Using AI to improve mRNA vaccine delivery by designing nanoparticles and predicting gene responses

To model transcriptional responses to mRNA vaccination, we generated a synthetic RNA-seq dataset based on biologically informed assumptions and guided by immunological response profiles simulated using the UISS platform in our previous work. The dataset comprised 300 genes measured across 10 samples (5 Control and 5 Post-Vaccination). A subset of genes was designed to simulate vaccine-induced immune activation: 30 genes were upregulated and 30 downregulated in the post-vaccination group relative to controls.

Additionally, well-established immune marker genes were included to represent specific compartments, B cells (CD19, MS4A1) (14), T cells (CD3D, CD8A, CD4) (15), plasma cells (IGHG1, IGHM, PRDM1) (16), and others, artificially upregulated to reflect canonical immune activation following antigen exposure.

Gene expression values were sampled from normal distributions, with mean shifts used to simulate differential regulation. To preserve biological plausibility, negative values, resulting from the statistical properties of normal distributions, were truncated to zero. This step ensures that all simulated expression values remain non-negative, reflecting the reality that gene expression levels, being measures of transcript abundance, cannot be less than zero. This approach enables the simulation of genes with no detectable expression while avoiding artifacts that could compromise downstream analysis.

This synthetic dataset serves a dual purpose. On one hand, it allows controlled benchmarking of the transcriptomic analysis pipeline, particularly in assessing its ability to recover known patterns of immune activation. On the other hand, it acts as a bridge to validate predictions generated by the Universal Immune System Simulator (UISS), a mechanistic, agent-based platform capable of modeling immune responses at multiple scales, from molecular signaling to cellular interactions and tissue-level dynamics (17, 18).

Specifically, UISS has been used to simulate host responses to mRNA vaccines, including the biodistribution of lipid nanoparticles (LNPs), antigen presentation, and subsequent activation of adaptive immunity (7). Based on its simulations, UISS produces immunological outputs, such as the expansion of specific immune cell subsets or the secretion of key cytokines, that can be mapped to gene expression patterns. While UISS does not generate RNA-seq data directly, these outcomes can be qualitatively and semi-quantitatively translated into gene expression profiles, enabling the construction of biologically plausible synthetic datasets.

By constructing a synthetic RNA-seq dataset that reflects these expected transcriptional signatures, we can assess whether downstream analysis methods (e.g., differential expression, immune risk indexing) can faithfully recapitulate the immune activation patterns originally predicted by UISS. This integration provides a robust framework for evaluating the predictive alignment between mechanistic modeling and transcriptomic data analytics in the context of rational vaccine design.

Differential gene expression analysis was performed using the DESeq2 package in R, employing negative binomial distribution modeling and Wald tests to identify significantly differentially expressed genes between the post-vaccination and control groups within the synthetic RNA-seq dataset (8). Gene-wise fold changes and adjusted p-values (Benjamini-Hochberg correction) were computed to isolate significantly upregulated immune-related genes (FDR < 0.05).

To infer the immunological profiles of each condition, marker genes characteristics of major immune compartments were selected based on established immunological literature. Specifically, we considered CD19 and MS4A1 for B cells (19), CD3D, CD8A, and CD4 for T cells (20) and IGHG1, IGHM, and PRDM1 (BLIMP-1) for plasma cells (21).

Based on prior immune simulation results, we introduced a compartment-specific risk index designed to quantitatively evaluate the potential for unintended immune activation (off-target effects). The immune risk index for each compartment was calculated by multiplying the simulated Delta_AUC (area under the curve representing cumulative immune activation over time, as established in previous immunological modeling studies (17)) by the count of significantly upregulated marker genes identified in the differential expression analysis for that immune compartment. This integrated approach combines functional simulation data with empirical transcriptomic profiles, providing a robust, interpretable, and semi-quantitative metric for assessing immune activation risks associated with vaccination or other therapeutic interventions.

