Finding relevant references to genes and proteins in Medline using a Bayesian approach

Area of Science:

• Biomedical Informatics
• Computational Biology
• Natural Language Processing

Background:

• Automated mining of biomedical literature for gene and protein references faces a precision-recall tradeoff.
• Accurate assessment of literature mining results is vital for balancing precision and recall.
• Hypothesis: Abstracts and titles discussing the same gene or protein employ similar word usage.

Purpose of the Study:

• To develop and validate a method for assessing the relevance of gene and protein references in biomedical literature.
• To balance precision and recall in literature mining for genes and proteins.
• To enable automated analysis of literature mining results.

Main Methods:

• A dictionary- and rule-based system was developed to mine Medline for gene and protein references.
• A Bayesian metric, estimated probability (EP), was used to score the relevance of each reference.
• Extensive analysis of Medline records from 1966-2001, including a large training set of known assignments.

Main Results:

• Analysis of Medline records (1966-2001) showed distinct distributions of EP values for good and bad gene/protein assignments.
• The Bayesian EP scoring method demonstrated effectiveness in balancing precision and recall.
• Achieved 100% recall at 61% precision, 63% recall at 88% precision, and 10% recall at 100% precision.

Conclusions:

• Medline entries discussing the same gene or protein exhibit similar word usage.
• The Bayesian EP-based method is a valid approach for assessing reference relevance and balancing precision/recall.
• This method allows for the determination of an EP cutoff value for accurate and reproducible automated literature analysis.