Accelerating Siloxane-Based Ionizable Lipid Design for LNPs with Data-Efficient Kolmogorov-Arnold Networks

Area of Science:

• Biomaterials Science
• Computational Chemistry
• Drug Delivery

Background:

• Ionizable lipids are crucial for lipid nanoparticle (LNP) efficacy, particularly in mRNA vaccines.
• Developing new ionizable lipids is challenging due to complex structure-property relationships and limited data.

Purpose of the Study:

• To develop a small-data-driven machine learning (ML) framework for accelerated discovery of novel siloxane-based ionizable lipids.
• To pioneer the use of Kolmogorov-Arnold networks (KANs) for predicting ionizable lipid performance.

Main Methods:

• Employed a small-data-driven framework utilizing Kolmogorov-Arnold networks (KANs), a symbolic regression ML approach.
• Integrated KANs with virtual screening and umbrella sampling simulations for candidate identification.
• Validated findings using molecular dynamics simulations to assess binding affinity to the endosomal membrane.

Main Results:

• The KAN model achieved high predictive accuracy (Qcv² = 0.710) for mRNA delivery efficiency using only 36 training samples.
• The KAN model outperformed conventional ML models and provided explicit mathematical formulas.
• Identified three candidate lipids, with the optimal one showing a 187% increase in binding affinity to the endosomal membrane.

Conclusions:

• The proposed framework offers a data-efficient paradigm for ML-guided ionizable lipid design.
• Successfully bridged symbolic regression with molecular dynamics validation for LNP therapeutics.
• Paved the way for designing next-generation LNP-based therapeutics, including mRNA vaccines.