Bone Metastasis Initiation Is Coupled with Bone Remodeling through Osteogenic Differentiation of NG2⁺ Cells

Bone is frequently affected by metastasis of various cancer types (1–4), and bone metastases may further spread to multiple other organs (5). The diagnosis of bone metastasis often relies on severe symptoms including pain and pathologic fractures, which are driven by the vicious cycle between metastatic cells and bone-resorbing cells, osteoclasts (6–8). Specifically, cancer cells cooperate with osteoblasts to activate osteoclasts through multiple mechanisms (2, 4, 9). The bone resorption by osteoclasts releases a number of different growth factors from the bone matrix, which in turn promote tumor growth and invasion. Based on this knowledge, current treatments target osteoclasts to slow down the vicious cycle and mitigate symptoms. However, further research is urgently needed to completely cure bone metastasis and prevent it from further disseminating to other organs.

Recent research has begun to unveil a prolonged period of the asymptomatic phase. Before osteoclasts are recruited and activated, the perivascular and osteogenic niches may interact with metastatic seeds and regulate the dormancy and proliferation of disseminated tumor cells (DTC), respectively (10–14). In particular, we have shown that the osteogenic cells and cancer cells can form heterotypic adherens junctions (hAJ) and gap junctions, which mediate the activation of mTOR signaling and calcium signaling in cancer cells, respectively (11, 12). Furthermore, the interaction with osteogenic cells also elicits global epigenomic reprogramming of cancer cells, leading to an increase of phenotypic plasticity and empowering them for tertiary metastasis (5, 15). However, the relationship between the perivascular niche and osteogenic niche remains elusive. One hypothesis is that they cooperate to initiate the progression from isolated DTCs to osteolytic metastases. Testing this hypothesis will shed light on the etiology of metastatic recurrences, which often happen years to decades after the removal of breast tumors.

Bone is a highly dynamic organ even in adults (16). Under both homeostatic and pathologic conditions, bone undergoes constant turnover (17). This process involves sequential migration, differentiation, proliferation, and cell–cell interactions of various cell types in the specialized bone remodeling compartments (BRC), including mesenchymal stem cells (MSC) and endothelial cells (18, 19). Interestingly, there is a close interaction between endothelial cells and MSCs in the bone. Perivascular cells are considered a major source of bone marrow MSCs (20), and endothelial cells appear to drive the differentiation of MSCs toward the osteoblastic lineages (21). After the initiation of bone remodeling, MSCs are recruited to BRCs via direct migration or circulation and coordinate both the resorption of old bone and formation of new bones through a precisely regulated differentiation process (18). Conceivably, those cellular activities lead to disturbance in the bone microenvironment (BME) and might consequently alter DTC fate and kinetics of bone metastasis colonization. Herein, we tested this hypothesis by examining the connection between fracture-incurred bone remodeling and metastasis initiation.

To investigate the impact of bone remodeling on the progression of DTCs, we adopted a spontaneous metastasis model based on the subcutaneous transplantation of Lewis lung carcinoma (LLC1) cells (22, 23). The C57BL/6 background of this model enabled usage of several syngeneic mouse strains in which different subsets of bone cells can be genetically ablated or modified. Eighty percent of tumor-bearing animals generated spontaneous bone metastases, which is substantially more frequent than other C57BL/6 cell line models we had examined (i.e., <20% for TRAMP-C1, PYMT-E, PYMT-M, and EO771 cells). Furthermore, expression of GFP and firefly luciferase in LLC1 cells did not result in immunogenic rejection of the cells or loss of these markers during tumor progression, which posed barriers in other models (24). Taken together, these advantages provided an unprecedented opportunity to examine how perturbation of specific BME components affects microscopic metastases that spontaneously occur in immunocompetent hosts.

We used two approaches to introduce pathologic bone fractures and stimulate consequent bone remodeling: drilling and bending, both of which have been used to study the wound-healing process of bone (25, 26). Source tumors were implanted subcutaneously and reached 1 cm in diameter and approximately the same weight (Supplementary Fig. S1A; numeric values of all figures are provided in Supplementary Raw Data). Tumor resection and fracturing of right femur bones were then sequentially carried out on the same day (Fig. 1A). Drilling and bending significantly increased metastasis frequency and tumor burden on the injured bone (Fig. 1B and C), but not the contralateral unwounded bones (Fig. 1C). The timing of bone injuries relative to resection of source tumors in these experiments indicated that the increased metastases were derived from DTCs already homing to bone.

Next, we examined the distribution of metastatic lesions along the femoral bone relative to the injury sites. Without pathologic bone fractures, metastases predominantly localize to metaphyseal regions at the two ends of long bones. However, the bone fractures significantly skewed the distribution toward the injury sites (Fig. 1B). Interestingly, although the less invasive drilling surgery only introduces an injury of 0.7 mm in diameter, the impact appeared to spread to adjacent regions across the entire bone, similar to the pattern observed in invasive bending models (Fig. 1D). This may be related to the fact that fractures often stimulate regional rather than local bone remodeling (27). Confocal microscopy and flow cytometry revealed an enrichment of GFP⁺ cancer cells within the fractured area (Fig. 1E–G; Supplementary Fig. S1B). An increased proportion of cancer cells at the fracture site exhibited positive Ki-67 staining, whereas tumor cells in other parts of bone remained negative for Ki-67 staining (Fig. 1F). Indeed, Ki-67⁺ cancer cells on average localized more closely to the fracture site (Supplementary Fig. S1C), and there is an inverse correlation between tumor burden and distance to the fracture site (Supplementary Fig. S1D), suggesting that fractures reprogram the BME to promote proliferation of metastatic cells.

Bone fracture induces inflammation and bone repair. It was reported that an inflammatory environment can promote the outgrowth of cancer cells (28). To test this possibility, we performed immune cell profiling on drilled or bent areas of fractured bones 17 days after the procedure but did not observe a significant change in the immune environment based on frequencies of major immune cell populations except the monocytes, B cells, and neutrophils in bending models (Fig. 1H; Supplementary Fig. S1E and S1F). In contrast, CD51⁺, PDGFRα⁺, and Sca-1⁺ cells with osteogenic potential were enriched after fracturing (Fig. 1I; Supplementary Fig. S1G). Although these data cannot rule out the effect of inflammation, we decided to first focus on cells of the osteogenic lineage. This decision was also based on our previous findings that the osteogenic niche plays critical roles in bone colonization under homeostatic conditions (11, 12, 15).

In order to characterize the roles of various cells with osteogenic potential in the BME, we crossed the ROSA26-LoxP-DTR (diphtheria toxin receptor) allele with Cre recombinase controlled by promoters of a number of widely studied markers of MSCs or skeletal stem cells, including NG2, Nestin, and Leptin receptor (LepR). We also included Tie2-Cre to examine the potential roles of endothelial cells and the perivascular niche. Administration of diphtheria toxin (DT) reduced cells with expression of the respective Cre recombinase (Supplementary Fig. S2A and S2B).

Drilling was used to introduce focal fractures in the femoral bones of animals of various strains. To more rapidly assess the ability of cancer cells to colonize the remodeling BME in multiple strains, we directly seeded cancer cells to the drilled sites and then monitored metastatic outgrowth by bioluminescence imaging (BLI; Fig. 2A). Right after the seeding, we administered DT to ablate the corresponding cell populations. Although the reduction of Nestin⁺ or Tie2⁺ cells generated little effect on bone colonization at the wounded sites (Fig. 2A), decrease of NG2⁺ and LepR⁺ cells impeded bone colonization. LepR-Cre is constitutively expressed during the development and its recombination activity labels both undifferentiated stem cells and differentiated lineages in adult animals, and depletion of LepR-Cre⁺ lineage by DT treatment leads to a profound increase of adipocytes and osteoblasts (25), which makes LepR-Cre not ideal for precise manipulation of bone marrow MSCs. On the other hand, NG2⁺ perivascular cells are important stem/progenitor cells in the osteogenic lineage (29), and the NG2-Cre strain used in this study is tamoxifen-inducible. Therefore, we decided to focus on NG2⁺ cells in our subsequent investigations.

Using the mice with inducible NG2⁺ lineage ablation as the hosts, we performed a spontaneous metastasis assay. Specifically, source tumors were resected when reaching approximately the same weight (Supplementary Fig. S2C), and resection and bone drilling were conducted on the same day (Fig. 2B). Again, this setting allows us to focus on metastatic cells that already arrived at the time of fracture. The reduction of NG2-Cre⁺ cells significantly decreased spontaneous metastasis to the injured bone, as shown by bioluminescence (Fig. 2B; Supplementary Fig. S2D) and flow cytometry of GFP⁺ cancer cells (Supplementary Fig. S2E). Interestingly, metastasis to the contralateral uninjured bones was also decreased (Supplementary Fig. S2D), suggesting that NG2⁺ cells may be important for metastasis under homeostatic conditions, which was tested in some later experiments. However, metastasis to lungs was not affected (Supplementary Fig. S2D), supporting the bone specificity of the role of NG2⁺ cells.

We examined the spatial distribution of spontaneous metastatic lesions more closely. The enrichment of metastatic tumors surrounding the drilled site was clearly diminished by the depletion of NG2⁺ cells (Fig. 2B–D). Importantly, NG2⁺ depletion did not alter the immune cell profiles in the bone marrow (Supplementary Fig. S2F). Taken together, these data implicate NG2⁺ MSCs as a major cell population driving bone colonization stimulated by remodeling.

