MIT researchers have come up with new ways to use artificial intelligence to design nanoparticles that can provide RNA vaccines and other types of RNA therapies more efficiently.
After training machine learning models to analyze thousands of existing delivery particles, researchers used it to predict new materials that would work even better. This model allowed researchers to identify particles that work well in different types of cells and discover ways to incorporate new types of material into the particles.
“What we did was to accelerate the identification of the optimal mix of ingredients of lipid nanoparticles and apply machine learning tools to help target different cell types or incorporate a variety of materials that are much faster than before.”
The approach could dramatically speed up the process of developing new RNA vaccines and therapies that can be used to treat obesity, diabetes and other metabolic disorders, researchers say.
Former MIT Postdoc Alvin Chan is currently an assistant professor at Nanyang Technological University and currently an assistant professor at the University of Minnesota, Ameya Kirtane is the lead author of the new open access research that is emerging today. Natural Nanotechnology.
Particle Prediction
RNA vaccines, such as the SARS-COV-2 vaccine, are usually packaged in lipid nanoparticles (LNPs) for delivery. These particles prevent mRNA from degradation in the body and help enter the cells after it is injected.
Creating particles that handle these jobs more efficiently can help researchers develop more effective vaccines. Additionally, better delivery vehicles can facilitate the development of mRNA therapies that encode genes for proteins that can be useful in treating a variety of diseases.
In 2024, Traverso’s lab launched a multi-year research program funded by the U.S. Advanced Research Projects Agency (ARPA-H) to develop new ingestible devices that can achieve oral delivery of RNA therapy and vaccines.
“Part of what we’re trying to do is develop ways to produce more protein, for example, for therapeutic applications. Maximizing efficiency is important to increase the amount of cells that can be produced,” says Traverso.
Typical LNPs are made up of four components: cholesterol, helper lipids, ionizable lipids, and lipids attached to polyethylene glycol (PEG). Various variations of each of these components can be exchanged to create a huge number of possible combinations. Modifying these formulations and testing each individually is very time-consuming, so Traverso, Chan, and their colleagues decided to resort to artificial intelligence to speed up the process.
“Most AI models of drug discovery focus on optimizing a single compound at once, but that approach does not work with lipid nanoparticles made of multiple interacting components,” says Chan. “To address this, we developed a new model called Comet, inspired by the same transformer architecture that runs large-scale language models like ChatGpt. Just as these models understand how words form meaning, Comet learns how different chemical components combine in nanoparticles and affect their properties.
To generate training data for machine learning models, researchers created a library of approximately 3,000 different LNP formulations. The team tested each of these 3,000 particles in the lab to see how efficiently the payload could be delivered to cells, and fed all this data to a machine learning model.
After the model was trained, the researchers asked to predict new formulations that would perform better than existing LNPs. They tested these predictions by using a new formulation to supply mRNA encoding fluorescent proteins to mouse skin cells grown in lab dishes. They found that the LNP predicted by the model actually performs better than the particles in the training data, and in some cases better than commercially used LNP formulations.
Accelerated development
Researchers began asking additional questions after showing that the model could accurately predict particles that would efficiently provide mRNA. First, they wondered whether they could train a model on nanoparticles that incorporates a fifth component known as branched polybeta aminoesters (PBAEs).
Research by Traverso and his colleagues showed that these polymers can effectively deliver nucleic acids themselves, so they wanted to investigate whether adding them to LNPs would improve the performance of LNPs. The MIT team created a set of approximately 300 LNPs containing these polymers and used them to train the models. The resulting model can predict additional formulations using PBAE, allowing for better behavior.
Next, the researchers began training models to make predictions about LNPs that function optimally in a variety of cells, including a type of cell called CACO-2 derived from colorectal cancer cells. Again, this model was able to predict LNPs that efficiently deliver mRNA to these cells.
Finally, the researchers used the model to predict which LNPs could withstand freeze drying. This is a lyophilization process that is often used to increase the shelf life of a drug.
“It’s a tool that can help you adapt to a completely different set of questions and accelerate development. We did a large training set in our model, but we did a more focused experiment and got an output that would help you with very different types of questions,” says Traverso.
He and his colleagues are currently working on incorporating some of these particles into potential treatments for diabetes and obesity. This is two of the main targets of the ARPA-H-funded project. Therapeutic agents that can be provided using this approach include GLP-1 mimetics that have similar effects to Ozempic.
This study was funded by the Koch Institute’s Go Nano Marble Center, Karl Van Tassel Career Development Professorship, MIT School of Mechanical Engineering, Brigham and Women’s Hospital, and ARPA-H.
