An improved interolog mapping-based computational prediction of protein-protein interactions with increased network coverage.

Automated and efficient methods that map ortholog interactions from several organisms and public databases (pDB) are needed to identify new interactions in an organism of interest (interolog mapping). When computational methods are applied to predict interactions, it is important that these methods be validated and their efficiency proven. In this study, we compare six Blast+ metrics over three datasets to identify the best metric for protein-protein interaction predictions. Using Blast+ to align the protein pairs, the ortholog interactions from DIP were mapped to String, Intact and Psibase pDBs. For each interaction mapped to each pDBs, we retrieved the alignment score, e-value, bitscore, similarity, identity and coverage. We evaluated these Blast+ values, and combinations thereof, with the Receiver Operating Characteristic (ROC) curves and computed the Area Under Curve (AUC). To validate these predictions, we used a subset of the Database of Interacting Proteins (DIP) composed of experimental interactions curated by the International Molecular Exchange (IMEx). The cut-off point for each metric/pDB was computed aiming to identify the best one that separates the true and false predicted interactions. In contrast to other methods that only compute the first Blast hit, we considered the first 20 hits, thus increasing the number of predicted interaction pairs. In addition, we identified the contribution of each individual pDB, as well as their combined contribution to the prediction. The best metric had an AUC of 0.96 for a single pDB and AUC of 0.93 for combined pDBs. Compared to other studies, with a cut-off point of 0.70 representing a specificity of 0.95 and a sensitivity of 0.90 for individual pDB, our method efficiently predicts protein-protein interactions.