Mining Protein Interactions from Text Using Convolution Kernels

As the sizes of biomedical literature databases increase, there is an urgent need to develop intelligent systems that automatically discover Protein-Protein interactions from text. Despite resource-intensive efforts to create manually curated interaction databases, the sheer volume of biological literature databases makes it impossible to achieve significant coverage. In this paper, we describe a scalable hierarchical Support Vector Machine(SVM) based framework to efficiently mine protein interactions with high precision. In addition, we describe a convolution tree-vector kernel based on syntactic similarity of natural language text to further enhance the mining process. By using the inherent syntactic similarity of interaction phrases as a kernel method, we are able to significantly improve the classification quality. Our hierarchical framework allows us to reduce the search space dramatically with each stage, while sustaining a high level of accuracy. We test our framework on a corpus of over 10000 manually annotated phrases gathered from various sources. The convolution kernel technique identifies sentences describing interactions with a precision of 95% and a recall of 92%, yielding significant improvements over previous machine learning techniques.