Spatial mapping of mitochondrial networks and bioenergetics in lung cancer

Mitochondria are critical to the governance of metabolism and bioenergetics in cancer cells1. The mitochondria form highly organized networks, in which their outer and inner membrane structures define their bioenergetic capacity2,3. However, in vivo studies delineating the relationship between the structural organization of mitochondrial networks and their bioenergetic activity have been limited. Here we present an in vivo structural and functional analysis of mitochondrial networks and bioenergetic phenotypes in non-small cell lung cancer (NSCLC) using an integrated platform consisting of positron emission tomography imaging, respirometry and three-dimensional scanning block-face electron microscopy. The diverse bioenergetic phenotypes and metabolic dependencies we identified in NSCLC tumours align with distinct structural organization of mitochondrial networks present. Further, we discovered that mitochondrial networks are organized into distinct compartments within tumour cells. In tumours with high rates of oxidative phosphorylation (OXPHOSHI) and fatty acid oxidation, we identified peri-droplet mitochondrial networks wherein mitochondria contact and surround lipid droplets. By contrast, we discovered that in tumours with low rates of OXPHOS (OXPHOSLO), high glucose flux regulated perinuclear localization of mitochondria, structural remodelling of cristae and mitochondrial respiratory capacity. Our findings suggest that in NSCLC, mitochondrial networks are compartmentalized into distinct subpopulations that govern the bioenergetic capacity of tumours.

Mitochondria are critical to the governance of metabolism and bioenergetics in cancer cells 1 . The mitochondria form highly organized networks, in which their outer and inner membrane structures define their bioenergetic capacity 2,3 . However, in vivo studies delineating the relationship between the structural organization of mitochondrial networks and their bioenergetic activity have been limited. Here we present an in vivo structural and functional analysis of mitochondrial networks and bioenergetic phenotypes in non-small cell lung cancer (NSCLC) using an integrated platform consisting of positron emission tomography imaging, respirometry and three-dimensional scanning block-face electron microscopy. The diverse bioenergetic phenotypes and metabolic dependencies we identified in NSCLC tumours align with distinct structural organization of mitochondrial networks present. Further, we discovered that mitochondrial networks are organized into distinct compartments within tumour cells. In tumours with high rates of oxidative phosphorylation (OXPHOS HI ) and fatty acid oxidation, we identified peri-droplet mitochondrial networks wherein mitochondria contact and surround lipid droplets. By contrast, we discovered that in tumours with low rates of OXPHOS (OXPHOS LO ), high glucose flux regulated perinuclear localization of mitochondria, structural remodelling of cristae and mitochondrial respiratory capacity. Our findings suggest that in NSCLC, mitochondrial networks are compartmentalized into distinct subpopulations that govern the bioenergetic capacity of tumours.
NSCLC is a heterogeneous disease at a histological, genetic and metabolic level 4 . Mitochondria are essential regulators of cellular energy and metabolism, playing a critical role in sustaining growth and survival of tumour cells 5 . The mitochondria organize into dynamic networks such that the structural architecture of their outer and inner membrane dictates cellular electron transport chain (ETC) activity and respiratory capacity 2,6 . However, our understanding of how mitochondrial networks are structurally and functionally regulated in NSCLC at an in vivo level is limited.
To better understand mitochondrial bioenergetics in NSCLC, we recently developed and validated a voltage-sensitive, positron emission tomography (PET) tracer known as [ 18 F]4-fluorobenzyl triphenylphosphonium ([ 18 F]FBnTP) 7,8 . This tracer allowed us to measure relative changes in the mitochondrial membrane potential (ΔΨ) in autochthonous KRAS-driven mouse models of NSCLC 9 . PET imaging of NSCLC tumours in KRAS-driven genetically engineered mouse models (GEMMs) identified that lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC) had distinctly different uptake values for the radiotracers [ 18 F]FBnTP and 18 F[FDG], suggestive of functionally distinct metabolic and bioenergetic phenotypes 9 . Therefore, we sought to determine whether [ 18 F]FBnTP uptake in NSCLC tumours directly correlated with OXPHOS activity in vivo.
The ETC generates a proton gradient to maintain ΔΨ and drive OXPHOS. Therefore, we examined whether uptake of [ 18 F]FBnTP correlated with the OXPHOS activity in lung tumours. Following PET imaging, ex vivo respirometry was carried out on snap-frozen lung tumours that allowed for the direct measurement of complex I and II maximal respiratory capacity (MRC) 12 Fig. 1e,f). Analysis of mitochondrial respiration in LUAD and LUSC tumours from a larger cohort of KPL and KL mice identified a significant increase in complex I and II MRC in LUAD cells versus LUSC cells (Extended Data Fig. 1f,g). Our analysis of mitochondrial activity in KPL and KL GEMMs showed that LUSC cells had significantly lower OXPHOS signatures compared with those of LUAD cells.
Next, we broadened our correlative examination of complex I and II MRC and [ 18 Fig. 2a,b). We identified a direct and significant correlation between [ 18 F]FBnTP uptake and complex I and II MRC ( Fig. 1c and Extended Data Fig. 2d,f). Clear histologic stratification was observed, as LUAD cells had higher [ 18

