AI has designed candidate drugs for antibiotic-resistant infections and genetic diseases . But efforts to incorporate AI into the design of lipid nanoparticles (LNPs), the revolutionary delivery vehicles behind mRNA therapies like the COVID-19 vaccines, have been much more limited.
Designing LNPs is especially challenging: Each formulation combines multiple lipid components whose ratios influence how the particle delivers genetic instructions inside cells. Scientists still lack a clear map connecting those chemical inputs to biological outcomes.
The reason? There simply isn’t enough data.
Now, engineers at the University of Pennsylvania have built LIBRIS, an automated microfluidic platform capable of generating LNP formulations at the speed and scale required to train predictive AI models. “This could accelerate lipid nanoparticle development by as much as 100-fold,” says Michael J. Mitchell , Associate Professor in Bioengineering (BE) and co-senior author of a study in ACS Nano .
The design space for LNPs is enormous, on the order of 10¹⁵ possible formulations, any one of which might be best suited for delivering a particular therapy. “If we want to explore that space with the help of AI,” adds Mitchell, “we need much more data than currently exists.”
Eventually, LIBRIS — short for “LIpid nanoparticle Batch production via Robotically Integrated Screening” — could even support the ‘rational design’ of LNPs, allowing researchers to specify a particle’s properties in advance, rather than generating variations and then determining their capabilities.
“This new microchip-based robotic approach is a major step in that direction,” says David Issadore , Professor in BE and co-senior author of the study. “AI excels at pattern recognition, but to find patterns that relate chemical structure to biological effect, we need enough data for those patterns to emerge.”
The Data Bottleneck
Generating new LNP formulations involves three basic steps: creating new ionizable lipids, whose chemistry largely determines the particle’s unique properties; preparing the formulation by combining those lipids with other ingredients; and then testing the resulting particles.
The first and last steps can already support the creation of massive datasets. “We can easily generate thousands of new ionizable lipids and simultaneously test thousands of LNP formulations,” says Andrew Hanna , a doctoral student in BE and the study’s first author. “But we can only formulate tens to hundreds of particle designs per hour.”
At present, there are two primary methods to formulate LNPs: mixing the ingredients by hand or combining them in a microfluidic device, essentially a tiny plastic chip with narrow channels that push the components together under pressure. “It’s a slow, time-consuming process,” says Hanna. “You can’t really formulate multiple LNP designs at the same time. After each run, you have to clean the equipment and start over.”
Even automation offers only limited relief. Robotic liquid handlers can prepare large libraries of lipid ingredients, but typically rely on inconsistent mixing methods that introduce variability from batch to batch. More controlled microfluidic systems produce consistent particles, but still operate largely in serial fashion and yield only small volumes at a time.
“Actually formulating the nanoparticles is the bottleneck,” Hanna says. “Until we can scale the process, we can’t generate the large, systematic datasets that machine learning models need.”
Automating LNP Formulation
The new machine resembles a tiny factory: tubes carrying different LNP components feed into a glass microfluidic chip encased in aluminum housing. Inside the chip, the components mix together in microscopic channels under precisely controlled pressure. A plastic well plate scurries about underneath the chip to collect the resulting streams of particles in solution.
Unlike conventional systems, the chip contains parallel channels that allow it to create up to eight distinct formulations simultaneously. Because those channels can be rapidly cleaned, the platform can operate more or less continuously, producing on the order of 1,000 LNP formulations per hour, roughly 100 times faster than manual microfluidic methods.
“If we can generate large, well-defined libraries of LNPs,” says Issadore, “then we can start producing the data sets needed to identify patterns that can unlock the full potential of LNP-based therapeutics.”
Toward Rational LNP Design
Until now, efforts to create new LNP formulations have largely relied on trial and error. Researchers generate libraries of related particles, test them in cells or animals and then analyze which variations perform best.
While this approach has yielded important breakthroughs — including the LNPs used in FDA-approved mRNA vaccines — it does not allow scientists to predict in advance how a new formulation will behave.
By dramatically accelerating LNP formulation output while maintaining precise control over particle composition, LIBRIS could help close that gap. The next step, the team says, is to begin mapping how specific chemical inputs influence biological outcomes.
“Our vision is to move from screening to design,” says Mitchell. “Instead of asking, ‘Which of these works best?’ we want to ask, ‘What properties do we want, and how do we build a nanoparticle to achieve them?’ This platform gives us the foundation to start answering that question.”
This study was conducted at the University of Pennsylvania School of Engineering and Applied Science and supported by a Burroughs Wellcome Fund Career Award at the Scientific Interface (CASI), a U.S. National Science Foundation (NSF) CAREER Award (CBET-2145491), an American Cancer Society Research Scholar Grant (RSG-22-122-01 ET), the Cystic Fibrosis Foundation (MITCHE24I0), the Wellcome Leap R3 program, NSF MRSEC Grant (DMR-2309034), NSF Biofoundry, the Center for Precision Engineering for Health at the University of Pennsylvania and the National Institute of Dental and Craniofacial Research of the U.S. National Institutes of Health (T90DE030854).
Additional co-authors include Sarah J. Shepherd, Gregory A. Datto, Rakan El-Mayta, Isabel B. Navarro, Adele S. Ricciardi, Marshall S. Padilla, Shuran Zhang, Hannah M. Yamagata and Nova Y. Meng of Penn Engineering; Neha Srikumar of Penn Engineering and Penn Medicine; Joshua R. Buser of Consonant Systems, LLC; and Taigh Anderson and Mohamad-Gabriel Alameh of the Children’s Hospital of Philadelphia.
ACS Nano
Experimental study
Cells
Automated and Parallelized Microfluidic Generation of Large and Precisely Defined Lipid Nanoparticle Libraries
26-Dec-2025
The authors declare the following competing financial interest(s): Andrew Hanna, Sarah Shepherd., Michael J. Mitchell., and David Issadore are inventors on a patent filed by the Trustees of the University of Pennsylvania (US Provisional Patent Application No. 63/717,568, filed November 7, 2024) describing the parallelized microfluidic generation of many nanoparticle formulations described herein.