Electron density-based GPT for optimization and suggestion of host-guest binders

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Host-guest chemistry involves molecular recognition between host molecules and guest molecules.
• Designing effective host-guest binders requires accurate prediction of molecular interactions and structures.
• Current methods for discovering new binders can be time-consuming and resource-intensive.

Purpose of the Study:

• To develop a novel machine learning model for the de novo production of host-guest binders.
• To accurately characterize generated molecules in 2D and 3D using electron density and electrostatic potentials.
• To apply the model to discover new guests for established molecular host systems.

Main Methods:

• A machine learning model was trained on electron density data.
• A variational autoencoder generated 3D representations of electron density and electrostatic potentials.
• Gradient descent was used to optimize guest generation.
• A transformer model converted generated guests into simplified molecular-input line-entry system (SMILES) format with >98% accuracy.

Main Results:

• The model successfully generated host-guest binders characterized by electron density and electrostatic potentials.
• Applied to cucurbit[n]uril (CB[6]), the model discovered 9 previously validated and 7 unreported guests.
• Applied to metal-organic cages ([Pd214]4+), the model discovered 4 unreported guests with varying association constants (Ka).

Conclusions:

• The developed machine learning model is effective for discovering novel host-guest binders.
• The model's ability to generate and characterize molecules accurately accelerates the discovery process.
• This approach holds significant potential for advancing host-guest chemistry and materials design.