PET-guided 3D SBEM imaging of NSCLC
The fact that we observed functionally distinct OXPHOS signatures in NSCLC tumour subtypes suggests that these tumours may have equally distinct structural organization of their mitochondrial networks. To investigate this, we developed a workflow that paired functional PET imaging with micro-computed tomography (microCT) and ultra-resolution three-dimensional serial block-face electron microscopy 13 (3D SBEM; Fig. 2a). The incorporation of microCT imaging allowed us to bridge gaps in resolution scales between whole-tumour imaging with PET and ultrastructure imaging with SBEM. We first carried out [ 18 F]FBnTP and [ 18 F]FDG PET-CT imaging on KRAS(G12D)-driven GEMMs to distinguish mitochondrial activity and glucose flux in tumours (Fig. 2b). We identified regions of high versus low radiotracer uptake within each tumour that were used to guide SBEM analysis ( Fig. 2b and Extended Data Fig. 3a-c). We then carried out microCT and histological analysis to: provide an overview of the tumour tissue density from the periphery to core regions; spatially orientate tumours in an x-y-z plane; and differentiate viable versus necrotic tissue (Fig. 2c, Extended Data Fig. 3d

Article
With this workflow in place, we next carried out high-resolution microCT scans on tumours to verify that the tumour sections had proper depth of penetration of heavy metal staining and contained viable, tumour-dense tissue (Fig. 2d, Extended Data Fig. 3f,g,i and Supplementary Video 2). Last, sequential SBEM images were acquired and compiled to generate 3D tumour volumes ( Fig. 2d and Supplementary Videos 3-5). SBEM imaging readily facilitates quantitative analysis on large-volume content-rich datasets based on the following three features: tomography (serial sectioning) and 3D rendering of tissue; a large field of view with content-rich images; and nanometre resolution.
The reduced [ 18 F]FBnTP uptake in LUSC raised the question as to whether these tumours had reduced vasculature as compared to LUAD. We quantified tumour volume and vasculature using microCT analysis and immunohistochemical staining for the endothelial marker CD34. The data showed no notable differences in vasculature densities (Extended Data Fig. 4a,b), or CD34 staining (Extended Data Fig. 4c,f), between OXPHOS HI LUAD and OXPHOS LO LUSC. By contrast, CD34 staining was significantly increased in normal tissue versus tumour (Extended Data Fig. 4d,e).
Having generated a 3D tumour volume, we next built a topographical map of the tumour's cellular landscape to identify specific cell types within the tumour. We carried out cell segmentation and labelling of each cell type (tumour versus immune versus endothelial) based on cellular morphological features (Fig. 2e, Extended Data Fig. 5a and Supplementary Video 6). In LUSC, the main immune cell types identified were neutrophils, whereas macrophages were most predominant in LUAD (Fig. 2e,f and Extended Data Fig. 5b-e). These results are consistent with those of previous work in mouse models and human NSCLC 14,15 , and demonstrate the development of an in vivo imaging workflow that enabled us to bridge non-invasive PET imaging of whole tumours with high-resolution microCT and ultra-resolution SBEM.

