SmuDGE

SmuDGE generates semantic embeddings of phenotypes to prioritize disease-associated genes by comparing phenotype sets and leveraging interaction-network-derived indirect associations.

Key Features:

• Semantic Similarity Measure: Utilizes a trainable semantic similarity measure to compare sets of phenotypes between entities such as disease-gene or disease-patient pairs.
• Feature Learning and Vector Representations: Produces vector-based phenotype representations that capture relationships from direct phenotype annotations and learned features.
• Exploitation of Interaction Networks: Exploits background knowledge from interaction networks, including protein-protein interactions, to derive phenotype representations for indirectly associated entities.
• Enhanced Coverage and Performance: Extends coverage to all genes within a connected interaction network and can match or outperform traditional semantic similarity methods in phenotype-based disease gene prioritization tasks.

Scientific Applications:

• Disease Gene Prioritization: Ranks and identifies candidate disease-associated genes, including cases with sparse or missing direct phenotype annotations in human and model organism studies.
• Phenotype-Based Comparisons: Compares phenotypic profiles between diseases, genes, or patients to support investigations of disease mechanisms and translational or personalized-medicine analyses.

Methodology:

SmuDGE uses Python scripts (e.g., smudge_E.py) to generate corpora and feature vectors by sampling nodes' surrounding environments in interaction networks, extracts phenotypic information from the PhenomNet ontology including associated and superclass terms, trains a semantic similarity measure, and creates vector-based semantic embeddings that capture direct and indirect phenotype associations.