AI-Validated Brain Targeted mRNA Lipid Nanoparticles with Neuronal Tropism

Abstract

Targeting therapeutic nanoparticles to the brain poses a challenge due to the restrictive nature of the blood-brain barrier (BBB). Here we report the development of mRNA-loaded lipid nanoparticles (LNPs) functionalized with BBB-interacting small molecules, thereby enhancing brain delivery and gene expression. Screening brain-targeted mRNA-LNPs in central nervous system (CNS) in vitro models and through intravenous administration in mice demonstrated that acetylcholine-conjugated LNPs achieved superior brain tropism and gene expression, outperforming LNP modifications with nicotine, glucose, memantine, cocaine, tryptophan, and other small molecules. An artificial intelligence (AI)-based model designed to predict the BBB permeability of small-molecule ligands showed strong alignment with our experimental results, providing in vivo validation of its predictive capacity. Cell-specific biodistribution analysis in Cre-reporter Ai9 mice showed that acetylcholine-functionalized LNPs preferentially transfected neurons and astrocytes following either intravenous or intracerebral administration. Mechanistic studies suggest that acetylcholine-LNP uptake is mediated by the functional engagement of acetylcholine receptors (AchRs) followed by endocytosis, which synergistically enhances intracellular mRNA delivery. Moreover, acetylcholine-LNPs successfully crossed a human BBB-on-a-chip model, enabling transgene expression in human iPSC-derived neurons. Their effective penetration and transfection in human brain organoids further support their potential activity in human-based systems. These findings establish a predictive and modular framework for engineering CNS-targeted LNPs, advancing precision gene delivery for brain disorders.

Keywords: Artificial Intelligence; Blood−Brain Barrier; Brain Targeting; Central Nervous System (CNS); Gene Delivery; Lipid Nanoparticles; mRNA.

1

Engineering a brain-targeted (BT) mRNA-loaded LNP library. (A) mRNA LNPs were formulated with a blood–brain barrier (BBB)-interacting small molecule-conjugated PEG-lipid using optimized microfluidic mixing to enable BBB crossing, after intravenous administration, and subsequent selective brain-cell transfection, including neurons. (B) BT-LNPs lipid phase composition [mol %]. DMG-PEG2000 was conjugated to derivatives of the presented small molecules, with tryptophan used in its original form. (C) The microfluidic mixing-based production method of LNPs was adjusted to facilitate the integration of small-molecule-conjugated PEG lipid. DMSO was used instead of EtOH to improve solubility, and process temperatures were optimized for best performance. (D) Physicochemical properties of BT-LNP library. Size, polydispersity index (PDI) (n ≥ 8 independent groups), and zeta potential (n = 3). Data are shown as mean ± SD. (E) Cryo-TEM imaging confirmed the uniformity, repeatability, and size distribution of BT-LNPs. Glucose-LNPs were used as a representative formulation for imaging. Scale bar = 100 nm. Illustrations were created in BioRender [Shklover, J. (2025) https://BioRender.com/fmeu0vq].

2

Predicting molecule–BBB interaction potential across species using a graph neural network-based AI model. (A) The AI model predicts the activity of a given molecule within a specific tissue of an organism. It begins by converting the molecule into a graph representation, which is encoded through a graph neural network (GNN) to generate a molecular embedding. In parallel, a curriculum learning strategy exposes the model to data from different organisms sequentially, and the organism information is encoded as an organism-specific tag. These embeddings are concatenated and processed by a classifier that predicts the molecule’s activity in the designated tissue. The model is trained for each tissue using all available organismal data, though not every tissue has data from all organisms, as shown in the figure. Illustration created in BioRender [Shklover, J. (2025) https://BioRender.com/fmeu0vq]. (B) The prediction of the molecule–BBB interaction potential (%) for the small-molecule library used in BT-LNPs was estimated using the developed AI model for both rat and human brains. Acetylcholine demonstrated the highest predicted probability of BBB interaction among all tested ligands. A methyl molecule was used to represent the untargeted condition.

3

In vitro and in vivo screening and evaluation of brain targeted (BT)-LNPs. (A) Cultures of hCMEC/D3 and differentiated SH-SY5Y cells were treated with firefly luciferase (FLuc)-encoding BT-LNPs to assess (i) the transfection efficiency (RLU = relative luminescence units) and (ii) the cytotoxicity. All treatments were normalized to the untreated control (n ≥ 12). ****p < 0.0001. (B) A high-throughput luminescence-based assay was used to evaluate the biodistribution profile and in vivo transfection efficiency of the FLuc-encoding BT-LNPs library following intravenous (IV) administration. (C) Transfection efficiency of BT-LNPs in various organs. Data in each graph are normalized to the untargeted control group within the same organ (n = 4–5 mice per treatment). *p ≤ 0.0357; **p = 0.0094; ***p = 0.0005; ****p < 0.0001. Results are presented as mean ± SD. One-way ANOVA with correction for multiple comparisons was used for statistical analysis. (D) (i) Schematic of the transwell-based BBB model. Human iPSC-derived brain microvascular endothelial cells (BMECs) were seeded on the upper side of the transwell insert, while human iPSC-derived neurons were cultured on the bottom (basolateral) side. Acetylcholine- or untargeted LNPs encoding mCherry mRNA were applied to the apical (BMEC-facing) compartment. (ii) Representative confocal images of the neuronal layer following 24 h of incubation. βTubIII (green) marks neuronal processes, mCherry (magenta) indicates successful transfection, and DAPI (blue) stains nuclei. Neurons exposed to acetylcholine-LNPs showed increased mCherry expression. n = 3 biological repetitions; scale bars = 100 μm and 50 μm for zoom-in images. Illustrations created in BioRender [Shklover, J. (2025) https://BioRender.com/fmeu0vq].

