Nanocarrier imaging at single-cell resolution across entire mouse bodies with deep learning

Materials

DLin-MC3-DMA (555308) was purchased from MedKoo Biosciences. DSPC (850365P), cholesterol (C8667) and DMG-PEG-2000 (880151P) were obtained from Sigma-Aldrich. Pur-A-Lyzer Midi Dialysis Kits (PURD35030, WMCO, 3.5 kDa) were purchased from Sigma-Aldrich. In vitro translation (IVT)-generated EGFP mRNA, spike protein mRNA and Alexa Fluor 647–labeled and Alexa Fluor 750–labeled EGFP mRNA, as well as FLuc mRNA, were purchased from RiboPro. Alexa Fluor 568–conjugated anti-GFP signal-enhancing nanobodies (gb2AF568-50) and Alexa Fluor 647–conjugated anti-GFP signal-enhancing nanobodies (gb2AF647) were purchased from ChromoTek. Propidium iodide was obtained from Sigma-Aldrich (P4864). VivoGlo luciferin was obtained from Promega. We used the following AAV constructs: Retro-AAV: (retro AAV-CAG-GFP, Addgene, 37825-AAVrg; titer ≥ 7 × 10¹² vg ml⁻¹; dilution 1:10 with 1× PBS) and PHP.eB-AAV: (PHP.eB-CAG-GFP, Addgene, 37825-PHPeB; titer ≥ 7 × 10¹³ vg ml⁻¹; dilution 1:10 in 1× PBS).

Animals

Mixed-gender C57BL/6 mice were obtained from Charles River Laboratories. The animals were housed under a 12-h light/dark cycle and had random access to food and water. The temperature was maintained at 18–23 °C, and humidity was at 40–60%. The animal experiments were conducted according to institutional guidelines of the Klinikum der Universität München/Ludwig Maximilian University of Munich and after approval of the Ethical Review Board of the Government of Upper Bavaria (Regierung von Oberbayern, Munich, Germany) and under European Directive 2010/63/EU for animal research. All experiments were performed in triplicates.

LNP formulation

MC3-based LNPs were prepared using the ethanol dilution method. All lipids with specified molar ratios (Dlin-MC3-DMA/DSPC/cholesterol/PEG-DMG at a 50:10:38.5:1.5 molar ratio) were dissolved in ethanol absolute (<99.8%), and mRNA was dissolved in 10 mM citrate buffer (pH 4.0). Those two solutions were mixed at an aqueous-to-ethanol volume ratio of 3:1 to make a final weight ratio of 1:10 (mRNA/total lipids) and then incubated for 10 min at room temperature. The final LNP formulations were dialyzed (Pur-A-Lyzer Midi Dialysis Kits, molecular weight cutoff (MWCO) 3.5 kDa) against PBS for 2 h to remove the ethanol and neutralize the pH. For Alexa Fluor 647–labeled EGFP mRNA LNPs, 20% non-labeled UTP was replaced by Alexa Fluor 647–labeled UTP^82,83. The hydrodynamic diameter and zeta potential of the LNPs were measured by DLS using disposable cuvettes in a Malvern Zetasizer. The encapsulation efficacy was examined using a Quanti-iT RiboGreen RNA Assay Kit (Thermo Fisher Scientific, Q33140)⁸⁴.

For other LNP formulations:

• Lung SORT (in molar ratios): cholesterol (19.3), DSPC (5), PEG-DMG (0.8), MC3 (25), DOTAP (50) (total lipids/mRNA = 40/1, wt/wt)

• SM-102 LNP formulations (in molar ratios): cholesterol (38.5), DSPC (10), PEG-DMG (1.5), SM-102 (50) (total lipids/mRNA = 10/1, wt/wt)

• ALC-0315 LNP formulations (in molar ratios): cholesterol (38.5), DSPC (10), PEG-DMG (1.5), ALC-0315 (50) (total lipids/mRNA = 10/1, wt/wt)

DNA origami production

The biotechnologically produced nanorod was prepared as described previously⁸⁵. In brief, the staple strands for the nanoobject are arranged as a pseudogene, interleaved by self-excising DNAzyme cassettes. The pseudogene is cloned into a phagemid backbone, a scaffold strand for the nanoobject. The fully cloned phagemid was introduced into Escherichia coli JM109 cells by transformation. Using helper phage rescue, the phage particles containing the phagemid single-stranded DNA are produced in a stirred-tank bioreactor, followed by phage and DNA purification to yield pure circular ssDNA. As a prerequisite for the in vivo experiments, the ssDNA was endotoxin purified as described by Hahn et al.⁸⁶, using Triton X-114. By adding zinc cations, the self-excising DNAzyme is activated and yields scaffold and staple strands for the nanorod. After a further DNA purification step with ethanol, the nanorod was assembled in a buffer containing 5 mM Tris, 5 mM NaCl, 1 mM EDTA and 10 mM MgCl₂ at a DNA concentration of 50 nM (15 min at 65 °C and then 60–45 °C at 1 h per °C). To assemble fluorescence-labeled DNA nanoobjects for the optimized DISCO clearing protocol, we purchased chemically synthesized single-stranded staple strands with Atto 550 or Atto 647N modification from Integrated DNA Technologies or Biomers. The folding reactions were set up as 100-µl or 2-ml reaction solutions with UV crosslinking for 5 min⁸⁷.

