Figure 1. Two models of PreQ1 class I type III riboswitch superimposed with the target structure (R1117, green). Model R1117TS119_3 (orange, panel B) correctly reproduces all canonical base pairs at the same time not reproducing any non-canonical interactions and having the RMSD of 9.90 Å. Model R1117TS287_2 (blue, panel A) misses 7% of canonical base pairs but is a much better model with the RMSD of 2.25 Å and non-canonical interactions reproduction rate of 39%. RMSD scores were computed on the 2-30 fragment.

Many RNA targets were derived from low-to-medium resolution (4-7 Å) cryo-EM maps. To remove bias and modeling uncertainties potentially present in the reference structures, we additionally evaluated predictions by directly comparing them to maps. After docking models into maps, predictions were scored with five different measures: CC_mask (cross-correlation masked by the area around the model), CC_peaks (cross-correlation masked by the model and the highest density peaks in the map), MI (mutual information), SMOC (segment-based Mander’s overlap coefficient), and AI (atomic inclusion)^27,28. It is worth noting that model-independent measures are known to overinflate scores for compressed models with unrealistic geometries. Thus, when using these scores, it is important to first filter out models with poor topology scores or, alternatively, combine these scores with geometric scores, such as Clashscore.