Animals
Male Uchl1-eGFP and Cd68-eGFP mice (aged 8 weeks) on the C57BL/6J background or wild-type C57BL/6J mice were fed either a chow diet or a high-fat diet (60% fat, D12492i from Research Diets) for 16–18 weeks ad libitum. Mice were maintained on a 12 h–12 h light–dark cycle. The set points in the animal room were adjusted to 20–24 °C temperature and 45–65% humidity. Body composition was determined using an EchoMRI-100H system (EchoMRI). For insulin-tolerance tests, mice were fasted for 6 h and intraperitoneally (i.p.) injected with 0.75 U kg−1 insulin. Blood glucose was measured from the tail vein at the indicated timepoints using glucose test stripes. Mice were euthanized after deep anaesthesia with a mix of ketamine and xylazine, followed by intracardiac perfusion with heparinized PBS (10 U ml−1 heparin) and by a perfusion with 4% paraformaldehyde (PFA). Mice were post-fixed overnight in 4% PFA and subsequently washed five times with PBS shaking (300 rpm) at room temperature for 1 h for each wash step. Animal experimentation was performed in accordance with the European Union directives and the German animal welfare act (Tierschutzgesetz). They have been approved by the state ethics committee and the government of Upper Bavaria (ROB-55.2-2532.Vet_02-21-133, ROB-55.2-2532.Vet_02-16-117, ROB-55.2-2532.Vet_02-17-49, ROB-55.2-2532.Vet_02-19-166).
Human participants
Trigeminal ganglion samples were dissected post-mortem from body donors at the Institute of Anatomy, University of Leipzig, Germany and fixed in 4% Histofix. Body donors gave their informed and written consent to explore the cadavers for research and educational purposes (ethical approval number 129/21-ck, Medizinische Fakultät Ethik-Kommission). The participants were divided into lean (BMI < 25) or obese (BMI > 30). Data on age and sex can be found in Supplementary Table 11. We dissected three regions of interest from each trigeminal ganglion per individual for proteomic profiling.
Whisker stimulation test
The whisker test paradigm was adapted from the methods described previously40,41,42,43 and the Neuroscore test44. To avoid introducing confounding variables, mice were kept in their original cages. A cotton swab with a wooden end was used to administer the test. Initially, the cotton swab was presented in front of the mouse’s head and allowed to touch it. This was followed by four consecutive strokes, first to the whiskers on the right side and then on the left side of the face. The response to the cotton swab stimulation was evaluated using a modified whisker score test. A normal behavioural response to the stimulation, such as turning the head towards or away from the cotton swab or initiating grooming, was assigned a score of one. A lack of response to the stimulation was assigned a score of zero. Both sides of the face were stimulated four times, and the scores were recorded by a blinded evaluator. The maximum whisker score was 8, in which mice would have responded to all stimuli. The total score was then averaged for both sides. High scores (3–4) indicated normal responses to the stimulation, while low scores (0–2) suggested a lack of reaction, consistent with sensory deficits.
vDISCO nanobody labelling and clearing
vDISCO was performed as previously described2,45 in combination with active pumping GFP-Nanobooster labelling (Atto647N-conjugated anti-GFP nanobooster Chromotek, gba647n-100) for 6 days and passive labelling for 3 days. This approach amplifies the endogenous eGFP signal in reporter mice and shifts it into the far-red spectrum, substantially improving signal-to-noise ratios throughout the tissue. Mice underwent DISCO clearing46 using a tetrahydrofuran (THF)/H2O series (50% THF, 70% THF twice, 90% THF, 100% THF) for 24 h per step followed by an incubation in dichloromethane for 6 h. Tissues were incubated in benzyl alcohol/benzyl benzoate (BABB, 1:2 (v/v)) until tissue transparency was reached (>48 h).
WildDISCO antibody labelling and clearing
WildDISCO antibody labelling was performed as previously described in combination with anti-UCHL1 (14730-1-AP1, Proteintech, 26 µl per 200 ml immunostaining buffer) and anti-CGRP (ab36001, Abcam, 26 µl per 200 ml immunostaining buffer)28. Mice underwent DISCO clearing as described above.
