A high-throughput approach to profile RNA structure

Delli Ponti, Riccardo; Marti, Stefanie; Armaos, Alexandros; Tartaglia, Gian Gaetano

From low-(top and bottom 50% of the PARS score distribution) to high-confidence (top and bottom 1%) experimental values, we found a consistent increase in the performances of all models (Supplementary Tables S6 and S7) , thus providing strong evidence on the reliability of CROSS predictions. For instance, the SHAPE-HIV model predicts the whole PARS-Human data set with an AUC of 0.70 and the top and bottom 1% of the scores are with an AUC of 0.80 (Supplementary Table S6 ). We note that very negligible overlap exists between yeast and human fragment sets (overlap: 0.001%; Jaccard index: 0.001; Supplementary Figure S6 ), which indicates that our method is not biased by specific sequences. On the same sets, approaches based on thermodynamic principles (15, 18) show lower performances (Yeast: accuracies in the range 0.72-0.74, Human: accuracies in the range 0.67-0.69) than CROSS (Yeast: 0.80 accuracy using PARS-Human model; Human: 0.81 accuracy using PARS-Yeast model; Supplementary Table S8), indicating that our method is particularly useful for predictions on high-throughput data sets.