Hermes: an ensemble machine learning architecture for protein secondary structure prediction

Abstract

AbstractMotivationProtein structure predictions, that combine theoretical chemistry and bioinformatics, are an increasingly important technique in biotechnology and biomedical research, for example in the design of novel enzymes and drugs. Here, we present a new ensemble bi-layered machine learning architecture, that directly builds on ten existing pipelines providing rapid, high accuracy, 3-State secondary structure prediction of proteins.ResultsAfter training on 1348 solved protein structures, we evaluated the model with four independent datasets: JPRED4 - compiled by the authors of the successful predictor with the same name, and CASP11, CASP12 & CASP13 - assembled by the Critical Assessment of protein Structure Prediction consortium who run biannual experiments focused on objective testing of predictors. These rigorous, pre-established protocols included 7-fold cross-validation and blind testing. This led to a mean Hermes accuracy of 95.5%, significantly (p<0.05) better than the ten previously published models analysed in this paper. Furthermore, Hermes yielded a reduction in standard deviation, lower boundary outliers, and reduced dependency on solved structures of homologous proteins, as measured by NEFF score. This architecture provides advantages over other pipelines, while remaining accessible to users at any level of bioinformatics experience.Availability and ImplementationThe source code for Hermes is freely available at: https://github.com/HermesPrediction/Hermes. This page also includes the cross-validation with corresponding models, and all training/testing data presented in this study with predictions and accuracy.