A detailed error analysis of 13 kernel methods for protein-protein interaction extraction

Area of Science:

• Bioinformatics
• Computational Biology
• Natural Language Processing

Background:

• Kernel-based classification is the leading method for protein-protein interaction (PPI) extraction from text.
• Existing methods vary in kernel functions, input representations, and feature sets.
• Instance-level performance comparisons between these methods are lacking.

Purpose of the Study:

• To analyze shared characteristics and differences among 13 current PPI extraction methods.
• To evaluate method performance on easy and difficult protein-protein interaction pairs.
• To identify avenues for future performance improvements in PPI extraction.

Main Methods:

• Comparative analysis of 13 state-of-the-art kernel-based PPI extraction methods.
• Evaluation across five protein-protein interaction corpora.
• Identification and analysis of correctly and incorrectly classified PPI instances.

Main Results:

• Identified distinct sets of easy and difficult protein-protein interaction (PPI) pairs across corpora.
• Methods using similar input representations showed comparable performance on specific PPI pairs.
• Ensemble methods combining dissimilar kernels yielded significant performance gains.
• Few shared characteristics were found among difficult PPI pairs, limiting breakthrough potential for incremental method improvements.

Conclusions:

• Current protein-protein interaction (PPI) extraction methods struggle to capture shared characteristics of positive PPI pairs.
• Corpus heterogeneity contributes to the limitations of existing PPI extraction methods.
• Future performance enhancements in PPI extraction are more likely to come from novel feature sets than from new kernel functions.