Fig. 2: CF-random outperforms other methods for predicting fold switching.

a CF-random performs more successfully and efficiently than SPEACH-AF, AF-cluster, and AF2 with no templates. b Most alternative conformations can be predicted successfully using very shallow sampling depths (those to the left of the dotted line comprise 72% of all predictions). In two cases, sampling at 1:2 produced homogenous well-folded structures similar to the dominant conformation, in which case single sequences were also tested (single). c Successful predictions of three fold-switching proteins not predicted by the other methods: human XCL1, human TRAP1N, and Escherichia coli (E.coli) RepE. Gray represents the reference PDB structures and single-folding regions; blue and pink indicate the fold-switching regions of dominant and alternative structures, respectively. Source data are provided as a Source Data file.