To support the development and evaluation of an AI-driven optimization pipeline for lipid nanoparticle (LNP) formulations, we generated a synthetic dataset consisting of diverse LNPs characterized by defined physicochemical parameters and corresponding immune activation scores (Delta_AUC). Each LNP formulation was parameterized based on four key physicochemical attributes known to significantly impact biodistribution, cellular uptake, and immunogenicity: particle size (ranging from 50 to 150 nm), which influences circulation time and tissue penetration (22); surface charge (−10 to +10 mV), affecting cellular interaction (23) and colloidal stability (24); PEGylation percentage (0.1 to 0.5 mol%), referring to the covalent attachment of polyethylene glycol (PEG) chains to the nanoparticle surface, a modification that confers a steric barrier against opsonization, reduces recognition and clearance by the mononuclear phagocyte system, prolongs systemic circulation time, and imparts a “stealth” property that enhances in vivo stability (24); and the presence or absence of active targeting ligands (binary encoded, where 0 represents untargeted and 1 represents targeted nanoparticles), enabling selective binding to specific cellular receptors (24, 25). A total of 100 distinct LNP formulations were systematically sampled across this four-dimensional parameter space, ensuring uniform representation and adequate coverage for robust AI model training. A summary of the main effects of these physicochemical parameters on biodistribution, cellular uptake, and immunogenicity, are summarized in Table 1:

Each formulation was assigned a Delta_AUC value, calculated using a biologically informed, non-linear scoring function explicitly designed to simulate realistic biodistribution and immunological response patterns observed experimentally:

In this formulation, ϵ represents Gaussian-distributed noise (mean = 0, standard deviation = 0.1), reflecting biological variability and measurement uncertainty typically encountered in experimental settings (26).

The scoring function for Delta_AUC was constructed to reflect biologically plausible relationships between key nanoparticle properties and delivery performance, based on known experimental trends. Specifically, the quadratic penalty terms for size and surface charge were introduced to model the existence of optimal values: nanoparticle diameters around 90 nm and near-neutral charges are experimentally associated with enhanced circulation times and improved biodistribution profiles. Therefore, the terms -0.01(Size - 90)² and -0.02(Charge)² penalize deviations from these optimal values, with the choice of coefficients scaling the relative importance of size and charge in the delivery performance.

Conversely, PEGylation and active targeting were modeled as linear contributors to performance. The positive coefficients (+0.5 for PEGylation and +1.5 for targeting) reflect the experimental evidence that moderate PEGylation improves nanoparticle stealth properties, and the presence of active targeting ligands substantially enhances cellular uptake by promoting receptor-mediated endocytosis.

Finally, Gaussian-distributed noise (ϵ, mean = 0, standard deviation = 0.1) was added to each Delta_AUC value to simulate biological variability and measurement uncertainty typically observed in vivo and in vitro assays. This biologically informed functional form allowed us to create a synthetic dataset, through an in-house R script, where optimal nanoparticle configurations (around 90 nm in size, with near-neutral surface charge, moderate PEGylation, and active targeting) systematically achieve higher Delta_AUC values, while suboptimal configurations are penalized. This design ensures that machine learning models trained on the dataset are exposed to realistic, non-linear, and multi-parametric optimization challenges, mimicking the complexity of real-world nanoparticle formulation tasks (27, 28).

This synthetic dataset was subsequently used to train and evaluate a supervised machine learning model, as described in the following section.

A supervised machine learning approach was employed to predict immune activation potential (ΔAUC) of lipid nanoparticle (LNP) formulations based on key physicochemical descriptors. A Random Forest regression model (29) was implemented using the randomForest package in R. Input features included particle size (nm), surface charge (mV), PEGylation percentage (mol%), and presence of targeting ligands (binary encoding).

The synthetic dataset described above, comprising 100 simulated LNP formulations generated by systematically varying key physicochemical parameters across biologically relevant ranges, was randomly partitioned into training (80%) and validation (20%) subsets. Model performance was assessed using root mean square error (RMSE) and the coefficient of determination (R²) on the validation set, providing quantitative estimates of predictive accuracy and generalizability (30). RMSE measures the average magnitude of the prediction errors, providing an estimate of how close the predicted values are to the actual ones: lower values indicate better accuracy. R² quantifies the proportion of variance in the observed data that is explained by the model, with values closer to 1 indicating higher predictive power and generalizability.

