LocText: relation extraction of protein localizations to assist database curation

Area of Science:

• Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Protein subcellular localization is crucial for function but experimental annotations are incomplete.
• Existing text-mining methods struggle to accurately extract protein-location relationships from literature.
• Database curators require efficient tools to supplement experimental data with literature-derived annotations.

Purpose of the Study:

• To develop a novel text-mining method, LocText, for extracting protein subcellular localization information.
• To improve the accuracy and efficiency of identifying protein localization annotations from scientific texts.
• To assist database curators in enriching protein localization data for model organisms.

Main Methods:

• LocText was developed using patterns from syntax parse trees and trained on the LocTextCorpus.
• An automatic named-entity recognizer was integrated with LocText for enhanced performance.
• The method was applied to mine recent publications for human, yeast, and thale cress protein localization data.

Main Results:

• LocText achieved high precision (86%±4) in extracting protein localization information.
• Validation of text-mined annotations showed high accuracy: 65% for human, 85% for yeast, and 80% for thale cress.
• Approximately 40% of validated annotations were novel, not previously present in Swiss-Prot.

Conclusions:

• LocText offers a cost-effective, semi-automated workflow for discovering novel protein localization annotations.
• The tool assists database curators in identifying potential annotations for expert verification.
• This approach enhances the quality and completeness of manually-curated protein databases.