How helpful are the protein-protein interaction databases and which ones?

Protein-protein interactions (PPIs) are critical, and so are the databases and tools (resources) concerning PPIs. But in absence of systematic comparisons, biologists/bioinformaticians may be forced to make a subjective selection among such protein interaction databases and tools. In fact, a comprehensive list of such bioinformatics resources has not been reported so far. For the first time, we compiled 375 PPI resources, short-listed and performed preliminary comparison of 125 important ones (both lists available publicly at startbioinfo.com), and then systematically compared human PPIs from 16 carefully-selected databases. General features have been first compared in detail. The coverage of 9experimentally verified9 vs. all PPIs, as well as those significant in case of disease-associated and other types of genes among the chosen databases has been compared quantitatively. This has been done in two ways: outputs manually obtained using web-interfaces, and all interactions downloaded from the databases. For the first approach, PPIs obtained in response to gene queries using the web interfaces were compared. As a query set, 108 genes associated with different tissues (specific to kidney, testis, and uterus, and ubiquitous) or diseases (breast cancer, lung cancer, Alzheimer9s, cystic fibrosis, diabetes, and cardiomyopathy) were chosen. PPI-coverage for well-studied genes was also compared with that of less-studied ones. For the second approach, the back-end-data from the databases was downloaded and compared. Based on the results, we recommend the use of STRING and UniHI for retrieving the majority of 9experimentally verified9 protein interactions, and hPRINT and STRING for obtaining maximum number of 9total9 (experimentally verified as well as predicted) PPIs. The analysis of experimentally verified PPIs found exclusively in each database revealed that STRING contributed about 71% of exclusive hits. Overall, hPRINT, STRING and IID together retrieved ~94% of 9total9 protein interactions available in the databases. The coverage of certain databases was skewed for some gene-types. The results also indicate that the database usage frequency may not correlate with their advantages, thereby justifying the need for more frequent studies of this nature.