High-throughput molecular dynamics (MD) simulations are a computational method consisting of using multiple short trajectories, instead of few long ones, to cover slow biological time scales. Compared to long trajectories this method offers the possibility to start the simulations in successive batches, building a knowledgeable model of the available data to inform subsequent new simulations iteratively. Here, we demonstrate an automatic, iterative, on-the-fly method for learning and sampling molecular simulations in the context of ligand binding for the case of trypsin–benzamidine binding. The method uses Markov state models to learn a simplified model of the simulations and decide where best to sample from, achieving a converged binding affinity in approximately one microsecond, 1 order of magnitude faster than classical sampling. This method demonstrates for the first time the potential of adaptive sampling schemes in the case of ligand binding.