Fluorescence light-sheet imaging
Light-sheet imaging for whole-mouse bodies was conducted using a dipping ×1.1 objective lens (Miltenyi BioTec) on an Ultramicroscope Blaze (Miltenyi BioTec) using the ImspectorPro (v.5.1) software. Tiling scans (×1) were acquired using two-sided illumination with 35% overlap, 100% sheet-width, 0.1 NA, 100 ms exposure and a 6 µm z-step size. The images were taken in 16 bit depth and at a nominal resolution of 5.9 μm per voxel on the xy axes. Stitching of tile scans was carried out using Fiji’s stitching plugin with the ‘Stitch Sequence of Grids of Images’ feature47 and custom Python scripts. Imaging of mouse bodies at higher resolution was conducted using a ×4 objective lens (Miltenyi BioTec) on the same system as described above but tiling scans were acquired with the LightSpeed Mode using a 20% overlap, 80% sheet-width, 0.35 NA and 5 ms exposure time and a 6 µm z-step size. The images were taken in 16 bit depth and at a nominal resolution of 1.62 μm per voxel on the xy axes.
3D reconstruction
Dorsal and ventral scans were fused as previously described2 using Arivis (v.3.0.1 and v.3.4) and the exported whole-body TIFF stacks were used for image analysis.
VR data annotation
Annotation of ground-truth data was performed in VR7 using the syGlass software (v.2.0.0) as previously described. To develop a robust and generalizable nerve segmentation model, a large and diverse dataset was curated from Uchl1-eGFP mouse scans imaged with the ×1.1 objective and annotated in VR. In total, the dataset comprised 1,217 patches (300 × 300 × 300 voxels) derived from 84 small subvolumes (300 × 300 × 300 voxels) and 8 larger subvolumes (~1,000 × 1,000 × 1,000 voxels). All large subvolumes were uniformly cropped into patches of 300 × 300 × 300 voxels to standardize the dataset. The training set incorporated 28 patches from Uchl1-eGFP volumes covering a range of anatomical contexts, 537 patches derived from 5 larger subvolumes of trigeminal nerves, and 118 patches from 1 larger subvolume of vertebral nerves. Together, these samples capture broad variations in nerve morphology and topological organization across the mouse body. To further enhance discriminative performance, particularly in regions susceptible to false-positive predictions, 29 negative sample patches containing structures such as adipocytes were included. For model evaluation, the testing set consisting of 7 patches from different parts of the mouse body, 478 trigeminal nerve patches cropped from 2 larger subvolumes of trigeminal nerves, 6 patches containing vertebral nerves and 14 negative patches. This design ensured thorough assessment of both segmentation accuracy and model generalizability across anatomical scales and tissue environments.
VR-annotation for Cd68-eGFP+ cells was performed in five 256 × 256 × 256 voxel patches from Cd68-eGFP whole-mouse scans, selected from representative regions of interest. Annotations were based on both the autofluorescence and Cd68-eGFP signal channels. These patches were further cropped down into 40 128 × 128 × 128 voxel patches that were used to train 3D networks for the segmentation of the markers of interest. Moreover, five 128 × 128 × 128 voxel patches were annotated as an independent test set used for evaluation.
For the development of the Tissue-Module, we annotated 27 organs of interest (Supplementary Table 7) in 12 downsampled (tenfold) mouse scans (6 from Cd68-eGFP and 6 from Uchl1-eGFP mice, with 6 chow-fed and 6 HFD-fed mice in total) using the autofluorescence and PI channels with the syGlass software. This approach was sufficient to distinguish all organs of interest. To generate reference annotations for the tissue segmentation, we annotated an initial dataset of three 1,024 × 1,024 × 1,024-voxel-sized patches in full resolution, containing 500 million voxels of fat (visceral, subcutaneous and brown), 145 million voxels of muscle, 16 million voxels of bone tissue and 8 million voxels of bone marrow. We iteratively increase the size of our annotated dataset through inference on unannotated patches, and manual correction of the wrongly segmented areas.
