Figure 1. Adaptive ensemble work diagrams.

Panel a schematizes an asynchronous replica exchange loop. Ensemble members are run asynchronously, so there is no global barrier before exchange or analysis. This is not per se an adaptive concern but is required for many efficient adaptive algorithms. An ensemble analysis then tests for convergence and either re-triggers the loop (perhaps with altered parameters) or writes a final output. Panel b schematizes more complex adaptive logic, where an initial simulation ensemble of protein-ligand interaction asynchronously triggers an analysis calculation (which could be clustering and Markov State Model construction). This analysis calculation either adaptively reseeds the ensemble simulation run or, if the run is converged, starts an ensemble free-energy-perturbation (FEP) calculation on a new ligand (lower branch). Depending on the result of this FEP calculation, it is either “accepted” and a new Markov State Model calculation started with the new ligand, or it is “rejected” and a new ligand tested. In all schemas, dark gray rectangles indicate ensemble simulations and light gray rectangles indicate analyses.