Dysregulated inter-mitochondrial crosstalk in glioblastoma cells revealed by in situ cryo-electron tomography

Significance

Glioblastomas (GBMs) are the most lethal primary brain tumors with limited survival. So far, it is hard to accurately discriminate between normal proliferating and distinctive tumor cells. Mitochondria are essential to GBMs and serve as potential therapeutical targets. Current knowledge about the ultrastructure of mitochondria from GBM cells is mainly derived from fluorescence and transmission electron microscopy, which suffers greatly from labeling or staining artifacts. Here, we utilize cryo-electron tomography, one of the state-of-the-art techniques, to quantitatively investigate nanoscale details of randomly sampled mitochondria in their native cellular context of GBM cells. Our findings decipher high-resolution inter-mitochondrial structural signatures and provide clues for diagnosis and therapeutic interventions for GBM and other mitochondria-related diseases.

Abstract

Glioblastomas (GBMs) are the most lethal primary brain tumors with limited survival, even under aggressive treatments. The current therapeutics for GBMs are flawed due to the failure to accurately discriminate between normal proliferating cells and distinctive tumor cells. Mitochondria are essential to GBMs and serve as potential therapeutical targets. Here, we utilize cryo-electron tomography to quantitatively investigate nanoscale details of randomly sampled mitochondria in their native cellular context of GBM cells. Our results show that compared with cancer-free brain cells, GBM cells own more inter-mitochondrial junctions of several types for communications. Furthermore, our tomograms unveil microtubule-dependent mitochondrial nanotunnel-like bridges in the GBM cells as another inter-mitochondrial structure. These quantified inter-mitochondrial features, together with other mitochondria-organelle and intra-mitochondrial ones, are sufficient to distinguish GBM cells from cancer-free brain cells under scrutiny with predictive modeling. Our findings decipher high-resolution inter-mitochondrial structural signatures and provide clues for diagnosis and therapeutic interventions for GBM and other mitochondria-related diseases.

Gliomas, the most prevalent solid tumors of the central nervous system, are rapidly fatal with an extremely decimal prognosis (1). Glioblastoma (GBM, World Heath Organization (WHO) grade IV), characterized by the highest aggressiveness, accounts for more than half of the entity (2). A long-appreciated hallmark of GBM is intra-tumor heterogeneity (3). The main cellular components of the bulk include heterogeneous glioblastoma stem cells (GSCs) and their differentiated counterparts. Remarkably, some GSCs are relatively quiescent and can evade treatments, including surgical resection followed by chemo-radiotherapy, probably due to their distinct genetic, metabolic, functional, and structural characteristics compared with proliferative GSCs and differentiated GBM cells (4). The proliferation of certain GSCs will propagate GBM growth after chemo-radiotherapy, resulting in the notoriously frequent recurrence of GBM (5). Over-resection or excessive chemo-radiotherapy treatment will inevitably harm normal proliferating cells, causing severe complications (6). Therefore, therapies with better precision based on variations among normal cells and different types of GBM cells are urgently needed.

The peripheral and protrusion regions, especially the latter, are probably invasive parts of GBM cells responsible for tumor proliferation, progression, communication, pathfinding, and therapy resistance (7, 8). Interestingly, mitochondria are shown being transported from brain cells, such as astrocytes and glial cells, to GBM cells via these regions as in vivo mechanisms to promote tumorigenesis (9), in addition to their well-documented transportation events between heterogeneous GBM cells (10, 11). Studies of the involvement of mitochondria in cancer biology date back to the 1950s, when Otto Warburg found that cancer cells prefer to utilize energy through glycolysis over oxidative phosphorylation, even in the presence of oxygen (12, 13). In addition to mitochondrial bioenergetics, other aspects of mitochondria, such as membrane dynamics, biogenesis and turnover, signaling events, and mitochondria-endoplasmic reticulum (ER) interactions, also have complex roles in tumorigenesis (14, 15). Mounting evidence suggests that mitochondria have been genetically altered, functionally dysregulated, and metabolically reprogrammed in GBM (16–22). However, the mechanisms underlying these variations remain poorly understood, and reliable direct evidence from a structural point of view is missing.

Current knowledge about the ultrastructure of mitochondria from GBM cells is mainly derived from fluorescence and transmission electron microscopy (TEM) (23, 24). Unfortunately, fluorescence microscopy primarily works on fixed cells, suffers from labeling artifacts, and lacks nanoscale details (25, 26). Traditional TEM yields mostly two-dimensional images and inevitably introduces structural artifacts through chemical fixation, dehydration, resin embedding, and heavy metal staining (24, 27, 28). As an emerging technique, cryo-electron tomography (cryo-ET) pristinely preserves native high-resolution mitochondrial structures through plunge freezing (29, 30). It allows direct visualization of organelles in their unperturbed physiological and pathological cellular context to extensive three-dimensional (3D) details (31–36).

