Ab initio gene prediction for protein coding regions

Abstract Motivation Ab initio gene prediction in non-model organisms is a difficult task. While many ab methods have been developed, their average accuracy over long segments of genome, and especially when assessed wide range species, generally yields results with sensitivity specificity levels the low 60% range. A common weakness most tendency to learn patterns that are species-specific varying degrees. The need exists for extract genetic features can distinguish coding non-coding regions not sensitive specific organism characteristics. Results new method based on neural network (NN) uses collection sensors create input presented. It shown accurate predictions achieved even trained significantly different phylogenetically than test organisms. consensus algorithm CoDing Sequence (CDS) subsequently applied first nucleotide level NN boosts through data driven procedure optimizes CDS/nonCDS threshold. An aggregate benchmark at shows this approach performs better existing methods, while requiring less training data. Availability https://github.com/BioMolecularPhysicsGroup-UNCC/MachineLearning Supplementary information available Bioinformatics online.