PathBinder--text empirics and automatic extraction of biomolecular interactions

Area of Science:

• Bioinformatics
• Computational Biology
• Biomedical Informatics

Background:

• The proliferation of online biological text necessitates automated methods for extracting biomolecular interactions.
• Machine learning relies on implicit text properties, which are often not explicitly stated in literature.
• An alternative strategy, text empirics, leverages explicit textual properties to aid in mining biomolecular interactions.

Purpose of the Study:

• To investigate specific properties of biological text passages to facilitate the text empirics approach.
• To develop and apply a system for mining biomolecular interactions from biomedical texts.
• To report empirical findings on text properties relevant to biomolecular interaction extraction.

Main Methods:

• Manual analysis of syntactic and semantic properties of sentences describing biomolecular interactions.
• Design of the PathBinder algorithm using empirical data for interaction extraction.
• Utilizing probabilistic methods to combine evidence from multiple PubMed sentences for interaction likelihood assessment.

Main Results:

• Empirical data on sentence properties were collected and analyzed.
• The PathBinder system was developed to search PubMed for interaction-describing sentences.
• A biomolecular interaction network was constructed based on probabilistic interaction likelihoods.

Conclusions:

• The text empirics approach provides a computationally efficient method for extracting biomolecular interactions.
• This approach achieves performance competitive with existing methods for automatic interaction extraction.
• The findings support the use of empirical text properties for mining biomedical literature.