However, these high-resolution 3D details from cryo-ET are at the expense of the small field of view during data collection. This results in biased observations from only a minuscule fraction of the large, intricately organized eukaryotic cell (37). Here, we greatly overcome the limitations of cryo-ET by randomly imaging hundreds of areas of the protrusion and peripheral regions of GBM cells and two types of cancer-free cells: astrocytes and microglia cells. Quantitative analyses of randomly collected tomograms demonstrate significant variations in mitochondria-centric organellar crosstalk features and intra-mitochondrial characteristics among cancer-free cell types and two primary GBM cell lines derived from two distinct patients. Our studies show the potential of utilizing cryo-ET as a biomarker technology to decipher unique high-resolution 3D structural signatures. This opens avenues for possible diagnosis and therapeutics for GBM and other mitochondria-related diseases.

Results

Cryo-ET on Mitochondria.

We first planted U251, a high-passage GBM cell line, on the Electron Microscopy (EM) grids and directly visualized their peripheral regions (SI Appendix, Fig. S1A). From our reconstructed 3D tomograms, we unambiguously identified intact double membrane-bound organelles with some cristae inside as mitochondria (SI Appendix, Fig. S1 B–D). Therefore, our approach of visualizing mitochondria in their native cellular context can faithfully capture their intact functioning structure without any noticeable detrimental artifacts that may damage organelle integrity and impair structural analysis. This approach did not need extra labeling or focused ion beam (FIB) milling, and thus dramatically increased the efficiency and facilitated large-scale data collection.

We then imaged the other five cell types with cryo-ET: G98 GSC (G98G), G98 differentiated cells (G98D), GBM8, normal human astrocytes (NHA), and microglial cells HMC3. G98G and GBM8 are patient-derived primary GSCs, which cooperate with differentiated GBM cells to promote tumor progression but have distinct transcriptomic profiles, proteomes and tumor microenvironment shaping mechanisms (38–41). For G98 cells, the G98G and serum-induced differentiated counterparts G98D were imaged to compare variations associated with GSC differentiation. HMC3, as a microglial cell line, is responsible for innate immune defense in the central nervous system (42). Besides immune functions, the microglial cells are also critical components of the GBM microenvironment, which supports GBM progression (43). NHA and HMC3 cells were imaged as cancer-free controls.

For each cell, ten to twenty tomograms were collected on randomly chosen areas on either the periphery or the protrusion region (SI Appendix, Figs. S2 A–F and S3), providing a relatively more comprehensive and less biased knowledge. In total, ~990 cellular areas were picked for imaging from ten NHA cells, ten HMC3 cells, ten G98G cells, ten G98D cells, ten GBM8 cells, and ten U251 cells. After discarding datasets that suffered from severe stage drift, the remaining 884 tilt series were reconstructed into 3D subcellular volumes called tomograms.

Characteristics of Intra-Mitochondria.

Out of 884 tomograms under investigation, 247 of them contained mitochondria. The distribution of mitochondrial numbers varied tremendously among tomograms (Fig. 1 A–D). The frequency of mitochondria-containing tomograms varied significantly among different cell types, with both primary GSCs G98G (37.2%) and GBM8 (34.1%) bearing substantially higher percentage of tomograms with full or partial mitochondria than the two cancer-free ones, NHA (16.3%) and HMC3 (17.6%) (Fig. 1E). Among the mitochondria-containing tomograms, G98G had markedly more percentage of tomograms with densely distributed mitochondria than NHA (Fig. 1F), suggesting more coupled local functions of mitochondria in G98G than NHA. Unlike G98G, GBM8 had no noticeable difference in mitochondrial distribution pattern compared with NHA, possibly due to different patient statistics between G98G and GBM8 (SI Appendix, Table S1). Furthermore, the two GSCs possessed increased global mitochondrial density and larger global mitochondrial volume fractions than the cancer-free ones (Fig. 1 G and H). This observation indicated a variable mitochondrial biogenesis and clearance equilibrium in GSCs compared with cancer-free cells. The distribution pattern of mitochondria didn’t vary much between G98G and G98D (Fig. 1 I and J), and mitochondrial global density and volume fraction slightly increased during serum-induced differentiation of G98 cells, suggesting no drastic changes in mitochondrial biogenesis during G98 differentiation (Fig. 1 K and L).

Fig. 1.

