Fig. 3. A 23S rRNA generative model using GTDB sequences and large ribosomal subunit structures.

a Graph Neural Network (GNN) model schematic. b GNN model architecture. Panels (a) and (b) are adapted from Ingraham et al⁴¹. to illustrate their use for RNA. For detailed model parameters and training data statistics refer to Supplementary Data 3. c Test perplexity of the GNN models plotted as a function of k-nearest neighbors, highlighting that the model does not significantly improve for k values greater than 50. The final perplexity of the model with hidden dimensions d = 128, and k = 50 was 1.751. d Histogram of inter-nucleotide distances sampled by selecting k nearest neighbors in the distance matrix for E.coli 23S rRNA structure (PDB ID: 7K00)⁴². Choosing k = 50 covers all distances less than 12 Å. e Comparison of the contact maps generated from the distance matrices, based either on the distance cutoff or the k nearest-neighbors criteria (see Methods). Top-right, the sum of the contact maps for 18 bacterial and archaeal ribosomal RNA structures, projected onto the MSA sequence alignment, and based on the 12 Å distance cutoff criterion. The number of contact maps that align for a given pair of nucleotides is color-coded in the color bar on the right. Bottom-left, contact map for E. coli 23S rRNA, based on selecting k = 50 nearest neighbors to each nucleotide. The two types of contact maps show high similarity. f Structure of the three stem-loops highlighted in (e). A 12 Å inter-helical packing contact is shown with a dashed line in (f), and with an arrow in (e).