Fig. 1

Overview of structure profiling data and PATTERNA. a, b Schematic representation of an RNA secondary structure with paired and unpaired nucleotides highlighted in blue and orange, respectively. Structure diagrams were obtained with Forna [100]. a SHAPE experiments entail structure-dependent formation of chemical adducts, indicated by black pins on the RNA, which are subsequently detected by sequencing and used to produce a reactivity for each nucleotide. High/low reactivities correspond to unpaired/paired nucleotides. b PARS experiments use two nucleases: RNAse S1 cleaves single-stranded RNA while RNAse V1 cleaves double-stranded RNA. Cleavage sites are detected by sequencing and summarized into a single score, where negative/positive scores indicate unpaired/paired nucleotides. c Cartoon overview of PATTERNA. PATTERNA is trained on input structure profiles using an iterative expectation-maximization algorithm that learns the statistical properties of nucleotide pairing states and the data distributions associated with each pairing state. The illustrated GMM model uses three Gaussian components per pairing state. Once trained, PATTERNA can be applied to the same transcripts used for training or to new transcripts. The scoring phase uses the structure profiling data and the trained model to infer the posterior probabilities of each pairing state, which are then used to score the state sequence that represents the motif. Motifs are scored across all starting nucleotides and input transcripts. Optionally, sequence constraints can be applied to restrict the search to regions that permit the formation of the motif’s base pairs. GMM Gaussian mixture model, P paired, U unpaired