Mitochondrial distribution in different cell types. (A) Slice of a typical tomogram from NHA without mitochondria. (B) Slice of a typical tomogram from HMC3 with a single mitochondrion. (C) Slice of a typical tomogram from G98G with 2 to 3 mitochondria. (D) Slice of a typical tomogram from GBM8 with ≥4 mitochondria. Mitochondria are highlighted with red arrows. (E) Frequencies of mitochondria-containing tomograms for two GSCs and two cancer-free cell types. Light colors represent mitochondria-free tomograms, and dark colors represent mitochondria-containing ones. Same for panel (I). (F) Mitochondrial distribution pattern for two GSCs and two cancer-free cell types. Light to dark colors represent single-mitochondrion tomograms, 2 to 3-mitochondria tomograms, and ≥4-mitochondria tomograms. Same for panel (J). (G) Global mitochondrial density for two GSCs and two cancer-free cell types. (H) Global mitochondrial volume fraction comparison for two GSCs and cancer-free cell types. (I) Frequencies of mitochondria-containing tomograms between G98G and G98D. (J) Mitochondrial distribution pattern for G98G and G98D. (K) Global mitochondrial density between G98G and G98D. (L) Global mitochondrial volume fraction comparison between G98G and G98D.

Next, we inspected the distributions of individual mitochondrial size, shape, and cristae levels (Fig. 2 A–D). Mitochondria from GSCs and HMC3 were larger in volume than those from astrocytes (Fig. 2E), indicating variations in volume-associated properties such as osmotic balance (44). We then measured central section area (CSA) and long-short axis ratio (LSAR), which served as biomarkers of cell health and indicators of mitochondrial shape, local signaling events, and biochemical reactions (45, 46). Our quantifications showed that two GSCs and HMC3 contained mitochondria with significantly larger CSA and LSAR than astrocytes (Fig. 2 F and G). Alterations in local Ca²⁺ signaling and mitochondrial depolarization may cause this difference (47). Interestingly, though similar in mitochondrial CSA and LSAR (Fig. 2 F and G), mitochondrial volumes from G98G cells were larger than those from HMC3 cells (Fig. 2E), which reflects their distinct dynamic status. Mitochondria from G98G, GBM8, and HMC3 were greater in size (volume and CSA) and more elongated in shape than NHA (Fig. 2 E–G), suggesting more frequent interactions between the mitochondria and their cellular surroundings.

Fig. 2.

Intra-mitochondrial characteristics. (A) Slices of tomograms with mitochondria of different volumes. The Left image is from HMC3; the Right image is from GBM8. (B) Slices of tomograms with mitochondria of different CSAs. The Left picture is from NHA; the Right image is from HMC3. (C) A slice of tomogram from GBM8 with mitochondria of different LSARs. The long axis and one of the short axes are depicted with red lines for the two elongated mitochondria. (D) Slices of tomograms with mitochondria bearing different cristae levels. The cristae level represents the percentage of mitochondrial volumes occupied by cristae. Tomograms of cristae levels 1 and 2 are from G98G. Tomogram with cristae level 3 is from HMC3. Tomogram with cristae level 4 is from G98D. Mitochondria are highlighted with red arrows in all slices of tomograms. (E) Mitochondrial volume statistics for two GSCs and two cancer-free cell types. Black stars denote the mean. The same scheme is used throughout all figures. (F) Mitochondrial CSA statistics for two GSCs and two cancer-free cell types. (G) Mitochondrial LSAR statistics for two GSCs and two cancer-free cell types. (H) Mitochondrial cristae level statistics for two GSCs and two cancer-free cell types. (I) Mitochondrial volume statistics for G98G and G98D. (J) Mitochondrial CSA statistics for G98G and G98D. (K) Mitochondrial LSAR statistics for G98G and G98D. (L) Mitochondrial cristae level statistics for G98G and G98D.

Cristae is another critical aspect of mitochondrial ultrastructure (48). Abnormal mitochondrial cristae indicate dysfunctions in mitochondrial proteins such as respiratory apparatus and optic atrophy 1 and are directly associated with diseases like dominant optic atrophy and Huntington’s disease (49–51). Compared with NHA, two GSCs, especially GBM8, had an apparent reduction in the percentage of the mitochondrial volume occupied by cristae (cristae level), and HMC3 had an even lower percentage (Fig. 2H). This observation suggested that mitochondria from two GSCs and HMC3 had a larger percentage of volume taken up by the mitochondrial matrix, indicating more active enzymatic reactions necessary for biosynthesis. Since two GSCs had higher mitochondrial density and significantly larger mitochondrial size than NHA (Figs. 1G and 2 E and F), the cristae volumes from both GSCs were still considerably greater than NHA despite their cristae levels being lower (Fig. 2H). This observation agrees with the previous reference, which indicates that GSCs primarily utilize oxidative phosphorylation, while astrocytes rely more on aerobic glycolysis (52). Our quantifications showed no significant differences between G98G and G98D in mitochondrial volume, CSA, LSAR, and cristae level (Fig. 2 I–L), meaning that these intra-mitochondrial features seem not ideal to serve as indicators for serum-induced G98 differentiation.

Inter-Mitochondrial Crosstalk.