Bone modeling is an ongoing process even under homeostatic conditions (17), albeit at a much lower rate compared with that during bone repair. We asked if NG2⁺ cells also influence bone metastasis under homeostatic bone remodeling. We used the same spontaneous bone metastasis setting—that is, DT treatment and source tumor resection were performed on the same day when implanted tumors reached a similar size (Fig. 2E; Supplementary Fig. S2G). Depletion of NG2⁺ cells significantly hindered bone metastasis (Fig. 2E–G) without affecting lung metastasis (Fig. 2H). Under this homeostatic condition, depletion of Nestin⁺ cells decreased the bone metastatic burden but not the frequency, whereas the reduction of spontaneous bone metastasis was not observed in other strains with depletion of LepR⁺ and Tie2⁺ cells (Supplementary Fig. S2H–S2P).

We next used an experimental bone metastasis model based on intrailiac artery (IIA) injection to directly deliver cancer cells to hind limb bones (30). This approach provides a definitive onset of bone colonization and is useful for characterizing the temporal course of this process (Fig. 2I). NG2 depletion again exhibited a significant decrease of metastatic burden in bone (Fig. 2I). We noticed that the impact of NG2 depletion appeared from a very early stage of bone colonization (Fig. 2I). To zoom into this phase, we carried out another experiment and administered DT prior to IIA implantation so that cancer cells immediately encountered an NG2-depleted BME upon arrival (Fig. 2J). Depletion of NG2⁺ cells showed no significant effects on the homing or survival of tumor cells to the bone, as determined by BLI and flow cytometry analysis of total, proliferative, or apoptotic tumor cells retrieved from bones 1 day after IIA injection (Supplementary Fig. S2Q–S2S). However, the lack of NG2⁺ lineage significantly impaired the ability of DTCs, especially the Ki-67⁺ fraction, to expand during the first 6 days of colonization (Fig. 2J; Supplementary Fig. S2T–S2V), confirming that the role of NG2⁺ cells is more pronounced in bone metastasis initiation.

Taken together, the results so far indicate that NG2⁺ cells play the most consistent role in various metastasis assays and among all cell populations examined. Furthermore, this role appears to be important in both pathologic and homeostatic conditions.

Given the roles of NG2⁺ cells in bone remodeling–induced metastasis initiation, we wondered about the normal functions of these cells in cancer-free bones. Although the NG2-CreER strain is often used to identify perivascular MSCs, there are other populations of MSCs that are characterized by other markers. Moreover, NG2 is also expressed to variable degrees by pericytes, chondrocytes, osteoblasts, osteocytes, smooth muscle cells, and peripheral nerve Schwann cells (20, 25). Therefore, the precise cellular identity, spatial location, and relative contribution of NG2-CreER⁺ cells to bone remodeling still need to be verified.

We first set out to analyze the differentiation potential of NG2-CreER⁺ cells. We used NG2-CreER;ROSA26-LoxP-TdTomato mice, in which cells expressing NG2-CreER and their descendant cells will express TdTomato permanently (designated as NG2-tdRED⁺ hereafter for brevity) and can then be purified from these mice for in vitro characterization (Fig. 3A). NG2-tdRED⁺ cells are more frequent in the endosteal bones and need to be extracted by enzymatic digestion (Supplementary Fig. S3A). In culture conditions favoring differentiation toward different lineages, NG2-tdRED⁺ cells became osteoblasts, adipocytes, and chondrocytes and therefore met the in vitro criteria of MSCs. However, compared with the NG2-tdRED⁻ counterparts, NG2-tdRED⁺ cells clearly exhibited a strong commitment toward osteolineage and a slightly increased adipogenic differentiation capacity (Fig. 3B; Supplementary Fig. S3B). PDGFRα⁺CD51⁺ stromal cells have been demonstrated to contribute to the majority of fibroblastic colony formation units (CFU) in bone marrow cells and represent a subset with greater self-renewal capacity in vivo (31). Nestin-GFP⁺ MSCs largely overlap with the PDGFRα⁺CD51⁺ subset and NG2 protein marks its perivascular subpopulation, which also contains most CFU-F of bone marrow stromal cells in mice (32). NG2-tdRED⁺ cells enriched PDGFRα⁺CD51⁺ fraction, possessed higher expression levels of self-renewal genes Pou5f1 and Sox2, and showed enhanced CFU-F activities compared with the NG2-tdRED⁻ population (Supplementary Fig. S3C–S3F).

We next performed a lineage-tracing experiment by combining the osteocalcin-GFP (OCN-GFP) allele with the NG2-CreER;ROSA26-LoxP-TdTomato strain. Theoretically, a short period of tamoxifen treatment could induce the recombinase activity mainly in the NG2⁺ progenitor cells and their progeny cells would inherit the expression of tdRED reporter. Osteocalcin is specifically expressed by mature osteoblasts, and therefore this lineage-tracing model allows the identification of NG2-CreER–derived osteoblasts (tdRED⁺GFP⁺) and other non–NG2-CreER–derived osteoblasts (tdRED⁻GFP⁺; ref. 33). We examined cells double positive for OCN-GFP and NG2-tdRED at different time points after a brief 5-day tamoxifen induction of NG2-CreER activity by either immunofluorescent staining or flow cytometry. Under the homeostatic condition, double-positive cells were rare 2 days after the tamoxifen induction (Fig. 3C and D; Supplementary Fig. S3G and S3H), suggesting minimal leakage of Cre activity in existing osteoblasts after induction. This population became apparent later and remained at a relatively constant level for more than 4 months (Fig. 3C and D; Supplementary Fig. S3G and S3H). The majority of NG2-tdRED⁺ cells remained negative for OCN-GFP, suggesting a potentially strong self-renewal capacity of NG2⁺ stromal cells (Supplementary Fig. S3G). Importantly, more than 20% to 50% of all the OCN-GFP⁺ osteoblasts were also positive for NG2-tdRED 4 weeks after induction (Fig. 3D; Supplementary Fig. S3I), indicating a substantial proportion of new osteoblasts are descendants of NG2⁺ MSCs. Furthermore, there was a high concordance between nondifferentiated NG2-tdRED⁺ cells and NG2 protein expression (Supplementary Fig. S3J), as well as another MSC marker, LepR (25), and pericyte marker PDGFRβ (ref. 34; Supplementary Fig. S3K and S3L). Interestingly, there was minimal overlap between NG2-tdRED and chondrocyte or adipocyte marker (Aggrecan or Perilipin, respectively; ref. 25; Supplementary Fig. S3M). In addition, about 30% of LepR-tdRED⁺ cells were also stained positively with NG2 antibody by flow cytometry (Supplementary Fig. S3N). Immunofluorescent staining showed that perivascular LepR-tdRED⁺ cells also express NG2 protein in endosteal and fracture regions but not the bone marrow cavity (Supplementary Fig. S3O). Taken together, the lineage-tracing results strongly support that NG2-tdRED⁺ cells are MSC-like cells committed for osteogenic differentiation and represent a major contributor to bone remodeling.

We then examined the impact of NG2⁺ cell depletion on bone remodeling. Using NG2-CreER;ROSA26-LoxP-DTR mice, administration of DT reduced the osteogenic lineages marked by either CD51⁺, PDGFRα⁺, or Sca-1⁺ cells (Supplementary Fig. S3P), suggesting a shrinkage of the cell reservoir with osteogenic potential. The ablation of NG2⁺ cells also led to decrease of both osteoblast and osteoclast activities (Fig. 3E–H) and a reduced rate of new bone formation (Fig. 3I and J). Thus, NG2⁺ MSCs play a critical role in bone remodeling under normal conditions.

Finally, we used the same lineage tracing and depletion systems to study the roles of NG2⁺ cells in repair of pathologic fractures. Indeed, it is evident that NG2-tdRED⁺ cells were recruited to callus and participated in the generation of new bones (Fig. 3K). It has been reported that in vivo–transplanted MSCs can engraft the bone and modulate bone-related pathogenesis and regeneration (35–37). To test whether NG2⁺ cells exert similar functions, we sorted out and injected NG2-tdRED⁺ bone stromal cells into the femoral cavity of wild-type (WT) mice, a procedure that also created a defect extending from the articular cartilage to the femoral medullary cavity. Ten days after surgery, the bones that received transplantation of NG2-tdRED⁺ cells exhibited a slight increase of new bone volume and a significant increase of new bone surface at the injured metaphyseal region (Supplementary Fig. S3Q–S3S) in comparison with the sham control femurs that were injected with saline. In line with previous reports (38, 39), NG2-tdRED⁺ cells were observed at the injured and other bone regions, and some of them also differentiated into mature osteoblasts (stained positively with osteocalcin; Supplementary Fig. S3T). Moreover, the loss of NG2⁺ cells by DT treatment resulted in significantly delayed bone repair in drilling models (Fig. 3L and M). Therefore, NG2⁺ MSCs also appear to mediate bone remodeling during the repair of pathologic fractures.

We asked if the early impact of NG2⁺ cells on bone metastasis is reflected by their spatial distributions relative to DTCs using NG2-CreER;ROSA26-LoxP-TdTomato mice. By adopting a tissue clarity approach (Fig. 4A), we were able to perform confocal microscopy and reconstruct 3D images of entire femur bones, and detect single DTCs and microscopic metastases that occur spontaneously (Fig. 4B) or experimentally introduced by IIA (Fig. 4C). In both cases, cancer cells were found more frequently in the endosteal region compared with central bone marrow and very often colocalized with NG2-tdRED⁺ cells (Fig. 4B and C). At a single-cell resolution, many cancer cells appeared to reside in the perivascular niche with or without direct contact with NG2-tdRED⁺ perivascular cells (Supplementary Fig. S4A). Some cancer cells even developed prolonged protrusions that connected NG2-tdRED⁺ cells (Supplementary Fig. S4A), resembling a unique cell–cell interaction we previously observed in vitro (40). We also used a computational approach to quantitate the distribution of DTCs and NG2-tdRED⁺ cells relative to each other in the early stage of synchronized bone colonization introduced by IIA. Compared with random simulated locations, the distance between the two cell populations was significantly shorter (Fig. 4D; Supplementary Fig. S4B). When the same analysis was applied to endothelial cells, we did not observe similar results (Supplementary Fig. S4C and S4D). As DTCs progress into microscopic metastases, there appeared to be an inverse correlation between the size of metastasis and distance to NG2-tdRED⁺ cells (Supplementary Fig. S4E). DTCs in NG2⁺ lineage–depleted animals were rare and mostly single cells (Supplementary Fig. S4F), but about 70% of them remained close to endothelial cells as compared with NG2-tdRED⁺ cells (Supplementary Fig. S4G–S4I), suggesting depletion of NG2⁺ cells does not alter the perivascular location of tumor cells. Finally, pathologic fractures by drilling or bending led to enrichment of both cancer cells and NG2-tdRED⁺ cells, as well as extensive direct cell–cell interactions between the two cell populations (Fig. 4E).

