Integrating information retrieval with distant supervision for gene ontology annotation

Area of Science:

• Bioinformatics
• Computational Biology
• Biomedical Informatics

Background:

• The Gene Ontology (GO) provides a standardized vocabulary for gene function annotation.
• Accurate GO term assignment is crucial for understanding gene roles in biological processes.
• Automating GO curation from full-text articles presents significant challenges.

Purpose of the Study:

• To participate in the BioCreative IV Gene Ontology Curation task.
• To develop and evaluate systems for identifying GO evidence sentences (GOESs).
• To develop and evaluate systems for predicting GO terms for genes.

Main Methods:

• For GOES identification (Subtask A): Logistic regression model trained on existing annotations, supplemented with external negative data, followed by a greedy gene-sentence association approach.
• For GO term prediction (Subtask B): Two system types were developed: (i) search-based systems using information retrieval and distant supervision on various text granularities, and (ii) a similarity-based system measuring word distance to GO terms/synonyms.

Main Results:

• Subtask A best system achieved an F1 score of 0.27 (exact match) and 0.387 (relaxed overlap match).
• Subtask B best system (search-based) achieved an F1 score of 0.075 (exact match) and 0.301 (hierarchical match).
• Search-based systems for Subtask B significantly outperformed the similarity-based system.

Conclusions:

• The developed systems demonstrate progress in automated GO curation tasks.
• Search-based approaches show promise for predicting GO terms from literature.
• Further refinement is needed to improve performance, particularly for exact match scenarios.