Research On Machine Learning-based Protein-Protein Interaction Extraction

With the explosive increment of biomedicine literature, how to extract the information from biomedicine literature is becoming an important research area in the field of bioinformatics. Because of the special significance of protein-protein interaction to life science, extracting protein-protein interactions (PPI) has become a hot topic.Due to the complexity and variability of biomedical texts, PPI extraction from them is a difficult task. This paper carries out in-depth research on the task of PPI extraction using machine learning methods, with the efforts and goals on:1. Feature-based PPI extraction methods, with the focus on how to generate surface features and structural features from free texts. Furthermore, the contributions of various features to PPI extraction are systematically analyzed.2. Convolution tree kernel-based PPI methods. We analyze the impact of different structural representations of relation instances on PPI extraction, thus laying a good foundation for further research.3. Composite kernel-based methods, which combine a feature-based kernel and a convolution tree kernel, aiming at capturing both the flat features and the structural features of relation instances.Experiments of PPI extraction on the AIMed corpus show that our composite kernel achieves the promising results, with the F-measure as high as 53.7. Thus, the achievement by this paper exhibits great reference value to the future research in PPI.