Peripheral nerve segmentation
We developed the Nerve-Module of MouseMapper for nerve segmentation by fine-tuning a pretrained foundation model, VesselFM14, using our curated dataset (described above). VesselFM was pretrained on a large-scale 3D vessel dataset and was originally designed for the broad task of 3D blood vessel segmentation. To adapt VesselFM to our nerve dataset, we used an incremental learning strategy, learning without forgetting (LwF)48. Using learning without forgetting, the model was fine-tuned on nerve-specific data while regularizing its outputs to general vessel-related structural knowledge represented in the pretrained weights, thereby reducing the risk of catastrophic forgetting49. This approach allows the model to efficiently leverage prior knowledge while ensuring stable convergence on nerve data.
The fine-tuning process was implemented using a patch size of 128 × 128 × 128, an initial learning rate of 1 × 10−3 with scheduled decay, the stochastic gradient descent (SGD) optimizer and a segmentation loss combining Cross Entropy loss and Dice loss50. Incorporating learning without forgetting, each training batch obtains two sets of outputs: predictions for the nerve segmentation task from the fine-tuning model and ‘soft targets’ from the fixed pretrained VesselFM model representing the original vessel segmentation task. The final loss is computed as the sum of the nerve segmentation loss and a distillation loss, Kullback–Leibler divergence, that penalizes deviations from the pretrained model’s outputs. A weighting factor balances the segmentation and distillation losses, controlling the trade-off between retaining prior knowledge and learning nerve-specific features. In our experiments, optimal performance was achieved with a weighting factor of 0.4 for the distillation loss. The model was trained for 1,250 epochs.
Before forwarding the patches into the network for training or testing, we performed sample-wise normalization. Specifically, during the training, for each group of patches, including patches from whole-body, trigeminal nerve, vertebral nerve and negative samples, we computed the 0.5th percentile and 99.5th percentile of all voxel intensity values to set the minimum and maximum thresholds. Intensity values below or above these thresholds were clipped accordingly, followed by min–max normalization. During the testing, the same normalization procedure was applied to the entire testing dataset. This normalization step enhanced image contrast by stretching the intensity range between the chosen percentiles and removing outliers, thereby emphasizing nerve regions to improve model performance.
We compared our nerve segmentation model with other advanced 3D image segmentation networks50 (Supplementary Table 1): VNet51, Attention U-Net52, nnFormer53, UNETR54, SwinUNETR55, nnU-Net56 and nnUNetRes56. Each network was trained on the same training dataset until full convergence, defined as no decrease in the training loss for ten consecutive epochs. Moreover, the original VesselFM was also included in the comparison, to demonstrate the impact of finetuning. To assess the generalizability of MouseMapper across different labelling strategies and species, we applied it to diverse external datasets, including Thy1-eGFP vDISCO-labelled mice, wildDISCO antibody-labelled samples (tyrosine hydroxylase, UCHL1 and CGRP) and a publicly available post-conception-week 7 human embryo stained for β3-tubulin (https://hudeca.com). The human dataset is licensed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).
Immune cell segmentation
For training the CD68 segmentation network (Immune-Module), we also fine-tune the VesselFM foundation model by freezing the encoder and finetuning the decoder. Thus, we can leverage the vast training of the foundation model by keeping its learned filters, but adapt the segmentation output through learning on our annotated dataset. As comparison baselines, we implemented the following architectures: 3D UNet57, V-Net51, Attention U-Net52, nnFormer53 and UNETR54. All were trained by using the nnU-Net pipeline56, with a patch size of 128 × 128 × 128 voxels, channel-wise z-score normalization, learning rate decay and SGD optimizer. The baselines were trained until convergence for 1,000 epochs and initial learning rates of [0.0001, 0.001, 0.01], whereas the best performance for VesselFM was achieved by finetuning the network’s decoder for 500 epochs with an initial learning rate of 0.01. We train using fivefold cross validation, and evaluate voxel Dice, instance Dice18 and report best score per architecture. Based on these metrics, we selected the finetuned VesselFM for carrying out our downstream quantifications.