4

Cellular-level examination and visualization of selected BT-LNPs in the mouse brain and human brain organoid. (A) Ai9 Cre reporter mice, which express tdTomato protein under the control of the CAG promoter upon Cre-mediated recombination, were used to study brain-cell selectivity and uptake of selected BT-LNP formulations. (B) Distribution (%) of each LNP formulation across major brain cell populations, shown as pie charts, illustrating cell-type preferences in vivo. (C) Mean fluorescence intensity (MFI) of tdTomato signal in each brain cell population: endothelial cells (i), microglia (ii), neurons (iii), and oligodendrocytes (iv) following LNP treatment. Astrocytes showed no significant positive signal compared to the control group (n = 3 mice per treatment). *p ≤ 0.0403. Data are presented as mean ± SD. One-way ANOVA with multiple comparisons adjusted p values were used for statistical analysis. (D) 48 h after acetylcholine-LNPs administration, brain sections show a widespread tdTomato expression, predominantly localized along vascular structures. A noninjected mouse served as a control. Images are shown at 4× magnification; scale bar = 1000 μm; zoom-in scale bar = 100 μm. (E) Human iPSC-derived cortical brain organoids were treated with acetylcholine-LNPs encapsulating mCherry mRNA (n = 2 independent groups). After 48 h, confocal imaging demonstrated mCherry expression (red) across both peripheral and deeper organoid regions, including rosette-like structures. Neurons are marked by βTubulin III (βTubIII, green) and nuclei by DAPI (blue); scale bars = 400 μm (top), 100 μm (bottom). Illustrations created in BioRender [Shklover, J. (2025) https://BioRender.com/fmeu0vq].

5

Mechanistic investigation of acetylcholine-conjugated LNPs (Ach-LNPs) cellular uptake. (A) (i), (ii) Luciferase mRNA-loaded Ach-LNPs were applied to BMECs (in a transwell BBB model) and SH-SY5Y neuronal-like cells. Cells were treated with Ach-LNPs alone or preincubated with acetylcholine receptor inhibitors: scopolamine (Scop, muscarinic antagonist) or mecamylamine (Mec, nicotinic antagonist). Untargeted LNPs served as controls. (B) To assess endocytosis involvement, SH-SY5Y cells were pretreated with methyl-β-cyclodextrin (MβCD), a caveolae-mediated endocytosis inhibitor. Data from A and B are presented as normalized luminescence units (RLU) relative to untargeted control (mean ± SD, n ≥ 3). *p = 0.0216; **p = 0.0011; ***p = 0.0010. (C) AchLightG sensor activation occurs upon Ach-LNP binding, inducing GFP fluorescence for real-time tracking. (D) (i) Representative images of HEK293 cells expressing AchLightG before and after 35 min of treatment with untargeted LNPs, free Ach (100 μM), or Ach-LNPs. Increased green fluorescence indicates sensor activation. Scale bar = 50 μm. (ii) Quantification of normalized GFP intensity before/after 35 min; both free Ach and Ach-LNPs significantly increased GFP signal (n ≥ 80). (iii) Time course of normalized GFP fluorescence shows significant increases (****p < 0.0001) after Ach or Ach-LNP treatment. The arrow shows the treatment application time. (E) (i) Time-lapse imaging of AchLightG-expressing HEK293 cells treated with mCherry mRNA-loaded Ach-LNPs shows progressive mCherry and GFP signal over 23 h. Scale bar = 50 μm. AchLightG sensor activation is shown in green (GFP) and mCherry expression on the fire scale. Scale bar = 50 μm. (ii) Quantification of GFP, mCherry, and mCherry/GFP intensity over time. (F) (i)–(iii) Quantification of normalized mCherry fluorescence up to 20 h post-LNP treatment (n = 60). Ach-LNPs yielded significantly higher final mCherry signal (**p = 0.0028). n = number of cells; error bars denote SEM; statistics by Student’s t test or one-way ANOVA with post hoc Tukey’s test. Experiments were independently repeated at least 3 times. Illustrations created in BioRender [Shklover, J. (2025) https://BioRender.com/fmeu0vq].

6

Cellular specificity of BT-LNPs following intracerebral (IC) injection. (A) IC injections of Cre-encoding acetylcholine and untargeted LNPs were conducted to compare their cellular specificity within the brain. Untargeted FLuc-mRNA LNPs were included as a negative control to set the basal tdTomato expression. Illustration created in BioRender [Shklover, J. (2025) https://BioRender.com/fmeu0vq]. (B) Representative images of brain sections that were fixed and immunostained for NeuN (neurons), Iba1 (microglia, (i)), and GFAP (astrocytes, (ii)). Cell nuclei were counterstained with DAPI. LNP uptake and endosomal escape were confirmed by the expression of tdTomato. Images were captured at 10× magnification; scale bar = 500 μm. (C) Representative higher-magnification images (20×) of cell staining in slides of both acetylcholine- and untargeted LNPs injected brains. Scale bar = 50 μm. (D) Image analysis of tdTomato colocalization ratio (%) in brain cells, including neurons (i), microglia (ii), and astrocytes (iii). Data are expressed as mean ± SD (n = 4 independent experiments, each with at least 40 replicates per group); unpaired t test p value; ****p < 0.0001.