Spatial and structural mapping of mitochondria
Having obtained SBEM datasets from OXPHOS HI LUAD and OXPHOS LO LUSC tumours, we carried out a quantitative structural and spatial analysis of mitochondrial networks within tumour cells. Mitochondria in 2D SBEM images appeared phenotypically similar in both LUAD and LUSC (Fig. 3a). These images accurately represented cross-sectional views of the mitochondria; however, they did not characterize the higher-order organization of mitochondrial networks. By contrast, 3D renderings of mitochondrial networks (shown in red) identified evident phenotypic differences in which LUAD cells had predominantly fused, elongated mitochondria whereas LUSC cells had predominantly fragmented mitochondria (Fig. 3b). These results demonstrated that qualitatively, 3D rendering of tumours identified differences in mitochondrial structure and spatial distribution between LUAD and LUSC cells.
Next we developed quantitative methods to analyse mitochondrial structure across our content-rich SBEM datasets. To achieve a comprehensive analysis of the large mitochondrial content within the SBEM volumes, we developed a deep learning convolutional neural network (CNN). The CNN achieved robust and accurate trinary segmentation that included partitioning of the image space of mitochondria, nucleus and background classes. We segmented the mitochondria and nuclei in about 200 LUAD and about 150 LUSC cells from each tumour volume and manually annotated nuclei and mitochondria in 13 2D SBEM images-10 for training and 3 for validation. Our network outputs class logits at each pixel, with the class yielding the highest logits denoting its label. The resulting segmentation yielded an average precision of 0.916 and a recall of 0.927 versus manual registration (Extended Data Fig. 6a,b). We used the CNN to infer the segmentation of 200-500 serial 2D SBEM images to generate 3D renderings of mitochondria and nuclei in lung tumours.
Mitochondria are dynamic organelles that continually remodel their networks resulting in pools that vary in both size and shape. Utilizing our deep learning CNN, we then analysed and quantified 20,000-50,000 mitochondria per tumour section with morphological measurements taken for length, total volume, sphericity (roundness) and spatial distribution in our SBEM tumour volumes. Mitochondria in OXPHOS LO LUSC were predominantly fragmented with increased sphericity versus OXPHOS HI LUAD cells (Fig. 3b,c). Density plots measuring total mitochondrial length and volume showed a narrow distribution in LUSC versus a broad Gaussian distribution of peaks in LUAD cells (Fig. 3d,f and Extended Data Fig. 7a-c). These results showed that mitochondria within OXPHOS LO LUSC cells were smaller and more fragmented than OXPHOS HI LUAD cells.
We followed these results with an analysis of mitochondrial dynamics in human tumour cell lines and found that the OXPHOS LO RH2 and Tu686 SCC lines had significantly higher mitochondrial fragmentation (as measured by mitochondrial circularity, aspect ratio and size) than OXPHOS HI H1975 and A549 LUAD cell lines (Extended Data Fig. 7d-g). Our analysis of human and mouse tumour cells identified a more uniform organization of mitochondrial dynamics in OXPHOS LO SCC cells than in OXPHOS HI LUAD cells. Moreover, we observed that mouse LUAD and LUSC had different spatial distribution of their mitochondria throughout the cell (Fig. 3b). To further evaluate, we developed another algorithm to measure the spatial distance between individual mitochondria and their corresponding nucleus in both SBEM and fluorescent images (Extended Data Fig. 8a-d). In SBEM images from OXPHOS HI LUAD cells, we discovered a broad spatial distribution of mitochondrial networks across the cytoplasm-spanning from the nucleus to the plasma membrane. By contrast, the mitochondria in OXPHOS LO LUSC were enriched at perinuclear regions of the cell (Fig. 3f, Extended Data Fig. 8e and Supplementary Videos 7 and 8). Next we analysed the spatial distribution of mitochondria relative to the nucleus in human tumour cell lines and found that mitochondrial networks in OXPHOS HI LUAD A549 and H1975 cells were localized predominantly throughout the cytoplasm, whereas mitochondrial localization was perinuclear in OXPHOS LO squamous RH2 and Tu686 cells (Extended Data Fig. 8f). These results indicate that OXPHOS LO SCC cells maintained a more spatially restricted and structurally homogeneous pool of fragmented perinuclear mitochondria (PNM) than OXPHOS HI LUAD cells.
The SBEM imaging at a resolution of 5-6 nm enabled us to investigate not only mitochondrial dynamics but also the organization of the ultrastructure of the cristae, which constitute the mitochondrial inner membrane. It is well established that the structural organization of the cristae directly impacts the bioenergetic capacity of the mitochondria 2,3,6 . The differences in OXPHOS signatures between NSCLC tumour subtypes suggested that OXPHOS HI LUAD cells may have distinctly different crista organization from OXPHOS LO LUSC cells. Among the OXPHOS HI LUAD and OXPHOS LO LUSC tumour volumes analysed, we identified three main crista structures that have been previously described in mammalian cells. These structures include: highly organized orthodox or lamellar cristae that we classified as type I (ref. 16 ); sparse and disorganized cristae that we classified as type II (ref. 17 ); and condensed cristae that we classified at type III (ref. 18  Orthodox, lamellar cristae support robust OXPHOS activity in cells, whereas disorganized and condensed cristae are associated with defects in cellular OXPHOS activity 16,[18][19][20] . We therefore analysed and compared crista density in mouse and human OXPHOS HI LUAD versus OXPHOS LO LUSC cells. The crista density was quantified using a Weka segmentation (ImageJ) program 21 (Extended Data Fig. 9g). Morphological analysis identified that LUAD cells had higher crista density and basal oxygen consumption rate (OCR) than LUSC cells (Extended Data Fig. 9h,i). Next we carried out an analysis of crista architecture in human and mouse OXPHOS HI LUAD and OXPHOS LO SCC cells. Quantification of crista types showed that OXPHOS HI LUAD cells had a mixed population of type I, II and III cristae. By contrast, OXPHOS LO SCC cells were enriched for condensed type III cristae with a significant reduction in type I cristae (Fig. 3j,k and Extended Data Fig. 9j). This corresponded to significantly lower basal OCR in squamous versus LUAD cells (Extended Data Fig. 9k). In sum, our structural analysis of mitochondria showed that human and mouse OXPHOS LO tumour cells consistently lacked organized type I cristae compared to OXPHOS HI tumour cells.