Mitochondria can form functionally coupled local or global networks in various cell types under physiological and pathological conditions (15, 53–57). Our above analyses showed that G98G and HMC3 have markedly more local cellular regions with densely distributed mitochondria than NHA (Fig. 1F). To more accurately describe how two mitochondria are close to each other, we measured the degree of mitochondrial clustering (DOMC) in 3D tomograms (SI Appendix, Fig. S4 A–E). Our results showed that two GSCs and HMC3, especially the G98G, exhibited generally more tightly packed patterns of neighboring mitochondria in their mitochondria-rich regions than NHA (SI Appendix, Fig. S4 A, B, and D), which helped explain differential distributions of Adenosine Triphosphate (ATP), reactive oxygen species, and Ca²⁺ signaling in these cellular regions with distinct local metabolic activities and functional capacities (55, 58). Regarding serum-induced G98G differentiation, the statistical analysis showed that G98G and G98D had no significant difference in DOMC (SI Appendix, Fig. S4E), suggesting no apparent changes in the mitochondrial clustering. Thus, we performed a more detailed mitochondria-wise analysis as follows.

Mitochondria from two GSCs and HMC3 had more neighboring mitochondria in their vicinity (<200 nm) (Fig. 3 A and B). Interestingly, mitochondria from G98D also had more neighboring mitochondria than G98G despite their DOMC staying roughly the same (Fig. 3C and SI Appendix, Fig. S4E), suggesting more inter-mitochondrial communications during GSC differentiation. When two mitochondria were spatially close enough, inter-mitochondrial junctions (IMJs), a homotypic membrane contact site, occurred in the tomograms as protein-mediated junctions (Fig. 3 D–G and Movie S1) (59, 60). Both GSCs had substantially more IMJs per unit volume than cancer-free cells (Fig. 3H), consistent with their higher mitochondrial densities (Fig. 1G) and more neighboring mitochondria (Fig. 3B). U251 and G98D had 1.5- and 1.6-fold increases in the IMJ density compared with G98G, respectively (Fig. 3I). Protein densities in GBM cells were recognized at IMJs as functional tethers connecting both mitochondrial outer membranes (MOMs) of appositional mitochondria (Fig. 3 C–F and Movie S1).

Fig. 3.

Mitochondrial connectivity analyses. (A) Definition of mitochondrial vicinity zone. (B) Statistics of the neighboring mitochondria in the mitochondrial vicinity zones for two GSCs and cancer-free cell types. (C) Statistics of neighboring mitochondrial numbers for G98G and G98D. (D) A slice of a tomogram from GBM8 shows a typical circular IMJ in the red square. The zoom-in view of the IMJ is shown in the Right image. (E) Annotation of D. Protein-like densities connect the two MOMs. (F) A tomogram slice from U251 shows a typical parallel IMJ in the red square. The zoom-in view of the IMJ is shown in the Right image. (G) Annotation of F. Configurations of the protein densities connecting both MOMs appear quite different from those shown in E. An NTLS is identified as connecting to another mitochondria. (H) The density of IMJs across two GSCs and two cancer-free cell types. (I) The density of IMJs among G98G, G98D, and U251. (J) Statistics of the IMJ distances for two GSCs and two cancer-free cell types. (K) Statistics of the IMJ distances for G98G and G98D. (L) Comparison of the ratio of parallel to circular IMJs. Light colors indicate circular IMJs, and dark colors indicate parallel IMJs. (M) A slice of tomogram from G98D showing a typical parallel IMJ (in the red square). (N) A slice of tomogram from G98G showing a typical circular IMJ (in the red square). Annotation colors: cyan, MOM; violet, MIM; blue, cristae membranes; yellow, ER; orange, single-membrane bound vesicles; white, mitochondrial metal cluster; gold, protein-like-densities; green, microtubules; pink, NTLS membranes; red, NTLS. The same annotation color strategy was utilized throughout the manuscript unless specified otherwise.

Similar protein densities were also observed at the mitochondria–ER (Mito-ER) and mitochondria–vesicle contact sites (Mito-vesicle or Mito-ves for abbreviation) (Fig. 3 D–G), as reported previously (59, 61–64). Both GSCs had shorter IMJ distances than NHA (Fig. 3J). Though there were no significant variations between G98G and G98D in the IMJ distance (Fig. 3K), there is a substantial difference in the type of IMJ between G98G and its differentiated counterpart (Fig. 3L). G98D more likely had IMJ of a “parallel” type where the two appositional OMMs had relatively larger membranous surface areas in proximity (Fig. 3M). In contrast, G98G more likely had “circular” IMJ, where only a tiny fraction of both OMMs were in close contact (Fig. 3N). Different IMJ types indicated different protein configurations between the two GBM cells (Fig. 3 D–G, M, and N), suggesting different communication mechanisms between their mitochondria.

Mitochondrial Nanotunnel-Like Structures (NTLS).

