Classification Systems for Bacterial Protein-Protein Interaction Document Retrieval

Introduction

Protein-protein interaction (PPI) network is essential to understand the fundamental processes governing cell biology. PPI network data for several organisms has already been generated by high-throughput studies and submitted to/or collected by various PPI interaction databases (Morrison et al., 2005). However, the majority of PPIs are reported in the scientific literature. To collect such data in a standardized way and to avoid duplication of efforts, IMEx¹ databases such as IntAct (http://www.jcvi.org/mpidb) conduct coordinated manual literature curation. Text mining system to prioritize articles for curators according to their PPI relevance can accelerate such curation processes significantly. For example, MPIDB curators scan a whole issue (20 to 50 articles) of the Journal of Bacteriology or Molecular Microbiology and find approximately 10% of these articles report interaction experiments. Thus, the curators spend roughly 90% of their time reading irrelevant articles. A text mining system to prioritize articles for curators can be developed using supervised classification algorithms that provide certain kinds of confidence scores during classification. In order to build such systems, a class-labeled corpus is needed where PPI-relevant documents are labeled as positive and those irrelevant as negative. In many real-world applications, it is common that positive instances are explicitly included in a designated database, but it is uncommon to also include negatives in the database (Elkan & Noto, 2008). In developing a PPI mining application, PPI-relevant documents can be retrieved from existing PPI knowledge bases and unlabeled documents are available in large literature repositories such as PubMed. Learning with only positively labeled documents has great importance in this application.

We consider learning with only positive labeled documents as learning from a noisy and imbalanced training set where unlabeled documents are considered as negatives with some mislabeled documents. We build a document retrieval system to assist the curation of MPIDB (Goll et al., 2008) and report our investigation of the stability of two document classification algorithms with respect to the ratio of positives and unlabeled documents in the training set and also of the impact of feature selection on the classification performance. We also propose to use different subsets of unlabeled documents and form an ensemble of classifiers.

In the following, we first describe the background of classification algorithms. The experimental methods are introduced next. We then present the results and discussion, and conclude our work.