Feature importance was assessed using the mean decrease in node impurity, a standard metric in Random Forest models that quantifies how much each variable contributes to improving decision tree splits. This analysis revealed that surface charge and PEGylation were the most influential predictors of ΔAUC, in line with their well-established roles in modulating nanoparticle biodistribution and immunogenicity. The trained model was subsequently used to predict ΔAUC values for all LNP candidates in the dataset. The top 10 formulations were selected based on their predicted scores and visualized accordingly, forming the basis for downstream optimization via genetic algorithms.

Full model and optimization settings, cross-validation protocol, and sensitivity analysis are reported inand. 1 1

Building on the predictive model trained on the synthetic LNP dataset, we used the optimized Random Forest as a surrogate fitness function within a genetic algorithm (GA) to search for new LNP formulations predicted to yield high Delta_AUC values, i.e., strong immune activation profiles. The GA was implemented using the GA library in R, which simulates an evolutionary process to solve optimization problems. We began with an initial population of 50 LNP formulations, randomly generated within biologically plausible parameter ranges (for size, charge, PEGylation, and targeting). Each formulation in the population was evaluated using the trained Random Forest model, which predicted its Delta_AUC score: this prediction served as the fitness value for the GA. The selection of individuals for reproduction was performed using a tournament strategy, where multiple candidates compete and the best is chosen for mating. To simulate genetic diversity and exploration of the solution space, we applied crossover (with a probability of 0.8) to exchange parameter values between formulations, and mutation (with a probability of 0.2) to introduce small random changes. This process was repeated over 100 generations. As the algorithm progressed, it increasingly favored formulations with higher predicted Delta_AUC, gradually converging towards optimal solutions. At the end of the run, we selected the top 10 formulations, those with the highest predicted Delta_AUC scores, for further analysis.

To quantify the robustness of the model predictions and the associated uncertainty, we performed statistical confidence and clustering analyses on both the immune risk index and the ΔAUC predictions.

For the immune risk index (Figure 2), 95% confidence intervals were estimated using a nonparametric bootstrap procedure (B = 1000 resamplings) applied to compartment-specific immune markers, weighted by their respective ΔAUC coefficients.

For the ΔAUC predictions (Figure 3), a bootstrap approach was applied to the random forest model, which was re-trained on 500 bootstrap samples of the synthetic LNP dataset to estimate prediction variability.

In addition, hierarchical clustering was incorporated into the heatmaps (Figure 4) to highlight parameter co-variation, and a correlation heatmap (Spearman’s ρ) was generated to visualize relationships among LNP physicochemical parameters and ΔAUC values.

All data preprocessing, statistical analyses, and initial visualizations were performed using R (v4.4.1) within RStudio 2024.04.2 + 764. Differential expression analysis was carried out with DESeq2 (v1.48.1), while predictive modeling and optimization were implemented using randomForest (v4.7.1.2) and GA (v3.2.4), respectively. Exploratory plots were generated with ggplot2 (v4.0.0) and pheatmap (v1.0.13).

To refine figure design and ensure visual consistency, selected key plots—such as ΔAUC comparisons, immune risk index distributions, and LNP ranking—were reproduced using Python (v3.13.2) in a dedicated virtual environment with matplotlib (v3.9.2) and seaborn (v0.13.2).

All analyses were executed on an iMac with Apple M3 chip (8-core CPU, 10-core GPU) equipped with 24 GB unified memory, running macOS Sequoia 15.6.1. This hybrid R/Python workflow ensured both graphical uniformity and full reproducibility across the study.

Differential gene expression analysis of the synthetic RNA-seq dataset accurately identified the simulated transcriptional changes. Among the 300 analyzed genes, all 30 genes designed to be upregulated, and the 30 genes designated as downregulated post-vaccination were correctly identified as significantly differentially expressed (FDR < 0.05), demonstrating the reliability and validity of the synthetic data generation methodology. Additionally, key immune marker genes representing distinct immune compartments, such as B cells (CD19, MS4A1), T cells (CD3D, CD8A, CD4), and plasma cells (IGHG1, IGHM, PRDM1), were significantly upregulated, consistent with expected immune activation patterns.