Organ and tissue segmentation
For the segmentation of internal organs, we used eight annotated mice (from the Cd68-eGFP and Uchl1-eGFP line) to train five different networks: 3D UNet57, V-Net51, Attention U-Net52, nnFormer53 and Swin UNETR55. All of the architectures were trained through the nnU-Net pipeline56 using z-score normalization of each channel, and foreground oversampling. The networks were trained with SGD optimizer, using a batch size of 2, patch size of 64 × 256 × 128 voxels, initial learning rates of [1 × 10−4, 1 × 10−3, 1 × 10−2] and learning rate decay, for a total of 1,000 epochs. The resulting networks were evaluated on two Cd68–eGFP and two Uchl1-eGFP mouse reconstructions. During training, we performed fivefold cross validation, and the final predictions were made by ensembling the five resulting networks. We report best voxel Dice scores in Supplementary Table 7 for each architecture. We identified the 3D UNet as the best performing network layout, with the following properties: 6 downsampling layers, 6 upsampling layers, 3 × 3 × 3 sized convolutional blocks and a maximum feature size of 320 in the bottleneck, trained with the initial learning rate of 0.01.
Second, we train a model to segment the soft tissues of mice, such as muscle and adipose tissue. We iteratively increased the size of our annotated dataset through inference on unannotated patches, and manual correction of the wrongly segmented areas. As a result, our final networks were trained on a dataset of 387 samples containing a total volume of 2 billion voxels of adipose tissue and 2 billion voxels of muscle. We then train on these patches the following neural network architectures: 3D UNet, V-Net, Attention UNet and UNETR. We trained using fivefold cross-validation and, for evaluation, we report and select based on the validation scores of the ensembles of the five resulting networks (Supplementary Table 8). The networks were trained with SGD optimizer, using a batch size of 2, patch size of 128 × 128 × 128 voxels, initial learning rate of [1 × 10−3, 1 × 10−2] and learning rate decay, for a total of 1,000 epochs. The convolutional 3D UNet performs best among the implemented baselines.
The final inference pipeline of the Tissue-Module first segments the organs and then the tissues. First, the autofluorescence and PI channels of the acquired LSFM stack are downsampled to a resolution of 59 × 59 × 60 μm per voxel and saved as a 3D NIfTI volume. This is then fed into the organ segmentation network. The result is a 3D volume containing the masks of the 27 organs of interest, which can be used downstream for localizing structures of interest within organs, or for the quantification of organ volumes. Next, the organ masks are upsampled, and a non-organ mask is calculated, which is applied to the original scan. Through this process, we obtain a mask of the mouse volume that does not contain internal organs. This can be applied to the full-resolution scans, on which sliding-window inference can be carried out with the tissue segmentation model, as described below, to obtain the tissue map. Lastly, by combining the organ maps and the tissue maps, we obtain a spatial segmentation of major organs and tissues in the mouse body.
Whole-body inference
To apply the nerve, immune and tissue segmentation network to whole-body scans in full resolution efficiently, we adapted the sliding window inference method previously used for segmentation tasks in medical images (MONAI)58 and the mouse brain (DELIVR)7. Our inference is implemented using the highly efficient ZARR file format and DASK parallel computing framework, enabling lazy loading and multiprocessing for data handling and writing tasks and, therefore, a rapid full-body analysis.
Before inference, we applied percentile normalization to each scan, similar to the model training stage. Given the significant imbalance between nerve/CD68+ voxels and background voxels in whole-body scans, we computed the 0.10th percentile and 99.9th percentile of all non-zero voxel intensity values to set the minimum and maximum thresholds, to effectively enhance the contrast between nerves and the background. During inference, we use the same patch normalization protocol as during network training, and patch size is selected to fit the memory resources available.
