Figure 3.

Prediction accuracy in terms of the average unsigned error (top) and Pearson correlation (bottom) with respect to convergence of the ΔΔG estimates. For the convergence analysis, we used the measure derived by Hahn and Then,²⁸ the application of which to the alchemical free energy calculations we have described previously.²⁹ The measure is defined in the range [−1;1], where the values close to 0 denote well-converged estimates; thus, the smaller values on the x-axis denote better convergence. It is important to note that this measure only reports on the convergence of the estimator, but does not include information on the potential lack of sampling in the relevant phase space regions for either of the physical end states. The data points depicted in the figure were calculated by considering subsets of points below a corresponding convergence threshold; for example, for a convergence value of 0.8, only those data points were considered that had an average convergence in the protein–ligand branch of the thermodynamic cycle less than or equal to 0.8. The numbers in the panels indicate how many ΔΔG values were considered given a corresponding convergence threshold. FEP+ results correspond to the Schindler et al.⁷ 5 ns simulation protocol calculated for the same ΔΔG values that have been identified for a corresponding convergence threshold for the simulations in the current work. The FEP+ curves in the panels differ due to the fact that disparate ligand pairs are considered at varying convergence levels for different force fields.