Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 63 of the 63 references it cites were found in a registry or an index; 25 of the 56 sources its claims cite were checked for support.

A deep-learning framework reveals whole-body perturbations at cell level

Open access

Abstract

Many diseases, including obesity, have systemic effects that perturb multiple organ systems throughout the body 1,2 . However, tools for comprehensive, high-resolution analysis of disease-associated changes at the whole-body scale have been lacking. Here we developed MouseMapper, a suite of foundation-model-based deep-learning algorithms enabling multi-system analysis of disease across the entire mouse body. MouseMapper enables whole-body quantitative analysis of nerves and immune cells, resolving fine axonal branches and immune-cell clusters while automatically segmenting 31 organs and tissues. We used MouseMapper to study diet-induced obesity, and identified structural alterations of the infraorbital branch of the trigeminal ganglia. This structural impairment in infraorbital nerves was associated with functional sensory deficits in whisker sensing. Furthermore, we identified proteomic changes in the trigeminal ganglion affecting axon remodelling and complement pathways both in mice and humans. MouseMapper also generated detailed three-dimensional inflammation maps by characterizing immune cell cluster compositions across tissues. The MouseMapper framework demonstrates robust generalizability across different imaging resolutions and datasets. Our study provides a powerful, scalable approach for identifying and quantifying systemic pathologies, bridging molecular insights from animal models to human conditions.

Overview

CoChat Summary

TL;DR

The authors present MouseMapper, a foundation-model-based deep-learning framework that segments nerves, immune cells, and 31 organs/tissues across whole mouse bodies. Applying it to diet-induced obesity reveals reduced nerve density and infraorbital nerve structural changes linked to whisker sensory deficits, plus conserved trigeminal ganglion proteomic alterations in mice and humans.

Key findings

  • MouseMapper automatically segments 31 organs and tissues, far exceeding previous methods such as AIMOS (six organs), and its Nerve-Module achieved a voxel Dice score of 0.7494, with cross-dataset Dice scores of 0.6916–0.7143 across different labelling strategies and resolutions.
  • In high-fat-diet-fed obese mice, whole-body nerve density decreased, including in adipose tissue and limbs; infraorbital nerve endings, edges, and vertices were reduced by 60.7%, 57.8%, and 57.6%, respectively, while nerve thickness was similar, and obese mice showed diminished whisker-stimulation responses.
  • Trigeminal ganglion proteomics identified more than 6,000 proteins per sample, with 230 differentially regulated proteins (67 upregulated, 163 downregulated) in HFD-fed mice; pathways included actin cytoskeleton, RHO GTPase effectors, axon guidance, complement/coagulation, ERBB signalling, and sphingolipid signalling.
  • Human post-mortem trigeminal ganglia from lean (BMI < 25) versus obese (BMI > 30) individuals showed conserved proteomic changes in axon guidance, neurodegeneration, and actin cytoskeleton regulation; the Immune-Module achieved voxel Dice 0.7878, outperforming existing methods (0.2140–0.5468).

Methods

  • MouseMapper is an ensemble of three modules: Nerve-Module, Immune-Module, and Tissue-Module, built by fine-tuning the VesselFM 3D foundation model with supervised VR annotations and sliding-window inference on whole-body light-sheet fluorescence microscopy datasets.
  • Male Uchl1-eGFP and Cd68-eGFP mice were fed chow or a 60% fat high-fat diet for 16–18 weeks; whole bodies were cleared with vDISCO or WildDISCO and imaged with ×1.1 (NA 0.1) or ×4 (NA 0.35) objectives. Human trigeminal ganglia were collected post-mortem from lean and obese body donors.
  • Analyses included Dice-based segmentation benchmarking, connected-component and graph extraction for nerve morphology, Cd68+ cluster-size quantification, and mass-spectrometry-based spatial proteomics with pathway enrichment.

Limitations

  • The paper states that whole-body light-sheet resolution limits resolving the smallest parenchymal nerve fibres and subcellular structures; the standard ×1.1 configuration has lateral resolution around 5.9 µm and cannot fully resolve thin axons, while the ×4 configuration (1.62 µm lateral resolution) increases data volumes up to 50 TB per mouse.
  • The paper states that full generalizability across disparate imaging modalities is not yet attainable, and MouseMapper may require additional fine-tuning for optimal performance on new datasets; Uchl1-eGFP mice do not tag endogenous UCHL1, and antibody labelling was less homogeneous in obese mice.
  • The paper states that no statistical method was used to predetermine sample size, and unless stated otherwise investigators were not blinded during experiments or data analysis; the text also makes evident that mouse experiments used male mice only.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41586-026-10535-2