Preparation of IgG–ssDNA conjugates

Oligonucleotides modified with 3′ thiol modification were purchased high-performance liquid chromatography (HPLC) purified and dried from Biomers. The oligos were dissolved in PBS (pH 7.2) with 5 mM TCEP and incubated for 1 h at room temperature. After three rounds of filter purification (10k Amicon Ultra 0–5-ml centrifugal filter), 10 nmol of the reduced thiol oligo was mixed with 10 equivalents of Sulfo-SMCC (sulfosuccinimidyl 4-(N-maleimidomethyl) cyclohexane 1-carboxylate; Thermo Fisher Scientific, dissolved in ddH₂O) for 15 min. After three rounds of filter purification (10k Amicon Ultra 0–5-ml centrifugal filter), including buffer change to PBS (pH 8), 100 µg of antibody in PBS (pH 8) was added. The reaction was incubated for 4 h at 4 °C at least. The conjugate was subsequently purified by ion-exchange chromatography (Dynamic Biosensors, proFIRE) using an NaCl gradient of 150–1,000 mM in PBS (pH 7.2). The purity of the oligo–antibody conjugates was analyzed by SDS-PAGE and agarose gel.

Assembly of antibody functionalized DNA nanorod

The assembled nanorod was purified with two rounds of PEG precipitation to remove excess staple strands, as described by Stahl et al.⁸⁸, with slightly higher PEG 8000 (8.25% wt/vol) and NaCl (275 mM) concentrations. The nanorod was equipped with up to four handle ssDNA strands (part of the original structure as assembled above) with complementary sequences to the IgG–ssDNA conjugate. The hybridization of the IgG–ssDNA conjugate to the nanorod was done in a buffer containing 5 mM Tris, 5 mM NaCl, 1 mM EDTA and 10 mM MgCl₂ at pH 8 overnight at 30 °C (Supplementary Fig. 16b). To concentrate the fully functionalized nanorod, we used another round of PEG purification, followed by dissolving the nanorod in Tris buffer (5 mM Tris, 5 mM NaCl, 1 mM EDTA and 5 mM MgCl₂) at a high concentration of 4 µM. We added PEG-polylysine (methoxy-poly(ethylene glycol)-block-poly(l-lysine hydrochloride, 10 lysine repeating units, 5-kDa molecular weight (MW) PEG, Alamanda polymers)) at a nitrogen-to-phosphate (N:P) ratio of 1:1 to stabilize the nanorod against low-salt conditions and nuclease activity in vivo. For the in vivo experiments, the nanorod was diluted to 0.5 µM or 2 µM using sterile PBS to reduce the magnesium concentration to 2 mM or less. Finally, the endotoxin concentration of the fully equipped nanorod was determined using Endosafe nexgen-PTS (Charles River Laboratories) to fulfill the FDA requirement of less than 36 EU ml⁻¹ for a 100-µl intravenous injection per day.

Different nanocarriers administered to mice and dose calculations

LNPs were administered via different routes, including intranasal, intramuscular, intravenous, oral gavage and intradermal. Intranasal delivery was performed on lightly anaesthetized mice by passively presenting the sample droplet by droplet with a pipette to the mouse and waiting until each droplet entered the nose during normal breathing. Intramuscular delivery was administered into the muscles of the hindlimb. Intravenous injections were done via the tail vein. Oral gavage administration was performed with an animal feeding needle on awake mice. Intradermal injection was administered to the back of anaesthetized mice onto a small, shaved area. For all injections, 30-gauge needles were used.

The injection doses varied between 0.0005 mg kg⁻¹ and 0.5 mg kg⁻¹ as described above. When translating human vaccine doses to animal models, it is crucial to account for differences in body surface area rather than just body weight (allometric scaling). According to the guidelines, to estimate the equivalent dose for a mouse, one should divide the human dose by a factor of 12.3. For example, if the human dose is 30 µg (approximate dose used in mRNA vaccines for humans), the corresponding mouse dose would be approximately 2.4 µg (30 µg divided by 12.3)⁵⁰. This conversion ensures that the administered dose elicits a similar pharmacodynamic response in mice^48,49.

