Random and natural non-coding RNA have similar structural motif patterns but can be distinguished by bulge, loop, and bond counts

An important question in evolutionary biology is whether and in what ways genotype-phenotype (GP) map biases can influence evolutionary trajectories. Untangling the relative roles of natural selection and biases (and other factors) in shaping phenotypes can be difficult. Because RNA secondary structure (SS) can be analysed in detail mathematically and computationally, is biologically relevant, and a wealth of bioinformatic data is available, it offers a good model system for studying the role of bias. For quite short RNA (length L [≤] 126), it has recently been shown that natural and random RNA are structurally very similar, suggesting that bias strongly constrains evolutionary dynamics. Here we extend these results with emphasis on much larger RNA with length up to 3000 nucleotides. By examining both abstract shapes and structural motif frequencies (ie the numbers of helices, bonds, bulges, junctions, and loops), we find that large natural and random structures are also very similar, especially when contrasted to typical structures sampled from the space of all possible RNA structures. Our motif frequency study yields another result, that the frequencies of different motifs can be used in machine learning algorithms to classify random and natural RNA with quite high accuracy, especially for longer RNA (eg ROC AUC 0.86 for L = 1000). The most important motifs for classification are found to be the number of bulges, loops, and bonds. This finding may be useful in using SS to detect candidates for functional RNA within junk DNA regions.