Taken together, our data reveal a spatial correlation between DTCs and NG2⁺ cells and uncover frequent and direct contact between these cell types during bone metastasis progression.

The frequent and direct contact between DTCs and NG2⁺ cells during bone colonization prompted us to inspect their interactions in a simplified system and for a broader scope of cancer models. We selected nine murine cancer cell lines including LLC1 used in previous experiments. These lines represent five different cancer types and two genetic backgrounds. We admixed these lines with primary NG2-tdRED⁺ bone mesenchymal stem/stromal cells (BMSC) freshly purified from syngeneic NG2-CreER;ROSA26-LoxP-TdTomato mice after tamoxifen induction. BMSCs that do not express NG2-tdRED from the same mice were also prepared for comparison. In 3D suspension cultures, the vast majority of cancer models formed heterotypic organoids with NG2-tdRED⁺ cells (Fig. 5A). Some are very similar to what we observed before using human cancer cells and MSCs (12), including 4T1.2, CMT-93, EMT6, LLC1, and PYMT-E cells. Most models benefited from interaction with both NG2-tdRED⁺ and NG2-tdRED⁻ BMSCs except EO771 cells (Fig. 5B). In four models (CMT-93, EMT6, LLC1, and PYMT-E), NG2-tdRED⁺ cells conferred a significantly stronger advantage compared with the NG2-tdRED⁻ counterparts (Fig. 5B). This difference appeared to be dependent on cell–cell contact, as the separation of BMSCs and cancer cells by Boyden Chambers largely diminished the extra cancer-promoting effects by NG2-tdRED⁺ cells (Supplementary Fig. S5A). The interaction with NG2-tdRED⁺ BMSCs promoted cancer cell proliferation, as indicated by an increased percentage of Ki-67⁺ proliferating tumor cells in organoids formed by LLC1 tumor cells and BMSCs (Fig. 5C and D). Real-time imaging in 2D cocultures also revealed that cancer cells rapidly established cell–cell contact with NG2-tdRED⁺ BMSCs and developed colonies surrounding these cells (Supplementary Video S1), and the size of colonies was significantly larger compared with cocultures with NG2-tdRED⁻ BMSCs and monocultures (Fig. 5E and F; Supplementary Videos S2 and S3). In fact, LLC1 cells could not grow without BMSCs in serum-free conditions. Furthermore, within the same coculture, colonies maintaining direct contact with NG2-tdRED⁺ BMSCs were significantly larger than those losing direct contact (Supplementary Fig. S5B and S5C), albeit the numbers of colonies did not differ between these two situations (Supplementary Fig. S5D).

The direct contact with NG2-tdRED⁺ BMSCs not only promotes the proliferation of cancer cells but also enhances their movement, as shown in a transwell migration assay. Specifically, GFP⁺ cancer cells were admixed with BMSCs in the upper level of the Boyden Chamber. The transwell migration of cancer cells appeared to be greatly enhanced by the presence of NG2-tdRED⁺ BMSCs as compared with NG2-tdRED⁻ BMSCs (Fig. 5G and H).

Finally, the tumor-promoting effects of NG2-tdRED⁺ BMSCs were even stronger when BMSCs were pretreated with medium favoring osteodifferentiation except for B16-F10 cells (Fig. 5I; Supplementary Fig. S5E and S5F). This extra effect was not observed for most tumor cells when coculturing with NG2-tdRED⁻ BMSCs (Fig. 5I) and was critically dependent on cell–cell contact (Supplementary Fig. S5G).

Taken together, the in vitro coculture experiments uncover a superior ability of NG2⁺ cells in promoting tumor progression compared with other BMSCs. Importantly, this effect appears to be reliant on direct cell–cell contact and becomes further strengthened upon osteodifferentiation.

Our previous studies demonstrated that cancer cells form hAJs with the osteogenic niche cells in early-stage bone colonization. E-cadherin from cancer cells and N-cadherin from osteogenic cells constitute hAJs (11, 12), which can also be observed in heterotypic organoids formed by E-cadherin⁺ mouse tumor cells (4T1.2, CMT-93, EMT6, LLC1, and PYMT-E) and NG2-tdRED⁺ BMSCs (Supplementary Fig. S5H). Consistently, four (CMT-93, EMT6, LLC1, and PYMT-E) of these tumor cells, which gained additional growth benefit, showed an elevated level of phosphorylated S6 kinase upon coculture with NG2⁺ BMSCs, indicating activation of mTOR signaling in these cells when interacting with NG2⁺ BMSCs (Supplementary Fig. S5I and S5J). Knockdown of E-cadherin on tumor cells by siRNAs significantly diminished the growth promotion effect of both NG2⁻ and NG2⁺ BMSCs in these cells under either normal medium or osteodifferentiation medium (Supplementary Fig. S6A). In this study, direct cell–cell contact again appeared to be critical in determining the roles of NG2⁺ cells, which enriches stem cells committed for osteogenesis. In support of this notion, transient knockdown of N-cadherin by siRNAs on both NG2⁻ and NG2⁺ BMSCs also blocked their growth promotion effects on E-cadherin⁺ tumor cells (Supplementary Fig. S6B). We found that N-cadherin was expressed at a higher level in NG2-tdRED⁺ cells compared with NG2-tdRED⁻ BMSCs (Supplementary Fig. S3D), and its expression was further increased during in vitro osteogenic differentiation in NG2⁺ cells (Supplementary Fig. S5F), both of which support an important role of N-cadherin in osteoblast differentiation and osteogenesis (41–43). Interestingly, a recent report suggested that N-cadherin⁺ stromal cells share a similar transcriptomic profile with NG2⁺ cells and are the main source of bone and marrow stromal progenitor cells (44), so we reasoned that knockout of N-cadherin expression in these cells may disrupt the interaction between NG2⁺ cells and cancer cells, thereby abolishing the effects of bone remodeling on metastasis initiation. To test this hypothesis, we bred NG2-CreER;LoxP-CDH2 mice to delete N-cadherin selectively in NG2-Cre⁺ cells in an inducible fashion (designated as NG2-Ncad^KO/KO hereafter; Supplementary Fig. S6C). Indeed, N-cadherin was significantly reduced in NG2-tdRED⁺ cells from NG2-Ncad^KO/KO mice (Supplementary Fig. S6D and S6E).

We first asked if depletion of N-cadherin also influences the normal function of NG2⁺ cells. In vitro differentiation assays revealed a notable decrease of osteodifferentiation as well as adipodifferentiation upon N-cadherin knockout (Supplementary Fig. S6F compared with Fig. 3B and Supplementary Fig. S3B). Loss of N-cadherin in NG2⁺ cells also decreased the bone mineralization rate (Fig. 6A and B) and slowed down the repair of bone fracture (Fig. 6C and D). These data together demonstrate the pivotal role of N-cadherin in homeostatic and pathologic bone remodeling.

We next examined whether loss of N-cadherin affects the cancer-promoting effects in NG2⁺ cells in cocultures. NG2-tdRED⁺ cells were also extracted from NG2-Ncad^KO/KO mice (referred to as KO BMSCs hereafter) and were admixed with a variety of cancer models in parallel with the previously isolated WT NG2-tdRED⁺ cells (hereafter referred to as WT BMSCs for simplicity) in the above-mentioned in vitro coculture experiments. Only the seven C57BL/6 cancer cells were tested because NG2-Ncad^KO/KO is available only in this genetic background. In five of the tested models, a significant reduction in tumor-promoting effect was observed (Fig. 6E). In addition, the levels of phosphorylated S6K were also reduced in PYMT-E, CMT-93, and LLC1 tumor cells cocultured with knockout cells (Supplementary Fig. S5J). Further experiments uncovered a decrease of Ki-67⁺ cells in the 3D heterotypic organoids (Fig. 6F and G) and an overall decrease of colony size in 2D cocultures using LLC1 cells cocultured with knockout cells (Fig. 6H and I; Supplementary Video S4). When the analysis was restricted to the cancer cells that maintained direct contact with NG2-tdRED⁺ cells, knockout of N-cadherin clearly diminished colony size (Supplementary Fig. S6G). However, this difference was not found for colonies that were not in firm contact with WT or knockout NG2-tdRED⁺ cells (Supplementary Fig. S6H). Loss of N-cadherin also impaired the migration of both cancer cells (Fig. 6J and K) and NG2-tdRED⁺ cells (Supplementary Fig. S6I). In addition, the excessive cancer-promoting effect stimulated by osteodifferentiation was also abolished by N-cadherin knockout, especially in the 3D cocultures (Fig. 6L; Supplementary Fig. S6J). Thus, N-cadherin in NG2⁺ cells is responsible for the superior ability of these cells in promoting cancer proliferation and migration in vitro.

