In this paper, we propose the neural network Adversarial Threshold Neural Computer (ATNC). ATNC is intended for the de novo design of novel small-molecule organic structures. ATNC is based on generative adversarial network (GAN) architecture and reinforcement learning (RL). ATNC uses a Differentiable Neural Computer (DNC) as a generator. A new specific block, called adversarial threshold (AT), is used as a filter. AT acts as a filter between the agent (generator) and the environment (discriminator plus objective reward functions). Furthermore, a new objective reward function named Internal Diversity Clustering (IDC) is introduced. In this work, ATNC is tested and compared to the ORGANIC model. Both models were trained on the SMILES string representation of the molecules, using four different objective reward functions. Results showed that ATNC outperformed the ORGANIC model. In fact, ATNC generated a larger proportion of valid and unique SMILES strings, with a better average length, than ORGANIC. Analysis of key molecular descriptors and chemical statistical features demonstrated that the molecules generated by ATNC elicited better drug-likeness properties than ORGANIC. We also performed an in vitro validation of the molecules generated by ATNC; results indicated that ATNC is an effective method for producing hit compounds.

Adversarial threshold neural computer for molecular <i>de novo</i> design

256th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.

Adversarial threshold neural computer for molecular de novo design