PreBIND and Textomy--mining the biomedical literature for protein-protein interactions using a support vector machine

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• Biomedical literature contains vast, unstructured molecular interaction data inaccessible to computational analysis.
• The Biomolecular Interaction Network Database (BIND) aims to compile this data in a machine-readable format.
• Manual curation of existing literature for BIND is a time-intensive process.

Purpose of the Study:

• To reduce the manual effort required for backfilling BIND with molecular interaction data.
• To develop and evaluate a Support Vector Machine-based system for extracting protein-protein interaction information from biomedical literature.

Main Methods:

• Utilized Support Vector Machine (SVM) technology for information extraction.
• Developed an information extraction system to locate protein-protein interaction data.
• Applied the system to a real-world curation task for BIND.

Main Results:

• The SVM achieved high performance in classifying abstracts: 92% precision, 90% accuracy, and 92% recall.
• The system is estimated to recall up to 60% of non-high-throughput interactions.
• System implementation reduced curation task duration by 70%, saving 176 days.

Conclusions:

• Machine learning methods, like SVMs, are valuable tools for enhancing biological database curation.
• Coupling machine learning with human review is crucial for realizing the full potential of these techniques.
• The PreBIND system is publicly available and supports searching for protein-interaction data in humans, mice, and yeast.