To assess potential off-target immune activation, we computed a compartment-specific immune risk index by multiplying the predicted ΔAUC values by the number of differentially expressed (DE) immune marker genes within each compartment, as shown in Figure 2:

The Ig compartment, representing antibody-producing plasma cells, displayed the highest risk index (~4.2), suggesting a strong activation of humoral responses, consistent with mRNA vaccine effects (31). The T cell compartment followed (~3.6), indicating robust T cell engagement. Memory T cells (Tmem) and cytotoxic T cells (TC) showed moderate risk levels (~2.2 and ~1.8, respectively), while B cells had a slightly lower activation (~1.6). Notably, the EP compartment, likely representing epithelial or non-immune cells, had the lowest index (~1.2), suggesting minimal off-target transcriptional activation. These results support the capacity of the simulated nanoparticle formulation to preferentially activate relevant immune pathways while sparing non-target tissues, aligning with the immune response patterns previously predicted by UISS.

The calculated immune risk index effectively quantified compartment-specific immune activation, clearly distinguishing between post-vaccination and control conditions. Specifically, the highest immune risk index values were observed in the T cell compartment, driven by strong upregulation of CD3D, CD8A, and CD4 genes, in alignment with simulated Delta_AUC scores derived from the UISS model. B cell and plasma cell compartments exhibited moderate immune risk scores, correlating with fewer significantly upregulated marker genes. Overall, the immune risk indexing method demonstrated strong correlation with simulated immune activation, offering a robust and interpretable approach for evaluating potential off-target immune responses.

Based on prior UISS simulations, immune compartments showed distinct activation patterns when comparing targeted and non-targeted mRNA vaccine delivery. Delta_AUC values were calculated to quantify the difference in immune activation between conditions. Compartments such as B cells and plasma cells (Ig) showed the highest differential activation, indicating preferential targeting and stronger immune engagement when delivery was optimized.

The difference in immune activation between targeted and non-targeted formulations (ΔAUC) was computed for each immune compartment.

As illustrated in Figure 3, the Ig compartment exhibited the highest increase in ΔAUC, followed by B and T cells, indicating a stronger activation under targeted delivery.

In contrast, epithelial (EP), cytotoxic (TC), and memory T (Tmem) compartments showed smaller ΔAUC values, suggesting that their activation is less affected by the delivery modality within the current simulation setup.

Using the synthetic dataset previously described, which uniformly sampled a four-dimensional physicochemical parameter space, we trained a Random Forest regression model to predict Delta_AUC values based on LNP characteristics. The model achieved strong predictive performance, with R² values exceeding 0.9 and low RMSE on the validation set, confirming its ability to capture non-linear relationships between input features and immune activation potential.

The model was then embedded as a surrogate fitness function within a genetic algorithm to identify LNP formulations predicted to maximize Delta_AUC. After 100 generations, the GA consistently converged toward optimal configurations, that is, nanoparticles around 90 nm in size, with near-neutral surface charge, moderate PEGylation, and active targeting, closely matching profiles known to enhance biodistribution and immunogenicity. Following model training and validation, ΔAUC values were predicted for the entire synthetic LNP dataset. After convergence, the genetic algorithm identified a set of top 10 LNP formulations that consistently exhibited superior predicted performance as shown in Table 1:

All selected candidates included active targeting ligands and exhibited particle sizes ranging from 88.8 to 93.9 nm, with a central tendency around 90–92 nm, aligning with theoretical optima for biodistribution. This outcome reflects the influence of the scoring function used during model training, which included a positive weighting for the presence of targeting ligands, thereby favoring formulations predicted to enhance receptor-mediated uptake and compartment-specific immune activation. Surface charges were consistently near-neutral, varying between −1.0 and −4.4 mV, and PEGylation percentages ranged from 0.26 to 0.34 mol%, centering around the biologically favorable 0.3 mol%. This near-neutral charge is known to minimize non-specific interactions with serum proteins and immune cells, thereby improving circulation time and reducing innate immune activation (32). Similarly, an optimal PEGylation density has been shown to balance nanoparticle stealth and cellular uptake, preventing rapid clearance while maintaining delivery efficiency.