We set out to determine the impact of NG2-specific knockout of N-cadherin on bone metastasis. Two different experiments were performed toward this end. In the first experiment, cancer cells were directly inoculated into drilled bones in either Ncad^WT/WT or NG2-Ncad^KO/KO mice. The cancer cell growth was significantly impaired in NG2-Ncad^KO/KO animals (Fig. 6M). In addition to bioluminescence quantitation, we also enumerated GFP⁺ cancer cells on the drilled site by flow cytometry and confirmed the difference in tumor progression between Ncad^WT/WT and NG2-Ncad^KO/KO animals (Supplementary Fig. S6K). Thus, N-cadherin in NG2⁺ cells is critical for the outgrowth of already-seeded cancer cells. The second experiment is a spontaneous bone metastasis assay. We carried out bone drilling on the same day of source tumor resection and monitored the development of spontaneous bone metastasis (Fig. 6N). We did not observe a difference in primary tumor growth (Supplementary Fig. S6L). However, N-cadherin knockout in NG2⁺ cells significantly decreased metastasis surrounding the drilled site (Fig. 6N–P). In contrast, there were no significant differences in metastasis to contralateral noninjured bones or the lung (Supplementary Fig. S6M). Taken together, these data strongly support a critical role of N-cadherin in NG2⁺ cells during bone remodeling–stimulated metastasis.

Finally, we examined if knockout of N-cadherin in NG2⁺ cells may affect drilling-associated inflammation. A flow cytometry–based characterization of the immune cell profile did not detect any significant alterations to the major populations (Supplementary Fig. S6N). Thus, although our data cannot directly rule out the influence of inflammation, the observed roles of NG2⁺ cells and N-cadherin in NG2⁺ cells do not seem to be mediated by the alteration of immune cells.

We examined the in situ protein expression of NG2 and N-cadherin in a small number of human bone metastases from various types of cancers, including breast, prostate, colon, and lung cancers, and confirmed that NG2 and N-cadherin were both expressed by cells surrounding metastatic cells (Fig. 7A). To gain deeper insight into the functional relevance of NG2⁺ cells in human bone metastasis, we compared NG2 expression (encoded by the gene CSPG4) in solid bone tumor tissues (tumor) versus matched liquid bone marrow involved in the metastases (involved) or in a different bone (distal), as well as bone marrow samples from tumor-free patients (benign) from a single-cell RNA sequencing (scRNA-seq) dataset of prostate bone metastases (45). A significantly higher level of NG2 expression was observed in tumor tissues (Fig. 7B). To determine whether the NG2 expression is specific to perivascular mesenchymal cells, we examined the coexpression of NG2 and other cell type–characteristic genes. Indeed, NG2⁺ cells also express RGS5, ACTA2, and PGF, the well-established pericyte markers (46), but not cancer cell–specific genes such as AR and KLK3 (Fig. 7C). Further analyses revealed that the frequency of NG2⁺ perivascular cells correlated with tumor proliferation index (ref. 47; Fig. 7D) and N-cadherin–related osteoblast differentiation signature (ref. 48; Fig. 7E), supporting our hypothesis that NG2⁺ cells drive both osteogenesis and tumor progression.

Because NG2 expression in bone metastasis is mainly contributed by perivascular mesenchymal cells, we set out to analyze a larger number of bone metastases in more cancer types using bulk RNA-seq data. In two datasets comprising metastases at different sites, we observed that NG2 expression is significantly higher in bone metastases of both breast and prostate cancers (Supplementary Fig. S7A). Moreover, among bone metastases, NG2 expression exhibits significant correlations with N-cadherin expression and a gene expression signature indicative of osteoblast differentiation (Supplementary Fig. S7B and S7C). Interestingly, such correlation was not observed in other metastases or primary tumors (Supplementary Fig. S7D). In breast cancer bone metastases, NG2 is also associated with an increase of EZH2, an increase of an embryonic stem cell signature overexpressed in aggressive tumors, and a decrease in ESR1 expression and related gene expression signatures (Fig. 7F). This is consistent with our previous finding that direct interaction with osteogenic cells promotes the phenotypic plasticity of cancer cells in the BME (5, 15). In prostate cancer bone metastases, non–AR-driven tumors exhibited a remarkably higher expression of NG2 (Fig. 7G), suggesting that cross-talk with osteogenic cells may activate alternative pathways to drive tumor progression. Taken together, our data support an important role of NG2⁺ perivascular cells in bone metastasis progression.

The BME is composed of many different cell types that are intricately organized (1). Even cells of seemingly the same type may be functionally and molecularly distinct depending on their geographic locations, cell of origin, and interacting cells. For example, Lin and colleagues have shown the critical role of endothelial cell–converted osteoblasts in the formation of osteoblastic metastases of prostate cancer (49). MSCs represent another great example of this kind. The exact identity and functions of MSCs have been intensively studied (20). Different subpopulations of MSCs may vary in their roles in homeostatic versus pathologic conditions, their differentiation capacity to different lineages in vitro and in vivo, and their location relative to vasculature and endosteum (50, 51). In this study, we identified a unique role of NG2-CreER⁺ cells, which are presumably a subset of MSCs. Previous studies have elucidated that NG2⁺ cells predominantly localize at the metaphysis and H-type vasculature in the bone, where osteogenesis and angiogenesis are coupled (29). On the other hand, most DTCs were found in the same area. Indeed, we and others both discovered colocalization between NG2⁺ cells and DTCs in the perivascular niche (13, 14). Moreover, a recent work by Yip and colleagues found that disseminated breast cancer cells preferentially colocalize with H-type vessels in bone, and tumor-derived G-CSF remodels the bone marrow vasculatures to resemble H-type vessels and therefore supports the metastatic growth in bone (52). Interestingly, NG2⁺ BMSCs appear to express a higher level of N-cadherin, a component of hAJ with cancer cells, than NG2⁻ BMSCs (11, 12). Thus, both geographic location and cancer-interacting molecule may distinguish NG2⁺ MSCs from other MSCs.

Our work highlights the dynamic nature of the BME. Various microenvironment niches have been implicated in metastasis, including the perivascular niche and the osteogenic niche. However, the potential connection between these niches remains elusive. NG2⁺ MSCs localize in a perivascular niche in the resting stage but can become mobilized and participate in osteogenesis during bone remodeling (19). Thus, it is conceivable that these cells may connect different niches. Our recent work suggested that cancer cells can migrate by tethering MSCs with a unique cellular protrusion (40). In this study, we also found that direct interaction with NG2⁺ cells stimulates cancer cell migration toward osteogenic signals. These data postulate an interesting hypothesis that cancer cells in the perivascular niche may “ride” NG2⁺ MSCs and be corecruited to sites of remodeling, thereby leaving one niche for another. This hypothesis will need to be tested in vivo ideally by real-time microscopy in future studies.

The dynamics of BME and NG2⁺ cells may also reconcile the seemingly contradictory results between this and a previous study. Nobre and colleagues showed that bone marrow NG2⁺Nestin⁺ perivascular MSCs enforce dormancy of EO771 cells through secretion of TGFβ2 (53). The biological contexts examined in the two studies are different. Whereas Nobre and colleagues focused on the role of resting NG2⁺Nestin⁺ cells in the perivascular niche, our study predominantly examined remodeling-activated NG2⁺ cells that couple bone metastasis progression with development of the osteogenic niche. In addition, the discrepancy may also partly result from intrinsic differences among cancer models. The EO771 cells do not express E-cadherin and therefore did not form heterotypic adhesion junctions with N-cadherin⁺ BMSCs and subsequently could not gain a proliferative advantage in cocultures with NG2⁺ cells in our experiments, which is in contrast to LLC1 and a few other models, suggesting different cell–cell interaction mechanisms.

The connection between bone turnover and bone metastasis can explain some circumstantial epidemiologic observations. For instance, Obi and colleagues observed an increased risk of bone metastasis related to fracture events in a large cohort of breast cancer survivors (54). There are also cases in which breast or lung cancer spreads to rare sites, such as the oral cavity and the jaws, in patients receiving dental implants (55–58), further indicating the potential connection between bone remodeling and the emergence of metastatic disease. Therefore, our discovery provides one potential mechanistic explanation for these clinical observations: Bone remodeling stimulates osteodifferentiation, and DTCs may take advantage of this process and become invigorated through direct interaction with NG2⁺ MSCs. Further epidemiology studies will be needed for a more in-depth investigation on links between bone metastasis and other life events that alter bone turnover rate.

Mouse melanoma cells (B16-F10; C57BL, cat. #CRL-6475, RRID: CVCL_0159), rectal carcinoma cells (CMT-93; C57BL, cat. #CCL-223, RRID: CVCL_1986), breast cancer cells EMT6 (BALB/c, cat. #CRL-2755, RRID: CVCL_1923) and EO771 (C57BL, cat. #CRL-3461, RRID: CVCL_GR23), Lewis lung carcinoma cells (LLC1; C57BL, cat. #CRL-1642, RRID: CVCL_4358), and prostate cancer cells (TRAMP-C1; C57BL, cat. #CRL-2730, RRID: CVCL_3614) were obtained directly from ATCC. 4T1.2 (BALB/c, RRID: CVCL_GR32) and AT-3 (C57BL, RRID:CVCL_VR89) cells were kindly provided by Dr. Robin Anderson (Olivia Newton-John Cancer Research Institute) and Dr. Ekrem Emrah Er (University of Illinois at Chicago), respectively. PYMT-E (C57BL, RRID: N/A) is a subline generated from MMTV-PYMT tumors in our lab as previously described (59). B16-F10, CMT-93, 4T1.2, LLC1, AT-3, and PYMT-E cells were maintained in DMEM high-glucose media (HyClone) supplemented with 10% FBS (Gibco); EMT6 cells were maintained in Waymouth's MB medium (Gibco) with 15% FBS; EO771 cells were cultured in RPMI 1640 medium (HyClone) supplemented with 10 mmol/L HEPES (Gibco) and 10% FBS; and TRAMP-C1 cells were cultured in DMEM high-glucose media supplemented with 5 mg/L bovine insulin (Sigma), 10 nmol/L dehydroisoandrosterone (ACROS Organics), and 10% FBS. All the media were supplemented with 1% penicillin–streptomycin (Lonza) and cells were cultured in a 5% CO₂ incubator. EO771, LLC1, and AT-3 cells were authenticated by short tandem repeat profiling provided by ATCC. Contamination of Mycoplasma was not detected in those cells using the PlasmoTest Mycoplasma Detection Kit (InvivoGen) at the time of cryopreservation. No cells were passaged for more than 2 months in vitro.

