NGT: Generative AI with Synthesizability Guarantees Discovers MC2R Inhibitors from a Tera-Scale Virtual Screen.

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery
• Medicinal chemistry and pharmacology

Background:

• Virtual compound libraries are essential for drug discovery, offering vast numbers of synthesizable molecules.
• Exponential growth of these libraries poses significant challenges for traditional compound searching methods.
• Efficient navigation of ultralarge virtual libraries is critical for accelerating drug discovery pipelines.

Purpose of the Study:

• To introduce NeuralGenThesis (NGT), a novel reinforcement learning approach for generating compounds from ultralarge virtual libraries.
• To enable the generation of compounds that satisfy multiple user-defined constraints simultaneously.
• To demonstrate the scalability and efficiency of NGT in identifying potential drug candidates.

Main Methods:

• Training a generative model on an ultralarge virtual library.
• Utilizing a normalizing flow to learn a latent space distribution for constraint satisfaction.
• Applying NGT to identify potent and selective inhibitors for the melanocortin-2 receptor (MC2R).

Main Results:

• NGT successfully generated compounds satisfying multiple constraints without pre-specifying property calculation methods.
• The approach enabled the identification of potent and selective MC2R inhibitors from a three trillion compound library.
• Demonstrated the capability of NGT to navigate and extract valuable molecules from massive chemical spaces.

Conclusions:

• NGT provides a powerful and scalable solution for exploring ultralarge virtual compound libraries.
• The reinforcement learning framework accelerates the in silico discovery and optimization phases of drug development.
• NGT facilitates efficient identification of drug candidates meeting specific therapeutic requirements.