and analysis of other biological relations beyond proteins and genes.

Methodology:

Applies knowledge graph embedding methods to structured axioms and deep convolutional neural models to textual definitions to produce integrated entity representations.