The protocols for collection and use of human bone metastasis samples were performed in accordance with the Declaration of Helsinki and approved by the Institutional Review Boards at Baylor College of Medicine (H-49396), The University of Texas MD Anderson Cancer Center (PA15-0225), and University of Texas Medical Branch (H-46675). All the patients have provided written informed consent on the use of their samples for research purposes when undergoing orthopedic surgery.

The in vivo procedures and usage of animal models were conducted in accordance with the protocol (AN-5734) approved by the Baylor College of Medicine Institutional Animal Care and Use Committee. C57BL/6J (B6, stock no. 000664, RRID:IMSR_JAX:000664), BALB/cJ (BALBc, stock no. 000651, RRID: IMSR_JAX:000651), B6.Cg-Gt(ROSA)26Sortm14(CAG-tdTomato)Hze/J (B6-tdRED, stock no. 007914, RRID: IMSR_JAX:007914), C57BL/6-Gt(ROSA)26Sortm1(HBEGF)Awai/J (B6-iDTR, stock no. 007900, RRID: IMSR_JAX:007900), B6.Cg-Tg(Cspg4-Cre/Esr1*)BAkik/J (NG2-CreERTM, stock no. 008538, RRID: IMSR_JAX: 008538), C57BL/6-Tg(Nes-Cre/ERT2)KEisc/J (Nes-Cre/ERT2, stock no. 016261, RRID: IMSR_JAX: 016261), B6.129(Cg)-Leprtm2(Cre)Rck/J (Lepr-Cre, stock no. 008320, RRID: IMSR_JAX: 008320), B6.Cg-Tg(Tek-Cre)1Ywa/J (Tie2-Cre, stock no. 008863, RRID: IMSR_JAX: 008863), and B6.129S6(SJL)-Cdh2tm1Glr/J (N-cadflox, stock no. 007611, RRID: IMSR_JAX: 007611) were purchased from The Jackson Laboratory. The OCN-GFP (RRID: N/A) strain was kindly provided by Dr. Dongsu Park (Baylor College of Medicine). To generate Cre^+/−;tdRED^+/+ reporter strains, the male Cre⁺ transgenic mice were bred with female tdRED^+/+ for two rounds to generate a Cre-positive strain with homozygous tdRED insertions. Then, the male Cre^+/−;tdRED^+/+ reporter mice were mated with female iDTR^+/+ mice to generate Cre^+/−;tdRED^+/−;iDTR^−/+ mice for lineage depletion models. For the in vivo tracing of osteogenic differentiation, the male NG2-CreERTM^+/−;tdRED^+/+ strain was crossed with the female Ocn-GFP^+/− strain to generate the NG2-CreERTM^+/−;tdRED^+/−;Ocn-GFP^+/− strain. For the N-cadherin knockout model, the male NG2-CreERTM^+/−;tdRED^+/+ strain was crossed with the female N-cad^flox/flox strain for two rounds to generate NG2-CreERTM^+/−;N-cad^flox/flox;tdRED^+/− or NG2-CreERTM^+/−;N-cad^flox/flox;tdRED^−/− mice. B6.NG2-CreERTM and B6.tdRED mice were backcrossed to BALB/cJ for more than 10 generations in our lab and then crossed to breed the BALBc-NG2-CreERTM^+/−;tdRED^+/+ strain.

The genomic DNA of ear tissues was purified by the MasterPure DNA Purification Kit (Epicentre, MC85200). PCR primers for genotyping of the Cre allele were as follows: Cre-F, 5′-GCG GTC TGG CAG TAA AAA CTA TC-3′; Cre-R, 5′-GTG AAA CAG CAT TGC TGT CAC TT-3′. Primers for genotyping of the Rosa26-tdRED allele were: tdRED-F, 5′-GGC ATT AAA GCA GCG TAT CC-3′; tdRED-R, 5′-CTG TTC CTG TAC GGC ATG G-3′; tRED-WT-F, 5′-AAG GGA GCT GCA GTG GAG TA-3′; and tRED-WT-R, 5′-CCG AAA ATC TGT GGG AAG TC-3′. Primers for genotyping of the Rosa26-iDTR allele were as follows: iDTR-Com, 5′-AAA GTC GCT CTG AGT TGT TAT-3′; iDTR-R, 5′-GCG AAG AGT TTG TCC TCA ACC-3′; iDTR-WT-R, 5′-GGA GCG GGA GAA ATG GAT ATG-3′. Primers for genotyping N-cad^flox/flox allele were: CDH2 KO-F, 5′-CCA AAG CTG AGT GTG ACT TG-3′; CDH2 KO-R, 5′-TAC AAG TTT GGG TGA CAA GC-3′. Primers for Ocn-GFP genotyping were: GFP-F, 5′-CTG GTC GAG CTG GAC GGC GAC GTA AC-3′; and GFP-R, 5′-ATT GAT CGC GCT TCT CGT TGG GG-3′.

If not specified otherwise, all the in vivo experiments were performed using female mice, and the Cre^−/− littermates were used as age-matched controls. To induce CreER activity, 5-week-old female mice were injected with 1 mg tamoxifen (Sigma) daily for 5 consecutive days (for lineage-tracing experiment and isolation of BMSCs) or two cycles of 5 consecutive days with a 2-day interval (for in vivo experiments). To deplete the Cre-expressing lineage, mice were intraperitoneally injected with 200 ng DT (20 ng for Tie2-Cre; Sigma-Aldrich, D0564) for 3 consecutive days in every 10-day cycle.

All the animal surgeries were performed using aseptic procedures and started with 7- to 8-week-old animals except for Fig. 2B (due to the institutional lockdown during the pandemic, this batch of experiments was started with about 3-month-old animals). If not specified, 20E4 LLC1 Fluc-EGFP cells were mixed with an equal volume of growth factor–reduced Matrigel matrix (Corning, cat. #356231 or R&D Systems, cat. #3433-010-01) and subcutaneously injected into the skin close to the right hindlimb. The tumors took about 18 days to reach 1.2 cm in diameter, and a tumor removal surgery was performed to completely remove the primary tumors. About 20% of subjects met the criteria of early euthanasia before day 32 due to lung metastasis and were not included in the final analysis. Most animals survived about 5 weeks after the implantation of primary tumors and were dissected and examined at days 32 to 35 for lung and bone metastasis.

For the drill-hole model, a 7-mm-long posterior skin incision was made along the fumural bones. Then, the muscle was carefully displaced to expose the femur shaft, and a defect was created on the central femoral bones by a 0.7-mm-diameter drill bit with the electric drill (Dremel 8050-N/18 Micro). For the bending model, an about 1-cm-long incision was made over the anterior skin from the knee to the proximal femur. To stabilize the fractured femurs, a sterile, 0.2-mm diameter minutien stainless steel pin (Fine Science Tools, cat. #26002-20) was inserted into the bone marrow cavity from the femoral plateau. The fracture was then created by 3-point bending. To facilitate the following bone sectioning, the steel pins were removed by another survival surgery 1 week later. For both models, the skin wound was closed with tissue glue and wound clips and mice were given postoperative analgesia for 6 days and monitored twice daily. If not specified, the wounded bones were collected about 17 days after the fracture surgeries.

To directly implant tumor cells in the wounded bone areas, the same 0.7-mm-diameter drill bit was used to carefully thin the femoral shaft without touching the bone marrow cavity. Then, 10 μL of tumor cell suspension with extracellular matrix was injected by a 28-G insulin syringe (Becton Dickinson, cat. #329461) through the thinned region. For the transplantation of BMSCs into the bone, a skin opening was made on the knee region, and the syringe was slowly inserted into the femoral cavity through the metaphyseal cartilage. The injection created an about 0.4-mm-diameter bone defect. Then, a new syringe with BMSCs was inserted along the hole into the bone marrow cavity and BMSCs were slowly injected. Syringes were held in place for 1 minute to allow the pressure to equilibrate and avoid leakage of cell suspension.

IIA injection of tumor cells was performed as previously described (12, 30). Briefly, animals were given preoperative analgesia, anesthetized, and positioned on a warming pad. The surgical area was cleaned and sterilized, and a 7- to 8-mm incision was made between the fourth and the fifth nipples. The iliac artery was carefully exposed by removing the adjacent tissues. Tumor cells (5E4) were suspended in 100 μL PBS and slowly injected into the iliac artery by a 31-G insulin syringe (Becton Dickinson, cat. #328418). After the equilibration of blood flow, the syringe was retrieved and cotton tips were gently applied to the surgical area to stop the bleeding before the wound was closed.

For the IIA model and intraosseous implantation model, the in vivo BLI was performed with IVIS Lumina II (PerkinElmer) at the indicated time points. Briefly, the hair on the posterior side of injected hindlimb was removed to allow the penetration of bioluminescence before animals were given 100 μL 15 mg/mL D-luciferin (Gold Biotechnology, cat. #LUCK-5G) via retro-orbital injection. For the spontaneous metastasis model, to eliminate the influence of remaining or recurred primary tumors, only the ex vivo BLI on the dissected tissues was performed. Animals were scanned immediately at D mode right after the administration of luciferin. The exposure time was manually adjusted between 1 and 120 seconds to avoid signal saturation. To quantify the total BLI intensity of a specified tissue, a fixed region of interest (ROI) was applied to all animals or dissected tissues, and the total bioluminescent counts were quantified. The values were then normalized to exposure time to ease the comparison between batches of scanning with different exposure lengths. The presence of metastases on the specific organ was defined as the detection of clustered BLI signals above 15 counts/pixel under a maximum 120-second exposure. To quantify the distribution of BLI signals along the femoral bones, the BLI images were exported as 8-bit gray graphs with log₁₀-transformed BLI spectrum ranging from 15 to 65,535 counts/pixel. Then, a rectangle ROI with 100 pixels (length) × 30 pixels (width) was drawn from the distal end to the proximal end of femoral bones. The distribution of transformed BLI intensities along this ROI was measured by the “Plot Profile” function of ImageJ.

