Fig. 9

Per-molecule optimization by REINVENT and Augmented Hill-Climb RL strategies for the transformer (Tr) and gated transformer (GTr) architecture against the DRD2 benchmark objectives. Tr is more unstable during RL by REINVENT which is stabilized by the GTr. In all cases Augmented Hill-Climb outperforms REINVENT at objective optimization. Although these transformer models are more prone to mode collapse than an RNN as observed by a drop in validity and uniqueness as shown in Additional file 1: Figures S18-19