Lipid nanoparticles (LNPs) have evolved from a specialised drug-delivery technology into a cornerstone of modern medicine. Their success became evident during the COVID-19 pandemic, enabling the delivery of the mRNA used in the Pfizer-BioNTech and Moderna vaccines. However, their applications extend far beyond vaccination, as LNPs can also deliver siRNA, plasmid DNA and gene-editing machinery, creating opportunities to treat genetic diseases, cancer and many other conditions.
Yet a fundamental challenge remains: how do we design an LNP that delivers the right payload, to the right cells, in the right place?
A typical LNP contains several components, including an ionizable lipid, helper phospholipids, cholesterol and polyethylene glycol (PEG)-functionalized lipids. Altering the chemical structure of a single lipid or even slightly changing the ratio between components can affect particle formation, stability, cellular uptake, biodistribution and ultimately therapeutic efficacy. The result is an enormous, multidimensional design space where traditional trial-and-error approaches can become expensive and time-consuming.
This is where data-driven machine learning (ML) and artificial intelligence (AI) approaches are beginning to transform how these drug delivery vectors are developed, fundamentally changing the design process for new therapeutics and vaccines.
From Screening to Prediction
At its core, ML provides a method for understanding the relationship between the inputs and outputs of an LNP. Rather than experimentally testing every possible combination of lipids and formulation parameters, researchers can train models on existing experimental data and use them to predict which unexplored set of formulation parameters and lipid chemistries are most promising.
This approach is particularly attractive for LNPs because their design space is inherently combinatorial. The challenge lies not just in identifying lipids with desirable properties, but in understanding how lipid chemistry, formulation composition and particle properties interact to influence biological activity.
In this context, ML serves as a bridge between chemical structure and biological function. Experimental data provide observations; the model learns patterns within those data; and those patterns can then be used to prioritize new experiments.
Letting AI Search the Chemical Space
One notable demonstration of this approach came in 2024, when Li et al. combined combinatorial chemistry, high-throughput screening and ML to discover new ionizable lipids for mRNA delivery1. The researchers synthesized a library of 584 ionizable lipids and measured the ability of the resulting LNPs to deliver mRNA. These data were then used to train ML models capable of predicting lipid performance.
The trained model was subsequently used to explore a virtual library containing approximately 40,000 candidate lipids. Rather than synthesizing and testing every candidate, the researchers focused on the top candidates predicted by the model. This led to the identification of a new ionizable lipid which outperformed established clinical benchmark lipids in transfecting muscle and immune cells across several tissues.
This study exemplifies one of the most valuable roles ML can play in drug discovery: the model does not replace the laboratory; it helps decide which experiments are worth doing.
An even larger-scale example was reported by Wang et al., who employed AI-driven virtual screening to explore nearly 20 million candidate lipids2. After successive rounds of computational filtering, selected candidates were synthesized and tested in mice. Six lipids identified during the second screening round performed at least as well as the benchmark lipid MC3, with one approaching the performance of the more potent clinical lipid SM-102.
From Better Lipids to Better Targeting
However, potency is only part of the LNP challenge as biodistribution is equally important. Conventional LNPs have a strong tendency to accumulate in the liver, limiting their application for targeting other organs. This limitation has spurred the development of selective organ targeting (SORT) LNPs, which can redirect delivery to specific organs, such as the lung and spleen, simply by modifying LNP composition (Fig. 1)3.
Figure 1. Schematic illustrating how different lipid chemistries can shift the in vivo delivery profile of LNPs from liver to lung and spleen using SORT lipids. Taken from Cheng et al.3
ML provides a natural framework for advancing this concept. Rather than manually testing formulations in the hope of finding one that favours a particular tissue, models can learn relationships between formulation composition and biodistribution, allowing researchers to systematically search for formulations with desired targeting properties.
Recently Witten at al. trained a deep-learning model on more than 9,000 LNP activity measurements and used it to screen approximately 1.6 million candidate lipids in silico4. The model identified previously unexplored lipids that demonstrated efficient mRNA delivery in the lung epithelium in ferrets following nebulization.
Another example comes from a 2025 study which combined design-of-experiments (DoE) approaches, high-throughput screening and ML to investigate LNP formulations with preferential immune cell targeting5. From 180 formulations spanning different lipid chemistries and compositions, the resulting models identified candidates predicted to improve immune cell selectivity while reducing hepatic delivery. Selected formulations were subsequently validated in vivo, where they demonstrated preferential expression in the spleen.
Both approaches illustrate that AI can move beyond predicting LNP performance in vitro to identifying new materials with promising delivery properties in more complex and clinically relevant models.
The Future: Closed-Loop LNP Discovery
Perhaps the most exciting possibility is therefore not simply using AI to make predictions but creating a closed loop discovery system.
In such a closed loop, a library of lipids is synthesized and screened with the help of liquid handling systems, with the resulting data feeding into an ML model that identifies chemical features, synthesis and formulation parameters associated with successful delivery. The model then proposes new candidates for synthesis and testing, creating a continuous cycle of learning and discovery (Fig. 2).
Design → Predict → Synthesize → Test → Learn → Design again.
The advantage is that each experimental cycle can focus on the most informative or promising regions of chemical space rather than relying on random or exhaustive screening. In other words, the objective is not simply to perform more experiments but to perform better experiments.
Figure 2. Diagram of the closed-loop optimization of lipid synthesis. Taken from Hanna et al.6
This approach also highlights an important distinction between AI-assisted discovery and the idea of an entirely automated laboratory. AI does not eliminate the need for experimental biology. LNP delivery is governed by complex phenomena, including self-assembly, protein corona formation, cellular uptake, biodistribution and immune interactions. Current datasets are also relatively small compared with those available for many conventional drug-discovery problems and models can struggle to generalize beyond the chemical and biological space on which they were trained. These limitations make experimental validation essential. They also make the design of the learning process itself critical: what should be measured, which formulations should be tested next, and how can models learn efficiently from relatively limited data?
The COVID-19 pandemic demonstrated the capabilities of LNPs. The next challenge is to enhance their precision, potency and safety while tailoring formulations to individual diseases and biological targets. As AI and ML continue to evolve, they will play a pivotal role in advancing LNP technology and drug delivery systems.
References
1. Li, B. et al. Accelerating ionizable lipid discovery for mRNA delivery using machine learning and combinatorial chemistry. Nat. Mater. 23, 1002–1008 (2024).
2. Wang, W. et al. Artificial intelligence-driven rational design of ionizable lipids for mRNA delivery. Nature Communications 15, (2024).
3. Cheng, Q. et al. Selective organ targeting (SORT) nanoparticles for tissue-specific mRNA delivery and CRISPR–Cas gene editing. Nat. Nanotechnol. 15, 313–320 (2020).
4. Witten, J. et al. Artificial intelligence-guided design of lipid nanoparticles for pulmonary gene therapy. Nat. Biotechnol. 43, 1790–1799 (2025).
5. Hanafy, B. I. et al. Advancing Cellular-Specific Delivery: Machine Learning Insights into Lipid Nanoparticles Design and Cellular Tropism. Adv. Healthc. Mater. 14, (2025).
6. Hanna, A. R., Issadore, D. A. & Mitchell, M. J. High-throughput platforms for machine learning-guided lipid nanoparticle design. Nat. Rev. Mater. 11, 50–64 (2026).