In addition to IMJs, our results showed mitochondria extended NTLS, probably for communicating with other mitochondria or cellular components over longer distances (Figs. 3 F and G and 4 A–H, SI Appendix, Fig. S5 A and B, and Movies S1 and S2). NTLS occurred at different frequencies (Fig. 4 I and J) in the various cell types we imaged. G98G had substantially more NTLS than NHA and HMC3 (Fig. 4I). Cancer cells, such as G98G, G98D, and U251, had comparable NTLS (Fig. 4J). Surprisingly, GBM8 had far less NTLS than the other cancer cells, including the other GSC G98G (Fig. 4 I and J). This variation may be due to the heterogeneous nature of GSC in general or caused by distinct statistics of these two GSCs derived from different patients (SI Appendix, Table S1).

Fig. 4.

Mitochondrial NTLSs across cell types. (A) A Slice of tomogram showing a typical NTLS-1 from NHA. NTLS-1 is shown in the red square. (B) Zoom-in view of NTLS-1 from A. A montage of central slices of the protruding mitochondrion is shown. (C) Annotation of the NTLS-1 from B. This NTLS-1 from NHA is an extension of MOM. (D) A slice of the tomogram shows a typical NTLS-2 in GBM8. (E) Annotation of D. Cristae structures are seen. (F) A slice of the tomogram shows a typical NTLS-3 from the U251 cell. (G) Annotation of F. (H) Zoom-in view of NTLS-3. Protein aggregates, rather than cristae, are found in NTLS-3. (I) The density of NTLSs across two GSCs and two cancer-free cell types. (J) The density of NTLSs for G98G, G98D, and U251. (K) Distribution of three different types of NTLS for two GSCs and two cancer-free cell types. (L) Distribution for G98G, G98D, and U251. (M) Central slices of tomogram from G98G with mitochondria extending two NTLS-3. This image is a montage of central slices from different Z-heights of the tomogram. (N) A slice of tomogram from G98D with two NTLS-3 and one connecting the two mitochondria. (O) A slice of the tomogram from HMC3 with two NTLS-2.

Besides frequency, NTLS were also identified in various forms (Fig. 4 K and L). NTLS Type-1 (NTLS-1) was merely simple extensions of the MOM, like the one shown in NHA (Fig. 4 A–C and Movie S3). NTLS Type-2 (NTLS-2) had many cristae structures in the nanotunnel (Fig. 4 D and E). The detailed analysis further showed that another typical NTLS in U251 (NTLS Type-3, NTLS-3) lacked cristae structures, only partially covered with membrane, and with protein-like densities accumulating in the nanotunnel (Fig. 4 F–H and Movies S1 and S2). The distributions of the three types of NTLS varied in different cells. GBM cells, such as G98G, G98D, and U251, primarily had NTLS-3, while HMC3 mainly had NTLS-2 (Fig. 4 K–O). NTLSs in different cell types may be different structures or the same structure at different cell stages.

Since NTLS-3 was the predominant form of NTLS in GBM cells (Fig. 4 K and L), we next tried to deduce the possible mechanisms associated with NTLS-3 in U251 cells (Fig. 5 A–C). Interestingly, the protrusion regions had a threefold increase in NTLS-3 per unit volume than peripheral areas (SI Appendix, Fig. S5C). Moreover, 21.8% of the mitochondria had NTLS-3 in the protrusion region, while only 8.2% of mitochondria in the periphery did (SI Appendix, Fig. S5D). This evidence suggested a possible direct involvement of NTLS-3 in extending protrusion regions. The average length of NTLS-3 is 543.9 ± 113.0 nm, significantly shorter than most of the nano-tunneling structures at lengths of approximately 1 to 30 µm from cardiomyocytes, skeletal muscles, and kidney cells (65, 66), suggesting that they are different structures. Indeed, 95.7% of NTLS-3-containing mitochondria were surrounded by microtubules within 200 nm. In comparison, only 55.5% of NTLS-free ones were (Fig. 5D). These microtubules piled up close to the nanotunnel regions more than the main body of the mitochondria (Fig. 5E). All these indicated a direct role of microtubules in the formation and extension of NTLS-3 (Movie S2).

Fig. 5.

Mitochondrial NTLS-3 in U251 cells. (A) A slice of the tomogram shows a typical NTLS-3 in U251. (B) Annotation of A. NTLS-3 is shown in the red square. (C) Zoom-in view of NTLS-3. Cytoskeletons and vesicles are removed for clarity. ER is in the vicinity of NTLS-3 but does not encircle it. (D) Comparison between mitochondria with and without NTLS-3 regarding microtubules in the mitochondrial vicinity. (E) Statistics of microtubule distribution in the mitochondrial vicinity. (F) A slice of a tomogram shows a mitochondrion with both NTLS-3 and IMJ. (G) Annotation of F. Both NTLS-3 and IMJ are shown in the red square. (H) Zoom-in view of the mitochondrion with both NTLS-3 and IMJ. Cytoskeletons, ER, and vesicles were removed for clarity. IMJ is in the red square. (I) Zoom-in view of IMJ. (J) Correlation between NTLS-3 and IMJ. (K) Correlation between NTLS-3 and mitochondrial CSA. (L) Correlation between NTLS-3 and mitochondrial volume. (M) Correlation between NTLS-3 and Mito-ER contacts. (N) Correlation between NTLS-3 and Mito-vesicle contacts. Annotation colors: light green, actin filaments.

