Over last two decades development of machine-learning algorithms has made quantitative structure-activity relation models increasingly more accessible. In particular, active learning, an iterative approach, has generated a lot of interest in the pharmaceutical industry, allowing for screening of large chemical libraries for active ligands with relatively few evaluations of ligand affinity. To illustrate the effectiveness of active learning, we apply it to lead-optimisation of phosphodiesterase 2 inhibitors. We use non-equilibrium free energy calculations to estimate ligand binding, which we use as the ground truth for training the models, allowing for a completely in silico approach. In the early iterations we focus on ligand variety to establish promising regions of chemical space. Later iterations search for more optimal ligands within those regions. This strategy yields several potent inhibitors of phosphodiesterase 2. We also quantify the efficiency of our protocol by applying it to a retrospective dataset with available experimental affinity measurements for the same protein.