Peri-droplet mitochondria enriched in LUAD
SBEM imaging of content-rich tumour volumes allowed for 3D mapping of not only mitochondrial networks, but also other cellular organelles. We discovered that OXPHOS HI LUAD cells contained a vast number of lipid droplets (LDs) that were nearly absent in OXPHOS LO LUSC cells (Fig. 4a,b). Integrated throughout these LDs was a subpopulation of mitochondria in OXPHOS HI LUAD cells that contacted single or clustered LDs to form peri-droplet mitochondria (PDM; Fig. 3a bottom left panel). PDM have been identified in brown adipose tissue 22 , heart 23 and skeletal muscle 24 but have not been described in lung cancer. In brown adipose tissue, PDM were characterized by elevated rates of mitochondrial respiration (that is, OXPHOS) versus cytoplasmic mitochondria (CM) 25 .
On the basis of the presence of PDM in OXPHOS HI LUAD cells, we asked whether mitochondria may be organized into discrete subpopulations within these tumours. Using SBEM images from mouse LUAD and LUSC  Article tumours, we grouped mitochondria into subpopulations based on their interaction with cellular organelles. These subpopulations include: PNM, localized to the nucleus; PDM that contacted LDs; and CM that were scattered throughout the cytoplasm and lacked both nuclear and LD contacts. OXPHOS HI LUAD cells had a similar distribution of PNM, PDM and CM populations compared to OXPHOS LO LUSC cells, which were nearly absent of LDs (Fig. 4a,c). To confirm our observations in the SBEM images, we measured LDs in vivo in both NSCLC xenografts and GEMMs by oil red O staining of tumours to label neutral lipids (triglycerides and diacylglycerols). We found a significant enrichment in lipids in OXPHOS HI LUAD cells that was absent in OXPHOS LO SCC cells (Fig. 4d-f and Extended Data Fig. 10a,b).
We next investigated whether PDM were present in human NSCLC tumour subtypes. Analysis of The Cancer Genome Atlases for human LUAD and LUSC identified that gene expression of DGAT1, a regulator of LD biogenesis, and that of PLIN5, which regulates LD formation 26,27 , were significantly upregulated in LUAD versus LUSC (Extended Data Fig. 10c,d). We then stained the OXPHOS HI LUAD cell lines H1975, H1651 and A549 and the OXPHOS LO SCC cell lines RH2 and Tu686 with BODIPY to identify LDs. Our results showed that the OXPHOS HI LUAD cell lines had a significant increase in LDs and PDM formation compared with OXPHOS LO squamous cell lines (Fig. 4g-l and Extended Data Fig. 10e,f). These results demonstrate that human and mouse OXPHOS HI LUAD cells had a significantly higher number of LDs and PDM compared with OXPHOS LO squamous tumours.
Given the nanometre resolution of our SBEM images, we next examined the crista architecture in PDM from OXPHOS HI LUAD cells. Representative 2D and 3D images from tumour volumes showed elongated mitochondria that were interlaced throughout LD clusters and contacted multiple LDs (Fig. 4j). Notably, PDM exhibited densely packed orthodox type I cristae that aligned vertically at the mitochondria-lipid contact sites ( Fig. 4k and Extended Data Fig. 10g). The enrichment of PDM for more OXPHOS-proficient type I cristae agrees with the data in Fig. 3, as well as previous studies in brown adipose tissue 25 .
We next examined both OXPHOS activity and nutrient preferences for glucose, glutamine or fatty acids in PDM-rich versus PDM-deficient human OXPHOS HI LUAD and OXPHOS LO SCC cell lines. We measured the percentage of change in basal and maximal OCR following treatment with inhibitors of pyruvate metabolism (UK5099), glutamine metabolism (BPTES) or fatty acid oxidation (etomoxir) in the H1975 and A549 versus RH2 and Tu686 cell lines. Our results showed that PDM-enriched LUAD cell lines utilize pyruvate, glutamine and fatty acid oxidation to support respiration, whereas PDM-deficient squamous cell lines were reliant on pyruvate and glutamine inhibition, but not fatty acid oxidation ( Fig. 4l and Extended Data Fig. 10h).
Subsequently, we sought to understand whether PDM-rich NSCLC cells relied on different nutrient sources compared to PDM-deficient ones. We examined whether restriction of nutrients (glucose, glutamine or free fatty acids would limit growth of H1975 or RH2 cells. Our results showed that restricting free fatty acids significantly inhibited growth of H1975 cells but not RH2 cells (Fig. 4m). Both cell types were highly dependent on glutamine, as previously described in lung tumour cells 28 . H1975 cells grew well in low-glucose conditions, whereas RH2 cells did not grow, agreeing with previous studies that identified LUSC are reliant on both glucose and glutamine for cell survival 11,29 . In sum, we discovered PDM as a mitochondrial subpopulation that were enriched in OXPHOS HI LUAD cells. Our data suggest that PDM-rich LUAD cells relied on a broad source of nutrients to support OXPHOS and growth whereas nutrient dependency in PDM-deficient squamous tumour cells was restricted to glucose and glutamine metabolism.  (Fig. 1e). We asked whether the perinuclear localization of mitochondria identified in hypermetabolic SCC cells was due to suppression of mitochondrial trafficking along the cytoskeleton as a result of high glucose flux (Fig. 3b and Extended Data Fig. 11a,b). We thus proposed a model in hypermetabolic SCC cells in which high glucose flux serves to confine mitochondria to the nucleus by repression of mitochondrial motility through activation of the hexosamine and OGT pathways (Fig. 5a). We measured mitochondrial motility at basal levels or following inhibition of glucose flux in hypermetabolic, glycolytic RH2 and Tu686 versus less glycolytic H1975 and A549 cell lines 11 (Fig. 1e). Mitochondrial motility was quantified by measuring the displacement of individual mitochondria over time, as previously described 33 . Basal glucose uptake and extracellular acidification rate was notably increased, whereas mitochondrial motility was significantly decreased in RH2 and Tu686 versus H1975 and A549 cell lines (Fig. 5b, Extended Data Fig. 11c-e and Supplementary Videos 9 and 10). We next measured mitochondrial motility following inhibition of glucose flux or the hexosamine pathway. Glucose uptake was restricted by modification of cell culture medium (low glucose (12 mM) or galactose), targeted inhibition of glucose transport through the pan GLUT1 and GLUT3 inhibitor KL-11743 (ref. 34 ) or RNA-mediated interference knockdown of the GLUT1 transporter that all led to a marked increase in mitochondrial motility (Extended Data Fig. 11f-h). The hexosamine pathway was inhibited using azaserine, which inhibits glutamine-fructose-6-phosphate transaminase [isomerizing] 2 (GFPT2), and the OGT inhibitor OSMI-1, as previously shown 35 . Inhibition of glucose uptake and the hexosamine pathway induced a significant increase in mitochondrial motility in RH2 and to a lesser extent H1975 cells (Fig. 5c-e and Extended Data Fig. 11i,j).

