Fig. 4

Learning of PF13354 out-of-equilibrium. a We show the three overlaps \(q_{ext}\), \(q_{int1}\) and \(q_{int2}\) of the sampled configurations used to estimate the model statistics as a function of the iterations (left axis) and the waiting time T_wait between two consecutive samples (right axis). At difference with the learning at equilibrium, T_wait is here kept constant during the training and the configurations are correlated as suggested by the differences between the distributions of \(q_{int1}\), \(q_{int2}\) and \(q_{ext}\). b The plot of the quality metrics used to estimate the goodness of the training: in blue we show the Pearson correlation coefficient between the two-site connected frequencies of the natural sequences and of the evolving model as a function of the iterations (blue markers, left axis) and the fitting errors (red, orange and brown markers, right axis) computed as \(\varepsilon _{c}\) for the two-site connected statistics and as \(\varepsilon _{f,s}\) for the one-site and two-site non-connected statistics. c, d We show the projections of the natural sequences and of the re-sampled sequences into the first two principal components of the natural sequences while in e we plot the positive predictive value curve associated with the contact map prediction (shown in f) for the Beta-lactamase2 domain