Fig. 1.

Experimental setup described by Renz and al. [12]. The initial dataset is split in two sets. The first split is used as a training set for the optimization model and the model-control model, and the second split for the data-control model. For a given molecule, the optimization (resp. model-control, data-control) score $S_{\mathit{opt}}$ (resp. $S_{\mathit{mc}}$, $S_{\mathit{dc}}$) is given by the optimization model’s (resp. model-control model’s, data-control model’s) predicted probability of being active The optimization score is used to guide goal-directed generation, and the evolution of control scores is also tracked during optimization. While the optimization score $S_{\mathit{opt}}$ grows throughout training, the control scores $S_{\mathit{mc}}$ and $S_{\mathit{dc}}$ stagnates and reaches much lower values