Cites

63 of 63 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Direct visualization of nerves and macrophages in obesity at the whole-body scale. a, Workflow to study obesity-induced changes at the whole-body scale. The diagram was created using BioRender; Kaltenecker, D. https://biorender.com/t34651c (2026). b,c, Representative 3D reconstructions of vDISCO-cleared and imaged chow-fed and HFD-fed mice, showing Uchl1-eGFP+ peripheral nerves (b) (n = 4 chow and n = 5 HFD-fed mice) and Cd68-eGFP+ immune cells (c) (n = 5 chow, n = 5 HFD-fed mice). Scale bars, 5,000 μm (b,c).
Fig. 1 Direct visualization of nerves and macrophages in obesity at the whole-body scale. a, Workflow to study obesity-induced changes at the whole-body scale. The diagram was created using BioRender; Kaltenecker, D. https://biorender.com/t34651c (2026). b,c, Representative 3D reconstructions of vDISCO-cleared and imaged chow-fed and HFD-fed mice, showing Uchl1-eGFP+ peripheral nerves (b) (n = 4 chow and n = 5 HFD-fed mice) and Cd68-eGFP+ immune cells (c) (n = 5 chow, n = 5 HFD-fed mice). Scale bars, 5,000 μm (b,c).CC BY · PubMed Central
Fig. 2 Development of an ensemble of deep-learning methods for automated segmentation of nerves, immune cells, organs and tissues. a, Mice that underwent vDISCO clearing, light-sheet imaging and 3D reconstruction were analysed using MouseMapper. MouseMapper comprises three modules: the Nerve-Module for deep-learning-based nerve segmentation, the Immune-Module for deep-learning-based immune cell detection and the Tissue-Module for automated organ and tissue segmentation. b, To train the Nerve-Module, nerves were annotated using VR. c, 3D qualitative evaluation of the network performance for the segmentation of nerves based on the volumetric Dice score. Areas that overlap with reference annotations (true positive, TP) are masked in green; areas with no overlap in reference annotations (false positive, FP) are masked in red. Undetected reference annotation areas (false negative, FN) are marked in blue. Scale bar, 100 μm. d, The Tissue-Module uses the raw image scan and segments organs (organ map) and tissues (tissue map), which can be combined to generate a whole-mouse map.
Fig. 2 Development of an ensemble of deep-learning methods for automated segmentation of nerves, immune cells, organs and tissues. a, Mice that underwent vDISCO clearing, light-sheet imaging and 3D reconstruction were analysed using MouseMapper. MouseMapper comprises three modules: the Nerve-Module for deep-learning-based nerve segmentation, the Immune-Module for deep-learning-based immune cell detection and the Tissue-Module for automated organ and tissue segmentation. b, To train the Nerve-Module, nerves were annotated using VR. c, 3D qualitative evaluation of the network performance for the segmentation of nerves based on the volumetric Dice score. Areas that overlap with reference annotations (true positive, TP) are masked in green; areas with no overlap in reference annotations (false positive, FP) are masked in red. Undetected reference annotation areas (false negative, FN) are marked in blue. Scale bar, 100 μm. d, The Tissue-Module uses the raw image scan and segments organs (organ map) and tissues (tissue map), which can be combined to generate a whole-mouse map.CC BY · PubMed Central
Fig. 3 Whole-body nerve segmentation reveals structural changes in the infraorbital nerve in obesity. a, Representative whole-body nerve segmentation in chow and HFD-fed mice. b, Quantification of AI-segmented nerve densities in chow- and HFD-fed Uchl1-eGFP mice in whole bodies and the indicated areas. For quantification of the nerve density in the head, we masked out the brain for the analysis (n = 4 chow, n = 5 HFD mice; whole-body: two-tailed t-test, P = 0.0012; fat: two-tailed t-test, P = 0.015; head: two-tailed t-test, P = 0.0488). c, Schematic of a mouse head, showing the trigeminal nerve with its three branches arising from the trigeminal ganglion. The diagram was created using BioRender; Kaltenecker, D. https://biorender.com/5v5wjam. d, Representative heads of chow and HFD-fed mice showing Uchl1-eGFP+ nerves and AI-based segmentation of these nerves. Structural changes in the infraorbital nerve (part of the maxillary branch of the trigeminal nerve) are indicated by the orange arrow. Scale bars, 1,500 μm. e–h, Quantification of characteristics (nerve endings (e), edges (f), vertices (g) and thickness (h)) of the infraorbital nerve after graph extractions (n = 6 infraorbital nerves from 3 chow mice, and 5 infraorbital nerves from 3 HFD-fed mice). Statistical analysis was performed using the two-tailed Mann–Whitney U-test, P = 0.0043 (e); two-tailed t-test, P = 0.0022 (f); and two-tailed t-test, P = 0.0022 (g). *P < 0.05, **P < 0.01. Data are mean ± s.e.m. Source data
Fig. 3 Whole-body nerve segmentation reveals structural changes in the infraorbital nerve in obesity. a, Representative whole-body nerve segmentation in chow and HFD-fed mice. b, Quantification of AI-segmented nerve densities in chow- and HFD-fed Uchl1-eGFP mice in whole bodies and the indicated areas. For quantification of the nerve density in the head, we masked out the brain for the analysis (n = 4 chow, n = 5 HFD mice; whole-body: two-tailed t-test, P = 0.0012; fat: two-tailed t-test, P = 0.015; head: two-tailed t-test, P = 0.0488). c, Schematic of a mouse head, showing the trigeminal nerve with its three branches arising from the trigeminal ganglion. The diagram was created using BioRender; Kaltenecker, D. https://biorender.com/5v5wjam. d, Representative heads of chow and HFD-fed mice showing Uchl1-eGFP+ nerves and AI-based segmentation of these nerves. Structural changes in the infraorbital nerve (part of the maxillary branch of the trigeminal nerve) are indicated by the orange arrow. Scale bars, 1,500 μm. e–h, Quantification of characteristics (nerve endings (e), edges (f), vertices (g) and thickness (h)) of the infraorbital nerve after graph extractions (n = 6 infraorbital nerves from 3 chow mice, and 5 infraorbital nerves from 3 HFD-fed mice). Statistical analysis was performed using the two-tailed Mann–Whitney U-test, P = 0.0043 (e); two-tailed t-test, P = 0.0022 (f); and two-tailed t-test, P = 0.0022 (g). *P < 0.05, **P < 0.01. Data are mean ± s.e.m. Source dataCC BY · PubMed Central