Mitochondria fission also showed a constricted tubular structure between two daughter mitochondria, and the assembly of this structure usually required extensive Mito-ER contacts and the involvement of branched actin filaments (67–70). However, ER was found neither encircling NTLS-3 nor forming extensive Mito-ER contacts close to NTLS-3 in U251 (Fig. 5 A–C and SI Appendix, Fig. S5B). Moreover, bundled or individual actin filaments, rather than branched ones, were abundant around NTLS-3 (Fig. 5 A, B, F, and G and SI Appendix, Fig. S5B). All these indicated that NTLS-3 in U251 may not be the tubular structure during mitochondria fission but a newly identified mitochondrial structure that may participate in tumorigenesis.

NTLS-3 and IMJ can occur on the same mitochondrion (Figs. 3 F and G and 5 F–I and Movie S1). Interestingly, mitochondria with NTLS-3 in U251 formed significantly fewer IMJs than those without NTLS-3 (Fig. 5J). The negative correlation between NTLS-3 and IMJ was unlikely the result of smaller mitochondrial surface areas, since no significant size variations were found between mitochondria with and without NTLS-3 (Fig. 5 K and L). To our surprise, the Mito-ER and Mito-vesicle contact numbers did not decrease on mitochondria with NTLS-3 (Fig. 5 M and N), which indicated a distinct mechanism of forming inter-mitochondrial crosstalk than establishing the Mito-ER and Mito-vesicle contacts.

Mito-ER and Mito-Vesicle Crosstalk.

Mito-ER contact in GBMs has been linked with the sensitivity of cancer cells to lymphocyte-mediated killing (71). We quantified the Mito-ER contact sites from our tomograms. For every mitochondrion, Mito-ER contacts were visually inspected and identified (Fig. 6 A and B) when a subdomain of the ER (technically, these structures are ER or ER-like membranous structures, but for clarity purposes, we called these membranes ER in this paper) was found close enough (<50 nm) to the MOM (72). Our results showed that mitochondria from two GSCs and HMC3 formed significantly more Mito-ER contact sites than those from the control astrocytes (Fig. 6C), in agreement with the previous findings where abundant mitochondria-associated membranes were observed in GBM cells (10, 23, 24). Moreover, G98G formed significantly more Mito-ER contact sites than GBM8 (Fig. 6C) and G98D (Fig. 6D). This suggested that compared with GBM8 or G98D, G98G might have distinct Mito-ER associated pathways such as lipid metabolism and Ca²⁺ signaling. They may also maintain and regulate their mitochondrial networks differently (73).

Fig. 6.

Mito-ER and Mito-vesicle contact sites. (A) Slice of tomogram from GBM8 cell. In the central image, a Mito-ER contact is shown in the yellow square, and a Mito-vesicle contact is shown in the orange square. Zoom-in views of both contact sites are shown in images on both sides, with the ellipsoidal circles indicating the positions of the contacts. (B) Annotations of A. (C) Statistics of the Mito-ER contact numbers for two GSCs and cancer-free cell types. (D) Statistics of the Mito-ER contact numbers for G98G and G98D. (E) Statistics of Mito-vesicle contact numbers for two GSCs and cancer-free cell types. (F) Statistics of the Mito-vesicle contact numbers for G98G and G98D. Annotation colors: gold, Mito-ER contact; light orange, Mito-vesicle contact.

In addition to the Mito-ER contacts, other single-membrane-bound organelles were shown to make contacts with mitochondria (Fig. 6 A and B), and these contacts also have implications for cancer (74, 75). HMC3 had significantly fewer Mito-vesicle contact sites compared with the other three cell types (Fig. 6E). Also, there were no significant changes in Mito-vesicle contacts between G98G and G98D (Fig. 6F). All these implied distinctive patterns of signal integrations and functional regulations between mitochondria and other single-membrane-bound organelles between HMC3 and the other three cell types.

Cell-Type Predictions.

