🕸🌱 Semantic Alignment of Gene and Trait Ontologies Using Vector Embeddings

This ongoing project bridges distributional semantics and biological ontologies to explore whether embedding models can recover and reason about relationships within and between genes (GO) and plant traits (PTO). By mapping biological concepts into a continuous vector space, we examine the feasibility of trait categorisation, gene–trait inference, and the future integration of language-based generalisation in bio-ontological knowledge systems.

Impact: This approach may enable scalable, zero-shot categorisation of novel traits and facilitate ontology-aware gene–phenotype predictions, contributing to both plant biology and semantic AI applications, such as knowledge graph completion, explainable trait discovery, and bio-NLP.

Key Analyses and Findings

• Ontology embeddings: We used sentence-transformers to embed 1,682 trait classes from PTO and 44,000+ gene function terms from GO.
• Trait clustering:
  • K-Means clustering on embeddings showed a low silhouette score (max ≈ 0.14), confirming that traits live in a semantically continuous space.
  • Hierarchical clustering provided a more interpretable taxonomy; visualised via TSNE and dendrograms.
    • Notably, traits with high semantic similarity (e.g., those containing terms like seed, leaf, sterility) clustered early with low dissimilarity values.
    • The resulting hierarchy is interpretable and biologically plausible, but further validation with domain experts is required.
• Gene–trait similarity:
  • Cosine similarity was computed between GO–PTO pairs.
  • A curated demo set showed no consistent distinction in similarity scores across biologically positive, neutral, or negative gene–trait pairs.
  • Even negatively correlated pairs (e.g., salt stress vs. shoot dry weight) yielded similarity scores as high or higher than positively associated ones.

Interim Conclusions

• Gene–trait semantic similarity shows no systematic signal using SBERT-based cosine similarity. Scores were inconsistent across hypothesized positive, neutral, and negative pairs.
• Surprisingly, some biologically unrelated or negatively correlated pairs received higher similarity scores than truly related ones, suggesting that semantic closeness does not imply biological relevance.
• The outcome likely reflects a limitation of distributional models, which capture contextual co-occurrence rather than mechanistic or causal relationships.
• Trait categorisation via hierarchical clustering produced an interpretable structure; for example, traits with related biological terms (such as leaf, seed, and sterility) clustered together early, validating a certain degree of semantic grouping.
• These findings emphasise the need to supplement language-based models with structured data (e.g., ontologies, knowledge graphs) and expert validation for applications in functional biology.

How to Use / Reproduce

1. Download data (.pkl) for ontologies and embeddings from the folder data (optionally, you can also generate data yourself by following the next step)

2. Run notebook in order (in the folder notebooks):

   00_read_ontology_files.ipynb: Read in ontologies (GO and PTO) and save them as .owl and .pkl

   01_ontology_to_embeddings.ipynb: Encode ontology labels from owl-files into vector embeddings and save

   02_compare_embedding_based_clustering_WIP.ipynb: Read PTO label embeddings for K-Means and Hierarchical Clustering

   03_predict_gene_phenotype_relationships_WIP.ipynb: Read PTO and GO label embeddings and compute similarities

Future Work

• Integrate graph-based models to improve biological specificity
• Explore hybrid embedding strategies using symbolic constraints
• Collaborate with plant science experts to curate and expand validation sets