Fig. 4 Structural changes in the infraorbital nerve are associated with changes in the trigeminal ganglion proteome. a, Functional assessment of the response after whisker stimulation. n = 7 (chow), n = 6 (HFD). Statistical analysis was performed using two-tailed t-tests; P = 0.0040. Data are mean ± s.e.m. b, Samples for spatial proteomics profiling were extracted from the trigeminal ganglia of Uchl1-eGFP mice. The diagram was created using BioRender; Kaltenecker, D. https://biorender.com/5v5wjam. c, Differentially regulated proteins in chow versus HFD-fed mice. n = 6 chow trigeminal ganglia (right and left, from 3 mice), n = 6 HFD trigeminal ganglia (right and left, from 3 mice). d, Pathway analysis showing differentially regulated pathways. The grey bar on the left of the plot represents the log10-transformed P value of each pathway, whereas the right side of the plot shows the number of proteins significantly different in each pathway, where red represents the number of upregulated proteins and blue represents the number of downregulated proteins. e, A subset of differentially regulated pathways and the corresponding proteins. f,g, The BMI of lean individuals (BMI < 25) and individuals with obesity (BMI > 30) (f; n = 5 (lean), n = 4 (obesity); two-tailed Mann–Whitney U-test, P = 0.0159) from which trigeminal ganglia were isolated post-mortem for proteomic profiling (g). h, Pathway analysis showing differentially regulated pathways in human trigeminal ganglia. n = 15 (lean; 3 regions of interest sampled per individual from 5 individuals total) and n = 12 (obesity; 3 regions of interest per individual from 4 individuals total). The colour scheme is as described for d. Source data
Fig. 4 Structural changes in the infraorbital nerve are associated with changes in the trigeminal ganglion proteome. a, Functional assessment of the response after whisker stimulation. n = 7 (chow), n = 6 (HFD). Statistical analysis was performed using two-tailed t-tests; P = 0.0040. Data are mean ± s.e.m. b, Samples for spatial proteomics profiling were extracted from the trigeminal ganglia of Uchl1-eGFP mice. The diagram was created using BioRender; Kaltenecker, D. https://biorender.com/5v5wjam. c, Differentially regulated proteins in chow versus HFD-fed mice. n = 6 chow trigeminal ganglia (right and left, from 3 mice), n = 6 HFD trigeminal ganglia (right and left, from 3 mice). d, Pathway analysis showing differentially regulated pathways. The grey bar on the left of the plot represents the log10-transformed P value of each pathway, whereas the right side of the plot shows the number of proteins significantly different in each pathway, where red represents the number of upregulated proteins and blue represents the number of downregulated proteins. e, A subset of differentially regulated pathways and the corresponding proteins. f,g, The BMI of lean individuals (BMI < 25) and individuals with obesity (BMI > 30) (f; n = 5 (lean), n = 4 (obesity); two-tailed Mann–Whitney U-test, P = 0.0159) from which trigeminal ganglia were isolated post-mortem for proteomic profiling (g). h, Pathway analysis showing differentially regulated pathways in human trigeminal ganglia. n = 15 (lean; 3 regions of interest sampled per individual from 5 individuals total) and n = 12 (obesity; 3 regions of interest per individual from 4 individuals total). The colour scheme is as described for d. Source dataCC BY · PubMed Central
Fig. 5 Whole-body-wide inflammation in obesity. a, Representative images of Cd68-eGFP mice; AI-based segmentation of the Cd68-eGFP signal is overlaid in red. n = 5 (chow), n = 5 (HFD). b, Representative z-projection (300 µm) view showing ViscAT with Cd68-eGFP (top) and AI-based segmentation (bottom) of an obese mouse. n = 5 (chow), n = 5 (HFD). c, Representative z-projection (300 µm) showing segmented cells that were grouped into three size clusters: small (blue), medium (green) and large (red). Clusters located at different depths along the projection axis can appear to overlap in the image, even though they are spatially distinct in 3D space. n = 5 (chow), n = 5 (HFD). d, Whole-body Cd68-eGFP segmentation results in representative chow and HFD-fed mice showing small, medium and large cluster densities. n = 5 (chow), n = 5 (HFD). e, Quantification of small, medium and large cluster proportions in the indicated organs and tissues. n = 5 (chow), n = 5 (HFD). Statistical analysis was performed using two-tailed t-tests for all comparisons except for adrenal glands and muscle big CD68+ clusters, for which a two-tailed Mann–Whitney U-test was used. Exact P values were as follows: small CD68+ clusters: P = 0.0015 (liver), 0.0341 (stomach), P = 0.026 (peyer patches), P = 0.0003 (ViscAT); medium CD68+ clusters: P = 0.001 (liver), P = 0.0006 (ViscAT); big CD68+ clusters: P = 0.0079 (adrenal glands), P = 0.0091 (stomach), P = 0.0007 (abdominal wall), P = 0.0029 (peyer patches), P = 0.0303 (vesicular gland), P = 0.0001 (ViscAT), P < 0.0001 (ScAT), 0.0079 (muscle). ***P < 0.001, ****P < 0.0001. Data are mean ± s.e.m. Scale bars, 5,000 μm (a), 1,000 μm (c, left), 500 μm (b) and 200 μm (c, right). Source data
