Machine-Learning Framework to Predict the Performance of Lipid Nanoparticles for Nucleic Acid Delivery.

Area of Science:

• Nanomedicine
• Computational Chemistry
• Biotechnology

Background:

• Lipid nanoparticles (LNPs) are crucial for gene therapies like mRNA and siRNA delivery.
• Quantitative structure-activity relationship (QSAR) modeling is vital for LNP formulation but faces challenges due to complex compositions and interactions.
• Predicting LNP performance requires sophisticated approaches to handle multicomponent systems and biological variability.

Purpose of the Study:

• To develop a machine learning (ML) framework for predicting LNP activity and cell viability in nucleic acid delivery.
• To overcome the limitations of traditional QSAR modeling for complex LNP formulations.
• To establish a scalable method for predicting LNP performance and guiding rational design.

Main Methods:

• Curated a dataset of 6454 LNP formulations from 21 studies.
• Employed 11 molecular featurization techniques and 6 ML algorithms for classification.
• Utilized scaffold-based 5-fold cross-validation and SHAP for feature attribution.
• Developed a transfer-learning strategy to bridge in vitro-in vivo prediction gaps.

Main Results:

• Achieved >90% accuracy in predicting LNP activity and cell viability using ML models.
• Descriptor-based features with ensemble models (random forest, extra trees) showed highest performance.
• Identified key physicochemical and compositional features driving LNP performance through SHAP analysis.
• Transfer-learning models achieved >82% accuracy for in vitro-to-in vivo prediction.

Conclusions:

• Interpretable ML frameworks can guide rational LNP design for gene therapy.
• The developed ML framework offers a scalable approach to QSAR modeling for nanomedicine.
• Synergistic effects of molecular features are critical for LNP performance.