The predicted ΔAUC values, calculated using the biologically informed non-linear scoring function described in the Methods section, ranged from 0.99 to 1.73. The highest score (1.73) was achieved by the top-performing formulation (Rank 2), while the lowest among the top 10 (Rank 4) was 0.99. Although the ΔAUC range was narrower than initially anticipated, the results highlight the genetic algorithm’s ability to finely discriminate between LNP designs with subtle yet functionally meaningful differences.

Notably, all top-ranked formulations exhibited overlapping physicochemical features: particle sizes around 90–92 nm, near-neutral surface charges, and PEGylation levels close to 0.3 mol%, indicating strong convergence toward a shared optimal profile. These findings not only validate the effectiveness of the GA in identifying high-performing candidates but also reinforce design patterns observed in earlier model-driven rankings. In particular, the convergence toward moderate PEGylation and near-neutral charge mirrors experimental literature that associates such profiles with optimal biodistribution and reduced innate immune activation.

The distribution of predicted ΔAUC scores for the top 10 LNP candidates is shown in Figure 4. Notably, LNP-2 achieved the highest predicted score, followed by a gradual decline among the subsequent formulations.

To complement the tabulated summary of physicochemical features (Table 2), we generated a heatmap (Figure 5A) to provide a visual overview of parameter distributions among the top 10 GA-optimized LNP candidates. As previously noted, the selected formulations exhibited broadly consistent trends across size, surface charge, PEGylation, and targeting, reflecting convergence toward a shared optimal physicochemical profile. The heatmap reinforces these findings, offering an intuitive depiction of the design space occupied by the top-performing nanoparticles.

The heatmap presents the z-score–normalized physicochemical parameters—PEG content, predicted ΔAUC, particle size, and surface charge—for the ten GA-optimized LNP formulations.

Two main patterns emerge:

1. Consistency in design parameters:

Most top-performing LNPs occupy a narrow region of the design space, showing moderate PEG percentages (~0.27–0.31%), near-neutral to slightly negative charges (−4 to −1 mV), and diameters close to 90–94 nm. This convergence indicates that the optimization process favored formulations with balanced stability and cellular uptake potential.

2. ΔAUC-driven clustering:

The ΔAUC column highlights subtle differences in predicted immunogenic performance across formulations. LNP-1 and LNP-8 exhibit the highest relative ΔAUC (lighter shades), while others form a compact cluster with lower but comparable predicted responses, reflecting minor variations around the optimal region.

Overall, the figure visually reinforces the model-driven convergence toward an optimal physicochemical profile, characterized by ~90 nm size, low PEG content, and slightly negative charge, consistent with literature-reported parameters for clinically validated mRNA-LNP systems.

To further explore interdependencies among physicochemical variables, a Spearman correlation matrix (Figure 5B) was computed using the top 10 GA-optimized LNP formulations.

Size and surface charge showed moderate positive correlation (ρ = 0.59), while PEG content was inversely correlated with both size and charge, indicating that formulations with lower PEG fractions tend to have slightly larger and less negatively charged particles.

Collectively, these findings illustrate the effectiveness of the machine learning–driven design strategy in prioritizing LNP formulations for further refinement and experimental validation. This approach provides a rational and scalable pathway for accelerating the development of safe and effective mRNA delivery systems.

Finally, to evaluate whether the simulated transcriptomic patterns and model-driven predictions align with experimentally observed vaccine responses, we performed an external validation using public RNA-seq data from COVID-19–vaccinated individuals (GSE171110↗).

The results of this comparative analysis are presented in the following section.

To assess the biological plausibility of the simulated immune response, we validated the synthetic transcriptomic signatures against a public RNA-seq dataset (GSE171110↗) profiling peripheral blood samples from COVID-19–vaccinated and healthy individuals.

This dataset was selected because it captures in vivo immune activation after SARS-CoV-2 vaccination, closely reflecting the biological processes represented in our simulation (e.g., B-cell, T-cell, and immunoglobulin upregulation).

