De novo generation of multi-target compounds using deep generative chemistry

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Polypharmacology drugs, inhibiting multiple proteins, offer therapeutic advantages but are challenging to design.
• Current drug design methods often focus on single targets, limiting the potential of polypharmacology.

Purpose of the Study:

• To develop a novel computational approach for designing polypharmacology drugs.
• To generate de novo molecular structures with predicted multi-target inhibition, drug-likeness, and synthetic accessibility.

Main Methods:

• Developed POLYGON, a generative reinforcement learning framework to explore chemical space.
• Trained POLYGON using binding data for over 100,000 compounds to predict polypharmacology interactions.
• Generated novel compounds targeting protein pairs with documented co-dependency.

Main Results:

• POLYGON achieved 82.5% accuracy in recognizing polypharmacology interactions from existing binding data.
• Docking analysis confirmed that generated compounds bind to dual targets with favorable free energies and similar orientations to known inhibitors.
• Synthesized 32 compounds targeting MEK1 and mTOR, with most demonstrating >50% reduction in protein activity and cell viability at 1-10 μM concentrations.

Conclusions:

• Generative modeling, exemplified by POLYGON, shows significant potential for accelerating the design of effective polypharmacology drugs.
• The synthesized MEK1/mTOR inhibitors demonstrate the practical utility of computationally designed polypharmacology agents.
• This work paves the way for more efficient development of multi-target therapeutics.