MolQuery: Prediction of Lipid Synthesizability Using Active Learning.

Area of Science:

• Molecular Design
• Drug Delivery Systems
• Artificial Intelligence in Chemistry

Background:

• Generative Artificial Intelligence (GenAI) generates vast molecular data, but synthesizability remains a challenge.
• Effective GenAI models for molecular systems require robust synthesizability assessment platforms.
• Lipid molecules are crucial for mRNA delivery via lipid nanoparticles (LNPs).

Purpose of the Study:

• To introduce MolQuery, a pipeline for accurate chemical synthesizability prediction.
• To integrate active learning (AL) for efficient model training with limited data.
• To improve the development of GenAI-based molecular design tools.

Main Methods:

• Developed MolQuery, a comprehensive pipeline integrating active learning (AL).
• Applied AL to enhance the accuracy of synthesizability predictions for lipid molecules.
• Trained machine learning models efficiently using small datasets.

Main Results:

• MolQuery demonstrated highly accurate predictions of lipid synthesizability.
• The pipeline effectively filters synthetic LNP datasets.
• Active learning significantly improved model performance compared to existing solutions.

Conclusions:

• MolQuery is a valuable tool for assessing and improving the synthesizability of designed molecules.
• The integration of AL enhances the efficiency and accuracy of molecular design pipelines.
• This work facilitates the practical application of GenAI in developing lipid-based drug delivery systems.