Fig. 1

Overviews of structure-constrained molecular generation and COMA. a Descriptions of three types of goal-directed molecular optimisation and examples. b Overviews of model architecture and training scheme of COMA. The generative model of COMA is based on a variational autoencoder, and the training process consists of following two steps: metric learning with contractive and margin losses to achieve the structural similarity constraint and reinforcement learning to produce molecules satisfying both constraints of structural similarity and property improvement. c Overview of experiments. The goals of four benchmark tasks are to enhance each molecular property score, and the goal of use case study for sorafenib resistance is to decrease an affinity score against the ABCG2 protein