A high-throughput approach to profile RNA structure

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

The model built on PARS-Human is able to predict SHAPE-HIV data with an AUC of 0.75 ( Figure 4A and B) . Increasing the confidence threshold of SHAPE data (from Figures S4 and S5) . The plot illustrates the performances of the icSHAPE-Mouse model tested on the SHAPE-HIV data set. (C) High-confidence predictions: the arrows connect the training and testing sets along withrelative accuracies (cross-validation on training sets are marked with circular arrows). We used the same number of nucleotides with high (doublestranded) and low (single-stranded) propensity scores for comparison with experimental data. Negligible overlap exists between training and testing sets (Jaccard index < 0.002 between each couple of sets analyzed; Supplementary Figure 4B ). We compared experimental and predicted values on fragments of 200 nucleotides, reporting an average correlation of 0.60 (peak of 0.86 in the region 3 800-4 000) for the HIV transcriptome ( Figure 4A and C) .