The animals were euthanized at 6 h (biodistribution) and 72 h (expression) after administration. Liposomes based on the FDA-approved drug Doxil deliver Atto 647 dye. The liposome formulation consists of hydrogenated soya phosphatidylcholine, cholesterol and PEG-modified phosphatidylethanolamine in a 55:40:5 molar ratio, with Atto 647 (AD 647, ATTO-TEC) as the cargo. These liposomes were produced using the thin film hydration method followed by size extrusion with a 200-µm filter to achieve uniformity⁸⁹. After production, the liposomes were concentrated using an Amicon 30-kDa filter. Branched PEI delivers single-stranded Alexa Fluor 647–labeled DNA (Alexa 647-AGACGTGTGCTCTTCCGATCTTCTGCCCTCCAG ATCGGAAGAGCGTCGTGTG). The particles were formulated using branched PEI with an MW of 25 kDa, combined with ssDNA labeled with 647 dye, at an N:P ratio of 15. These particles were produced through a self-assembly process, described in ref. ⁹⁰. Liposomes and polyplexes were injected intramuscularly at a concentration of 0.1 mg kg⁻¹. Different DNA origamis (untargeted origami and Cx3cr1 antibody-targeted origami) were injected intravenously at a dose of 50.725 mg kg⁻¹, with various injection timepoints ranging from 20 min to 20 h. Retro-AAV and PHP.eB-AAV were injected intramuscularly and intravenously, respectively (10¹³ GC ml⁻¹ in 25 μl)^67,68,69. After 2 weeks, the mice were perfused and fixed.

Protein corona preparation

Blood from wild-type Balb/c mice was collected retro-orbitally to obtain either plasma (in microcentrifuge tubes coated with EDTA, collected on ice, spun down for 15 min at 2,000 r.p.m. at 4 °C to retrieve plasma (supernatant)) or serum (in non-coated microcentrifuge tubes, blood left to clot at room temperature for 30 min, spun down for 15 min at 2,000 r.p.m. at 4 °C to retrieve serum (supernatant)). LNPs (50 µl each of the final preparations as described above, encapsulating 1 µg of unlabeled mRNA per formulation) were incubated with mouse plasm or serum for 15 min at 37 °C at a 31:1 volumetric ratio. A 0.7-M sucrose solution was prepared by dissolving solid sucrose in MilliQ water. The LNP/plasma mixture was loaded onto a 300-µl sucrose cushion centrifuged at 15,300g and 4 °C for 1 h. The supernatant was removed, and the pellet was washed with 1× PBS. Next, the pellet was centrifuged at 15,300g and 4 °C for 5 min, and the supernatant was removed. Washing was performed twice more. The protein-covered LNPs were kept dry at −20 °C upon mass spectrometry analysis.

Perfusion and tissue preparation

Mice were deeply anesthetized using a combination of ketamine and xylazine (concentration, 2 ml per 1 kg intramuscularly). Afterwards, the chest cavity of mice was opened for intracardial perfusion with heparinized PBS (10–25 U ml⁻¹ heparin dissolved in 0.01 M PBS) using 100–125-mmHg pressure for 5–10 min at room temperature until the blood was washed out. (The color of the liver changes from red to light yellow). Next, the mice were perfused with 4% paraformaldehyde (PFA) to fix the entire mouse body for 5 min. Finally, the skin was carefully removed, and the bodies were fixed in 4% PFA overnight at 4 °C and transferred to PBS for later immunostaining or directly processed for clearing as described below.

Optimized DISCO technique to visualize dye-labeled LNP-mRNA

As we injected LNP with Alexa Fluor 647–labeled mRNA into mice, we developed a modified DISCO technique to render the mice transparent while retaining the fluorescent signal within the mouse body. A simplified transcardial circulatory system was established using a peristaltic pump (ISMATEC, REGLO Digital MS-4/8 ISM 834, reference tubing, SC0266). One reference tubing was connected by two connectors (Omnilab, 5434482) end to end and extended by additional PVC tubing (Omnilab, 5437920). The head part from a 1-ml syringe (Braun, 9166017V) was cut and inserted into the outflow PVC tubing as a connector for the perfusion needle (Leica, 39471024). Next, a PBS-perfused and PFA-fixed mouse body was placed into a 250-ml glass chamber (Omnilab, 5163279). The inflow tubing of the transcardial circulatory system was fixed below the surface of PBS in the glass chamber using adhesive tape, and the air bubbles were completely removed from the circulation tubing. After fixation, the mice were intracardially perfused with the decolorization solution (1:4 dilution of CUBIC reagent, consisting of 25 wt% N,N,N,N′-tetrakis (2-hydroxypropyl) ethylenediamine (from Sigma-Aldrich) and 15 wt% Triton X-100) for 12 h. After this, the mouse bodies were washed with PBS for 12 h, repeated three times, and then perfused with a decalcification solution (10 wt/vol% EDTA, pH 8.00, from Carl Roth) for 2 d. After another three transcardially administered PBS washes lasting 12 h each, the mice were transferred to passive clearing (no transcardial pumping) under a hood. We proceeded to incubate the mice in 70 vol% THF, 80 vol% THF and 2 × 100 vol% THF. This was followed by a 20-min treatment with DCM and then BABB until the tissues were fully transparent.

