Leveraging artificial intelligence, researchers at MIT have devised an innovative method for crafting nanoparticles that can more effectively transport RNA vaccines and various RNA-based therapies.

After training a machine-learning algorithm to scrutinize thousands of current delivery particles, the researchers utilized it to forecast new materials that would enhance efficacy. The model also allowed the researchers to pinpoint particles suitable for diverse cell types and to discover methods for integrating new types of materials into the nanoparticles.

“Our approach involved employing machine-learning tools to expedite the identification of ideal ingredient combinations in lipid nanoparticles to target distinct cell types or facilitate the incorporation of diverse materials, significantly quicker than previously feasible,” states Giovanni Traverso, an associate professor of mechanical engineering at MIT, a gastroenterologist at Brigham and Women’s Hospital, and the lead author of the study.

This method could substantially accelerate the development of new RNA vaccines, as well as therapies that may address obesity, diabetes, and other metabolic conditions, according to the researchers.

Alvin Chan, a former MIT postdoctoral researcher and currently an assistant professor at Nanyang Technological University, along with Ameya Kirtane, another former MIT postdoc now serving as an assistant professor at the University of Minnesota, are the primary authors of this new study, published today in Nature Nanotechnology.

Particle forecasts

RNA vaccines, including those for SARS-CoV-2, are generally enclosed in lipid nanoparticles (LNPs) for delivery. These nanoparticles safeguard mRNA from degradation within the body and facilitate its entry into cells post-injection.

Developing particles that perform these functions more efficiently could aid researchers in creating even more potent vaccines. Enhanced delivery systems might also streamline the development of mRNA therapies that encode genes for proteins aimed at treating various diseases.

In 2024, Traverso’s laboratory initiated a multi-year research initiative, financed by the U.S. Advanced Research Projects Agency for Health (ARPA-H), to create new ingestible devices capable of facilitating oral delivery of RNA treatments and vaccines.

“One of our objectives is to devise methods for producing more protein for therapeutic purposes. Enhancing efficiency is crucial to maximize cellular production capabilities,” Traverso explains.

A standard LNP comprises four elements — cholesterol, a helper lipid, an ionizable lipid, and a lipid bonded to polyethylene glycol (PEG). Various variants of these components can be interchanged to generate a vast array of potential combinations. Altering these formulations and evaluating each one separately is incredibly time-consuming; thus, Traverso, Chan, and their colleagues opted to harness artificial intelligence to expedite the process.

“The majority of AI models in drug discovery concentrate on optimizing a single compound at a time, but this approach fails for lipid nanoparticles, which consist of multiple interacting components,” Chan notes. “To counter this, we developed a new model named COMET, inspired by the same transformer architecture that underpins large language models like ChatGPT. Just as those models grasp how words combine to convey meaning, COMET learns how different chemical components unite within a nanoparticle to affect its characteristics — such as its ability to deliver RNA into cells.”

To produce training data for their machine-learning model, the researchers assembled a library of approximately 3,000 unique LNP formulations. The team evaluated each of these 3,000 particles in the lab to determine how effectively they could deliver their payload to cells, then input all of this information into the machine-learning model.

Once the model was trained, the researchers requested it to predict new formulations that might outperform existing LNPs. They validated those predictions by utilizing the new formulations to transport mRNA encoding a fluorescent protein to mouse skin cells cultured in a lab dish. They discovered that the LNPs projected by the model indeed outperformed the particles in the training data, and in some instances, even surpassed commercially available LNP formulations.

Accelerated advancement

After demonstrating that the model could reliably predict particles capable of effectively delivering mRNA, the researchers began exploring additional inquiries. They first contemplated whether they could train the model on nanoparticles incorporating a fifth ingredient: a type of polymer known as branched poly beta amino esters (PBAEs).

Research by Traverso and his team has indicated that these polymers can efficiently deliver nucleic acids independently, so they aimed to investigate whether their inclusion in LNPs could enhance performance. The MIT team created a set of around 300 LNPs that also contained these polymers, which they utilized to train the model. Consequently, the resulting model could predict additional formulations with PBAEs that would yield superior results.

Subsequently, the researchers endeavored to train the model to forecast which LNPs would function best in various cell types, including Caco-2 cells, derived from colorectal cancer cells. Once more, the model successfully predicted LNPs that would efficiently transport mRNA to these cells.

Finally, the researchers employed the model to identify which LNPs could best endure lyophilization — a freeze-drying technique commonly used to prolong the shelf life of medications.

“This serves as a tool that enables us to adapt it to an entirely different set of questions and facilitates accelerated development. We executed a comprehensive training set that contributed to the model, but afterwards, you can conduct more targeted experiments and obtain outputs that are beneficial for a variety of questions,” Traverso states.

He and his collaborators are currently focused on incorporating some of these particles into potential treatments for diabetes and obesity, which represent two of the primary objectives of the ARPA-H funded project. Therapeutics that could be delivered via this method include GLP-1 mimetics with effects similar to Ozempic.

This work was supported by the GO Nano Marble Center at the Koch Institute, the Karl van Tassel Career Development Professorship, the MIT Department of Mechanical Engineering, Brigham and Women’s Hospital, and ARPA-H.