Fig. 5From: Deep clustering of protein folding simulationsCVAE can identify intermediate, partially folded states from Fs-peptide simulations. a We project the conformations from the simulations onto a latent space spanned by the three dimensions of the CVAE (z1−z3). Each conformation is then painted with the corresponding RMSD to the native state. As observed from the figure, there is a gradation from unfolded states (i.e., higher RMSD values) towards more folded states (lower RMSD values). To allow easy interpretation, we use t-SNE to project the conformations. b t-SNE projection painted with RMSD values and c fraction of native contacts shows clear separation between folded conformations, labeled (vii) along with partially folded states, namely (ii), (iii), (iv), and (vi), along with unfolded states, labeled (v). It is notable that state (viii) has lower RMSD to the folded state (vii), where as shows very few of the native contacts that are found in the folded state, implying that this state may be a misfolded state in the ensemble simulations. d A simple histogram representation of the projected t-SNE space allows us to characterize the conformational states observed within the simulations. The labels provide a relative description of the location of the various conformational statesBack to article page