Capability of the inferred couplings to reproduce frequencies using different methods (Boltzmann learning, pseudolikelihood, and mean-field DCA). The validation is done running a parallel MC simulation on 20 sequences and calculating the root-mean-square deviation (RMSD) between the obtained frequencies and the empirical ones. We report a positive control (statistical error due to the finite number of sequence), a negative control (RMSD between empirical sequences and a random sequence), and the RMSD from the ensemble of the 20 sequences used as a starting point of Boltzmann learning simulations. Families are labeled using the PDB code of the representative crystallographic structure. Average RMSD is also reported.