Glucose regulation of mitochondrial motility
Next we investigated the role of OGT in the regulation of mitochondrial motility in NSCLC cells. In RH2 cells, glucose restriction, via treatment with KL-11743, azaserine, OSMI-1 or RNAi-mediated OGT knockdown, led to a decrease in total cellular O-GlcNAcylation ( Fig. 5f and Extended Data Fig. 11k-m). Additionally, OGT knockdown significantly increased mitochondrial motility in RH2 and H1975 cells (Fig. 5g). In sum, these results demonstrate that glucose flux regulated mitochondrial motility in NSCLC tumour cells through the hexosamine pathway and OGT.
We then explored whether the increased mitochondrial motility as a result of glucose flux inhibition induced remodelling of mitochondrial networks. In RH2 cells, KL-11743 treatment induced a redistribution of mitochondria from nuclear to cytoplasmic localization (Extended Data Fig. 11b,n,o). We discovered that inhibition of glucose flux induced remodelling of crista architecture marked by a significant increase in type I cristae and a concomitant decrease in type III cristae in RH2 cells ( Fig. 5h and Extended Data Fig. 11p,q). Glucose restriction, inhibition of the hexosamine pathway and OGT knockdown in RH2 and/or H1975 cells yielded similar results to KL-11743 treatment (Extended Data Fig. 11r-u).
The enrichment of well-formed orthodox type I cristae in NSCLC cells following crista remodelling suggested that OXPHOS activity may be upregulated following inhibition of glucose flux and/or the hexosamine-OGT pathway. We measured the maximal OCR in tumour cells following inhibition of glucose flux, the hexosamine pathway and OGT and these all induced a significant increase in basal and maximum OCR, as well as the complex I and II MRC in RH2, and to a lesser extent in H1975 cells (Fig. 5i and Extended Data Fig. 11v-z).   (Fig. 5j). Ex vivo respirometry carried out on tumours showed that KL-11743 treatment induced a significant increase in complex I and II MRC versus vehicle (Fig. 5k). These results demonstrate that inhibition of glucose flux and the hexosamine-OGT pathway led to significantly increased mitochondrial motility, enrichment of type I cristae and the upregulation of OXPHOS activity in glycolytic squamous tumours and to a lesser extend in LUAD cells. Last, we investigated the role of microtubules, actin and vimentin in the regulation of mitochondrial motility in H1975 and RH2 NSCLC cell lines. Disruption of both microtubule and actin networks by treatment with latrunculin A or nocodazole, respectively, led to a notable decrease in mitochondrial motility in both cell lines (Extended Data Fig. 12a,b). By contrast, siRNA-mediated knockdown of vimentin led to a significant increase in mitochondrial motility (Extended Data Fig. 12c,d). We next confirmed that both latrunculin A and nocodazole treatment induced a significant decrease in the distance of mitochondria from the nucleus resulting in an enrichment of PNM, whereas knockdown of vimentin led to a significant increase (Extended Data Fig. 12e-h). Last, we measured basal OCR after disruption of the components of the cytoskeleton and identified that latrunculin A and nocodazole treatment induced a notable decrease in basal OCR whereas vimentin knockdown induced a significant increase (Extended Data Fig. 12i-l).
These results indicate that both microtubules and active networks support mitochondrial motility, cytoplasmic distribution of mitochondria and respiratory capacity whereas vimentin functions in an antagonistic and suppressive manner.

Discussion
In summary, we present an in vivo structural and functional analysis of mitochondrial networks and bioenergetic phenotypes across mouse and human NSCLC tumours. The ex vivo respirometry we carried out on NSCLC tumours identified significant increases in complex I and complex II respiratory activity in [ 18 F]FBnTP-positive LUAD cells, whereas glycolytic LUSC cells with low [18F]FBnTP uptake had persistently low complex I activity. Supporting this, it was recently shown that ubiquinol oxidization is required for lung tumour growth, emphasizing a critical role of complex III and the ETC in lung tumorigenesis 36 . These results underscore the diversity of bioenergetic activity among NSCLC tumour subtypes. The functional diversity we discovered between LUAD cells and LUSC cells accurately predicted distinct structural mitochondrial phenotypes among lung tumour subtypes. PET-guided 3D SBEM, combined with CNN machine learning analysis, facilitated the large-scale mapping of the structural and spatial distribution of mitochondrial networks within [ 18 F]FBnTP-positive LUAD cells and [ 18 F]FBnTP-negative LUSC cells. The broad diversity of mitochondrial structures identified in LUAD and LUSC tumours correlated with equally diverse bioenergetic profiles and metabolic dependencies in these histological subtypes. In addition, we identified a previously unrecognized compartmentalization of mitochondrial subpopulation in LUAD cells in which PDM populations were enriched in these tumours.
The closely regulated and uniform organization of mitochondrial networks we profiled in LUSC cells corresponded with reduced metabolic flexibility compared with LUAD cells. LUSC were more reliant on glucose and glutamine metabolism to support OXPHOS and less so on fatty acid oxidation, whereas LUAD cells utilized glucose, glutamine and fatty acids to support cellular respiration. These results agree with previous studies that demonstrated LUSC cells rely on glucose and glutamine to support tumour metabolism and are selectively sensitive to inhibition of these pathways 11,28 . Structure-function studies defining the relationship between mitochondrial architecture and metabolic dependencies may hold promise as an emerging diagnostic and therapeutic strategy that can be leveraged to exploit bioenergetic and metabolic liabilities unique to lung cancer subtypes. We anticipate that coupling PET imaging with 3D SBEM will have dynamic applications beyond that of lung cancer and enrich our understanding of how mitochondrial bioenergetics impact human disease.

Online content
Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-023-05793-3. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Cell culture
Cells were cultured in Dulbecco's modified Eagle's medium (Thermo Fisher Scientific) supplemented with 10% fetal bovine serum (Hyclone) and 1% penicillin/streptomycin (Gibco). A549 cells and H1975 cells were obtained from ATCC. The lung squamous cell line (human) RH2 were established in the laboratory of S.M.D. (UCLA). The head and neck squamous cell line (human) Tu686 was a gift from the laboratory of M.S.J. (UCLA). The mouse lung squamous cell line derived from mouse 5 (LPP) and mouse LUAD cell line derived from mouse 4 (LPP) were established in our laboratory. After resecting the tumour tissues, they were minced with razors and digested with collagenase/dispase (Sigma). Lysate was filtered through a 70-μm cell strainer. Dissociated single cells were centrifuged and resuspended in Dulbecco's modified Eagle's medium (10% fetal bovine serum). The cells were plated in tissue culture dishes and the medium was changed until there were enough colonies to expand the culture. All cells were grown at 37 °C in 5% CO 2 in a humidified incubator, and a test for mycoplasma was carried out using the LookOut Mycoplasma PCR Detection Kit (Sigma). Cell identities were confirmed by Laragen Inc. The endoribonuclease-prepared siRNA used in this study was: GLUT1 (SLC2A1) (EHU028011, Sigma); OGT (EHU082301, Sigma).