Differential gene expression analysis was performed using DESeq2 on both datasets with identical thresholds (|log2FC| > 0.5, FDR < 0.1). Comparative validation metrics were then computed between the two sets of differentially expressed genes (DEGs), including overlap significance (Fisher’s exact test), directionality concordance, and Pearson correlation of log2 fold-changes.

These metrics are derived from the comparative analysis between simulated and real datasets and do not represent raw biological measurements.

As shown in Table 3, although the overlap between simulated and experimental DEGs was modest (8 shared genes, Fisher’s p = 0.0707), 62.5 % of them displayed concordant regulation direction, and the overall fold-change correlation (r = 0.22) indicated a positive trend in expression magnitude, supporting the biological plausibility of the simulated immune response.

These results confirm that the simulated immune activation patterns, particularly those involving B-cell and plasma-cell markers, exhibit partial but consistent agreement with experimental vaccine transcriptomics. The positive correlation and directional concordance demonstrate that the synthetic simulation preserves biologically plausible immune activation trends without overfitting to specific datasets.

This validation step provides an important bridge between in silico predictions and experimental evidence, reinforcing the translational relevance of the proposed computational framework.

To contextualize the optimized LNPs generated by the in silico framework, their physicochemical characteristics were compared with those reported for clinically validated mRNA–LNP formulations, such as those used in the authorized COVID-19 mRNA vaccines. The parameter space explored in this study (particle size 50–150 nm, surface charge −10 to +10 mV, PEGylation 0.1–0.5 mol % and targeting presence/absence) was designed to represent generic LNPs carriers before mRNA encapsulation.

Publicly available data indicate vaccine LNPs to be small (80–100 nm), slightly negative (~ −5 mV), to contain PEG-lipids around 1.5–2 mol%, and to lack active targeting. Our optimized LNPs converge to the same size window (~90–92 nm) and to a similarly neutral/slightly negative charge, but to a lower PEGylation (~0.30 mol%) and to the presence of targeting ligands (33, 34).

These parameters are summarized in Table 4, together with the corresponding optimized values obtained from the top 10 genetic-algorithm candidates. The ideal LNP identified in this study falls within the experimentally observed range of vaccine-like LNPs, while exhibiting slightly more neutral surface charge, lower PEG-lipid content, and active targeting features predicted to enhance biodistribution and reduce off-target immune activation.

Despite being generated from a pre-encapsulation design space, the optimized LNPs fall within the clinically observed ranges for size and surface charge. Two systematic differences emerge: (i) the optimized candidates feature a lower PEG fraction (~0.30 mol%) than marketed vaccines (1.5–2 mol%), and (ii) they all include targeting ligands, while current products do not. The first difference reflects that our simulations considered PEGylation as an adjustable parameter within a simplified lipid mixture; extending the PEG dimension to 0–2 mol% in future simulations would be straightforward and would not require changing the optimization logic. The second difference reflects the objective function used here, which rewarded predicted improvements in delivery specificity and reduced off-target immune activation; this is consistent with next-generation LNPs but not yet with first-generation COVID-19 products.

The incorporation of mRNA is known to slightly alter these physicochemical properties, generally increasing particle size by 5–15 nm and shifting the surface charge neutrality, while maintaining values within the same overall range (35).

To provide an experimental reference for the biological effects associated with these physicochemical parameters, Table 5 summarizes how the key response variables, such as biodistribution, cellular uptake, and immunogenicity, are typically evaluated in mRNA–LNPs vaccines.

These variables are quantified through established experimental methods, such as in vivo imaging or qPCR for biodistribution, flow cytometry or confocal microscopy for cellular uptake, and immunoassays (ELISA, ELISpot, cytokine profiling) for immunogenicity (34, 36).

The reported experimental ranges highlight consistent biological behaviors across LNP-based vaccine systems, supporting the predictive validity and translational relevance of the optimized in silico framework.

A more detailed comparison between the optimized in silico parameters and experimental data from recent literature is provided in. 1

This comparison indicates that the in silico search was conducted within clinically realistic physicochemical boundaries, while deliberately extending the design space toward targeted and lower-PEGylation to explore safer delivery profiles.