A sentence sliding window approach to extract protein annotations from biomedical articles

Area of Science:

• Biomedical Informatics
• Computational Biology
• Natural Language Processing

Background:

• Text mining and statistical natural language processing (NLP) are rapidly advancing in biomedical research.
• A need exists for comparative assessments and standardized evaluation criteria for these methods.
• The Critical Assessment of Text Mining Methods in Molecular Biology (BioCreative) contest addresses these needs.

Purpose of the Study:

• To evaluate text mining system performance on biomedical texts.
• To assess tools for recognizing named entities like genes and proteins.
• To evaluate tools for automatic protein annotation extraction.

Main Methods:

• Development and application of a "sentence sliding window" approach.
• Extraction of text fragments from full-text articles for protein annotations.
• Analysis of the accuracy of extracted individual entities (proteins, GO terms) and complete annotations (protein-function relations).

Main Results:

• The sentence sliding window approach demonstrated high efficiency in extracting protein annotations.
• This method yielded the highest number of correctly predicted annotations.
• Correct extraction of individual entities (proteins, GO terms) significantly surpassed the extraction of complete protein-function relations.

Conclusions:

• Averaging sentence sliding windows show promise for information extraction, particularly without conventional training data.
• Combining this approach with advanced statistical estimators and machine learning can enhance future annotation extraction in biomedical text mining.