Plant proteins that are secreted without a classical signal peptide leader sequence are termed leaderless secretory proteins (LSPs) and are implicated in both plant development and (a)biotic stress responses. In plant proteomics experimental workflows, identification of LSPs is hindered by the possibility of contamination from other subcellar compartments upon purification of the secretome. Applying machine learning algorithms to predict LSPs in plants is also challenging due to the rarity of experimentally validated examples for training purposes. This work attempts to address this issue by establishing criteria for identifying potential plant LSPs based on experimental observations and training random forest classifiers on the putative datasets. The resultant plant protein database LSPDB and bioinformatic prediction tools LSPpred and SPLpred are available at lsppred.lspdb.org. The LSPpred and SPLpred modules are internally validated on the training dataset, with false positives controlled at 5%, and are also able to classify the limited number of established plant LSPs (SPLpred (3/4, LSPpred 4/4). Until such time as a larger set of bona fide (independently experimentally validated) LSPs is established using imaging technologies (light/fluorescence/electron microscopy) to confirm sub-cellular location, these tools represent a bridging method for predicting and identifying plant putative LSPs for subsequent experimental validation.
Read full abstract