GEMMs of lung tumour
We used five GEMMs in this study: (1)  or LentiCre (Kerafast) delivered intranasally as described previously 37 . Tumour growth was routinely monitored by bioluminescence imaging using an IVIS imager (PerkinElmer). All animal experiments were approved by UCLA's Animal Research Committee (ARC) and carried out following ARC protocols and requirements. The tumour burden endpoints (morbidity; weight loss no greater than 20%; laboured breathing; impingement of animal's ability to ambulate, eat and drink) allowed by our Institutional Animal Care and Use Committee were not exceeded. Lung tumours from different GEMM mice were collected and snap frozen in liquid nitrogen. Snap-frozen samples were stored at −80 °C until respirometry assay and western blotting analysis.

Subcutaneous implantation in NSG mice
A549, H1975, A549 Rho, RH2 and Tu686 cells were cultured in vitro under the conditions described above. Cells were collected and suspended in PBS, then mixed with Matrigel Membrane Matrix (Corning) and implanted subcutaneously on the flanks of NSG mice (2-4 × 10 6 cells per flank). All animal experiments were approved by UCLA's ARC and carried out following ARC protocols and requirements. The tumour burden endpoints (tumour volume no greater than 2,000 mm 3 ; morbidity; weight loss no greater than 20%; laboured breathing; impingement of animal's ability to ambulate, eat and drink) allowed by our Institutional Animal Care and Use Committee were not exceeded. For the treatment study of the GLUT1 and GLUT3 inhibitor KL-11743 (C6), mice were treated with C6 (100 mg kg −1 ) delivered by oral gavage for 8 days. Subcutaneous tumours were either snap frozen in liquid nitrogen or fixed in 10% formalin overnight. Snap-frozen samples were stored at −80 °C until respirometry assay. Formalin-fixed samples were sent to the Translational Pathology Core Laboratory at UCLA for embedding and sectioning.

PET-CT imaging
PET-CT imaging and analysis were carried out on GNEX as previously described 37,39 . PET signals were measured as percentage of injected dose per gram after 1 h uptake and normalized to heart signal. Tumours with a tumour-to-heart [ 18 26 ).

Ex vivo respirometry analysis on frozen tissues
Tumour tissues were isolated from GEMMs of lung cancer and snap frozen using liquid nitrogen. Frozen samples were stored in −80 °C until use in the Seahorse experiments. Frozen tissues were thawed on ice and homogenized in MAS buffer (70 mM sucrose, 220 mM mannitol, 5 mM KH 2 PO 4 , 5 mM MgCl 2 , 1 mM EGTA, 2 mM HEPES pH 7.4) with protease inhibitor cocktail (Roche). Homogenates were centrifuged at 1,000g for 10 min at 4 °C and supernatant was collected. Protein concentrations were determined by BCA assay kit (Thermo Fisher). Homogenates (12 μg per well for human samples and 6 μg per well for mouse samples) were loaded into a Seahorse XF96 microplate in MAS buffer (20 μl each well) and centrifuged at 2,000g for 5 min at 4 °C. After centrifugation, the volume was increased to 150 μl by adding 130 μl MAS containing cytochrome c (10 μg ml−1). At port A, substrates of NADH (1 mM) were injected to determine the respiratory capacity of mitochondrial complex I; succinate (5 mM) + rotenone (2 μM) were injected to determine the respiratory capacity of mitochondrial complex II. The following compounds were injected so that final concentrations were as followsport B: rotenone (2 μM) + antimycin (4 μM); port C: TMPD (0.5 mM) + ascorbic acid (1 mM); port D: azide (50 mM). OCR rates were measured using a Seahorse XF96 Extracellular Flux Analyzer (Agilent Technologies) and normalized to mitochondrial content quantified by MTDR. Homogenates were stained with 500 nM MTDR for 10 min followed by two wash steps to remove the dye (Thermo Fisher). MTDR fluorescence was read on a Tecan Spark plate reader (Ex: 633 nm; Em: 678 nm).

In vitro respirometry analysis on cultured cells
Cells were seeded into a Seahorse XF96 microplate before the assay and maintained in a tissue culture incubator (37 °C in 5% CO 2 ) to reach 90-100% density before running the Seahorse assay. To measure complex I and complex II respiratory capacity, cells were permeabilized with 4 nM XF plasma membrane permeabilizer (Agilent). Permeabilized cells were started in state 3 with substrates of pyruvate (5 mM) + malate (0.5 mM) and 4 mM ADP for respiration-driven through complex I; succinate (5 mM) + rotenone (2 μM) and 4 mM ADP for complex II. Following state 3 measurements, injections included the following-port A: oligomycin (2 μM); ports B and C: FCCP (B: 0.75 μM and C: 1.35 μM); port D: rotenone (1 μM) + antimycin (2 μM). To measure OCR and extracellular acidification rate in intact cells, cells were washed twice and incubated with freshly prepared assay medium (Seahorse XF Base Medium + 2 mM l-glutamine + 1 mM pyruvate + 10 mM glucose) for 30 min. The following compounds were injected in the order of oligomycin (2 μM), FCCP (0.75/1.35 μM), rotenone (1 μM) + antimycin A (2 μM). Cell count (per well) was determined by the number of nuclei stained with Hoechst (10 μg ml −1 ) and quantified by an Operetta High-Content Imaging System (PerkinElmer). OCR and extracellular acidification rate were normalized to cell count per well. For nutrient-dependent respirometry analysis, conditions of UK5099 (5 μM), BPTES (3 μM) and etomoxir (3 μM) were applied.

