Exploring Low-Toxicity Chemical Space with Deep Learning for Molecular Generation.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Drug discovery

Background:

• Drug discovery faces challenges in developing compounds with optimal pharmacology and low toxicity.
• Existing methods often struggle to balance desired properties with safety profiles.

Purpose of the Study:

• To develop a conditional generative model for creating novel drug-like molecules with minimized toxicity.
• To enhance structural diversity and pharmacological properties in generated compounds.

Main Methods:

• Combined a semisupervised variational autoencoder (SSVAE) with an MGA toxicity predictor.
• Implemented hierarchical constraints on the toxicity space for multiobjective optimization.

Main Results:

• The model demonstrated effectiveness, novelty, and diversity in generated molecules.
• Generated molecules predominantly fall within low-toxicity regions, indicating efficient constraint of toxic structures.
• The approach successfully generated diverse, low-toxicity molecules, unlike post-generation filtering.

Conclusions:

• The developed generative model efficiently produces drug-like molecules with low toxicity and high diversity.
• This strategy can significantly improve the quality of generated molecules in target-based drug discovery.