Cd68-eGFP segmentation quantification
The binary masks obtained after CD68 marker segmentation were split into components by using the cc3d library59 for connected component analysis on subregions of the full-resolution scans. Subsequently, each individual detected connected component was post-processed by storing its location, volume, centre of mass, and shape27. Based on the location of the centre of mass, we automatically assign each segmented Cd68-eGFP+ cluster to the internal organs or the segmented tissues, with blobs not located in any of these being discarded as false positives. We further discard components of which the shape was elongated (string-like) as false positives, as these can often be artifacts, representing high-contrast blood vessels or nerves. Lastly, we grouped the detected Cd68-eGFP clusters into three discrete categories, based on their volume (amount of segmented voxels within a component): small (smaller than 50 voxels), medium (between 50 and 500 voxels) and large (over 500 voxels). We chose these categories based on the observation that, when considering the total spatial volume of all clusters, each of these three categories would represent a similar amount (approximately 30%) of the total CD68+ segmented volume. Then, for each mouse and for each organ or tissue, we studied the percentage composition of each of these categories, and analysed the differences between the chow and HFD groups.
While applying the CD68 segmentation network to the whole-mouse bodies, we observed that it displays zero-shot transfer learning abilities in the limited setting of applying the model in inference to certain novel tissues, where we observed positive detections. Thus, to validate any reported changes, we performed (1) visual analysis of the resulting segmentation, and (2) a VR-based annotation of a representative test patch in the tissue of interest. We compared the result of the automatic segmentation against the manual annotation in order to evaluate the network’s transfer learning abilities. We only considered valid quantifications where the network passed with a Dice score >60%.
Uchl1-eGFP segmentation quantification
After inference, we obtained the whole-body nerve segmentation of Uchl1-eGFP mice. We then performed connected component analysis to post-process the segmentation results, eliminating large false-positive segments caused by high-intensity regions within the mouse body. Subsequently, we quantified the nerve voxels and density from three perspectives: the entire body, individual tissues and specific organs.
To quantify nerves in the entire body, the organ and tissue segmentations from the Tissue-Module were combined to form a binary mask of major organs and tissues in the mouse body. By dilating this binary mask, we created a whole-body mask that covers the entire mouse body, enabling us to compute the nerve voxels and density within. For tissue-wise quantification, the tissue segmentation from the Tissue-Module was used to calculate the nerve voxels and density in fat and muscle tissues. To further distinguish adipose compartments, we manually separated the visceral and subcutaneous fat within the tissue mask by referencing the abdominal wall mask derived from organ segmentation network. For quantifying nerves for specific organs, we accounted for structures in the immediate vicinity of the organs by extending the organ segmentation by 500 µm to calculate the organ wise statistics. Notably, to create the head mask, we overlaid the dilated brain masked with whole-body mask followed by minor manual refinement, resulting in a precise mask for quantifying the nerve voxels inside. For the 4× scan analysis, we included the limbs as regions of interest. Using the whole-body mask and referencing the head and heart masks as anatomical landmarks, we defined an initial limb region, which is located lateral to the head and superior to the top of the heart. This preliminary region was then manually refined to generate the final limb mask used for nerve voxel and density analysis.
Graph extraction
Graph extraction was performed as previously described21,60. Similarly we extracted the skeletonization, depth map and extracted a graph of the resulting skeleton. All small, isolated subgraphs were pruned from the graph. As the resulting image data were too large to fit into a reasonable amount of RAM, we separated the whole image into sub-blocks using nibabel. We next extracted the graphs from each sub-block and merged them together. We fused all nodes together on the border between two blocks where the Euclidian distance between nodes was less than a given threshold by introducing a new edge between the nodes. We quantify the thickness of each node and each edge using the depth map, the degree of each node and the number of leaf nodes (nodes with degree = 1).