Whole-animal perfusion and tissue fixation
Fixative solution (2% paraformaldehyde, 2.5% glutaradehyde, 0.15 M cacodylate and 2 mM Ca 2+ ) was freshly prepared every time. Animals were anaesthetized by ketamine (200 mg kg −1 ) and xylazine (10 mg kg −1 ) delivered by intraperitoneal injection. Anaesthesia took effect after several minutes. The depth of anaesthesia was tested using tail pinch and paw prick, and it was ensured that the breathing did not stop. The body cavity was cut with scissors, and cut up the midline to the sternum. The heart was exposed and a needle was placed into the left ventricle and then the right atrium was snipped with iridectomy scissors. The animal was perfused for 30 s at the flow speed of 8 ml min −1 with Ringer's solution supplemented with 2% xylocaine and 1,000 U herapin. Then the valve on the pump was switched on to fixative solution and perfusion was carried out for 5-8 min. After continuous perfusion, mouse lung tumour was collected, placed in ice-cold fixative solution and fixed overnight in the fridge. Fixed tissues were washed three times with 0.15 M cacodylate buffer (2 mM Ca 2+ ) and stored in the same buffer at 4 °C until further processing.

Sample preparation for SBEM imaging
Post-fixed tissues of mouse lung tumour were washed in 0.15 M cacodylate buffer (2 mM Ca 2+ ). Tissues were stained for 1 h with 2% osmium and 1.5% potassium ferrocyanide in 0.15 M cacodylate buffer (2 mM Ca 2+ ). Tissues were washed five times (5 min each time) with double-distilled (dd)H 2 O, and then placed in filtered TCH buffer (0.05 g thiocarbohydrazide in 10 ml ddH 2 O) for 20 min at room temperature. Tissues were washed five times (5 min each time) with ddH 2 O and then stained with 2% osmium in ddH 2 O for 30 min at room temperature. Next, tissues were stained with 2% uranyl acetate overnight at room temperature and lead aspartate solution (0.66% (w/v) lead in 0.03 M aspartic acid) for 30 min in a 60 °C oven. Tissues were dehydrated in serial ice-cold ethanol (70%, 90% and 100%) and ice-cold acetone after washing in ddH 2 O. Tissues were then embedded in serial Durcupan resin (50%, 75%, 100%) and solidified in a 60 °C oven for 2 days.

SBEM
SBEM volumes were collected on a Zeiss Gemini 300 microscope equipped with a Gatan 3View 2XP microtome system. The volumes were collected at 2.5 kV using a 30-μm aperture, the gun in analytic mode, and the beam in high-current mode. Focal charge compensation with nitrogen gas was used to mitigate charging artefacts. The dwell time was 1 μs. The pixel size was either 5 or 6 nm, and the Z step size was always 50 nm. Following data collection, the images were converted to .mrc format and cross-correlation was used for rigid image alignment of the slices using the IMOD image processing package 40 .

MicroCT
MicroCT imaging was carried out on a Zeiss Versa 510 microscope. The wet lung specimens were imaged with no contrast staining in buffer at 40 kV, using 360° of specimen rotation and 801 projection images were collected. The pixel size was 11.5 μm. Specimens stained for SBEM and embedded in epoxy were imaged at 80 kV, using 360° of specimen rotation and 1,601 projection images were collected. The pixel size was between 5 and 8 μm.

3D visualization and analysis of SBEM images
Individual cells were segmented using Amira software (Thermo Scientific). Sequential SBEM images were processed with median filtering and contrast adjustment to enhance the appearance of intracellular space. Interactive thresholding was applied to distinguish cellular and extracellular space and generate binary images. Binary images were processed with the separate object algorithm to fragment cells without clear extracellular boundaries. The connected components algorithm was applied to create individual labels for each identified cell. Segmented individual cells in each stack were visualized with a voxelization rendering.

Machine-learning-based segmentation of nucleus and mitochondria in SBEM images
Our method segments nuclei (class 1) and mitochondria (class 2) from the background (class 0) in a 3D SBEM volume. The method takes individual slices from the volume (2D images) as input and predicts the segmentation for the slices independently. To carry out the segmentation, a deep CNN was used, which is a function that takes an image as input and outputs logits (or softmax confidence) corresponding to each class. The class with the highest response is chosen for each pixel to yield the segmentation map for each slice. Specifically, an encoderdecoder architecture based on U-Net 41 , with three modifications was chosen. First, ResNet 42 blocks (with 32, 64, 128, 256 and 256 filters) was used instead of standard convolutional blocks. Second, as mitochondria are small objects, we chose to limit spatial downsampling to one-eighth of the original image. To do so, the max-pooling operation was removed after the third layer. Third, an atrous spatial pyramid pooling 43 layer was used with strides 6, 12, 18 and 24 as the last layer of our encoder to increase the field of view of our filters. All weights use random uniform initialization, and we used leaky-ReLU activations. Our model is trained using the standard cross-entropy loss and is optimized using Adam 44 with β 1 = 0.9 and β 2 = 0.999. We used an initial learning rate of 1 × 10 −4 and decreased it to 1 × 10 −5 after 150 epochs for a total of 200 epochs. We use a batch size of 8 and resize each image to 768 × 768. We carry out random horizontal and vertical flips, rotations between [−20°, 20°], zero-mean additive Gaussian noise with standard deviation of 0.08, and crop sizes up to 90% of the image height and width as data augmentation. Each augmentation was applied with a 50% probability. Our model was trained on ten densely labelled images, and was validated on three additional labeled images that are not included in the training set. Training takes about 8 h on an Nvidia GTX 1080 GPU, and inference takes about 11 ms per 2D image. Our segmentation model was evaluated using the Sorensen-Dice index (DICE), the Jaccard index (also referred to intersection-over-union, IoU), precision and recall. DICE and IoU are metrics to gauge the similarity of two sets, commonly used to evaluate segmentation results. Our segmentation model achieves an average of 0.92 DICE score, 0.86 IoU, 0.916 precision and 0.927 recall, as compared to that of manual registration.

