A high-throughput approach to profile RNA structure

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

CROSS is able to identify sequence patterns that cannot be captured by a position weight matrix approach. We searched the positive and negative fragment sets extracted from icSHAPE-Mouse and PARS-Human data for sequence patterns (Supplementary Table S9 ) using DREME (Materials and Methods: Sequence patterns) (19) . In icSHAPE-Mouse sequences, the G/GC/ACGU/GC motif occurs with frequencies of 63% and 43% in the positive (556 645 fragments) and negative (355 740 fragments) sets (Supplementary Table S9 ), indicating poor discrimination. As for PARS-Human, the top motif in the positive fragment set GCU/GC/AG/G (71% frequency) is also non-specific (frequency of 47% in the negative set). This observation indicates that the neural network approach is particularly suitable to identifying complex patterns in biological sequences, which is key to discover trends in large data sets (20) . We also note that CROSS models are sensitive to single point mutations: the signal drops progressively upon insertion of random mutations in the original sequences (PARS-Yeast; Supplementary Figure S7 ). As expected, mutations in the central position of the fragment produce the most dramatic reduction in the predictive power of the method (Supplementary Figure S8 ).