Navigating chemical space: multi-level Bayesian optimization with hierarchical coarse-graining.

Area of Science:

• Computational chemistry
• Molecular modeling
• Drug discovery

Background:

• Vast chemical space poses challenges for molecular discovery.
• Conventional screening methods lack scalability.
• Efficient exploration of chemical space is crucial.

Purpose of the Study:

• Develop an active learning method for effective chemical space exploration.
• Utilize transferable coarse-grained models to compress chemical space.
• Optimize molecules for enhanced phase separation in phospholipid bilayers.

Main Methods:

• Employing transferable coarse-grained models at multiple resolutions.
• Transforming discrete molecular spaces into smooth latent representations.
• Performing Bayesian optimization within latent spaces using molecular dynamics simulations to calculate free energies.

Main Results:

• Successfully balanced combinatorial complexity and chemical detail using multi-level representations.
• Demonstrated effective exploration and exploitation via funnel-like strategy.
• Identified optimal compounds and provided insights into relevant chemical space neighborhoods.

Conclusions:

• The developed multi-level active learning approach efficiently navigates large chemical spaces.
• Lower-resolution neighborhood information effectively guides higher-resolution optimization.
• This method facilitates free energy-based molecular optimization for targeted applications.