Mitochondrial motility and mitochondrial crista analysis
A total of 2 × 10 5 cells were plated per well into CELLview 4-compartment glass-bottom tissue culture dishes (Greiner Bio-One) 48 h before the imaging section, and maintained in a tissue culture incubator (37 °C in 5% CO 2 ). Cells were stained with TMRE (15 nM, Thermo Fisher) for 1 h to analyse mitochondrial motility, and with 10-N-nonyl acridine orange (100 nM, Thermo Fisher) for 1-3 h to analyse mitochondrial cristae. Live-cell imaging was carried out on the Zeiss LSM 880 with Airyscan using the alpha Plan-Apochromat 100×/1.46 Oil DIC M27 objectives. Images were deconvolved by Airyscan Processing in ZEN software. Image analysis was carried out using ImageJ (Fiji) in the order of: background subtraction, crop region of interest, adjust thresholds and measure parameters. Mitochondrial motility was analysed using a program developed in the laboratory of O.S.S. To measure mitochondrial motility, each field of view was imaged every 20 s, 10 times (yielding a total of 10 frames). Mitochondrial placement in frame 1 of each time series is identified as the reference point. The following 9 frames (n = 2, 3, 4…10) in the image set were then overlaid on the reference image, and the overlapping area between frame n and frame 1 was calculated and denoted as overlap_area(n). The non-overlapping mitochondria area between frame (n) and the reference image was denoted as travel_ area(n) and calculated with the formula: travel_area(n) = mitochondria_ area(n) − overlap_area(n). Mitochondrial motility was then calculated as the displacement ratio for each frame, ratio(n) = travel_area(n)/ overlap_area(n). We then fitted a linear regression to the ratio for each image frame (n > 1) with the time intervals (20 s, 40 s…180 s), and the regression line is set in the form of ratio = a (slope) × time intervals + b (intercept). The value of "a (slope)" for each time series of image sets is our readout of mitochondrial motility. Macros were designed for automated segmentation and quantification of mitochondrial crista structure and density according to previously described methods 19,21 .

Article
The original images used for classification of mitochondrial cristae were 2D images, and the consensus for classification was obtained from two.

Quantification of distance between mitochondria and nucleus in fluorescent images
Given the centre of a mitochondria, we want to find its minimum distance to the surface of a nucleus, which is segmented as an ellipse. The result can be obtained by solving a constrained optimization problem using Lagrange multipliers. The original ellipse of the nucleus is tilted by a degree and is centred at an arbitrary coordinate. To ease the computation, we translate and rotate the original ellipse to a standard ellipse that is centred at (0,0) with the major axis aligned with the x axis and the minor axis aligned with the y axis. The centre of the mitochondria is mapped to the new coordinate system accordingly. Through substitution, the gradient of the objective function that involves only y is a fourth-degree polynomial, which has at most four possible roots that satisfy the criteria. We then choose the coordinate (x,y) on the ellipse that has the minimal distance to the target nucleus.

Oil red O staining
The oil red O staining was carried out following the protocols of a commercial kit (StatLab, KTOROPT). Frozen tissues were fixed by O.C.T compound (Fisher HealthCare) and sectioned at the Translational Pathology Core Laboratory (UCLA). The frozen tissue sections were fixed in 10% formalin for 5 min, and then rinsed in distilled water. Slides were immersed in propylene glycol for 2 min, followed by immersion in preheated oil red O solution at 60 °C for 6 min. Slides were then washed in 85% propylene glycol for 1 min and running distilled water for 1 min. The nuclei were stained with modified Mayer's haematoxylin for 1 min and rinsed in distilled water for 1 min. The slides were mounted with aqueous medium and covered with a coverslip.

Statistics and reproducibility
Statistical analyses were carried out on GraphPad Prism 9 or R studio. Differences between groups were determined using unpaired two-tailed t-test or one-way ANOVA if more than two groups were compared. For treatment studies, Dunnett's test was used to compare every mean to a control mean. For non-treatment studies, Tukey's test was used to compare every mean to every other mean. Data are presented as mean ± s.e.m. unless specified otherwise. Numbers of biological replicates are indicated in the figure legends. All experiments were repeated in at least duplicate. No statistical methods were used to predetermine sample size.

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

Data availability
Source data for western blots are provided in Supplementary Fig. 1. Data that support the findings of this study have been deposited in the Cell Image Library (http://cellimagelibrary.org/groups/54862) or are available from the corresponding author upon reasonable request. Source data are provided with this paper.

Code availability
The code for measuring mitochondrial parameters in SBEM image sets and confocal fluorescence image sets have been deposited in Github (https://github.com/tiard/mito-networks-3d).