Optimized DISCO technique for whole body immunostaining

The protein encoded by the cargo mRNA was expressed after injecting LNP-mRNA into the whole mouse body. We stained the expressed protein using signal-enhancing nanobodies to confirm the mRNA functional integrity. Decolorization and decalcification were performed as above. Then, nanobodies were dissolved in a permeabilization solution containing 0.5% Triton X-100, 1.5% goat serum (Gibco, 16210072), 0.5 mM methyl-beta-cyclodextrin (Sigma-Aldrich, 332615) and 0.2% trans-1-acetyl-4-hydroxy-L-proline (Sigma-Aldrich), and the solution was actively pumped transcardially for 6 d. After the immunostaining step, the mouse bodies passively cleared at room temperature with gentle shaking under a fume hood by incubation in 70 vol% THF, 80 vol% THF, 100 vol% THF and again 100 vol% THF for 12 h each. Then, the mouse bodies were treated with DCM for 20 min and immersed BABB until the tissue was rendered completely transparent.

The antibodies, nanobodies and dye used for mouse body immunostaining were rat anti-CD45 (BD Biosciences, 14-0451-82), SARS-CoV-2 spike antibody (GeneTex, GTX135356), Alexa Fluor 568–conjugated anti-GFP signal-enhancing nanobodies (ChromoTek, gb2AF568-50), Alexa Fluor 647–conjugated anti-GFP signal-enhancing nanobodies (ChromoTek, gb2AF647) and propidium iodide (Sigma-Aldrich, P4864).

Pre-adipocyte differentiation and receptor blocking

Brown pre-adipocyte differentiation

Immortalized brown pre-adipocytes from wild-type C57Bl/6 mice⁹¹ were counted and seeded on chamber slides (Nunc Lab-Tek II chambered cover glass, Z734853). The fully confluent pre-adipocytes were induced with induction medium cocktail containing IBMX (0.5 mM), insulin (100 nM), indomethacin (5 µM), dexamethasome (5 µm) and rosiglitazone (1 µM) in DMEM 10% FBS for 2 d, and, afterwards, the medium was changed every 2 d to differentiation medium (DMEM, 10% FBS) containing 100 nM insulin, T3 and rosiglitazone.

Receptor blocking and immunostaining

On day 7 of differentiation, cells were either treated with PBS (control) or infected with AAV. In the blocking experiments, the cells were pre-treated with heparin (500 µg ml⁻¹; Merck, H4784-250MG), anti-FGFR1 (20 µg ml⁻¹; Invitrogen, PA5-86349), anti-integrin αVβ5 (20 µg ml⁻¹; Bioss, BS-1356R) or anti-AAVR (20 µg ml⁻¹; Proteintech, 21016-1-AP) antibodies for 1 h before Retro-AAV application (Addgene, 37825-AAVrg). On day 9, cells were washed with PBS and fixed in 4% PFA (pH 7.4) for 10 min at room temperature and blocked and permeabilized using 3% BSA and 0.3% Tween 20. Cells were then incubated with anti-perilipin-1 antibody (1:500; Cell Signaling Technology, 3470) overnight at 4 °C, followed by three times wash with PBS, and then incubated with GFP-Booster Alexa Fluor 568 (1:1,000; Thermo Fisher Scientific, 17363333) and goat anti-rabbit IgG Alexa Fluor 647 (1:500; Thermo Fisher Scientific, A-21244). After washing three times, slides were mounted using Dako antifade mounting medium and imaged on a Leica TCS SP8 microscope.

Rehydration and immunostaining of cleared tissue

To confirm that the LNP entered specific cells in mouse hearts, we used rat anti-CD45 (BD Biosciences, 14-0451-82; diluted 1:500), Alexa Fluor 568 goat anti-rat IgG (H+L) antibody (Thermo Fisher Scientific, A-11077; diluted 1:500), mouse anti-troponin T (Abcam, ab8295; diluted 1:500), Alexa Fluor 568 goat anti-mouse IgG (H+L) antibody (Thermo Fisher Scientific, A-11004; diluted 1:500), mouse anti-podocalyxin (R&D Systems, MAB1556; diluted 1:500), rabbit anti-αSMA (Abcam, ab5694; diluted 1:500) and Alexa Fluor 568 goat anti-rabbit IgG antibody (Thermo Fisher Scientific, A-11036; diluted 1:500). The cleared mouse heart was rehydrated with the following steps at room temperature with gentle shaking: 100% THF, 90% THF, 70% THF, 50% THF and 0.01 M PBS. Afterwards, the heart tissue was dissected, cut into 10-µm sections and incubated in 0.01 M PBS containing 0.2% gelatin, 0.5% Triton X-100 and 5% goat serum for 1 d at 37 °C. The sections were then incubated with the primary antibodies diluted in the same solution overnight, washed twice in PBS, incubated with secondary antibodies for 4 h and washed in PBS three times.

Confocal imaging

The heart slices, stained to differentiate cell types, were imaged using an inverted laser scanning confocal microscopy system equipped with a ×40 oil immersion lens (Zeiss, EC Plan-Neofluar ×40/1.30 oil DIC M27) and a ×25 water immersion long working distance (WD) objective lens (Leica, numerical aperture (NA) 0.95, WD = 2.5 mm), employing a z-step size of 0.3 μm.

