Fig. 4: Performance of all competing methods in recovering protein conformational distributions and residue contact probabilities. | Nature Machine Intelligence

Fig. 4: Performance of all competing methods in recovering protein conformational distributions and residue contact probabilities.

From: Conditional diffusion with locality-aware modal alignment for generating diverse protein conformational ensembles

Fig. 4

a, Conformations generated by MD simulations and seven generative models, projected onto the top two TICA components for α3D, Protein B and Homeodomain. Each dot represents a conformation and is colour coded by its density (the darker the point, the higher the density). The JS-TIC lower bound from intratrajectory MD is shown in the MD ensemble as a reference. b, Absolute differences in residue contact probability maps for the WW domain and Homeodomain, comparing seven competing methods with the MD-derived reference. Sparser or lighter pixels indicate smaller differences. Contact patterns of the β-sheets in the WW domain (boxed in orange and magenta) and the α-helix–α-helix interaction in the Homeodomain (boxed in red) are shown in the upper triangles. The folding structures at the top are marked with the corresponding colours.

Back to article page