Adversarial Threshold Neural Computer for Molecular de Novo Design.

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Drug Discovery
• Machine Learning for Molecular Design

Background:

• De novo design of novel small-molecule organic structures is crucial for drug discovery.
• Existing generative models face challenges in generating valid, diverse, and druglike molecules.
• Generative Adversarial Networks (GANs) and reinforcement learning offer promising avenues for molecular design.

Purpose of the Study:

• To introduce and evaluate the Adversarial Threshold Neural Computer (ATNC) model for de novo small-molecule design.
• To enhance molecular diversity using a novel objective reward function, Internal Diversity Clustering (IDC).
• To compare the performance of ATNC against the ORGANIC model in generating druglike molecules.

Main Methods:

• Developed the ATNC model, integrating a Differentiable Neural Computer generator with a novel adversarial threshold (AT) block.
• Employed generative adversarial network architecture and reinforcement learning.
• Trained models on SMILES representations of 15K druglike molecules, using objective functions including internal similarity, druglikeness filters, and IDC.

Main Results:

• ATNC significantly outperformed the ORGANIC model in generating valid (72% vs. 7%) and unique (77% vs. 86% - note: ORGANIC had higher unique rate but lower valid rate) SMILES strings when combined with IDC.
• Analysis of molecular descriptors and chemical features indicated superior druglikeness properties for ATNC-generated molecules.
• In vitro validation confirmed ATNC's effectiveness in producing hit compounds.

Conclusions:

• The ATNC model represents an advancement in deep learning for de novo molecular design.
• The adversarial threshold block and IDC reward function enhance the generation of valid, diverse, and druglike molecules.
• ATNC shows strong potential for accelerating the discovery of novel hit compounds.