Retro Drug Design: From Target Properties to Molecular Structures.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Medicinal chemistry

Background:

• Accelerating drug discovery is a key goal in pharmaceutical research.
• Artificial intelligence (AI) and computational methods offer new possibilities for drug development.
• Existing methods face challenges in efficiently generating novel drug candidates with specific properties.

Purpose of the Study:

• To introduce a novel strategy, retro drug design (RDD), for *de novo* small-molecule drug creation.
• To design drug candidates that simultaneously satisfy biological activity and multiple physicochemical/ADMET property requirements.
• To demonstrate the feasibility and effectiveness of RDD in generating novel kinase inhibitors.

Main Methods:

• Molecular structures represented using the optATP descriptor system and transformed via principal component analysis.
• Predictive models trained on experimental data using optATP and shallow machine learning.
• Monte Carlo sampling to identify target properties in the loading vector space.
• Deep learning model to decode molecular structures from identified solutions.

Main Results:

• RDD successfully generated novel kinase inhibitors with high novelty (Tanimoto similarity < 0.50).
• Out of 3,040 compounds meeting all criteria, 20 were synthesized and tested.
• Fifteen compounds showed inhibitory activity, with eight designated as strong hits, five possessing excellent ADMET properties.

Conclusions:

• Retro drug design (RDD) is a powerful AI-driven strategy for *de novo* drug discovery.
• RDD can efficiently generate novel small-molecule drug candidates with optimized biological activity and ADMET profiles.
• This approach holds significant potential to enhance and accelerate the current drug discovery pipeline.