Human-in-the-loop assisted de novo molecular design

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Cheminformatics

Background:

• De novo molecular design utilizes reinforcement learning to explore chemical space.
• Integrating human feedback into molecular optimization remains a challenge.
• Current methods require manual, iterative tuning of multi-parameter optimization (MPO) scoring functions.

Purpose of the Study:

• To develop a principled human-in-the-loop machine learning approach for adapting MPO scoring functions.
• To enable chemists to directly guide molecular optimization using their implicit knowledge.
• To reduce the trial-and-error involved in defining optimization goals.

Main Methods:

• A probabilistic model captures user input and uncertainty regarding the scoring function.
• Active learning is employed to interact efficiently with the user.
• Two case studies demonstrate learning MPO parameters and incorporating non-parametric domain knowledge.

Main Results:

• The method effectively learns scoring functions from user feedback during molecule browsing.
• Significant improvements were achieved in under 200 feedback queries in simulated cases.
• Demonstrated success in optimizing for a high Quantitative Drug-likeness (QED) score and identifying DRD2 receptor ligands.

Conclusions:

• The proposed human-in-the-loop system significantly enhances the efficiency of de novo molecular design.
• It allows for rapid adaptation of scoring functions to match complex medicinal chemistry goals.
• This approach streamlines the drug discovery process by integrating expert knowledge seamlessly.