Light sheet imaging

We used the Miltenyi Biotec UltraMicroscope Blaze light sheet imaging system. For the full-scale mouse body, we used ×1.1 and ×4 magnification objectives (LaVision BioTec MI PLAN ×1.1/0.1 NA (WD = 17 mm) and Olympus XLFLUOR ×4 corrected/0.28 NA (WD = 10 mm), respectively). High-magnification tile scans were acquired using 25% overlap, and the light sheet width was reduced to obtain maximum illumination in the field of view. The z-step size was 6 μm, and the exposure time was 100 ms. Data collection was performed using ImSpector (Miltenyi Biotec, version 7.3.2).

Intravital imaging

For multiphoton imaging to determine non-targeting DNA origami half-life in blood, we used an upright Zeiss LSM710 confocal microscope equipped with a Ti:Sa laser (Coherent, Chameleon Vision II) and two external photomultiplier detectors for red and green fluorescence. Eight-week-old C56BL/6N mice obtained from Charles River Laboratories were anesthetized intraperitoneally with a combination of medetomidine (0.5 mg kg⁻¹), fentanyl (0.05 mg kg⁻¹) and midazolam (5 mg kg⁻¹) (MMF). The body temperature was monitored and maintained throughout the experiment using a rectal probe connected to a feedback-controlled heating pad. A catheter was placed in the femoral artery to administer fluorescent dye or nanoparticles. A rectangular 4 × 4-mm cranial window was drilled over the right frontoparietal cortex under continuous cooling with saline, as described⁹². The mouse was injected with FITC-dextran at 3 μl g⁻¹ to identify the brain vessels and obtain a baseline image. Afterwards, the animal was positioned on the multiphoton microscope adapted for intravital imaging of small animals. Non-targeting DNA origami, conjugated with Atto 550, was injected at a concentration of 1 μM and a volume of 120 μl. Scanning was performed with time series at an 80-μm depth, using 10% laser power at 800 nm, a GAASP detector with an LP < 570-nm filter and a master gain of 600 for the FITC channel. For the nanoparticles channel, an LP > 570-nm filter with a master gain of 600 was used. In total, the imaging time did not exceed 70 min. Fluorescence analysis was conducted using FIJI software.

Bioluminescence imaging

Different doses of LNP FLuc mRNA (0.5, 0.05, 0.005 and 0.0005 mg/kg⁻¹; lipid-to-mRNA ratio, 5:1) were injected intravenously into mice. Six hours after the injection of LNP-Luc mRNA, an aqueous solution of L-luciferin (250 μl, 1.6 mg; BD Biosciences) was administered intraperitoneally. The bioluminescence signal was measured 10 min after injection and was then quantified using Living Image Software version 4.2 (Caliper Life Sciences).

Protein corona measurement of DNA origami

DNA origami was dissolved in HEPES buffer or serum-containing media with graded concentrations (70%, 90% and 99%). The folded capillary cell containing the samples was then measured three times, with six sub-runs each, at a temperature of 25 °C, which was measured by fluorescence correlation spectroscopy.

Viral injection procedure

C57BL/6J mice (all male, aged 6–8 weeks) from Charles River Laboratories were used for AAV injections. The virus was delivered intravenously through the tail vein and intramuscularly in the biceps femoris using a 31-gauge insulin syringe (BD Biosciences, 328438). AAV-PHP.eB-CAG-GFP (Addgene, 37825-PHPeB; titer ≥ 1 × 10¹³ vg ml⁻¹; diluted 1:10 with 1× PBS) was used for intravenous injections, whereas AAV-retro-CAG-GFP (Addgene, 37825-AAVrg; titer ≥ 7 × 10¹² vg ml⁻¹; diluted 1:10 with 1× PBS) was used for intramuscular injections. Both intravenous and intramuscular injections were administered at a volume of 30 µl.

Before both injections, mice were anesthetized with isoflurane (5% for induction and 1% for maintenance in oxygen, 0.5 L min⁻¹) and placed on a heating pad to maintain a body temperature of 37 °C. After injection, mice received 7.5 mg kg⁻¹ carprofen subcutaneously.

Deep learning for LNP nanocarrier detection

Data annotation

Annotation was performed in VR using SyGlass Annotation software (version 1.7.2). In total, nanoparticle-targeted cells were manually labeled in 31 patches of varying sizes from organs of interest throughout the body, namely from the head, heart, lungs, liver, spleen and kidneys. To determine the optimal architecture and hyperparameters, we split the VR-annotated data into two subsets: a training/validation subset and a test set. The training and validation subset comprises 21 patches containing 13,927 annotated nanoparticle instances. The test subset has 10 patches (at least one per organ of interest) with 6,424 annotated particles. This split was manually selected to be able to evaluate the trained methods on a test set that is representative of all annotated organ types.