Computational load of MouseMapper
The experiments presented in this work were carried out using a cohort of 19 mice (10 HFD fed, 9 chow fed). The nine 1× Uchl1-eGFP whole-body scans (4 chow and 5 HFD) generated a total of 105,948 2D z-slices and 10.926 trillion voxels, occupying 9.42 TB after ZARR compression. Moreover, nine 4× UCHL1 ventral scans from the same mice, together with two 4× dorsal scans from two of them (1 chow and 1 HFD) produced another 64,968 2D z-slices and 54.888 trillion voxels, occupying 56.1 TB after compression. The ten 1× Cd68-eGFP whole-body scans of 5 chow-fed and 5 HFD-fed mice generated a total of 112,515 z-slices and 10.779 trillion voxels, occupying 7.48 TB after compression. Two additional 4× Cd68-eGFP ventral scans from two mice (1 chow fed and 1 HFD fed) produced another 11,706 2D z-slices and 9.726 trillion voxels, occupying 9.8 TB after compression. To accurately quantify these data, our annotation efforts resulted in significantly ample datasets. For the Nerve-Module, we manually annotated 72 GB of data. While building the Immune-Module, we annotated 350 MB of data from representative areas in visceral and subcutaneous fat, as well as in the peritoneum. The organ segmentation network of the Tissue Module was trained using 10 GB of downsampled organ data, whereas the tissue segmentation network (for fat, muscle, bone and bone marrow) was trained using 46 GB of full-resolution tissue annotations, built as a mixture of manual and automatic annotations. To train the networks building our MouseMapper pipeline, as well as to run the predictions and quantifications presented in this paper, the High-Performance Computing cluster of Helmholtz Zentrum Munich was used. Thus, the processes could be parallelized and carried out more efficiently.
We estimate approximately 500 GPU hours for multiple model training and evaluation, 265 GPU hours for 1× segmentations of scans (nerves, CD68+ cells, tissues), about 550 GPU hours per 4× scan and 0.1 GPU hours for organ inference per scan. For nerve voxel and density calculations, we estimate roughly 330 CPU hours for all 1× scans and about 60 CPU hours per 4× scan. CD68+ blob post-processing used approximately 100 CPU-hours for all 1× scans. Graph extraction and postprocessing is performed solely on CPU. We estimate 216 CPU hours for 1× graph extraction.
Spatial proteomics sample preparation
For spatial proteomics of trigeminal ganglia of Uchl1-eGFP mice, 18G needle punches were prepared from rehydrated trigeminal ganglia and subsequently used for proteomics sample preparations as described previously11. In brief, the samples were resuspended in 6% SDS buffer, heat denatured at 95 °C for 45 min at 600 rpm in a thermoshaker, sonicated in high mode for 30 cycles (30 s off, 30 s on) (Bioruptor Plus, Diagenode) and then precipitated using 80% acetone overnight at −20 °C. The next day, these samples were centrifuged and the pellet was resuspended in SDC lysis buffer (2% SDC, 100 mM Tris-HCl pH 8.5). The samples in the SDC buffer were sonicated in high mode for 15 cycles (30 s off, 30 s on) (Bioruptor Plus, Diagenode). The samples were again heated to 95 °C at 600 rpm in a thermoshaker for 45 min. The protein samples were digested overnight with trypsin and LysC (1:50, protease:protein ratio) at 37 °C with 1,000 rpm shaking. The resulting peptides were acidified with 1% trifluoroacetic acid (TFA)/99% isopropanol at a 1:1 volume-to-volume ratio, vortexed and centrifuged to pellet residual particles. The supernatant was transferred to fresh tubes and subjected to an in-house built StageTip clean-up consisting of three layers of styrene divinylbenzene reversed-phase sulfonate (SDB-RPS; 3 M Empore) membranes. Peptides were loaded onto the activated (100% acetonitrile, 1% TFA in 30% methanol and 0.2% TFA, respectively) StageTips, run through the SDB-RPS membranes and washed by ethyl acetate including 1% TFA, isopropanol including 1% TFA and 0.2% TFA, respectively. Peptides were then eluted from the membranes through 60 µl elution buffer (80% acetonitrile, 1.25% NH4OH) and dried using a vacuum centrifuge (40 min at 45 °C). Finally, peptides were reconstituted in 8–10 µl of loading buffer (2% acetonitrile, 0.1% TFA) and stored at −80 °C until further use.
For proteomics profiling of human trigeminal ganglia, samples were reduced and denatured in lysis buffer (2% SDC, 10 mM TCEP and 100 mM Tris-HCl pH 8.5, 40 mM chloroacetamide) at 95 °C for 45 min in a PCR thermocyler, sonicated in high mode for 30 cycles (30 s off, 30 s on; Bioruptor Plus; Diagenode) and heated again to 95 °C for 45 min. Contaminants and detergents were removed using SP3-based precipitation and washing on 5 µl magnetic beads61. In brief, 100 µl ethanol was used for precipitation, 50 µl ethanol was used for washing and proteins were then dried in the SpeedVac before adding trypsin and LysC proteases. A second overnight digestion step was added to improve digestion efficiency as previously described11.
