RESUMO
BACKGROUND: RNA molecules play critical roles in the cells of organisms, including roles in gene regulation, catalysis, and synthesis of proteins. Since RNA function depends in large part on its folded structures, much effort has been invested in developing accurate methods for prediction of RNA secondary structure from the base sequence. Minimum free energy (MFE) predictions are widely used, based on nearest neighbor thermodynamic parameters of Mathews, Turner et al. or those of Andronescu et al. Some recently proposed alternatives that leverage partition function calculations find the structure with maximum expected accuracy (MEA) or pseudo-expected accuracy (pseudo-MEA) methods. Advances in prediction methods are typically benchmarked using sensitivity, positive predictive value and their harmonic mean, namely F-measure, on datasets of known reference structures. Since such benchmarks document progress in improving accuracy of computational prediction methods, it is important to understand how measures of accuracy vary as a function of the reference datasets and whether advances in algorithms or thermodynamic parameters yield statistically significant improvements. Our work advances such understanding for the MFE and (pseudo-)MEA-based methods, with respect to the latest datasets and energy parameters. RESULTS: We present three main findings. First, using the bootstrap percentile method, we show that the average F-measure accuracy of the MFE and (pseudo-)MEA-based algorithms, as measured on our largest datasets with over 2000 RNAs from diverse families, is a reliable estimate (within a 2% range with high confidence) of the accuracy of a population of RNA molecules represented by this set. However, average accuracy on smaller classes of RNAs such as a class of 89 Group I introns used previously in benchmarking algorithm accuracy is not reliable enough to draw meaningful conclusions about the relative merits of the MFE and MEA-based algorithms. Second, on our large datasets, the algorithm with best overall accuracy is a pseudo MEA-based algorithm of Hamada et al. that uses a generalized centroid estimator of base pairs. However, between MFE and other MEA-based methods, there is no clear winner in the sense that the relative accuracy of the MFE versus MEA-based algorithms changes depending on the underlying energy parameters. Third, of the four parameter sets we considered, the best accuracy for the MFE-, MEA-based, and pseudo-MEA-based methods is 0.686, 0.680, and 0.711, respectively (on a scale from 0 to 1 with 1 meaning perfect structure predictions) and is obtained with a thermodynamic parameter set obtained by Andronescu et al. called BL* (named after the Boltzmann likelihood method by which the parameters were derived). CONCLUSIONS: Large datasets should be used to obtain reliable measures of the accuracy of RNA structure prediction algorithms, and average accuracies on specific classes (such as Group I introns and Transfer RNAs) should be interpreted with caution, considering the relatively small size of currently available datasets for such classes. The accuracy of the MEA-based methods is significantly higher when using the BL* parameter set of Andronescu et al. than when using the parameters of Mathews and Turner, and there is no significant difference between the accuracy of MEA-based methods and MFE when using the BL* parameters. The pseudo-MEA-based method of Hamada et al. with the BL* parameter set significantly outperforms all other MFE and MEA-based algorithms on our large data sets.