Fig. 2: Comparison of performance between RMSF-net and other relevant approaches.

a Overview of Occ2RMSF-net. In the first stage, the cryo-EM density (40³) is input into Occ-net to predict the probabilities (40³) of structure occupancy on the voxels, with P_u denoting the probabilities of voxels being occupied by the protein structure and P_o denoting that of not being occupied by the structure. Then in the second stage, the two-channel probabilities are input into RMSF-net to predict the RMSF on the center 10³ voxels. b Test performance of Occ2net. For six classification thresholds from 0.3 to 0.8, the precisions, recalls and F1-scores of the positive class (structure occupied) on the test set were computed and are shown in the plot. c Comparision of RMSF prediction performance between Occ2RMSF-net and RMSF-net_cryo on the dataset. CC is an abbreviation for correlation coefficient. d Count distribution of test correlation coefficients for DEFMap, RMSF-net_cryo, and RMSF-net on the dataset. e Data distribution of correlation coefficients for RMSF-net_cryo and RMSF-net_pdb relative to RMSF-net on the dataset. f Count distribution of test correlation coefficients for RMSF-net_pdb, RMSF-net_pdb01, and RMSF-net on the dataset. g Data distribution of correlation coefficients for RMSF-net and RMSF-net_pdb relative to RMSF-net_cryo on data points where the test correlation coefficients with RMSF-net_cryo are above 0.4. The color for each method in d, e, f and g is shown in the legend.