Semantic Similarity of GO Terms

Two fundamental paradigms exist for computing GO term similarity. Graph-based methods rely purely on the topology of the ontology—measuring edge distances, shortest path lengths, or the depth of the lowest common ancestor (LCA). Information content (IC) methods weight each term by its specificity: rare, specific terms carry high IC, while broad, general terms carry low IC.

• Resnik's measure: Uses only the IC of the LCA
• Lin's measure: Normalizes by the IC of both query terms
• Jiang & Conrath: Incorporates edge distance with IC weighting
• Wang's measure: Uses a graph-based local topology weighting without requiring corpus statistics

The lowest common ancestor is the most specific term that subsumes both query terms in the ontology hierarchy. Its identification is the critical computational step in most similarity algorithms. In the GO directed acyclic graph (DAG), a term pair may have multiple common ancestors due to multiple parentage—the LCA is the one with maximum depth or information content.

• Multiple inheritance in GO complicates LCA selection
• The true LCA may not be unique; disambiguation strategies vary
• GraSM and AIC methods average across all disjunctive common ancestors to avoid overestimating similarity

Information content values are not intrinsic to the ontology—they depend on the reference corpus used for annotation frequency calculation. A term's IC is typically computed as the negative log of its annotation probability: IC(t) = -log(p(t)).

• Corpus choice matters: IC from UniProt differs from IC from a species-specific annotation set
• Intrinsic IC methods (e.g., topology-based IC) avoid corpus bias by deriving specificity from the ontology structure alone
• Annotation propagation rules (transitive closure) must be applied before frequency counting to avoid underestimating term occurrence

Semantic similarity operates on individual GO terms, but biological queries typically involve gene products annotated with multiple terms. Aggregation strategies convert term-pair similarities into gene-level scores.

• Best-Match Average (BMA): Averages the maximum similarity for each term in set A against all terms in set B, and vice versa
• Maximum strategy: Takes the single highest term-pair similarity—sensitive but noisy
• Average strategy: Averages all pairwise similarities—dilutes signal from multi-domain proteins
• SimGIC: A Jaccard-like set overlap measure weighted by term IC, avoiding pairwise decomposition entirely

Many gene products are incompletely annotated, with known functions missing from the GO database. This creates systematic underestimation of true functional similarity. Additionally, two proteins may share a deep functional similarity that is not captured by their explicit annotations because the relevant GO term has not yet been created.

• Annotation depth bias: Well-studied genes appear artificially more similar to each other
• Propagation strategies: Evidence code filtering (excluding IEA) can reduce but not eliminate bias
• Semantic drift: As the ontology evolves, similarity scores computed on older ontology versions may diverge from current biological understanding

The GO comprises three disjoint sub-ontologies: Biological Process (BP), Molecular Function (MF), and Cellular Component (CC). Standard semantic similarity measures should not be applied across sub-ontology boundaries because the root nodes are distinct and the semantic spaces are incommensurable.

• Comparing a BP term to an MF term via a shared ancestor requires traversing to the root, yielding near-zero similarity
• Cross-ontology normalization is an active research problem for integrative analyses
• Practical pipelines compute separate similarity matrices per sub-ontology and aggregate results downstream

A fundamental metric quantifying the specificity of a GO term. It is defined as the negative logarithm of the probability of a term occurring in a reference corpus.

• High IC: Rare, specific terms (e.g., 'pyrimidine dimer repair').
• Low IC: Common, generic terms (e.g., 'cellular process').

IC is the core currency for modern similarity measures like Resnik, Lin, and Schlicker, where shared ancestors with high IC indicate strong functional relatedness.

The shared parent term in the Gene Ontology graph that possesses the highest Information Content. Semantic similarity algorithms do not just look for any shared ancestor; they specifically identify the MICA to ensure the comparison is based on the most specific biological context shared by the two query terms, avoiding the dilution of similarity by overly generic root terms.

A foundational node-based semantic similarity metric defined simply as the Information Content of the MICA. It is mathematically straightforward but does not account for the specificity of the query terms themselves.

• Formula: Sim(t1, t2) = IC(MICA)
• Limitation: Two pairs of terms with the same MICA will yield identical similarity scores, even if one pair is much more specific than the other.

An information-theoretic metric that normalizes the IC of the MICA by the sum of the ICs of the two query terms. This generates a value between 0 and 1, correcting Resnik's bias.

• Formula: Sim(t1, t2) = (2 * IC(MICA)) / (IC(t1) + IC(t2))
• Advantage: Provides a relative measure that accounts for the depth of the query terms, making it more robust for comparing terms at different specificity levels.

A network-based visualization method that organizes enriched gene sets into a similarity network. Nodes represent gene sets, and edges represent the mutual overlap of their leading-edge genes. Semantic similarity of GO terms is often used as an alternative to gene overlap to define edges, creating a functional map that clusters redundant biological themes and simplifies the interpretation of complex enrichment results.