A revolutionary breakthrough in medical technology comes from the prestigious Massachusetts Institute of Technology (MIT), where engineers, using the power of artificial intelligence, have developed a method that could dramatically accelerate the development of a new generation of RNA vaccines and therapies. At the heart of this discovery is an advanced machine learning model that designs nanoparticles with incredible precision for more efficient delivery of sensitive RNA molecules directly into human cells. This approach not only promises a faster path to new vaccines but also opens the door to innovative treatments for some of today's greatest health challenges, including obesity, diabetes, and other metabolic disorders.

A team of researchers from MIT has managed to "teach" an AI model to analyze and understand the interactions of thousands of existing drug delivery particles. After intensive training, the model was able to independently predict entirely new combinations of materials with improved properties. But its capabilities go even further. Artificial intelligence has enabled scientists to identify specific nanoparticle formulations that are optimal for different cell types and to explore how to integrate new, previously unused materials into existing systems. Giovanni Traverso, an associate professor of mechanical engineering at MIT and a gastroenterologist at Brigham and Women's Hospital, who is also the senior author of the study, points out: "We have applied machine learning tools to accelerate the identification of optimal ingredient mixtures in lipid nanoparticles. This allows us to target different cell types or incorporate new materials much faster than was previously possible."

A Revolution in Drug Delivery: How Do Lipid Nanoparticles Work?

Messenger RNA (mRNA) technology became globally known thanks to the vaccines against the SARS-CoV-2 virus. The key to the success of these vaccines lies in lipid nanoparticles (LNPs), microscopic fat bubbles that serve as transport vehicles. Their role is twofold: first, they protect the fragile mRNA molecule from degradation in the body as soon as it is injected, and second, they help it pass through the cell membrane and reach its destination inside the cell, where the desired protein is then produced (in the case of vaccines, the viral protein that stimulates an immune response).

Creating more effective particles is crucial for the development of even more potent vaccines and therapies. Better delivery vehicles could mean lower doses, fewer side effects, and a stronger therapeutic effect. This is particularly important for the development of mRNA therapies that encode genes for proteins that could treat a wide range of diseases. A standard lipid nanoparticle consists of four key components: an ionizable lipid that binds RNA, cholesterol that provides structural stability, a helper lipid that aids in releasing the payload inside the cell, and a lipid linked to polyethylene glycol (PEG) that prevents particle aggregation and prolongs their circulation in the bloodstream. Each of these components can have numerous variants, creating an astronomical number of possible combinations. The traditional approach, which relies on testing each formulation individually, is extremely slow, expensive, and inefficient. It was this challenge that prompted the MIT team to turn to artificial intelligence.

COMET: Artificial Intelligence Inspired by Language Models

To solve this complex problem, the researchers developed a completely new model called COMET (Composition-Oriented Transformer). Alvin Chan, a former postdoctoral fellow at MIT and one of the lead authors of the study published in the journal Nature Nanotechnology, explains the innovativeness of their approach. "Most AI models in drug discovery focus on optimizing a single compound, but that approach doesn't work for lipid nanoparticles, which consist of multiple interconnected components," says Chan. "That's why we developed COMET, inspired by the same transformer architecture that powers large language models like ChatGPT. Just as these models understand how words combine to create meaning, COMET learns how different chemical components come together in a nanoparticle to influence its properties—such as how well it can deliver RNA into cells."

For the model to "learn" this complex chemical language, the researchers first had to create an extensive dataset for training. They generated a library of approximately 3,000 different LNP formulations. The team rigorously tested each of these 3,000 particles in the lab to measure their efficiency in delivering cargo to cells. All the collected data, both successful and unsuccessful attempts, were fed into the COMET model. After the model was trained on this massive dataset, the researchers gave it a key task: predict new formulations that would be superior to all existing ones in the training set. The results were extraordinary.

Confirmation of Efficacy and Pushing the Boundaries

To verify the predictions of the artificial intelligence, the scientists synthesized the new formulations that COMET proposed. They then used these nanoparticles to deliver mRNA encoding a fluorescent protein into mouse skin cells grown under laboratory conditions. By measuring the intensity of the fluorescence, they could precisely quantify how much mRNA was successfully delivered and translated into protein. It turned out that the lipid nanoparticles predicted by the model were indeed significantly more effective not only than the particles from the original training set but in some cases also outperformed commercially available LNP formulations currently used in clinical practice.

This success confirmed that the model not only recognizes patterns in the data but also possesses a kind of "creativity" in proposing new, optimized solutions, thereby dramatically shortening the development cycle that would otherwise take years.

Expanding Possibilities: From New Materials to Targeted Cells

After proving that the model could accurately predict effective particles for mRNA delivery, the researchers decided to test its versatility by asking additional, more complex questions. First, they were interested in whether the model could be adapted to design particles containing a fifth component: a polymer known as branched poly(beta-amino ester) (PBAE). Previous research by Traverso's team had shown that these polymers themselves can efficiently deliver nucleic acids, so the question arose whether their addition to LNPs could further improve performance. The team created a new set of about 300 hybrid LNPs that also contained these polymers and used them for additional model training. The resulting model was able to predict new, even better hybrid formulations.

The next step was to train the model for specific applications—predicting LNPs that would work best in different types of cells. They focused on Caco-2 cells, which are derived from colon cancer cells and are often used as a model for studying drug absorption in the intestines. In this case as well, the model successfully identified formulations optimized for efficient mRNA delivery specifically to these cells, opening the way for targeted therapies. Finally, the scientists used the model to solve a practical problem in the pharmaceutical industry: drug stability. They tasked it with predicting which LNP formulations would best withstand lyophilization—a freeze-drying process often used to extend the shelf life of drugs. The ability to predict such properties could significantly reduce the costs and logistical challenges associated with the distribution and storage of RNA drugs.

The Future of Personalized Medicine and New Therapeutic Targets

Professor Traverso emphasizes that COMET is more than a tool for one-time optimization. "This is a platform that allows us to adapt it to a whole range of different questions and help accelerate development. Although we started with a large training dataset, we can now conduct much more focused experiments and get useful results for very different problems," he explains. The implementation of such AI tools represents a paradigm shift in pharmaceutical research and development.

His team is now actively working on applying some of these newly designed particles in the development of potential treatments for diabetes and obesity. These goals are part of a multi-year research program funded by the U.S. Advanced Research Projects Agency for Health (ARPA-H), one of whose ambitious goals is the development of oral forms of RNA therapies, i.e., drugs that could be taken in tablet form. Therapies that could be delivered with this approach also include GLP-1 mimetics, a class of drugs similar to the popular Ozempic, which would encourage the body to produce therapeutic proteins itself, offering a longer-lasting and potentially safer solution for treating metabolic diseases.