To collect tissues for immunostaining, animals were preperfused with 30 mL 10 IU/mL heparin PBS solution to completely remove the blood cells before dissection. The dissected tissues were then fixed overnight in 4% PFA. Bone tissues were decalcified with 0.5 M pH 7.4 EDTA solution overnight except those for bone histomorphometry assay and new bone formation assay. Then, tissues were cryopreserved in 30% sucrose PBS solution and embedded in OCT. The CryoJane tape-transfer system (Leica) was used to collect high-quality 10-μm thick sections from frozen bone tissues. The frozen sections were kept in −80°C freezer until further staining. The Breast Center Pathology Core at Baylor College of Medicine assisted with the preparation and sectioning of paraffin-embedded tissues and cells.

To determine the rate of new bone formation, animals were pretreated with tamoxifen to induce CreER activity at the age of 5 weeks and later subjected to two rounds of DT treatment from week 7 if needed. On week 10, mice were injected with 20 mg/kg Calcein dissolved in 2% sodium bicarbonate solution (4 mg/mL stock solution; Sigma, cat. #C0875-5G) via retro-orbital venous. Five days later, mice were given another dose of 40 mg/kg Alizarin Red S (Acros Organics, cat. #AC400480250) in PBS (8 mg/mL stock solution) through retro-orbital injection. On day 7, the mice were euthanized, perfused, and dissected. Hindlimb bones were collected, fixed, and embedded immediately in OCT. The nondecalcified bone was sectioned by a Leica CM3050S Cryostat installed with the CryoJane tape-transfer system with low-profile microtome blades. Bone sections were then mounted with Prolong Gold Antifade Mountant with DAPI (Invitrogen, cat. #P36935), and three different parts of each femur bone were randomly imaged by confocal microscope. The distance between Calcein and Alizarin Red S–positive bands was calculated by Zen software (Zeiss).

For bone histomorphometry analysis, mice from the same litter were treated with tamoxifen to induce Cre activity and later subjected to two rounds of DT treatment. Femur bones were collected 3 weeks later, fixed in 4% PFA overnight, and transferred to the Research Histology Core at The University of Texas MD Anderson Cancer Center in 70% ethanol PBS. There, bone samples were embedded and sectioned in a methylmethacrylate block. The Goldner's Trichrome and TRAP enzymatic staining were performed by Bone Histomorphometry Core at The University of Texas MD Anderson Cancer Center and analyzed by Leah Guerra.

Femur bones were dissected, cleaned, fixed in 70% ethanol PBS solution, and scanned by a Bruker Skyscan 1272 scanner (Bruker-MicroCT) at 50 kVp, 200 μA X-ray energy. The image resolution was 6 μm. Rotation angle (0.60°) and averaging of three scannings were applied at each step for a total 360° rotation. Three-dimensional bone images were reconstructed by CTvox (Bruker-MicroCT, v3.0.0), and the cross-sectional images were generated by NRecon (Bruker-MicroCT, v1.6.9.8) and DataViewer (Bruker-MicroCT, v1.5.6.2). To quantify new bone formation after fracture surgery, a fixed ROI was applied to extract the same volume of cortical bone around the defect area with minimal adjacent nonwounded bone tissue. Specifically, for drill-hole experiments, a circle ROI was drawn on the 200 axial cortical bone slides around the drill site, whereas for the intragrowth plate transplantation of BMSCs, a rectangular area with 50 coronal trabecular bone slides below the damaged growth plate was chosen for further analysis. Bone volumes were then analyzed via CT Analyzer (Bruker-MicroCT, v1.15.4.0).

For paraffin-embedded sections, slides were baked at 55°C overnight, dewaxed, and rehydrated using the standard protocol. Antigen retrieval was performed using a pressure cooker at 125°C and 25 psi for 5 minutes with pH 9.0 EDTA-Tris solution. For frozen sections, the slides were taken out of the freezer and warmed at room temperature for 10 minutes before being rinsed by PBS. All the slides were treated with 0.1M NH₄Cl solution for 10 minutes to reduce the autofluorescence and blocked in 10% donkey serum in PBS-GT (2% Gelatin, 0.5% TritonX-100) for 1 hour at room temperature. If the mouse primary antibodies were used on mouse tissues, M.O.M. Blocking Reagent (Vector Laboratories) was used for additional blocking. The slides were then incubated with primary antibodies at 4°C overnight and then stained with corresponding secondary antibodies for 2 hours. The primary antibodies used in this study were as follows: chicken anti-GFP (Abcam, cat. #ab13970, RRID: AB_300798, 10 mg/mL, 1:500); rabbit anti-mRFP (Rockland, cat. #600-401-379, RRID: AB_2209751, 1 mg/mL, 1:500); goat anti-mouse VE-Cadherin (R&D Systems, cat. #AF1002, RRID: AB_2077789; 1 mg/mL, 1:200); goat anti-mouse CD31(R&D Systems, cat. #AF3628, RRID: AB_2161028, 1 mg/mL, 1:200); rat anti-mouse Endomucin (Santa Cruz, cat. #sc-65495, RRID: AB_2100037, 200μg/mL, 1:100); rabbit anti-mouse Ki-67 (Abcam, cat. #ab15580, RRID: AB_443209, 1 mg/mL, 1:100, for frozen sections); rat anti-mouse Ki-67 (eBioscience, cat. #14-5698-82, RRID: AB_10854564, 0.5 mg/mL, 1:100, for paraffin-embedded sections); rabbit anti-mouse NG2 (EMD Millipore, cat. #AB5320, RRID: AB_11213678, 1 mg/mL, 1:200); goat anti-mouse LepR (R&D Systems, cat. #AF497, RRID: AB_2281270, 1 mg/mL, 1:100); rabbit anti-mouse PDGFRβ (Abcam, cat. #ab32570, RRID: AB_7771650.162 mg/mL, 1:100); rabbit anti-mouse Aggrecan (Millipore Sigma, cat. #AB1031, RRID: AB_90460, 0.5 mg/mL, 1:100); rabbit anti-mouse Perilipin A/B (Sigma, cat. #P1873, RRID: AB_532267, 1.0–1.4 mg/mL, 1:100); rabbit anti-human NG2 (Cell Signaling Technology, cat. #43916S, RRID: N/A, 18 μg/mL, 1:50); mouse anti-human/mouse N-cadherin (BD Biosciences, cat. #610921, RRID: AB_398236, 250 μg/mL, 1:25); mouse anti-human/mouse N-cadherin (Invitrogen, cat. #33-3900, RRID: AB_2313779, 0.5 mg/mL, 1:25); rat anti-human cytokeratin-8 (DSHB, cat. #TROMA-I, RRID: AB_531826, 1:200); rat anti-human cytokeratin-19 (DSHB, cat. #TROMA-III, RRID: AB_2133570, 1:200); mouse anti-DsRed Alexa Fluor 594 (Santa Cruz, cat. #sc-390909 AF594, RRID:AB_2801575, 200 μg/mL, 1:100); rabbit anti-Osteocalcin (Abcam, cat. #ab93876, RRID:AB_10675660, 1 mg/mL, 1:100); rabbit anti-S6K1 (phospho T389+T412; Abcam, cat. #ab60948, RRID:AB_944606, 1 mg/mL, 1:100); and goat anti-mouse E-cadherin (R&D Systems, cat. #AF748, RRID:AB_355568, 0.2 mg/mL, 1:50). The secondary antibodies used in this study were as follows: donkey anti-chicken Alexa Fluor 488 (Jackson ImmunoResearch, cat. #703-545-155, RRID: AB_2340375, 2 mg/mL, 1:500); donkey anti-mouse Alexa Fluor 488 (Jackson ImmunoResearch, cat. #715–545-151, RRID: AB_2341099, 2 mg/mL, 1:500); Donkey anti-rabbit Alexa Fluor 488 (Jackson ImmunoResearch, cat. #711-546-152, RRID:AB_2340619, 2 mg/mL, 1:500); donkey anti-goat Alexa Fluor 488 (Jackson ImmunoResearch, cat. #705-546-147, RRID:AB_2340430, 2 mg/mL, 1:500); donkey anti-rabbit Alexa Fluor 555 (Invitrogen, cat. #A-31572, RRID: AB_162543, 2 mg/mL, 1:500); donkey anti-rat Brilliant Violet 480 (Jackson ImmunoResearch, cat. #712-685-153, RRID: AB_2651113, 1:200); donkey anti-rat Alexa Fluor 555 (Abcam, cat. #ab150154, RRID:AB_2813834, 2 mg/mL, 1:500); donkey anti-rat Alexa Fluor 647 (Jackson ImmunoResearch, cat. #712-605-153, RRID: AB_2340694, 2 mg/mL, 1:500); donkey anti-goat Alexa Fluor 555 (Invitrogen, cat. #A-21432, RRID:AB_2535853, 2 mg/mL, 1:500); donkey anti-chicken Alexa Fluor 647 (Jackson ImmunoResearch, cat. #703-606-155, RRID:AB_2340380, 2 mg/mL, 1:500); donkey anti-goat Alexa Fluor 647 (Jackson ImmunoResearch, cat. #705-605-147, RRID: AB_2340437, 2 mg/mL, 1:500); and donkey anti-rabbit Brilliant Violet 421 (Jackson ImmunoResearch, cat. #711-675-152, RRID: AB_2651108, 1:100). TROMA-I and TROMA-III were deposited to the Developmental Studies Hybridoma Bank (DSHB) by P. Brulet/R. Kemler (DSHB Hybridoma Product TROMA-I/TROMA-III). The tdTomato protein maintains fluorescent activity in PFA-fixed frozen tissues and therefore does not require additional staining if other rabbit-derived primary antibodies have to be used on the same slides. Whenever compatible, Hoechst 33342 (Thermo Fisher Scientific, cat. #62249, 1 μg/mL) was used for nucleus staining. The stained slides were mounted with Prolong Gold Antifade Mountant (Invitrogen, cat. #P36934) and scanned by either a Zeiss LSM880 confocal microscope or a Zeiss AxioScan.Z1 slide scanner. To acquire images of a large field, tile scanning was performed with a 5% overlap. The fluorescent images were exported in TIFF format by Zen software and further analyzed by ImageJ. Briefly, the threshold was readjusted to filter the background and increase the brightness for each channel. The “Despeckle” function of ImageJ was applied to the images to reduce the noise. To quantify the expression of phosphorylated S6K1 in tumor cells, the GFP⁺ area was segmented, and the mean cytoplasmic intensity of pS6K1 was quantified by CellProfiler (Broad Institute, v4.1.3; ref. 60).