Fig. 5 Whole-body-wide inflammation in obesity. a, Representative images of Cd68-eGFP mice; AI-based segmentation of the Cd68-eGFP signal is overlaid in red. n = 5 (chow), n = 5 (HFD). b, Representative z-projection (300 µm) view showing ViscAT with Cd68-eGFP (top) and AI-based segmentation (bottom) of an obese mouse. n = 5 (chow), n = 5 (HFD). c, Representative z-projection (300 µm) showing segmented cells that were grouped into three size clusters: small (blue), medium (green) and large (red). Clusters located at different depths along the projection axis can appear to overlap in the image, even though they are spatially distinct in 3D space. n = 5 (chow), n = 5 (HFD). d, Whole-body Cd68-eGFP segmentation results in representative chow and HFD-fed mice showing small, medium and large cluster densities. n = 5 (chow), n = 5 (HFD). e, Quantification of small, medium and large cluster proportions in the indicated organs and tissues. n = 5 (chow), n = 5 (HFD). Statistical analysis was performed using two-tailed t-tests for all comparisons except for adrenal glands and muscle big CD68+ clusters, for which a two-tailed Mann–Whitney U-test was used. Exact P values were as follows: small CD68+ clusters: P = 0.0015 (liver), 0.0341 (stomach), P = 0.026 (peyer patches), P = 0.0003 (ViscAT); medium CD68+ clusters: P = 0.001 (liver), P = 0.0006 (ViscAT); big CD68+ clusters: P = 0.0079 (adrenal glands), P = 0.0091 (stomach), P = 0.0007 (abdominal wall), P = 0.0029 (peyer patches), P = 0.0303 (vesicular gland), P = 0.0001 (ViscAT), P < 0.0001 (ScAT), 0.0079 (muscle). ***P < 0.001, ****P < 0.0001. Data are mean ± s.e.m. Scale bars, 5,000 μm (a), 1,000 μm (c, left), 500 μm (b) and 200 μm (c, right). Source dataCC BY · PubMed Central
Extended Data Fig. 1 HFD-feeding associated with increased body weight and macrophage infiltration. a,b, Body composition analysis using EchoMRI of Uchl1, (n = 3/group, two-tailed t-test, p-values: 0.0002 (BW), <0.0001 (Fat)) and Cd68-eGFP mice (n = 3/group, two-tailed t-test, p-values: 0.0007 (BW), 0.0019 (Fat)). BW: Body weight. c,d Insulin tolerance test of Uchl1 (n = 4 chow, n = 5 HFD, two way ANOVA with Sídak’s multiple comparisons test, p-values: two-tailed t-test, p-values: 0.021 (30 min), 0.0022 (45 min), 0.0002 (60 min), 0.0389 (90 min)) and Cd68-eGFP mice (n = 5 chow, n = 5 HFD, two way ANOVA with Sídak’s multiple comparisons test, p-values: two-tailed t-test, p-values: 0.0002 (15 min), 0.0013 (45 min), 0.0069 (60 min)). *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, data are represented as mean ± standard error. e, Representative images (1x objective, 0.1NA) of indicated areas in chow and HFD-fed mice showing Cd68-eGFP+ cells in cyan (n = 5 chow, n = 5 HFD). f, Representative light-sheet images of indicated areas that were acquired using a 1.1x (0.1 Numerical aperture (NA) and a 4x objective (0.35 NA) (n = 4 chow, n = 5 HFD). Source data
Extended Data Fig. 1 HFD-feeding associated with increased body weight and macrophage infiltration. a,b, Body composition analysis using EchoMRI of Uchl1, (n = 3/group, two-tailed t-test, p-values: 0.0002 (BW), <0.0001 (Fat)) and Cd68-eGFP mice (n = 3/group, two-tailed t-test, p-values: 0.0007 (BW), 0.0019 (Fat)). BW: Body weight. c,d Insulin tolerance test of Uchl1 (n = 4 chow, n = 5 HFD, two way ANOVA with Sídak’s multiple comparisons test, p-values: two-tailed t-test, p-values: 0.021 (30 min), 0.0022 (45 min), 0.0002 (60 min), 0.0389 (90 min)) and Cd68-eGFP mice (n = 5 chow, n = 5 HFD, two way ANOVA with Sídak’s multiple comparisons test, p-values: two-tailed t-test, p-values: 0.0002 (15 min), 0.0013 (45 min), 0.0069 (60 min)). *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, data are represented as mean ± standard error. e, Representative images (1x objective, 0.1NA) of indicated areas in chow and HFD-fed mice showing Cd68-eGFP+ cells in cyan (n = 5 chow, n = 5 HFD). f, Representative light-sheet images of indicated areas that were acquired using a 1.1x (0.1 Numerical aperture (NA) and a 4x objective (0.35 NA) (n = 4 chow, n = 5 HFD). Source dataCC BY · PubMed Central
Extended Data Fig. 2 4x imaging allows to see finer structural details. a-d, Representative 3D reconstruction of the lower belly area (a) and projections of adipose tissue at different zoom in levels of a HFD-fed Uchl1-eGFP animal using the 4x objective (0.35 NA) to showcase expression of Uchl1-eGFP+ nerves shown in green in all panels and PI showing the nucleus in red (c,d). LN: Lymph node. e, Representative 3D reconstruction of a HFD-fed Uchl1-eGFP animal to showcase expression of Uchl1-eGFP+ nerves shown in green in panels e-i. f-i, Zoomed in views of dashed regions in (f) showing the heart demonstrating the nerves in the heart with yellow arrows, lungs in (g) nerves with blue arrows, spleen in (h), nerves with white arrows, abdominal cavity with the kidneys and a part of the spleen in (i), nerves in the visceral fat are shown here with pink arrows, brown adipose tissue (BAT) in (j) different zoomed in regions show the innervation with orange arrows. The thickness of the MIPs are as follows: 500 µm in (f), 100 µm in (g), 230 µm for left and 500 µm for right panel in (h), 1500 µm in (i), 600 µm in (j). k-m, Representative images of tissues from obese Cd68–eGFP mice. Dashed boxes indicate regions shown at higher zoom-ins in the adjacent Images in Z-projections. Images were acquired with a 4x objective (0.35NA) and are shown using different brightness and contrast settings in different organs (n = 5 chow, n = 5 HFD).
Extended Data Fig. 2 4x imaging allows to see finer structural details. a-d, Representative 3D reconstruction of the lower belly area (a) and projections of adipose tissue at different zoom in levels of a HFD-fed Uchl1-eGFP animal using the 4x objective (0.35 NA) to showcase expression of Uchl1-eGFP+ nerves shown in green in all panels and PI showing the nucleus in red (c,d). LN: Lymph node. e, Representative 3D reconstruction of a HFD-fed Uchl1-eGFP animal to showcase expression of Uchl1-eGFP+ nerves shown in green in panels e-i. f-i, Zoomed in views of dashed regions in (f) showing the heart demonstrating the nerves in the heart with yellow arrows, lungs in (g) nerves with blue arrows, spleen in (h), nerves with white arrows, abdominal cavity with the kidneys and a part of the spleen in (i), nerves in the visceral fat are shown here with pink arrows, brown adipose tissue (BAT) in (j) different zoomed in regions show the innervation with orange arrows. The thickness of the MIPs are as follows: 500 µm in (f), 100 µm in (g), 230 µm for left and 500 µm for right panel in (h), 1500 µm in (i), 600 µm in (j). k-m, Representative images of tissues from obese Cd68–eGFP mice. Dashed boxes indicate regions shown at higher zoom-ins in the adjacent Images in Z-projections. Images were acquired with a 4x objective (0.35NA) and are shown using different brightness and contrast settings in different organs (n = 5 chow, n = 5 HFD).CC BY · PubMed Central
