Classification of protein-protein interaction full-text documents using text and citation network features

Area of Science:

• Biomedical text mining
• Bibliome informatics
• Bioinformatics

Background:

• Protein-protein interactions (PPIs) are crucial in biological processes.
• Accurate classification of scientific literature is essential for extracting PPI information.
• Previous methods often focused on abstracts, necessitating approaches for full-text documents.

Purpose of the Study:

• To classify full-text articles for relevance to protein-protein interactions in the Biocreative II.5 Challenge.
• To evaluate the performance of a lightweight linear classifier (VTT) and a novel citation network-based classifier.
• To explore new avenues in bibliome informatics for biomedical literature analysis.

Main Methods:

• Utilized two distinct classifiers: Variable Trigonometric Threshold (VTT) linear classifier and a Naive Bayes classifier.
• Incorporated features from the citation network for the Naive Bayes classifier.
• Augmented the provided training data with full-text documents from the MIPS database.

Main Results:

• The lightweight VTT classifier was a top-performing submission in the full-text classification task.
• Performance was evaluated using Area Under the Curve, Accuracy, Balanced F-Score, and Matthew's Correlation Coefficient.
• The citation network classifier performed above average, indicating potential for future development.

Conclusions:

• The VTT classifier demonstrates high competitiveness for full-text article classification in the PPI domain.
• Citation network analysis presents a promising, albeit nascent, approach for biomedical text mining.
• Further investigation into citation network features could enhance bibliome informatics tools.