Fig. 5.

Further refined workflow for part of Fig. 4 on training a Restricted Boltzmann Machine (RBM) to detect contact patterns. Specific steps to extract the contact weights from RNA tertiary structure are as follows. a Prepare contact training set. A contact map of a given RNA is constructed from its nucleotide-nucleotide distance matrix. Two nucleotides are considered in contact if a pair of heavy atoms, one from each nucleotide, is less than 8 Å apart. The lower triangles of the contact map are maintained and then converted to a one-dimensional array as the input to RBM. b Training set. The training set consists of all contact maps of riboswitch structures but with the testing homologous riboswitch structure removed. c RBM model. Parameters in RBM are trained by the Contrastive Divergence (CD) algorithm. d Gibbs sampling. We run Gibbs sampling to generate new contact maps using RBM model. The last 50,000 samples are maintained for contact weight calculation. e Contact weight. The Gibbs sampling results are normalized into one contact matrix representing nucleotide-nucleotide contact weights for a typical riboswitch structure