Since statistically significant variations were obtained regarding mitochondria-focused intra- and inter-organellar features among GSCs, cancer-free cells, and differentiated glioma cells under scrutiny, we decided to determine whether accurate cell-type predictions can be achieved based on our quantified tomographic features. We collapsed several mitochondrial crosstalk features (Figs. 1 F, G, J, and K and 3 B and C and SI Appendix, Fig. S6 A–D) into mitochondrial connectivity score (MitoCS) (SI Appendix, Fig. S6 E–G). We then combined features depicting NTLS and microtubules in the mitochondrial vicinity (Fig. 4 I–L and SI Appendix, Fig. S7 A–E) into mitochondrial nanotunnel score (MitoNS) (SI Appendix, Fig. S7 F–H). In addition to the above mentioned, mitochondrial volume (Fig. 2 A, E, and I), CSA (Fig. 2 B, F, and J), LSAR (Fig. 2 C, G, and K), cristae level (Fig. 2 D, H, and L) and mitochondrial metal cluster (MC) (SI Appendix, Fig. S8 A–D) were also considered in the predictive modeling.

We divided quantified tomographic features into two categories. The first category focused on the intra-mitochondria characteristics. We performed multilayer perceptron (MLP)-based machine learning analyses with these intra-mitochondria features (76) and achieved a 75.6% accuracy in the cell-type prediction (cancer or cancer-free) in one of the analyses (Fig. 7A). We then switched to the second category of inter-mitochondrial/organellar features for prediction purpose. An overall 93.9% accuracy was achieved in one of the analyses (Fig. 7B). After this, we combined both intra-mitochondrial and inter-organellar features to perform predictions, and a 100% prediction accuracy was achieved in one of the analyses (Fig. 7C). The MLP-based neural network (Fig. 7D), the receiver operating characteristic (ROC) curve (Fig. 7E), and ranking of feature importance (Fig. 7F) were also shown.

Fig. 7.

Predictive modeling based on quantified mitochondrial features. (A) The prediction results with intra-mitochondrial characteristics only (intra- features). This model is trained with the square root of the mitochondrial volume, the square root of CSA, LSAR, cristae level, and MC. The overall prediction accuracy is 75.6%. (B) The prediction results with mitochondrial connectomic features (inter- features). This model is trained with MitoCS, MitoNS, Mito-ER, Mito-vesicle, and Mito-microtubule. The overall prediction accuracy is 93.9%. (C) The prediction results with both intra-mitochondrial and inter-organellar measurements (both features). The overall prediction accuracy is 100%. (D) A standard supervised MLP-based neural network trained with both intra-mitochondrial and inter-organellar features. (E) A typical ROC curve for the model shown in D. (F) The predictors are ranked from most to least important for the model in D. (G) Comparison of prediction accuracy for intra-, inter-, and both features.

Since “cancer or cancer-free” predictions were successfully made, we then performed similar MLP-based analyses on G98 mitochondrial features to make “GSC or differentiated cells” predictions (SI Appendix, Fig. S9 A–G). With intra-, inter-, and both features, prediction accuracies of 66.0%, 73.6%, and 87.2% were achieved in one of the analyses, respectively (SI Appendix, Fig. S9 A–C). Specifically, MitoCS and MitoNS contributed significantly to the cancer or cancer-free cell-type predictions (Fig. 7F), and MitoCS and MitoMC contributed substantially to the GSC or differentiated cells predictions (SI Appendix, Fig. S9F). Interestingly, the inter-features behaved better than the intra-features in the cell-type predictions (Fig. 7G and SI Appendix, Fig. S9G). Overall, the result of predictive modeling proved the necessity to quantitatively investigate nanoscale details of mitochondria in their native cellular context.

Discussion

GBM is an incurable disease in dire need of novel, reliable classification methods for its heterogeneous cancer cells to improve both therapeutics and diagnosis. Here, we applied in situ quantitative cryo-ET to study a collection of randomly sampled mitochondria in their native cellular environment from protrusion and peripheral regions of different cancer cells to nanoscale details. Our results show significant differences in mitochondrial features not only between GSCs and cancer-free brain cells but also between G98G and its differentiated counterparts. Furthermore, we reveal substantial variations in the connectomic features between mitochondria and other organelles (including mitochondria themselves), and these variations can be used to construct machine learning models with reasonably good classification and predictive capacity.

The IMJ structures are proposed to be part of the dynamics of mitochondria and possibly contribute to the pathologies of cells (60, 77), however, their involvement in GBM has never been explored. Previous studies indicated that mitoNEET, a 2Fe-2S outer mitochondrial membrane protein, directly formed IMJs and may contribute to cancer progression (77, 78). Our results show that IMJs can occur in different morphologies characterized by different distances, types (the parallel IMJ and the circular IMJ) and various protein configurations in different GBM cells, possibly involving mitoNEET and other unidentified proteins. Besides IMJs, our results show that the frequency of contacts between mitochondria and other cellular membranes, such as ER and vesicles, also varies dramatically between cancer-free brain cells and GSCs, indicating an alteration in general mitochondrial contactology in GBM cells. Our work opens avenues to explore the high-resolution in situ structures of the protein complexes of IMJs, such as mitoNEET, and the detailed mechanisms underlying IMJs in GBM in the future.

