HPOLabeler

HPOLabeler predicts associations between human proteins and Human Phenotype Ontology (HPO) terms to prioritize protein-phenotype links.

Key Features:

• Learning to Rank (LTR) framework: applies an LTR framework to combine component model scores for large-scale multi-label ranking of HPO terms.
• Protein-protein interaction (PPI) networks: uses PPI network evidence and identifies PPI as a highly informative data source for prediction.
• Gene Ontology annotations: incorporates Gene Ontology annotations as an evidence type for scoring candidate HPO terms.
• InterPro domain data: uses InterPro domain data as an evidence feature for candidate term scoring.
• Trigram frequency analysis: computes trigram frequency features as part of the evidence set.
• HPO term frequencies: integrates HPO term frequency information as an input feature.
• Component learning models: trains logistic regression, nearest neighbor algorithms, and a Naive method as component predictors.
• Input score generation: generates input scores for candidate HPO terms from the diverse evidence types via the component models.
• Ranked HPO term output: produces ranked lists of HPO terms for each input protein.
• Empirical evaluation: assesses performance using cross-validation and temporal validation experiments.
• Annotation incompleteness observation: reports that incomplete annotation of newly discovered proteins can reduce temporal validation performance.

Scientific Applications:

• Protein-HPO association prediction: prioritizes HPO terms for proteins lacking curated HPO annotations.
• Genetic disease phenotype analysis: supports identification of candidate phenotypic abnormalities linked to disease-associated proteins.
• Functional annotation inference: leverages GO and InterPro evidence to infer phenotype associations from functional and domain data.
• Therapeutic hypothesis generation: informs genotype-phenotype link hypotheses relevant to targeted therapeutic research.

Methodology:

Integrates PPI networks, Gene Ontology annotations, InterPro domain data, trigram frequencies, and HPO term frequencies to generate input scores; trains component models (logistic regression, nearest neighbor, Naive) and applies a Learning to Rank framework to combine scores and produce ranked HPO term lists; evaluates performance via cross-validation and temporal validation.