All-paths graph kernel for protein-protein interaction extraction with evaluation of cross-corpus learning

Area of Science:

• Biomedical text mining
• Bioinformatics
• Natural Language Processing

Background:

• Automated extraction of protein-protein interactions (PPI) is crucial in biomedical text mining.
• Existing methods often have limitations in utilizing comprehensive sentence structure information.
• A graph kernel-based approach is proposed to address these limitations.

Purpose of the Study:

• To introduce and evaluate an all-paths graph kernel for automated PPI extraction.
• To provide a comprehensive evaluation of the proposed method on multiple PPI corpora.
• To investigate the generalizability of machine learning models across different datasets.

Main Methods:

• Utilizing an all-paths graph kernel that leverages full dependency graphs of sentences.
• Performing extensive evaluations on five publicly available PPI corpora.
• Conducting detailed analyses of cross-corpus training and testing effects.

Main Results:

• The proposed graph kernel method achieves state-of-the-art performance.
• Achieved 56.4 F-score and 84.8 AUC on the AImed corpus in comparable evaluations.
• Demonstrated insights into the challenges of applying models to unseen data.

Conclusions:

• The graph kernel approach demonstrates state-of-the-art performance in PPI extraction.
• The study highlights the importance of proper evaluation strategies to ensure comparability and validity.
• Findings offer insights into model generalization and potential extensions to complex interaction extraction.