Efficient multi-objective molecular optimization in a continuous latent space.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Drug discovery

Background:

• Small molecule drug discovery faces challenges in identifying novel compounds with optimal properties.
• Existing methods may not efficiently explore chemical space for desired molecular characteristics.

Purpose of the Study:

• To develop a novel computational method for generating new drug-like molecules with improved predicted properties.
• To accelerate the lead optimization process in drug discovery.

Main Methods:

• Combining in silico prediction models (biological activity, pharmacokinetics) with Particle Swarm Optimization (PSO).
• Navigating a machine-learned continuous representation of chemical space guided by a multi-component objective function.
• The objective function integrates prediction models, desirability ranges, and substructure constraints.

Main Results:

• The proposed method consistently identified more desirable molecules for the studied tasks.
• The optimization process was achieved in a relatively short timeframe.
• Demonstrated successful navigation of a learned chemical space.

Conclusions:

• The novel in silico approach effectively aids in the discovery of improved small molecules.
• This method has the potential to significantly accelerate and enhance the lead optimization phase for medicinal chemists.