Evaluating eukaryotic secreted protein prediction

Open Access
Publisher Website
Google Scholar
Abstract
Background: Improvements in protein sequence annotation and an increase in the number of annotated protein databases has fueled development of an increasing number of software tools to predict secreted proteins. Six software programs capable of high throughput and employing a wide range of prediction methods, SignalP 3.0, SignalP 2.0, TargetP 1.01, PrediSi, Phobius, and ProtComp 6.0, are evaluated. Results: Prediction accuracies were evaluated using 372 unbiased, eukaryotic, SwissProt protein sequences. TargetP, SignalP 3.0 maximum S-score and SignalP 3.0 D-score were the most accurate single scores (90–91% accurate). The combination of a positive TargetP prediction, SignalP 2.0 maximum Y-score, and SignalP 3.0 maximum S-score increased accuracy by six percent. Conclusion: Single predictive scores could be highly accurate, but almost all accuracies were slightly less than those reported by program authors. Predictive accuracy could be substantially improved by combining scores from multiple methods into a single composite prediction.