Extended Data Fig. 3 Visualization of the WildDISCO antibody labelling in obese mice. a,b, 3D representation of ventral light-sheet fluorescence microscopy scan of wildtype mice that were antibody labelled with the pan-nerve marker UCHL1 in chow (a) and HFD-fed (b) mice (n = 3/group).
Extended Data Fig. 3 Visualization of the WildDISCO antibody labelling in obese mice. a,b, 3D representation of ventral light-sheet fluorescence microscopy scan of wildtype mice that were antibody labelled with the pan-nerve marker UCHL1 in chow (a) and HFD-fed (b) mice (n = 3/group).CC BY · PubMed Central
Extended Data Fig. 4 Visualization of the WildDISCO double labelling. a,b, 3D representation of ventral light-sheet fluorescence microscopy scan of a wildtype mouse (n=1) that was antibody labelled with the pan-nerve marker UCHL1 and a sensory nerve marker (CGRP). Approximately 64% of UCHL1+ nerves were positive for CGRP.
Extended Data Fig. 4 Visualization of the WildDISCO double labelling. a,b, 3D representation of ventral light-sheet fluorescence microscopy scan of a wildtype mouse (n=1) that was antibody labelled with the pan-nerve marker UCHL1 and a sensory nerve marker (CGRP). Approximately 64% of UCHL1+ nerves were positive for CGRP.CC BY · PubMed Central
Extended Data Fig. 5 Generalizability of the MouseMapper Nerve-Module and development of the Immune Module. a-d, Qualitative evaluation of segmentation quality of the Nerve-Module on diverse data sets: antibody labelled mouse tissue of TH (a, 516 µm maximum intensity projection (MIP)), UCHL1/PGP9.5 (b, 3932 µm MIP), CGRP (c, 1396 µm MIP) and Thy1-EGFP (d, 11994 µm MIP). e, Post-conception week 7 human embryo head antibody labelled with β tubulin (10024 µm sum of slices projection). f, To train the Immune-Module, Cd68-eGFP+ cells were annotated using virtual reality (VR). Annotations were used to train a deep neural network, which generates the segmentation of Cd68-eGFP+ cells as network output. g,h, Qualitative evaluation of the network performance for the segmentation of Cd68-eGFP+ cells (g) or antibody labelled liver tissue of CD45 (h) based on instance Dice. Areas that overlap with reference annotations: True positives (TP); no overlap in reference annotations: false positives (FP); undetected reference annotation areas: false negatives (FN). Source data
Extended Data Fig. 5 Generalizability of the MouseMapper Nerve-Module and development of the Immune Module. a-d, Qualitative evaluation of segmentation quality of the Nerve-Module on diverse data sets: antibody labelled mouse tissue of TH (a, 516 µm maximum intensity projection (MIP)), UCHL1/PGP9.5 (b, 3932 µm MIP), CGRP (c, 1396 µm MIP) and Thy1-EGFP (d, 11994 µm MIP). e, Post-conception week 7 human embryo head antibody labelled with β tubulin (10024 µm sum of slices projection). f, To train the Immune-Module, Cd68-eGFP+ cells were annotated using virtual reality (VR). Annotations were used to train a deep neural network, which generates the segmentation of Cd68-eGFP+ cells as network output. g,h, Qualitative evaluation of the network performance for the segmentation of Cd68-eGFP+ cells (g) or antibody labelled liver tissue of CD45 (h) based on instance Dice. Areas that overlap with reference annotations: True positives (TP); no overlap in reference annotations: false positives (FP); undetected reference annotation areas: false negatives (FN). Source dataCC BY · PubMed Central
Extended Data Fig. 6 AI-based segmentation of organs and tissues with using the Tissue-Module of MouseMapper. a, Pipeline depicting the workflow of the organ and tissue segmentation model that was used in Cd68-eGFP and Uchl1-eGFP mouse lines. b,c, Representative HFD-fed mouse showing AI-segmented organs and tissue. Each colour represents a different organ or tissue segmented displaying b, the ventral, c, the sagittal view of the body. d-f, Head from a representative chow animal in rostral view with AI-segmented organs and tissues in different z-planes. g,h, Organ (g) and tissue volumes (h) segmented in Uchl1-eGFP mice with the pipeline shown in a (n = 4 chow, n = 5 HFD animals; two-tailed t-test, p-values: 0.0064 (Kidneys), 0.0118 (Lymph nodes), <0.0001 (AT, adipose tissue), 0.0002 (ViscAT), <0.0001 (ScAT)). i,j, Organ (i) and tissue (j) volumes segmented in Cd68-eGFP mice with the pipeline shown in a (n = 5 chow, n = 5 HFD animals; two-tailed t-test, p-values: <0.0001 (Liver), 0.0136 (Lymph nodes), <0.0001 (AT, adipose tissue), <0.0001 (ViscAT), <0.0001 (ScAT)), *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, data are represented as mean ± standard error.
Extended Data Fig. 6 AI-based segmentation of organs and tissues with using the Tissue-Module of MouseMapper. a, Pipeline depicting the workflow of the organ and tissue segmentation model that was used in Cd68-eGFP and Uchl1-eGFP mouse lines. b,c, Representative HFD-fed mouse showing AI-segmented organs and tissue. Each colour represents a different organ or tissue segmented displaying b, the ventral, c, the sagittal view of the body. d-f, Head from a representative chow animal in rostral view with AI-segmented organs and tissues in different z-planes. g,h, Organ (g) and tissue volumes (h) segmented in Uchl1-eGFP mice with the pipeline shown in a (n = 4 chow, n = 5 HFD animals; two-tailed t-test, p-values: 0.0064 (Kidneys), 0.0118 (Lymph nodes), <0.0001 (AT, adipose tissue), 0.0002 (ViscAT), <0.0001 (ScAT)). i,j, Organ (i) and tissue (j) volumes segmented in Cd68-eGFP mice with the pipeline shown in a (n = 5 chow, n = 5 HFD animals; two-tailed t-test, p-values: <0.0001 (Liver), 0.0136 (Lymph nodes), <0.0001 (AT, adipose tissue), <0.0001 (ViscAT), <0.0001 (ScAT)), *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, data are represented as mean ± standard error.CC BY · PubMed Central
