Evaluation of Database Annotation to Determine Human Mitochondrial Proteins

Abstract

Subcellular localization can be a helpful indication of the function of an unknown protein. Among the reported human mitochondrial proteins, hundreds of proteins have still not been functionally confirmed. To date, several databases for such proteins have been developed; however, their annotations overlap incompletely. A key issue in the completion of a reliable catalog of mitochondrial proteins is the integration of all this information, and the evaluation of the influence of different forms of evidence is also important. Here, we integrated various pieces of evidence (features) from both experimental and computational analyses. Linear and nonlinear prediction models were examined to predict human mitochondrial proteins. By employing a random forest model, an F-score of 0.929 was achieved by cross validation. The contributions of individual features toward the accurate prediction of localization were evaluated. We found only minor differences in importance among different features, with accurate prediction requiring the combination of many features; however, evidence from mass spectrometry experiments emerged as a prominent feature. Focusing on human mitochondrial proteins, we have constructed a high-accuracy prediction model that utilizes many weak features. Evaluation of the importance of individual features provides insights into what information is most valuable for the confirmation of protein localization.