Learning from positive and unlabeled documents for retrieval of bacterial protein-protein interaction literature

With the advance of high-throughput genomics and proteomics technologies, it becomes critical to mine and curate protein-protein interaction (PPI) networks from biological research literature. Several PPI knowledge bases have been curated by domain experts but they are far from comprehensive. Observing that PPI-relevant documents can be obtained from PPI knowledge bases recording literature evidences and also that a large number of unlabeled documents (mostly negative) are freely available, we investigated learning from positive and unlabeled data (LPU) and developed an automated system for the retrieval of PPI-relevant articles aiming at assisting the curation of a bacterial PPI knowledge base, MPIDB. Two different approaches of obtaining unlabeled documents were used: one based on PubMed MeSH term search and the other based on an existing knowledge base, UniProtKB. We found unlabeled documents obtained from UniProtKB tend to yield better document classifiers for PPI curation purposes. Our study shows that LPU is a possible scenario for the development of an automated system to retrieve PPI-relevant articles, where there is no requirement for extra annotation effort. Selection of machine learning algorithms and that of unlabeled documents would be critical in constructing an effective LPU-based system.