Extended Data Fig. 7 Quantification of segmented nerves and 3D visualization of Uchl1-eGFP mouse heads after vDISCO clearing, imaging, and reconstruction. a, Total nerve voxels of segmented nerves in the whole body, fat and muscle (n = 4 chow, n = 5 HFD animals; two-tailed t-test, p-value: 0.0002 (Fat). b-d, Quantification of nerve density in indicated organs/tissues. (n = 4 chow, n = 5 HFD animals; two-tailed t-test, p-value: 0.0051 (heart), 0.0111 (thymus), 0.0067 (lymph nodes), 0.0147 (sublingual gland), 0.0034 (ScAT), 0.004 (submandibular gland); *p < 0.05, **p < 0.01, ***p < 0.001. e, Representative whole-body nerve graphs extracted from segmentations colour-coded by local radius. For visualization, all radii below 5.6 µm were grouped into the 5.6 µm category. f, Histogram of nerve radii extracted of 1× graphs from all mice (n = 4 chow, n = 5 HFD animals). g, Average nerve-segment radius from 1× datasets (n = 4 chow, n = 5 HFD animals; two-tailed t-test, p-value = 0.0054). h, Histogram of nerve radii extracted from 4× datasets from one chow and one HFD mouse scan. i, Limb nerve density based on PGP9.5 antibody staining (n = 3/group; two-tailed t-test, p-value = 0.0386). j, Dorsal view of 3D reconstructions of heads of chow- and HFD-fed mice showing Uchl1-eGFP+ nerves. White arrows indicate structural changes in the infraorbital nerve. k, Sagittal view of 3D reconstructions of heads of chow- and HFD-fed mice showing Uchl1-eGFP+ nerves with zoomed-in views of the intricate nerve structures. The left side of the head was clipped away for optimizing visualization. Data are represented as mean ± standard error. Source data
Extended Data Fig. 7 Quantification of segmented nerves and 3D visualization of Uchl1-eGFP mouse heads after vDISCO clearing, imaging, and reconstruction. a, Total nerve voxels of segmented nerves in the whole body, fat and muscle (n = 4 chow, n = 5 HFD animals; two-tailed t-test, p-value: 0.0002 (Fat). b-d, Quantification of nerve density in indicated organs/tissues. (n = 4 chow, n = 5 HFD animals; two-tailed t-test, p-value: 0.0051 (heart), 0.0111 (thymus), 0.0067 (lymph nodes), 0.0147 (sublingual gland), 0.0034 (ScAT), 0.004 (submandibular gland); *p < 0.05, **p < 0.01, ***p < 0.001. e, Representative whole-body nerve graphs extracted from segmentations colour-coded by local radius. For visualization, all radii below 5.6 µm were grouped into the 5.6 µm category. f, Histogram of nerve radii extracted of 1× graphs from all mice (n = 4 chow, n = 5 HFD animals). g, Average nerve-segment radius from 1× datasets (n = 4 chow, n = 5 HFD animals; two-tailed t-test, p-value = 0.0054). h, Histogram of nerve radii extracted from 4× datasets from one chow and one HFD mouse scan. i, Limb nerve density based on PGP9.5 antibody staining (n = 3/group; two-tailed t-test, p-value = 0.0386). j, Dorsal view of 3D reconstructions of heads of chow- and HFD-fed mice showing Uchl1-eGFP+ nerves. White arrows indicate structural changes in the infraorbital nerve. k, Sagittal view of 3D reconstructions of heads of chow- and HFD-fed mice showing Uchl1-eGFP+ nerves with zoomed-in views of the intricate nerve structures. The left side of the head was clipped away for optimizing visualization. Data are represented as mean ± standard error. Source dataCC BY · PubMed Central
Extended Data Fig. 8 Trigeminal ganglion proteome differences in chow vs high-fat diet fed animals. a, Dissected trigeminal nerve from Uchl1-eGFP mouse before (left) and after (right) puncture with the 18 G needle. Red circle indicates the excised region that was subjected to proteomic analysis. b, Number of protein groups detected in each proteomic sample. Cho stands for chow group and HFD stands for high-fat diet-fed group. L stands for the left trigeminal ganglion and R stands for the right trigeminal ganglion. c, Pathway analysis from the TG of chow vs HFD is depicted. The pathways that are affected are shown on the y-axis of the plot. Grey bar on the left side of the plot represents the log10 of p-value of each pathway whereas the right side of the plot depicts the number of proteins significantly different in each pathway, red representing the number of up-regulated proteins and blue representing the number of down-regulated proteins. d, Western blot of protein lysates from trigeminal ganglia of chow and HFD-fed mice to validate proteins that were differentially expressed in the proteomic analysis: SEPTIN7, SERPINA1, p-ERK, and ERK proteins using western blot. Vinculin is shown as representative loading control (n = 3/group). e, Quantification of SEPTIN7, SERPINA1 and pERK/ERK protein levels detected in chow vs HFD groups (n = 3/group, two-tailed t-test, p-values: 0.0229 (SEPTIN7), 0.0355 (SERPINA1), 0.0357 (p-ERK/ERK). For source data, see Supplementary Fig. 6. f, Volcano plot showing differentially regulated proteins in the trigeminal ganglion of lean and obese humans (n = 15 lean (3 regions of interest sampled per individual from 5 individuals total); n = 12 obese (3 regions of interest per individual from 4 individuals total)). Data are represented as mean ± standard error. Source data
Extended Data Fig. 8 Trigeminal ganglion proteome differences in chow vs high-fat diet fed animals. a, Dissected trigeminal nerve from Uchl1-eGFP mouse before (left) and after (right) puncture with the 18 G needle. Red circle indicates the excised region that was subjected to proteomic analysis. b, Number of protein groups detected in each proteomic sample. Cho stands for chow group and HFD stands for high-fat diet-fed group. L stands for the left trigeminal ganglion and R stands for the right trigeminal ganglion. c, Pathway analysis from the TG of chow vs HFD is depicted. The pathways that are affected are shown on the y-axis of the plot. Grey bar on the left side of the plot represents the log10 of p-value of each pathway whereas the right side of the plot depicts the number of proteins significantly different in each pathway, red representing the number of up-regulated proteins and blue representing the number of down-regulated proteins. d, Western blot of protein lysates from trigeminal ganglia of chow and HFD-fed mice to validate proteins that were differentially expressed in the proteomic analysis: SEPTIN7, SERPINA1, p-ERK, and ERK proteins using western blot. Vinculin is shown as representative loading control (n = 3/group). e, Quantification of SEPTIN7, SERPINA1 and pERK/ERK protein levels detected in chow vs HFD groups (n = 3/group, two-tailed t-test, p-values: 0.0229 (SEPTIN7), 0.0355 (SERPINA1), 0.0357 (p-ERK/ERK). For source data, see Supplementary Fig. 6. f, Volcano plot showing differentially regulated proteins in the trigeminal ganglion of lean and obese humans (n = 15 lean (3 regions of interest sampled per individual from 5 individuals total); n = 12 obese (3 regions of interest per individual from 4 individuals total)). Data are represented as mean ± standard error. Source dataCC BY · PubMed Central
