An artificial intelligence model predicts how brain immune cells react to RNA and DNA nanoparticles, helping scientists design safer and more effective nucleic acid therapies faster.
(Nanowerk Spotlight) Nucleic acid therapeutics have moved from simple gene-silencing strands to programmable RNA and DNA assemblies known as nucleic acid nanoparticles (NANPs). These molecular constructs can deliver drugs, regulate genes, or modulate immune activity with precision built directly into their sequence.
We explored this shift in a previous Nanowerk Spotlight article (“Smart RNA nanodevices reprogram themselves to silence genes in cancer cells”). That article described reconfigurable RNA nanoparticles that remain inactive in healthy tissue but activate inside tumor cells, releasing gene-silencing molecules when they detect cancer-specific RNA signals. These programmable designs show how nucleic acid nanotechnology is evolving toward adaptive systems that respond selectively to their biological environment.
As these systems grow more complex, predicting how the immune system will interpret them becomes more difficult. The same cellular receptors that detect viral RNA can mistake therapeutic nanoparticles for pathogens, causing inflammation or neutralizing their effect.
Laboratory screening in human immune cells can reveal these reactions, but it is slow and resource intensive. A faster approach would predict immune outcomes directly from the nucleotide sequences that define each nanoparticle.
A new study published in Small (“From Sequence to Response: AI-Guided Prediction of Nucleic Acid Nanoparticles Immune Recognitions”), applies this idea to the brain’s immune environment. It introduces a transformer based model, a form of artificial intelligence that detects patterns in long sequences, to predict how human microglia, the immune sentinels of the central nervous system, respond to custom nucleic acid nanoparticles using sequence data alone.
“Our comprehensive study details the characterization and immune profiling of 176 distinct NANPs that we individually prepared and analyzed,” Prof. Kirill A. Afonin at the University of North Carolina at Charlotte, tells Nanowerk. “To the best of my knowledge, this represents the largest library of NANPs ever screened at once, and the first study of this scale involving human microglia.”
Schematic of the experimental flow. The upper panel illustrates the connectivity principles of selected NANP architectures. The lower panels depict the experimental workflow, including the production of a diverse NANP library, and side-by-side physicochemical characterization of all NANPs with further assessment of their immunorecognition using in vitro biological assays, application ofmachine learning to model immune responses based on structural and compositional features, and validation of the developed model. (Image: Reprinted from DOI:10.1002/smll.202509459, CC BY) (click on image to enlarge)
The researchers built a systematic library of 176 nanoparticles, each composed of short nucleic acid strands that self-assembled into defined shapes. The library included flat polygons such as triangles with four strands, squares with five, and pentagons with six, as well as compact three-dimensional cubes with six strands. Each was made in DNA only, RNA only, and mixed DNA RNA forms.
Physical characterization confirmed correct assembly and consistent quality. Measurements showed average diameters near 15 nanometers with limited variation and clear trends in stability. RNA rich constructs had higher melting temperatures, meaning they held their structure better at heat, while DNA only particles were less stable. In serum degradation tests, mixed DNA RNA particles lasted longest, DNA only forms showed moderate stability, and RNA only versions degraded fastest.
To test immune activity, the team used cultured human microglial cells. These were exposed to each nanoparticle at equal concentration, and the secretion of two cytokines was measured. Interferon beta signals antiviral defense, and interleukin 6 promotes inflammation.
Results showed that three dimensional cubes produced the strongest cytokine release, reaching about 1,000 picograms per milliliter for interferon beta and up to 8,000 for interleukin 6. Flat polygons produced lower values. RNA heavy particles induced stronger responses, while DNA only constructs triggered little or none. These findings align with how innate sensors detect RNA more readily than DNA.
The modeling approach followed the logic of structure activity relationships but used sequence alone. Each strand was split into overlapping triplets of nucleotides, or 3-mers, to capture short sequence patterns. Because assemblies contain multiple strands that can be ordered in different ways, all possible strand arrangements were included in training so the model would focus on content rather than order.
A transformer network processed the tokens through layers of self-attention that detect relationships between distant parts of the sequence and produced predicted cytokine levels. The dataset was divided into training and test subsets, and the procedure repeated multiple times for reliability.
The model’s predictive accuracy was high. Across validation runs, the coefficient of determination reached about 0.96 for interferon beta and 0.97 for interleukin 6. When the model was tested on a separate group of nanoparticles that were not used during training, it still achieved accuracy scores of 0.91 and 0.85, with low error.
Comparisons with Random Forest and recurrent neural network baselines showed clear improvement. These results confirm that enough information about immune triggering is encoded in the sequence itself, even without structural input.
Afonin explains that the goal was not to replace experiments but to provide a reliable computational filter for early-stage screening. “We wanted to create a system that learns directly from sequence and predicts immune activity before the first wet-lab test,” he said. “This allows scientists to design nanoparticles with specific biological behaviors in mind rather than discovering them by trial and error.”
Further analysis linked immune patterns to physical properties. Among two-dimensional shapes, cytokine output increased with RNA content and thermal stability. For cubes, interleukin 6 responses correlated with particle size, while interferon beta correlated with both size and melting temperature.
This suggests that the immune system senses a combination of sequence composition and structural robustness. By capturing these relationships computationally, designers can anticipate how changes in sequence might alter immune signaling before synthesis or testing.
The group also updated the online AI Cell platform to make the tool accessible to the research community. Users can input nucleotide sequences and receive predicted cytokine levels for microglia within seconds. The interface includes reference particles that represent low, medium, and high immune activation to help calibrate results.
Afonin emphasized the reason for making the system public: “We want other researchers to use this platform as a design aid, to test new ideas safely and efficiently. Sharing the data and models lets the community build on them rather than starting from scratch.”
The methods section provides full experimental and computational detail. It describes nanoparticle synthesis, purification, and stability assays, and the enzyme linked immunosorbent assays used for cytokine measurement. They list all nucleotide sequences and share code and model files to ensure reproducibility.
Afonin and his collaborators acknowledge that the work’s scope is limited. Predictions apply to one cell type, two cytokine endpoints, and a defined set of particle architectures. Future studies could include other immune markers and nanoparticle classes. Yet the current work shows that deep learning can extract immune relevant information directly from sequence and reach accuracy sufficient for guiding design.
By combining an extensive experimental dataset with a sequence learning model, this work closes a long-standing gap between nanoparticle design and immune evaluation. It translates the structural and compositional diversity of nucleic acid nanoparticles into a predictive framework that supports rational design.
For developers of brain targeted RNA or DNA therapeutics, this capability could shorten iteration cycles and reduce the risk of unintended immune activation. The transformer model does not replace experiments but acts as an early filter, linking design and biology through data driven prediction.
For authors and communications departmentsclick to open
Lay summary
Prefilled posts
ORCID information
Kirill A. Afonin (University of North Carolina at Charlotte)
, 0000-0002-6917-3183 first author
Nanowerk Newsletter
Get our Nanotechnology Spotlight updates to your inbox!
Thank you!
You have successfully joined our subscriber list.
Become a Spotlight guest author! Join our large and growing group of guest contributors. Have you just published a scientific paper or have other exciting developments to share with the nanotechnology community? Here is how to publish on nanowerk.com.