The OCT-embedded, intact hindlimbs were sectioned on both sides by cryostat to expose the bone marrow cavity and become about 500-μm thick sections. The sections were then rinsed by PBS to completely remove the embedding material and incubated with 1 mg/mL sodium borohydride solution for 30 minutes to reduce the autofluorescence. Samples were then transferred into a 1.5-mL Eppendorf tube with 1 mL blocking buffer [10% DMSO, 5% donkey serum, 0.5% IGEPAL-CA630, 1% anti-mouse CD16/32 Fc-Blocking antibody (Tonbo Bio, 70-0161-M001), and 1% BlokHen (Aves Labs)] and rotated in 4°C overnight. On the second day, samples were stained with primary antibodies in 10% DMSO, 5% donkey serum, 0.5% IGEPAL-CA630 staining buffer, including chicken anti-GFP (Abcam, cat. #ab13970, RRID:AB_300798, 10 mg/mL, 1:100), rabbit anti-mRFP (Rockland, cat. #600-401-379, RRID:AB_2209751, 1 mg/mL, 1:100), goat anti-mouse VE-Cadherin (R&D Systems, cat. #AF1002, RRID:AB_2077789, 1 mg/mL, 1:100), and goat anti-mouse CD31 (R&D Systems, cat. #AF3628, RRID:AB_2161028, 1 mg/mL, 1:100), for 3 days in 4°C with constant rotation. After several times of PBS washing for 1 day, samples were stained with donkey anti-chicken Alexa Fluor 488 (Jackson ImmunoResearch, cat. #703-546-155, RRID: AB_2340376, 2 mg/mL, 1:100), donkey anti-rabbit Alexa Fluor 555 (Invitrogen, cat. #A-31572, RRID: AB_162543, 2 mg/mL, 1:100), and donkey anti-goat Alexa Fluor 647 (Jackson ImmunoResearch, cat. #705–606–147, RRID: AB_2340438, 2 mg/mL, 1:100) in staining buffer for 3 days in 4°C with rotation. After 1-day PBS washing, samples were sequentially dehydrated in 30%, 50%, 70%, 90%, 100%, and 100% methanol per hour and then incubated with 100% methanol overnight. After dehydration, bone samples were transferred to a 5-mL Eppendorf tube with fresh BABB clearing reagent (1:2 benzoic acid:benzyl benzoate) and incubated at room temperature with gentle shaking for 2 days. The cleared bone samples were later mounted with BABB in a customized glass cassette with a depth-adjustable spacer (Sunjin Lab, cat. #IS002). The deep imaging was performed immediately with an Olympus FV1200 MPE confocal microscope. Specifically, a 10× water-immersed objective with an 8 mm working distance was used to obtain high-resolution 3D image stacks. Meanwhile, the laser power and PMT gain were scaled on different depths to generate images with uniform intensity. The mosaic scanning of the whole bone area was achieved with the motorized XY stage. The images were then stitched and exported as 8-bit TIFFs by Fluoview (Olympus), and the whole-view projection images were created by ImarisViewer (Oxford Instruments, v9.5.1).

The spatial distribution of tumor cells relative to NG2-tdRED⁺ cells or bone marrow vasculature was performed with a customized pipeline by MATLAB (MathWorks, v9.6). Basically, the ROI (bone marrow canal hereafter) was manually chosen from each image slide to remove the exterior periosteum and connective tissues, which typically contain saturated signals. Geometric active contour (GAC) was used to segment the cells and vessels from original images. GAC is a form of contour model that adjusts the smooth curve established in the Euclidean plan by moving the curve's points perpendicularly. The points move at a rate proportionate to the curvature of the image's region. The geometric flow of the curve, which encompasses both internal and external geometric measures in the ROI, and the recognition of items in the image were used to characterize contours. In the process of detecting items in an image, a geometric replacement for snakes was utilized. For the segmentation of tumor cells, a threshold of 50 μm² in 2D images or 100 voxels for 3D images was set to remove the cell debris or false-positive signals. For the analysis of 2D fracture bone images, the new bones were manually labeled based on the presence of Hoechst⁺GFP⁻Endomucin⁻Ki-67⁻ osteocytes. After the segmentation, a matrix with the spatial positions of cancer cells (with or without Ki-67 status), blood vessels, and NG2-tdRED⁺ cells was created and further analyzed. For random spot simulation, 1,000 spots with 10-μm diameter were randomly inserted in the image stacks after filtering the intravascular and bone matrix space, and their distances to the closest target were computed as described above.

A customized protocol was used in the lab to collect cells from both bone marrow cavity and endosteal compartments of bones. Fresh long bones were dissected immediately from the euthanized animals and placed in a 24-well plate containing 2 mL PBS with 2% FBS. The muscle and connective tissues were then carefully removed, and the hindlimb was broken off from the joint area to separate the femur and tibia bones. Bones were then transferred to a new well with 2% FBS PBS, and both metaphyseal ends were cut off to expose the bone marrow canal. For the quantification of cells at the fracture areas, both ends of femur bones were cut off and only the central shaft bones were subjected to the following procedures. Bone marrow was flushed out by a 26-G syringe with 2% FBS PBS into the wells, and the bone marrow cavity was thoroughly washed until the bones became pale. Bone marrow plugs were gently aspirated to become a single-cell suspension in red blood cell (RBC) lysis buffer (Tonbo Biosciences, cat. #TNB-4300-L100). The remaining bones were moved to a 12-well plate and excised into 1- to 2-mm bone chips by scissors. To release bone-attaching cells, fragmented bones were digested with 2 mL DMEM containing 1 mg/mL collagenase I (Sigma, cat. #C0130), 1 mg/mL collagenase II (Thermo Fisher Scientific, cat. #17101015), 4 mg/mL Dispase II (Sigma, cat. #D4693), 1 mg/mL BSA, 0.1 mg/mL DNase I (Sigma, cat. #DN25), 10 mmol/L HEPES, and 1 mmol/L EDTA at 37°C in 5% CO₂ for 45 minutes. The digested bones were then washed with 10 mL 10% FBS PBS and then PBS. The released cells from enzyme digestion and PBS washing were pooled and centrifuged. The endosteum-derived cells and the bone marrow cells were then processed separately (Supplementary Fig. S3A) or combined (others) for flow cytometry. To enable the detection of rare tumor cells in bones 1 day after IIA injection (Supplementary Fig. S2Q–S2S), the released bone cells were subjected to MACS sorting to deplete the majority of immune cells before flow cytometry.