Evotip PURE clean-up of human samples
A total of 1 µg of each tissue digest was desalted per Evotip. The Evotip PURE protocol was adjusted for offline C18 clean-up in a 96-well format as described previously10. Initially, Evotip PURE tips were rinsed with 20 µl of buffer B (comprising 80% acetonitrile, 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 a 1 min centrifugation at 100g and additional 1 min at 400g to empty the Evotips. The PURE Evotips were washed and equilibrated in 200 µl of buffer A (0.1% formic acid). The samples were acidified in 5% TFA, and the Evotip PURE was emptied by centrifuging at 800g for 1 min. The acidified samples were loaded onto the PURE Evotips and centrifuged at 800g for 1 min. The samples were washed with 200 µl of buffer A and spun down at 800g for 1 min. Elutions were collected in PCR strips by eluting with 20 µl of buffer B by centrifuging at 100g for 1 min followed by 450g 1 min. The peptides were dried in a SpeedVac and resuspended in 40 μl of 0.1% TFA supplemented with 0.015% DDM for MS analyses. Up to 2 μl (or 50 ng peptides) was injected per MS analysis.
LC–MS
The MS data for mouse samples was acquired in data-independent acquisition (DIA) mode. The liquid chromatography–tandem mass spectrometry (LC–MS/MS) analysis was carried out using the EASY nanoLC 1200 (Thermo Fisher Scientific) system coupled to a trapped ion mobility spectrometry quadrupole time-of-flight single-cell proteomics mass spectrometer (timsTOF SCP, Bruker Daltonik) through a CaptiveSpray nano-electrospray ion source. Peptides (50 ng) were loaded onto a 25 cm Aurora Series UHPLC column with CaptiveSpray insert (75 μm inner diameter, 1.6 μm C18) at 50 °C and separated using a 50 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% acetonitrile:water:formic acid, respectively. MS 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. The MS data for human samples were acquired using similar as previously described using a 5.5 cm long mPAC HT column to minimize carryover and to accelerate column cleaning and maintenance between different sample types10.
Proteomics data processing
For mouse data, diaPASEF raw files were searched against the mouse uniport database using DIA-NN62. Peptides length range from seven amino acids were considered for the search including N-terminal acetylation. Oxidation of methionine was set as a variable modification and cysteine carbamidomethylation as fixed modification. Enzyme specificity was set to trypsin/P with 2 missed cleavages. The FASTA digest for library-free search was enabled for predicting the library generation. The FDR was set to 1% at precursor and global protein level. The Match between runs feature was enabled and 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. For human data, analogous searches using DIANN v.2.0 were performed.
Proteomics data analysis
Mouse data were analysed using scanpy (v.1.10.1) and anndata (v.0.8.0) in Python v.3.10. In total, 12 independent samples were analysed from each group (high-fat diet and chow) from three animals with samples from both right and left trigeminal ganglia. All proteins expressed in less than half of the samples in each group were filtered out, resulting in 6,686 proteins used for downstream analyses. The data were log-transformed and normalized per sample. The missing values were input using KNNImputer (n_neighbors=5) from the sklearn package (v.1.2.1). Using scanpy’s dendrogram function, scipy’s hierarchical linkage clustering was calculated on a Pearson correlation matrix over groups which was calculated for 50 averaged principal components. To identify differentially regulated proteins across two groups (HFD and chow), we combined samples from the right and the left trigeminal ganglia. Differential expression analysis was conducted using Scanpy’s method ‘rank_genes_groups’ with method set to ‘t-test’. We applied a threshold of P < 0.05 and |log[fold change]| > 0.5 to identify differentially expressed proteins (Supplementary Table 13). These differentially expressed proteins were subsequently visualized using volcano plots. Pathway enrichment analysis was performed on the combined upregulated and downregulated proteins using the KEGG and Reactome databases. The most relevant pathways were highlighted, displaying the differentially expressed proteins involved in each pathway. For human data, the significance of differences in protein abundances between the obese and lean groups was determined using Excel (v.2016) with a two-tailed Student’s t-test, applied to proteins that were identified in both groups at least three times. A FC threshold was set to proteins that were identified with significant changes: a decrease in abundance (log2[FC] < −0.5) or an increase in abundance (log2[FC] > 0.5), with a significance cut-off of P < 0.05. Significantly altered protein groups were subsequently searched in the KEGG and Reactome databases using the DAVID annotation tool (v.Dec. 2021, Knowledgebase v.2023q4). Pathway analyses were performed by filtering for significant pathways (P < 0.05; Supplementary Table 12). A subset of pathways displayed in Fig. 4h were manually selected for visualization using Python (v.3.8).
