Skip to main content
. 2024 Dec 5;15:10627. doi: 10.1038/s41467-024-54812-y

Fig. 5. 23S rRNA sequences generated by GNN and GPT-like RNA models.

Fig. 5

a–d Cmsearch scores for sequences generated from the pretrained GNN model (a), finetuned GNN model (b), pretrained RNA LM (c), and finetuned RNA LM (d) trained on 23S rRNA sequences at generation temperature T = 0.5 compared to naturally occurring 23S rRNAs in GARNET. For the GARNET reference distributions, random subsets of 1000 bacterial sequences and 1000 archaeal sequences were used. e–h 23S rRNA sequences generated from the pretrained GNN model (e), finetuned GNN model (f), pretrained RNA LM (g), and finetuned RNA LM (h) according to the fraction of disrupted canonical base pairs (i.e. Watson-Crick-Franklin and G-U) relative to the Rfam RF02541 consensus secondary structure (denoted non-canonical base pairs) in the generated sequences compared to naturally-occuring 23S rRNAs.