As a pre-processing step before training, every patch was normalized to a 0–1 range by the minimum and maximum values of the patch.

Segmentation model development

To achieve the best segmentation performance for nanocarriers, we tested several state-of-the-art segmentation networks on our annotated dataset, including VNet³², U-Net++³³, Attention U-Net³⁴, UNETR³⁵, SwinUNETR³⁶, nnFormer³⁶ and 3D U-Net³⁷. The best-performing model was a 3D U-Net with six encoding layers, five decoding layers and leaky ReLU activation function, trained following the parameters and training protocol described by Isensee et at.⁹³. We named the best-performing model SCP-Nano. The model was trained following a five-fold cross-validation scheme with a patch size of 128 × 128 × 128 voxels subsampled from the original patches during training time. The batch size during training was 2. The applied loss function was an average combination of Dice and cross-entropy loss. The stochastic gradient descent (SGD) optimizer was used for training at a learning rate of 0.0001. The model was trained for 1,000 epochs, and the checkpoint with the lowest loss value on the validation subset was kept as the best model for every fold. All best models trained in the five-fold cross validation scheme were ensembled for downstream evaluation and inference. We performed manual organ annotations using VR.

Inference and analysis

A single whole body mouse LNP scan can be up to 30,000 × 10,500 × 2,000 voxels. As the whole scan is too large for the typical RAM and VRAM of even dedicated servers, we cut up the whole body images into patches up to 500 × 500 × 500 voxels. These patches were fed into the trained segmentation model to get corresponding nanoparticle segmentations. The nanoparticle segmentation maps of the whole body scan were obtained by stitching the patch segmentations.

For quantitative analysis, we have VR to annotate six target organs in every whole body scan, including the brain, heart, lung, liver, spleen and kidneys. By having defined volumes of interest for each organ, we can overlap this information with the previously obtained nanoparticle segmentation to obtain organ-specific quantitative data. For this, we adopted the connected component analysis method from the cc3d library (https://github.com/seung-lab/connected-components-3d) to label and measure every segmented nanoparticle dot in each organ. For LNP quantitative analysis, we took into account local brightness levels, as this gives information about quantity of LNPs taken up by the cells. Specifically, the relative intensity contrast of every segmented LNP dot compared to the background intensity was calculated. The background intensity was estimated per organ by the average intensity of voxels inside the organ while excluding voxels belonging to segmented dots. Then, for every segmented LNP dot, we summed all voxels’ relative intensity contrast values inside. Finally, relative contrast values of all segmented dots were summed up to reflect the amount of LNP in a specific area or organ.

Next, we created a cellular density map based on nanoparticle segmentation to visualize nanoparticle distribution in different organs. A sliding window strategy was employed to compute segmented nanoparticles’ local relative intensity contrast values. A window size of 16 × 16 × 4 voxels was deployed in our experiments. Every voxel in this cube was then assigned with the sum of relative contrast values. Finally, using a 3D Gaussian filter, we obtained a smooth and full-resolution density map.

The SCP-Nano pipeline code is publicly available at https://github.com/erturklab/SCP-Nano.

Sample preparation for mass spectrometry analysis

We divided the study into four groups: spike LNP, EGFP LNP, no-cargo LNP and PBS. Observing the LNP distribution in the heart, we found substantial LNP distribution in the middle wall between the left and right ventricles, whereas the other areas showed less LNP presence. Therefore, we used 20-gauge needles to manually extract three tissue samples from each heart (left, right and middle parts), totaling 45 samples.

Samples were prepared for mass spectrometry injections as described previously⁵¹. In brief, the tissues were washed with PBS and resuspended in SDS-lysis buffer (6% sodium dodecyl sulfate, 500 mM Tris-HCl, pH 8.5). This was followed by heating at 95 °C for 45 min at 1,000 r.p.m. in a thermomixer. The samples were then subjected to ultrasonication using a Bioruptor Pico sonication device operated at high frequency for 30 s on and 30 s off for 30 cycles. After ultrasonication, the samples were again heated at 95 °C for 45 min at 1,000 r.p.m. in a thermomixer. This was followed by protein precipitation in ice-cold acetone (80% vol/vol) overnight at −80 °C, followed by centrifugation for 15 min at 4 °C. For reduction and alkylation, the proteins were resuspended in SDC buffer and heated at 95 °C for 10 min at 1,000 r.p.m. Trypsin and LysC digestion were carried out at an enzyme/substrate ratio of 1:50, and the samples were incubated at 37 °C overnight at 1,000 r.p.m. in a thermomixer. Next, peptides were acidified using 1% trifluoroacetic acid (TFA) in 99% isopropanol in a 1:1 vol/vol ratio. The peptides were subjected to in-house-built StageTips consisting of two layers of styrene-divinylbenzene reversed-phase sulfonate (SDB-RPS; 3M Empore) membranes. Peptides were loaded on the activated (100% acetonitrile (ACN), 1% TFA in 30% methanol, 0.2% TFA, respectively) StageTips, run through the SDB-RPS membranes and washed by ethyl acetate (EtOAc) including 1% TFA, isopropanol including 1% TFA and 0.2% TFA, respectively. Peptide elution was carried out in 60 μl of 1.25% ammonia and 80% ACN (acetonitrile) and dried for 45 min at 45 °C in a SpeedVac (Eppendorf, Concentrator plus). The dried peptides were reconstituted in 8 μl of 2% ACN/0.1% TFA, and peptide concentration was estimated using Pierce Quantitative Colorimetric Peptide Assay.