Western blot
Protein lysates from trigeminal ganglia were prepared by homogenizing frozen tissue in RIPA buffer with freshly added inhibitors (1× EDTA-free protease inhibitor and 1× PhosSTOP) using the Tissuelyzer II (Qiagen). The samples were centrifuged at 13,000g and 4 °C for 30 min. The protein content of cleared lysates was determined using the Pierce BCA Kit (Thermo Fisher Scientific, 23225). Protein lysates were incubated with 6× Laemmli buffer at 95 °C for 5 min before loading it onto an SDS–PAGE gel (Novex WedgeWell, Tris-Glycine Mini Gels; Thermo Fisher Scientific, or Mini-PROTEAN Precast Gels; Bio-Rad Laboratories). Gels were run at 100–120 V and subsequently transferred to a nitrocellulose membrane (Bio-Rad). The membranes were blocked with 5% skimmed milk in TBS-T and incubated with primary antibodies against SEPTIN7 (Proteintech, 13818-1-AP), SERPINA1 (Proteintech, 16382-1-AP, p-ERK (Cell Signaling, phospho-p44/42 MAPK (Thr202/Tyr204), 9101) ERK (p44/42 MAPK, 9102) diluted 1:1,000 in 5% BSA V5. Vinculin (EPR8185, Abcam, ab129002) was diluted 1:10,000 in 5% BSA V5. Anti-rabbit IgG coupled to horseradish peroxidase were used at a dilution of 1:10,000 in 5% milk as secondary antibodies, and immunoreactive proteins were determined by chemiluminescence using the ChemiDoc MP System (Bio-Rad).
Multiplex antibody labelling and analysis
The MACSima Imaging Cyclic Staining technology from Miltenyi Biotec was performed according to the manufacturer’s protocol on paraffin-embedded tissue sections of epididymal white adipose tissue of obese mice63. Image acquisition was performed automatically by the MACSima instrument in seven regions of interest (ROIs). The following antibodies from Milteny Biotech were used at a dilution of 1:10: NK1.1 (REA1162), CD3 (REA641), F4/80 (REA126), MHC-II (REA813), CD31 (REAL260) and CD138 (REA104). Images were thresholded individually for each marker to optimize visualization and analysed using the spatiomic package64 (v.0.8.0). For each ROI a 30 × 30 self-organizing map (SOM) was trained on a subset of 1 million pixels. The SOM compresses the pixel data into a smaller set of representative prototypes, which were subsequently clustered using the Leiden algorithm65. The Leiden clusters were mapped to each pixel according to their SOM mapping. The clusters were annotated with cell types based on mean intensities in each cluster. The vicinity cluster compositions for each cluster were retrieved per ROI using spatiomic’s ‘vicinity_composition’ function. The ROIs were combined by summing the compositions of each ROI. The vicinity graph was retrieved from these combined compositions.
Statistical analysis
Results from biological replicates were expressed as mean ± s.e.m. Statistical analysis was performed using GraphPad Prism (v.9). To compare two conditions, unpaired Student’s t-tests or Mann–Whitney U-tests were performed. Insulin-tolerance tests were analysed using two-way ANOVA with Šídák’s multiple-comparison test. Proteomics data analysis was performed as described above. No statistical method was used to predetermine sample size. Mice were randomly assigned to chow or HFD groups. Unless stated otherwise, investigators were not blinded during experiments or data analysis.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
