Retro Drug Design: From Target Properties to Molecular Structures

Area of Science:

• Computational chemistry and drug discovery
• Artificial intelligence in medicine
• Pharmacology and medicinal chemistry

Background:

• Generating drug molecules with specific properties computationally is a major goal in pharmaceutical research.
• Existing methods often lack the ability to design novel molecules from scratch that meet multiple predefined criteria.

Approach:

• Developed Retro Drug Design (RDD), an AI strategy using predictive models and Seq2Seq deep learning to generate novel small molecules.
• Trained predictive models on experimental data using an atom typing based molecular descriptor system (ATP).
• Employed Monte Carlo sampling and Seq2Seq deep learning to decode molecular structures from target property spaces.

Key Points:

• RDD generated 180,000 chemical structures, with 78% being chemically valid and 31% meeting defined MOR activity and BBB permeability criteria.
• Only 267 of the 42,000 valid structures were commercially available, highlighting the novelty of AI-generated compounds.
• Assayed 96 compounds, with 25 identified as MOR agonists possessing excellent BBB penetration.

Conclusions:

• RDD demonstrates potential to revolutionize drug discovery by creating novel structures with desired biological functions and ADMET properties.
• AI-enabled drug discovery offers a fast track essential for addressing public health threats like pandemics.