A methodology for optimizing the e-value threshold in alignment-based gene ontology prediction using the ROC curve

Abstract

The prediction of gene ontology (GO) terms is an important field of study in computational biology. With the advent of high-throughput experimental technologies, large quantities of sequenced proteins have emerged and, consequently, the number of computational tools that are used for analysing such data has also increased. In this field, methods based on sequence alignments like BLASTP are the most commonly used tools by biologists and bioinformaticians. However, an incorrect choice for the e-value threshold advised to identify homology and subsequently spread GO terms, may originate predictors with very low sensitivities and thus achieve poor prediction performances. In this work, a new methodology for optimizing the e-value threshold used in alignment-based predictors is proposed. The methodology is based on selecting a neighborhood for creating a veto scheme among proteins with similar e-values. © Springer International Publishing Switzerland 2014.