Extended Data Fig. 9 High resolution Cd68-EGFP organ comparison and quantification across mapped tissue. a-h, Representative 3D reconstruction and 2D 60 µm thick maximum intensity projections (MIP) of (a) lung, (b) liver, (c) spleen, (d) heart, (e) scWAT, (f) eWAT, (g) thymus and (h) accessory axillary lymph node from the Cd68-eGFP mouse line after vDISCO and light sheet imaging for chow and high fat diet fed animals. The left two columns represent the 3D reconstruction of the organs and the right two columns represent the MIPs. For each representation chow fed animals are shown on the left column and the high-fat diet fed animals are shown in the right column. The images are acquired with a 4x objective and are visualized using the same brightness and contrast settings. i-k, Quantification of small, medium and big cluster proportions in indicated organs and tissues (n = 5/group, small CD68+ clusters: two-tailed t-test, p-values: 0.0101 (pancreas), 0.0108 (spinal cord); medium CD68+ clusters: two-tailed t-test, p-values: 0.0135 (pancreas), 0.0088 (spinal cord); large CD68+ clusters: two-tailed Mann-Whitney test, p = 0.0079 (pancreas), two-tailed t-test, p = 0.0322 (diaphragm)), *p < 0.05, **p < 0.01, data are represented as mean ± standard error. Source data
Extended Data Fig. 9 High resolution Cd68-EGFP organ comparison and quantification across mapped tissue. a-h, Representative 3D reconstruction and 2D 60 µm thick maximum intensity projections (MIP) of (a) lung, (b) liver, (c) spleen, (d) heart, (e) scWAT, (f) eWAT, (g) thymus and (h) accessory axillary lymph node from the Cd68-eGFP mouse line after vDISCO and light sheet imaging for chow and high fat diet fed animals. The left two columns represent the 3D reconstruction of the organs and the right two columns represent the MIPs. For each representation chow fed animals are shown on the left column and the high-fat diet fed animals are shown in the right column. The images are acquired with a 4x objective and are visualized using the same brightness and contrast settings. i-k, Quantification of small, medium and big cluster proportions in indicated organs and tissues (n = 5/group, small CD68+ clusters: two-tailed t-test, p-values: 0.0101 (pancreas), 0.0108 (spinal cord); medium CD68+ clusters: two-tailed t-test, p-values: 0.0135 (pancreas), 0.0088 (spinal cord); large CD68+ clusters: two-tailed Mann-Whitney test, p = 0.0079 (pancreas), two-tailed t-test, p = 0.0322 (diaphragm)), *p < 0.05, **p < 0.01, data are represented as mean ± standard error. Source dataCC BY · PubMed Central
Extended Data Fig. 10 Multiplex imaging and pixel-level analysis of immune and endothelial architecture. a, Representative multiplex immunofluorescence staining showing one region of interest with indicated markers in a section of obese epididymal white adipose tissue (n = 7 region of interests (ROI)). b,c, Zoom-in view of the boxed region in (a) highlighting multicellular interfaces and local tissue architecture (b) and individual marker channels (c). d, Self-organizing map (SOM) and Leiden clustering-based pixel classification illustrating segmentation according to marker combinations in an example from same field as in b), example 2 (containing a prominent vessel) and example 3 (big cell accumulation) from different ROIs. e, Mean fluorescence intensities in each pixel class across all channels used for pixel classification scaled by the maximum mean intensity, demonstrating marker-specific segregation. f, Quantitative pixel composition analysis from seven independent regions of interest (ROI1–ROI7) showing proportional contributions of endothelial and immune cell classes to total image area. g, Spatial proximity network derived from pixel-level co-occurrence analysis showing frequent interactions between analysed cell types. The thickness of each arrow indicates how frequently pixels of the target class (arrow tip) occur next to pixels of the source class (arrow origin).
Extended Data Fig. 10 Multiplex imaging and pixel-level analysis of immune and endothelial architecture. a, Representative multiplex immunofluorescence staining showing one region of interest with indicated markers in a section of obese epididymal white adipose tissue (n = 7 region of interests (ROI)). b,c, Zoom-in view of the boxed region in (a) highlighting multicellular interfaces and local tissue architecture (b) and individual marker channels (c). d, Self-organizing map (SOM) and Leiden clustering-based pixel classification illustrating segmentation according to marker combinations in an example from same field as in b), example 2 (containing a prominent vessel) and example 3 (big cell accumulation) from different ROIs. e, Mean fluorescence intensities in each pixel class across all channels used for pixel classification scaled by the maximum mean intensity, demonstrating marker-specific segregation. f, Quantitative pixel composition analysis from seven independent regions of interest (ROI1–ROI7) showing proportional contributions of endothelial and immune cell classes to total image area. g, Spatial proximity network derived from pixel-level co-occurrence analysis showing frequent interactions between analysed cell types. The thickness of each arrow indicates how frequently pixels of the target class (arrow tip) occur next to pixels of the source class (arrow origin).CC BY · PubMed Central

