Table 1. Benchmark dataset statistics and training alignment information content.

Family	Ntrain	Ntest	Consensus Length	1° info (bits)	2° info (bits)	3° info (bits)
tRNA	1357	30	72	40.3	26.7	1.2
Twister	1005	40	65	57.5	4.5	1.3
SAM	192	8	108	121.8	11.1	0.0

Hand-curated MSAs are split into training and test sets based on [45]. For each training MSA, information content in the primary sequence (in bits) is calculated [39], while information in secondary structure (nested base pairs) and tertiary structure (all other disjoint pairwise interactions between sites) is estimated using mutual information [6]. Each family’s consensus structure is inferred using CaCoFold and R-scape on the training alignment [41, 42]. Though R-Scape identifies no tertiary structure using the SAM riboswitch training alignment, a four-base pair pseudoknot has been observed experimentally [46]. This lack of pseudoknot detection is a characteristic of our SAM training alignment; R-scape predicts the pseudoknot when analyzing the RF00162 seed alignment.