The distributions and structural details of the nanotunnels are successfully utilized to classify different cell types, proving their diagnostic potential. The NTLS, especially NTLS-3 we found mainly in G98G, G98D, and U251, appear strikingly different from any mitochondrial nanotunnels reported previously. NTLS-3 contains protein aggregate-like densities, which may result from phase separation. They are partially covered by membranes and do not have visible cristae structures. Further analyses need to be conducted to determine the molecular identities and structures of these NTLS-3 components. Moreover, NTLS-3 appears to be strongly correlated with the presence of nearby microtubule pile-ups, which concentrate more in the vicinity of the nanotunnel than the main body of the extending mitochondria. Whether microtubules are the causal factor for NTLS-3 formation needs further validation.

Neuron cells, including NHA, HMC3, GBM8, G98, and U251, are ideal for directly growing on EM grids to visualize their protrusion and peripheral regions. However, whether the protrusion regions identified on the EM grids share the same functionalities as the invasive tunneling nanotubes and tumor microtubules requires further substantiation. Furthermore, to obtain a more in-depth understanding of the observed structural abnormalities of mitochondria, functional studies of the mitochondria in GBM cells are essential and need to be performed in the future.

Although quantitative analyses of mitochondria from the protrusion regions of the cells are informative enough to draw some important conclusions, future attempts with the cryo-FIB milling method to intensively inspect mitochondria from central cellular regions in an unbiased fashion are equally essential to show the complete picture. Our in situ quantitative cryo-ET of randomly sampled mitochondria has the potential of becoming a routine method to derive accurate high-resolution 3D models of mitochondria and mitochondrial connectomics in their native cellular context, providing clues not only to the therapeutics and diagnosis of GBM but also to other mitochondria-related diseases.

Materials and Methods

For full details, please refer to SI Appendix, Materials and Methods.

Cell Culture.

GBM8 and G98G cells were patient-derived GSCs isolated directly from GBM tumor tissue. G98G and GBM8 cells were cultured strictly in a serum-free medium. G98D cells were derived from G98G cells by culturing in a different medium comprising Dulbecco's Modified Eagle Medium F12 (DMEM/F12) with 10% fetal bovine serum for 4 to 7 d.

Electron Microscopy Grid Preparation, Data Collection, and Reconstruction.

Cells were plated on EM grids (Quantifoil Au R2/1, 200 meshes) for 48 h before being plunge-frozen into liquid ethane. Tilt series were acquired at a 3° interval from −60° to +60° and aligned and reconstructed using IMOD (79, 80), followed by binning by two in EMAN2 (81) and denoising using the nonlinear anisotropic diffusion tool in IMOD (82).

Mitochondrial Tomogram Segmentation and Quantification.

Visualization and segmentation of the binned and denoised mitochondrial tomograms were performed carefully, mainly using Amira software (Thermo Fisher Scientific, USA) (83). Quantifications of mitochondrial features were mostly performed using Amira software.

Statistical Analyses

Statistical analyses were performed with IBM SPSS Statistics (Version 25). Chi-square tests, nonparametric Kruskal–Wallis tests, one-way ANOVAs, Student t tests, and nonparametric Mann–Whitney tests were performed for different variables describing mitochondrial structural features in different cell types.

Predictive Modeling.

We built MLP-based supervised machine learning models in IBM SPSS Statistics (Version 25) to predict cell types (76). In the first set of predictions, each mitochondrion has a cell-type label of cancer-free (NHA, HMC3) or cancer (G98G, G98D). In the second set of predictions, each mitochondrion has a cell-type label of stem cell (G98G) or differentiated cell (G98D). For both sets of predictions, the intra-mitochondrial and inter-organellar features were used as predictors. ROC curves were generated, and the importance of predictors was ranked.

Supplementary Material

Appendix 01 (PDF)

Movie S1.

3D tomogram and annotation of densely distributed mitochondria from U251 cell. IMJ and mitochondrial nanotunnel are observed.

Movie S2.

3D tomogram and annotation of mitochondrial nanotunnel from U251 cell. Microtubule filaments are observed in the vicinity of the nanotunnel.

Movie S3.

3D tomogram and annotation of mitochondrial nanotunnel from NHA.

Data, Materials, and Software Availability

Representative tomograms (binned by 2 or 4) of mitochondria from different cell types have been deposited in the Electron Microscopy Data Bank as entries: EMD-39024 (84) (Fig. 2C), EMD-39023 85 (Fig. 3M), EMD-39012 (86) (Fig. 3N), EMD-39021 (87) (Fig. 4 A–C), EMD-39015 (88) (Fig. 4O), EMD-39019 (89) (Fig. 5 F–I). All other data needed to evaluate the conclusions in the paper are present in the paper and supporting information.