Verification

Claims the paper makes

10 Valid · 1 Not valid · 16 Not checked · 135 Own results and interpretations

74 more claims

Create a free account to see all 162 claims and to verify the 31 references still open.

See all claims — free account
  • It may state two separate claims as one. 56
  • Probably not a substantial claim. 41
  • The quote only partly states this claim. 2

Create a free account to see these claims.

Citations

Cites

Show all 62 references

Research works printed but not linked: 3

  1. Paetzold, J. C. et al. Whole Brain Vessel Graphs: a dataset and benchmark for graph learning and neuroscience. In Proc. Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2021).
  2. Chen, Z. & Liu, B. Lifelong Machine Learning 55–75 (Springer, 2022).
  3. Oktay, O. et al. Attention U-Net: learning where to look for the pancreas. In 1st Conference on Medical Imaging with Deep Learning (OpenReview, 2018).

Cited by

See all 11

Metadata

Authors
Kaltenecker, D., Horvath, I., Al-Maskari, R., Chen, Y., Kolabas, Z. I., Hoeher, L., Todorov, M., Minde, D.-P., Kapoor, S., Turhan, S. G., Kuemmerle, L. B., Steinke, H., Wohlgemuth, T., Ali, M., Kofler, F., Morigny, P., Geppert, J., Jeridi, D., Wittmann, B., Luo, J., Shit, S., Cigankova, C., Kolenic, V. M., Gür, N., Aydeniz, E., Yücecan, A., Ertürk, M., Simons, L. H. A., Pan, C., Piraud, M., Rueckert, D., Rohm, M., Hellal, F., Elsner, M., Bhatia, H. S., Bechmann, I., Menze, B. H., Herzig, S., Paetzold, J. C., Diaz, M. B., & Ertürk, A.
Affiliations
  • University Hospital Heidelberg (DE)
  • Helmholtz Munich (DE)
  • Institut für Schlaganfall- und Demenzforschung (DE)
  • Leipzig University (DE)
  • University of Zurich (CH)
  • Purdue University West Lafayette (US)
  • Chinese Academy of Sciences (CN)
  • TUM Klinikum (DE)
  • Koç University (TR)
Corresponding author
Ertürk, Ali
Venue
Nature
Published
2026-05-20
Volume
655
Issue
8124
Pages
1016-1026
ISSN
  • 0028-0836 (ISSN-L)
  • 1476-4687
OpenAlex
W7161822303
Cited by
5
References
62
Citations by year
2026: 5
FWCI
13.59
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Biophysics › Cell Image Analysis Techniques
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Medical Image Segmentation Techniques
Keywords
foundation models, multi-organ segmentation, trigeminal ganglion, infraorbital nerve, diet-induced obesity, complement pathway, immune cell clusters, proteome changes, mouse model
MeSH terms
Deep Learning, Animals, Axons, Diet, Female, Humans, Immune System, Inflammation, Male, Mice, Inbred C57BL, Obesity, Trigeminal Ganglion, Vibrissae, Imaging, Three-Dimensional, Proteomics, Mice, Neuroimaging

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-28, under version 3 of the claim definition (prompt extract-v3).

A second model labelled each claim on 2026-09-28 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: A deep-learning framework reveals whole-body perturbations at cell level. https://app.cochat.ai/papers/10.1038/s41586-026-10535-2
The paper
Kaltenecker, D., Horvath, I., Al-Maskari, R., Chen, Y., Kolabas, Z. I., Hoeher, L., Todorov, M., Minde, D.-P., Kapoor, S., Turhan, S. G., Kuemmerle, L. B., Steinke, H., Wohlgemuth, T., Ali, M., Kofler, F., Morigny, P., Geppert, J., Jeridi, D., Wittmann, B., . . . Ertürk, A. (2026). A deep-learning framework reveals whole-body perturbations at cell level. Nature. https://doi.org/10.1038/s41586-026-10535-2

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.