Figure 3.

Top: sensitivity plot for contact prediction via parameters inferred on PSE-1 and AAC6 [29] datasets. Blue curve: the score is computed as the Frobenius norm of the couplings inferred with the AMaLa method. Orange curve: the score is computed as the Frobenius norm of the couplings inferred with the standard pseudo-likelihood maximization approach. In panel (a), we have the result for PSE-1. At the $L/2$-th ranked residue pair, AMaLa provides AUC$(L/2)=0.71$, PPV$(L/2)=0.58$, whereas PlmDCA yields AUC$(L/2)=0.72$, PPV$(L/2)=0.61$. Panel (b) shows the sensitivity plot for AAC6. In this case, AMaLa yields at half of the length AUC$(L/2)=0.51$, PPV$(L/2)=0.51$, whereas for PlmDCA, we have AUC$(L/2)=0.34$, PPV$(L/2)=0.31$. Bottom: contact maps up to $L/2$ predictions. In the upper-right half, the results related to AMaLa are reported, whereas in the lower-left, the prediction provided by PlmDCA is reported. Correctly predicted contacts are colored in green/blue, while wrong prediction are reported in red/orange for PlmDCA/AMaLa, respectively. Panel (c) reports the result for PSE-1. Even if DCA provides both higher AUC and PPV, AMaLa seems to predict more long range contacts. A similar outcome, although less pronounced, can be appreciated in panel (d), which shows the contact map related to AAC6.