Fig. 3: Secondary structure prediction.
From: RiNALMo: general-purpose RNA language models can generalize well on structure prediction tasks
a RNAs fold into various shapes according to their function and while doing so, many of their nucleotides pair up using a hydrogen bond. These pairings are crucial for structural stability and form structural motifs such as hairpin loops and bulges. b, RiNALMo produces nucleotide embeddings for the given RNA sequence. Nucleotide pair embeddings are constructed by applying outer concatenation to RiNALMo’s outputs. Finally, pair representations are fed into the convolutional bottleneck residual neural network (ResNet) which produces base pairing probabilities that are then converted into the final secondary structure prediction. c, Precision, recall and F1 performance of different deep learning models on the TS0 evaluation dataset. d Distribution of F1 scores for predictions of different models on the TS0 dataset (sample size n = 1305). e Precision, recall and F1 performance of different structure prediction tools on the TestSetB evaluation dataset. f Distribution of F1 scores for predictions of different structure prediction tools on the TestSetB dataset (n = 430). Cfold denotes CONTRAFold and RNAstruct denotes RNAstructure. g A target RNA from the TS0 evaluation dataset and its predictions from different deep learning models. In (c, e), the best result for each metric is shown in bold. In (d, f), Box plots show the median (center line), 25th and 75th percentiles (bounds of box), whiskers extending to the smallest and largest values within 1.5× the interquartile range, and individual outliers beyond the whiskers.