LNP corona samples preparation

Nanoparticle samples were resuspended in 25 µl of 4% SDS-Tris-HCl buffer (comprising 4% SDS and 10 mM TCEP, pH 8.5) in PCR strips. The LNP samples were heated at 95 °C for 5 min, followed by sonication for 15 min at 4 °C. Subsequently, 25 µl of 2% SDC buffer (also containing 10 mM TCEP and 40 mM CAA) was added, and the mixture was heated again at 95 °C for 10 min. A 3-µl aliquot of a 1:1 mix of SpeedBeads was introduced, and proteins were precipitated using ethanol to a final concentration of 75% under stirring for 5 min. The beads were captured on a magnetic 96-well rack and washed twice with 100 µl of 80% ethanol. Residual ethanol was evaporated in a SpeedVac. The samples were then digested with 0.4 µg of Trypsin-LysC in 50 µl of 50 mM TEAB with 0.02% LMNG at 37 °C overnight.

Evotip PURE clean-up of LNP corona samples

The Evotip PURE protocol was modified to ensure highly reproducible and standardized offline C18 clean-up in a 96-well format. Initially, Evotip PURE tips were rinsed with 20 µl of Buffer B (comprising 80% ACN, water and 0.1% formic acid) and spun down at 800g for 60 s. The Evotips were conditioned with 10 µl of isopropanol, followed by an impulse spin at 100g, a 1-min incubation and an additional 4 min at 100g to empty the Evotips. The PURE Evotips were equilibrated in 20 µl of Buffer A (0.1% formic acid) and impulse spun at 800g for storage until the acidified samples were ready to load. Samples were acidified in 0.4% TFA, and the Evotip PURE was emptied by centrifuging at 800g for 1 min. The acidified samples were loaded onto the PURE Evotips and spun at 800g for 1 min. The samples were washed twice with 20 µl of Buffer A and spun down at 800g for 1 min. Elutions were collected in PCR strips by eluting with 20 µl of 45% Buffer B (containing 45% ACN, water and 0.1% TFA) at 450g. The peptides were dried in a SpeedVac and resuspended in 0.1% TFA supplemented with 0.015% DDM for mass spectrometry analyses.

Liquid chromatography with tandem mass spectrometry

The mass spectrometry data were generated in data-independent acquisition (DIA) modes. The liquid chromatography with tandem mass spectrometry (LC−MS/MS) analysis was carried out using an EASY nanoLC 1200 (Thermo Fisher Scientific) coupled with a trapped ion mobility spectrometry quadrupole time-of-flight single-cell proteomics mass spectrometer (timsTOF SCP; Bruker Daltonik) via a CaptiveSpray nano-electrospray ion source. Peptides (100 ng) were loaded onto a 15-cm Aurora Elite UHPLC column with CaptiveSpray insert (75-μm ID, 1.7-μm C18) at 60 °C and separated using a 60-min gradient (5–20% Buffer B in 30 min, 20–29% Buffer B in 9 min, 29–45% in 6 min, 45–95% in 5 min, wash with 95% Buffer B for 5 min, 95–5% Buffer B in 5 min) at a flow rate of 300 nl min⁻¹. Buffers A and B were water with 0.1 vol% formic acid and 80:20:0.1 vol% ACN:water:formic acid, respectively. Mass spectrometry data were acquired in single-shot library-free DIA mode, and the timsTOF SCP was operated in DIA/parallel accumulation serial fragmentation (PASEF) using the high-sensitivity detection-low sample amount mode. The ion accumulation and ramp time were set to 100 ms each to achieve nearly 100% duty cycle. The collision energy was ramped linearly as a function of the mobility from 59 eV at 1/K0 = 1.6 Vs cm⁻² to 20 eV at 1/K0 = 0.6 Vs cm⁻². The isolation windows were defined as 24 × 25 Th from m/z 400 to 1,000.

LNP corona samples were analyzed using modified chromatography. In brief, 50 ng of peptides per injection was loaded on a 5.5-cm High Throughput μPAC Neo HPLC Column (Thermo Fisher Scientific) and analyzed using an 80-min active gradient method at a flow rate of 250 nl min⁻¹.

