Genome-wide prediction of bacterial effectors across six secretion system types using a feature-based supervised learning framework

Abstract

AbstractGram-negative bacteria are responsible for hundreds of millions infections worldwide, including the emerging hospital-acquired infections and neglected tropical diseases in the third-world countries. Finding a fast and cheap way to understand the molecular mechanisms behind the bacterial infections is critical for efficient diagnostics and treatment. An important step towards understanding these mechanisms is discovering bacterial effectors, the proteins secreted into the host through one of the six common secretion system types. Unfortunately, current effector prediction methods are designed to specifically target one of three secretion systems, and no accurate “secretion system-agnostic” method is available.Here, we present PREFFECTOR, a computational feature-based approach to discover effectors in Gram-negative bacteria without prior knowledge on bacterial secretion system(s) or cryptic secretion signals. Our approach was first evaluated using several assessment protocols on a manually curated, balanced dataset of experimentally determined effectors across all six secretion systems as well as non-effector proteins. The evaluation revealed high accuracy of the top performing classifiers in PREFFECTOR, with the small false positive discovery rate across all six secretion systems. Our method was also applied to four bacteria that had limited knowledge on virulence factors or secreted effectors. PREFFECTOR web-server is freely available at: http://korkinlab.org/preffector.