RNA therapies, including vaccines, hold huge promise for treating a wide range of diseases by teaching cells to produce therapeutic proteins. But RNA molecules are fragile and difficult to deliver safely and efficiently to the right cells in the body. To protect and transport them, researchers use LNPs—tiny carriers made from a mixture of lipids.
Designing these nanoparticles, however, is no simple task. Even small changes to their ingredients can have a major impact on delivery performance. Traditionally, scientists have relied on a slow, manual process of testing thousands of combinations in the lab—a time-consuming bottleneck that has limited the pace of RNA therapeutic development.
While AI has been used in drug discovery, most tools focus on optimizing single molecules. That approach doesn’t work well for multi-component systems like LNPs, where ingredient interactions are complex and unpredictable.
The Solution: COMET, a Machine-Learning Model for Nanoparticles
To tackle this, the team at MIT created COMET, a machine-learning model specifically designed to handle the complexity of multi-component nanoparticles.
Drawing inspiration from the transformer architecture behind language models like ChatGPT, COMET is trained to understand how different ingredients, such as cholesterol, helper lipids, ionizable lipids, and PEG lipids, work together to affect an LNP’s ability to deliver RNA into cells.
Training the model required building a massive dataset. The team synthesized roughly 3000 unique LNPs, each loaded with mRNA encoding a fluorescent protein, and tested how well each formulation delivered its payload. That data was used as part of COMET’s learning process, allowing it to identify patterns between composition and performance.
The researchers then challenged the model to generate new LNP formulations predicted to outperform the original set. When tested, these AI-designed nanoparticles showed markedly better RNA delivery—outpacing both the baseline designs and even some formulations currently used in commercial applications. The results confirmed COMET’s accuracy and practical value.
Broader Applications and Future Impact
What sets COMET apart is its flexibility. The researchers showed that it could be adapted for a wide range of goals—not just boosting performance, but also tailoring formulations for specific targets or constraints.
For example:
- New materials: The team introduced a fifth component—branched poly(beta-amino esters), or PBAEs—into the LNP design. COMET quickly identified hybrid nanoparticles combining the strengths of both traditional lipids and PBAEs, resulting in more effective carriers.
- Cell-specific delivery: By retraining the model on data from colorectal cancer-derived Caco-2 cells, the team enabled COMET to predict LNPs best suited for that cell type—a key step toward developing ingestible RNA therapies.
- Shelf-life optimization: COMET also helped identify formulations that withstand lyophilization, a freeze-drying process essential for distributing RNA medicines globally, especially to regions lacking cold-chain infrastructure.
These applications underscore COMET’s potential as a platform technology for RNA delivery. It’s already being used in an ARPA-H-funded project focused on oral RNA treatments for metabolic diseases like diabetes and obesity, including therapies that mimic GLP-1-based drugs such as Ozempic.
Conclusion
By combining deep experimental data with machine learning, MIT’s COMET model offers a smarter, faster way to design lipid nanoparticles for RNA delivery. It eliminates the need for exhaustive lab screening, while allowing for rapid customization—whether that means optimizing for new materials, targeting specific cells, or enhancing stability.
Journal Reference
Chan, A., Kirtane, A.R., Qu, Q.R. et al. Designing lipid nanoparticles using a transformer-based neural network. Nat. Nanotechnol. (2025). DOI: 10.1038/s41565-025-01975-4
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.