Proteomics data processing

diaPASEF raw files were searched against the mouse UniProt database using DIA-NN⁹⁴. A peptide length range of seven amino acids was considered for the search, including N-terminal acetylation. Methionine oxidation was set as variable modifications and cysteine carbamidomethylation as fixed modification. Enzyme specificity was set to Trypsin/P with two missed cleavages. The FASTA digest for the library-free search was enabled to predict library generation. The false discovery rate was set to 1% at precursor and global protein levels. The match-between-runs (MBR) feature was enabled, and the quantification mode was set to ‘Robust LC (high precision)’. The Protein Group column in DIA-NN’s report was used to identify the protein group, and PG.MaxLFQ was used to calculate the differential expression.

Proteomics downstream data analysis

Proteomics analysis of MC3 in the heart tissue

1. 1.

   Proteomics data analysis

   Scanpy (version 1.10.2) and anndata (version 0.10.8) packages in Python 3.10 were used to implement the downstream analysis pipeline to analyze the LC–MS samples. Nine independent samples were analyzed from each group of three animals, except for PBS, where we had eight independent samples.

2. 2.

   Quality control

   All proteins expressed in less than half of the samples in each group were filtered out.

3. 3.

   Data processing

   The data were log transformed and normalized per sample. The missing values were input using KNNImputer (n_neighbors = 5) from the sklearn package (version 1.5.1). To correct the batch effect across the different runs, the data further underwent batch correction using ComBat from scib tools (https://github.com/theislab/scib) (version 1.1.3).

4. 4.

   Dendrograms

   With Scanpy’s dendrogram function, hierarchical linkage clustering was calculated on a Pearson correlation matrix over groups for 50 averaged principal components.

5. 5.

   Differential expression tests

To identify differentially regulated proteins across two groups (for example, spike LNP versus no-cargo LNP), Scanpy’s method that ranks gene groups using t-tests was used. The maximal P value and minimal log fold change were used to identify DEPs. The thresholds are P < 0.05 and |log fold change| > 0.5. These DEPs were further used to plot volcano plots. Using the Reactome database and GProfiler, we performed enrichment analysis on genes with DEPs. Visualization was focused on the broadest categories of biological pathways.

The vascular-related genes were manually curated using related pathways from a publicly available database (Reactome). The vascular-related gene sets were tested individually using a normalized mean expression score for the different groups.

Proteomics analysis of SM-102 and Lung SORT formulations in the heart tissue

1. 1.

   Proteomics data analysis

   Mouse samples were analyzed using Scanpy (version 1.10.2) and anndata (version 0.10.8) in Python 3.10. The study included two different formulations: SM-102 and Lung SORT. Samples were categorized into three groups: no-cargo LNP, EGFP and spike protein. Each group comprised samples from three animals, resulting in nine samples per group for each formulation. Additionally, nine control samples consisting of PBS were analyzed from three animals.

2. 2.

   Quality control

   All proteins expressed in less than half of the samples in each group were filtered out, resulting in 2,191 proteins used for downstream analyses.

3. 3.

   Data processing

   The data were log transformed and normalized per sample. The missing values were input using KNNImputer (n_neighbors = 5) from the sklearn package (version 1.5.1).

4. 4.

   Differential expression tests

To identify differentially regulated proteins across two groups (for example, SM-102 spike versus PBS), Scanpy’s method that ranks gene groups using t-tests was used. The maximal P value and minimal log fold change were used to identify DEPs. The thresholds are P < 0.05 and |log fold change| > 0.5. These DEPs were further used to plot volcano plots.

The normalized intensity of selected DEPs was plotted for each formulation and group. Additionally, a Student’s t-test was conducted to assess the statistical significance of differences compared to the control group (PBS).

Protein corona of LNP formulations

1. 1.

   Proteomics data analysis

   Mouse samples were analyzed using Scanpy (version 1.10.2) and anndata (version 0.10.8) in Python 3.10. The study included plasma from four different LNP formulations: MC3, SM-102, BioNTech and Lung SORT. Each group comprised samples from three animals. Additionally, control samples consisting of PBS were analyzed from three animals.

2. 2.

   Quality control

   All proteins expressed in less than half of the samples in each group were filtered out.

3. 3.

   Data processing

   The data were log transformed and normalized per sample. The missing values were input using KNNImputer (n_neighbors = 5) from the sklearn package (version 1.5.1).

4. 4.

   Differential expression tests

To identify differentially regulated proteins across two groups (for example, MC3 versus PBS), Scanpy’s method that ranks gene groups using t-tests was used. The maximal P value and minimal log fold change were used to identify DEPs. The thresholds are P < 0.05 and |log fold change| > 0.5. These DEPs were further used to plot volcano plots. The DEPs of each of the formulations with comparison to the control PBS were determined. To investigate the potential protein corona, we identified common differentially binding proteins through pairwise comparisons between each formulation and the PBS control. Subsequently, the normalized intensities of common differentially binding proteins were plotted, and a Student’s t-test was performed to determine the statistical significance of differences relative to the control group (PBS). Potential binders for these proteins were determined using the StringDB Python package and manual curation from UniProt.

Statistics and reproducibility

All experiments were independently repeated at least three times with similar results. Relevant statistical tests are described in the figure legends or in the relevant sections above. Representative data, including micrographs, are presented where applicable.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.