Detection of Protein Subcellular Localization Based on Syntactic Dependency Paths

Abstract

Despite the widespread use of computers in biological research, the end result of almost all scientific experiments is a publication in the form of texts. So, many studies have sought to extract information from biomolecular text using natural language processing technology. Previous studies have asserted that linguistic information is useful for information extraction from text. In particular, syntactic relations among linguistic information are good for detection of protein interactions, protein subcellular localization, and so on. However, previous systems in detecting protein subcellular localization use only shallow syntactic parsers, and they did not show good performance. To improve performance in detecting protein subcellular localization, this paper proposes a three-step method based on a full syntactic dependency parser. In the first step, we construct syntactic dependency paths from each protein to its location candidate. In the second step, we retrieve root information of the syntactic paths. In the last step, we extract syntactic patterns of protein sub-trees and those of location sub-trees. According to the root information and syntactic patterns of sub-trees, we extract correct (protein, localization) pairs. Even without biomolecular knowledge, our method shows reasonable performance. In the experimental results using Medline abstracts data, our proposed method gave an F-measure of 76% for training data and 60% for test data, significantly outperforming previous methods. We also describe the contributions of root information and syntactic patterns of sub-trees to the performance.