The bone cell suspension was first passed through 70-μm cell strainer to completely remove bone debris and aggregated cells and then resuspended in 10 mL RBC lysis buffer to get rid of most blood cells. After 10-minute incubation at room temperature, cells were centrifuged and resuspended with 1 mL 2% FBS PBS. Samples were then aliquoted for staining of different panels (250 μL per panel) and blocked with anti-CD16/32 antibody (Tonbo, cat. #40-0161, RRID: AB_2621443, 0.2 mg/mL, 1:100) for 10 minutes on ice, followed by staining of fluorescent dye–conjugated primary antibodies or isotype control antibodies on ice for 20 minutes. The antibodies used in this study were as follows: immune cell panel, CD45-VF450 (Tonbo, cat. #75–0451, RRID: AB_2621947, 0.2 mg/mL, 1:200), CD11b-APC/Cy7 (Tonbo, cat. #25–0112, RRID: AB_2621625, 0.2 mg/mL, 1:200), Ly6G-Percp/Cy5.5 (Tonbo, cat. #65–1276, RRID: AB_2621899, 0.2 mg/mL, 1:200), Ly6C-PE/Cy7 (BioLegend, cat. #128018, RRID: AB_1732082, 0.2 mg/mL, 1:200), CD3e-PE (Tonbo, cat. #50-0031, RRID: AB_2621730, 0.2 mg/mL, 1:200), CD4-APC (Tonbo, cat. #20–0041, RRID: AB_2621543, 0.2 mg/mL, 1:200), CD8a-FITC (Tonbo, cat. #35–0081, RRID: AB_2621671, 0.5 mg/mL, 1:200), F4/80-BV605 (BioLegend, cat. #123133, RRID: AB_2562305, 0.1 mg/mL, 1:200), and B220-BV711 (BioLegend, cat. #103255, RRID: AB_2563491, 0.2 mg/mL, 1:200); stromal cell panel, CD45-BV605 (BioLegend, cat. #103140, RRID: AB_2562342, 0.2 mg/mL, 1:200), Ter119-VF450 (Tonbo, cat. #75-5921, RRID: AB_2621967, 0.2 mg/mL, 1:200), CD31-AF647 (BioLegend, cat. #102506, RRID: AB_312913, 0.5 mg/mL, 1:200), Sca-1-Percp/Cy5.5 (eBioscience, cat. #45–5981-82, RRID: AB_914372, 0.2 mg/mL, 1:200), CD51-PE (BioLegend, cat. #104106, RRID: AB_2129493, 0.2 mg/mL, 1:200), CD140α-PE/Cy7 (BioLegend, cat. #135912, RRID: AB_2715974, 0.2 mg/mL, 1:200); regulatory T-cell and Ki-67 panel, CD45-VF450 (Tonbo, cat. #75–0451, RRID: AB_2621947, 0.2 mg/mL, 1:200), CD3e-Percp/Cy5.5 (Tonbo, cat. #65–0031, RRID: AB_2621872, 0.2 mg/mL, 1:200), CD4-PE/Cy7 (Tonbo, cat. #60-0041, RRID: AB_2621828, 0.2 mg/mL, 1:200), CD8a-BV711 (BioLegend, cat. #100759, RRID: AB_2563510, 0.2 mg/mL, 1:200), Foxp3-PE (Tonbo, cat. #50-5773, RRID: AB_2621797, 0.2 mg/mL, 1:200), and Ki-67-APC (BioLegend, cat. #652406, RRID: AB_2561930, 0.2 mg/mL, 1:200); and apoptotic cells, Annexin V-APC/7-AAD Kit (Tonbo, cat. #20-6410-KIT, 5 μL for Annexin V, 025 μg 7-AAD). The staining of nuclear proteins Foxp3 and Ki-67 was performed with the Foxp3/Transcription Factor Fixation/Permeabilization Kit (eBioscience, cat. #00-5521-00). After the staining, the live cells were resuspended in 200 μL FACS buffer with a drop of DAPI (Invitrogen, cat. #R37606). To quantify the absolute number of tumor cells per bone sample, a determined amount of counting beads (BD Biosciences, cat. #335925) was added. The absolute number was calculated as (gated events)/(gated counting bead events) × 10 × 250 (μL) × 4. The flow cytometry was run on a BD LSRFortessa flow cytometer and further analyzed by FlowJo software (BD, v10.7.1). The markers for gating each population were as follows: B cells, DAPI⁻CD45⁺CD11b⁻B220⁺; T cells, DAPI⁻CD45⁺CD11b⁻CD3e⁺; CD4 T cells, DAPI⁻CD45⁺CD11b⁻CD3e⁺CD4⁺; CD8 T cells, DAPI⁻CD45⁺CD11b⁻CD3e⁺CD8a⁺; Ly6C^hi monocytes, DAPI⁻CD45⁺CD11b⁺Ly6C⁺Ly6G⁻; neutrophils, DAPI⁻CD45⁺CD11b⁺Ly6C⁺Ly6G⁺; Ly6C^lo monocytes, DAPI⁻CD45⁺CD11b⁺Ly6C^lo/−Ly6G⁻; macrophages, DAPI⁻CD45⁺CD11b⁺Ly6C⁻Ly6G⁻F4/80⁺; regulatory T cells, CD45⁺CD3e⁺CD4⁺CD8⁻Foxp3⁺; tumor cells, DAPI⁻CD45⁻GFP⁺; tdRED⁺ BMSCs, DAPI⁻CD45⁻Ter119⁻CD31⁻tdRED⁺; Tie2-tdRED⁺ cells, DAPI⁻CD45⁻Ter119⁻tdRED⁺; and PDGFRα⁺, CD51⁺, or Sca-1⁺ BMSCs, DAPI⁻CD45⁻Ter119⁻CD31⁻PDGFRα⁺(or CD51⁺, Sca-1⁺). The exact gating strategies are summarized in Supplementary Fig. S1.

To isolate enough NG2⁺ BMSCs for transplantation, the whole bone cells were prepared and pooled from the bones (femur, tibia, calvarium, and sternum bones) of nine NG2-CreERTM^+/−;tdRED^+/+ female mice. To facilitate cell sorting, the majority of immune cells were predepleted by MACS. Briefly, whole bone cells were incubated with CD45-biotin, CD11b-biotin, CD3e-biotin, Ly-6C/G-biotin, and TER-119-biotin (BD Biosciences, cat. #559971) in PBS with 2% serum and 1% antibiotics for 15 minutes at 4°C. Cells were washed and then resuspended and incubated with streptavidin-bound magnetic beads (BD Biosciences, cat. #557812) for 15 minutes at 4°C. Then cells were rinsed twice, and the biotin-negative cells were collected by EasySep Magnet (StemCell, cat. #18000). The CD45⁻CD31⁻TER119⁻DAPI⁻NG2-tdRED⁺ cells were immediately sorted from the enriched population by a BD Aria II cell sorter with a 100-μm nozzle. About 120,000 cells in total were collected and then resuspended in 60 μL PBS. Purified cells were directly transplanted into the femur bones of five 8-week-old female C57BL/6 mice (10 μL per animal) through intrafemoral injection with a 28-G BD insulin syringe. As sham controls, 5 age-matched female C57BL/6 mice were injected with 10 μL PBS.

The isolation and in vitro expansion of bone-derived mesenchymal stromal cells were performed as previously described (61) in an aseptic environment. Briefly, 3-week-old female NG2-CreERTM^+/−;tdRED^+/+ mice were treated with tamoxifen for 3 days to induce Cre activity and tdRED expression in NG2⁺ cells. One week later, the mice were euthanized and immediately immersed in 70% ethanol for 3 minutes. Hindlimb bones were dissected using sterile scissors and forceps, and the attached tissues were carefully rubbed away with autoclaved tissue papers. Femurs and tibias were then disconnected from the joint and placed in a sterile cell culture dish with 10 mL α-MEM supplemented with 2% FBS and antibiotics. Bone canals were exposed by cutting off both ends of the marrow cavity, and the bone marrow was flushed away with the medium by a 26-G syringe. The cleaned bones were then transferred to a 35-mm cell culture dish, crashed and fragmented into 1- to 2-mm chips with sterile bone pliers and scissors, and digested with 3 mL α-MEM medium containing 1 mg/mL Collagenase II (Thermo Fisher Scientific, cat. #17101015), 10% FBS, and antibiotics at 37°C on a shaker with 200 r.p.m for 1 hour. The digested bones were then collected, rinsed with α-MEM medium, and placed into a 10-cm culture dish with 10 mL α-MEM medium supplemented with 10% mesenchymal stem cell–qualified FBS (Gibco, cat. #12662029) and 1% antibiotics. On the second culture day, the medium was changed to remove the floating cells, and later the medium was changed every 3 days. After 7 days in culture, the attached cells were released by digestion of 2 mL MSC specialized Trypsin/EDTA (Lonza, cat. #CC-3232) for 3 minutes at 37°C. The detached stromal cells were collected, filtered, and washed through the cell strainer. Cells were resuspended with 0.4 mL FACS buffer and then sorted by a BD Aria II cell sorter with a 100-μm nozzle. The tdRED⁺ and tdRED⁻ cells were cultured in the α-MEM medium supplemented with 10% mesenchymal stem cell–qualified FBS, passaged every week, and used within 5 passages after isolation.

The CFU-F assay was performed by seeding 2,000 NG2-tdRED⁺ or control NG2-tdRED⁻ BMSCs in a 6-well plate. The cells were then maintained in α-MEM medium with 10% FBS at 37°C in 5%CO₂. After 7 days, the plates were washed twice with PBS and briefly fixed with 10% formalin solution at room temperature for 10 minutes. After fixation, cells were stained with 0.5% crystal violet solution (Sigma, cat. #C0775) for 15 minutes and rinsed with distilled water to remove the unbound dye. After drying, the plates were scanned with Cytation 5 (Biotek), and colonies with more than 50 cells were counted.

For the chondrogenic differentiation assay, the isolated mesenchymal stromal cells were suspended in the medium at a density of 10⁷ cells/mL and seeded as a 5 μL micromass drop on the bottom of a 24-well plate. Cells were allowed to attach to the plate for 2 hours in the cell culture incubator. Then, 1 mL of StemPro Chondrogenesis differentiation medium (Invitrogen, cat. #A1007101) was added to each well. Cells were then maintained at 37°C in 5% CO₂ for a minimum of 14 days. The differentiation medium was replenished every 3 days. At the endpoint, cells were rinsed with PBS three times and then fixed with 4% PFA for 30 minutes. After PBS washing, the fixed cells were stained with 1% Alcian blue (8GX) solution (Thermo Scientific, cat. #J60122.14) for 30 minutes, rinsed three times with 0.1N HCl to remove unbound dyes, and finally rinsed with PBS three times. For the osteogenic differentiation assay, 4E4 cells were plated in each well of the 96-well plate, and the osteogenic differentiation was induced by the StemPro Osteogenic Differentiation Kit (Invitrogen, cat. #A1007201) for 3 weeks. The differentiation medium was changed every 3 days. Then, after 10 minutes of fixation by 10% formalin, cells were stained with fresh 2% Alizarin Red S Solution (Thermo Scientific, cat. #400480250) at room temperature in the dark for about 45 minutes. The stained plates were then rinsed in distilled water to completely remove unbound dyes. For the adipogenic differentiation assay, 4E4 cells were seeded in a 96-well plate with the α-MEM complete growth medium supplemented with 2 μmol/L rosiglitazone (AdipoGen, cat. #AGCR13570M010), 500 μmol/L IBMX (Millipore Sigma, cat. #I5879), 1 μmol/L dexamethasone (MP Biomedicals, cat. #0219004025), and 10 μg/mL insulin (Millipore Sigma, cat. #I9278) for 3 days. Then, the differentiated cells were maintained in the α-MEM complete growth medium with 10 μg/mL insulin for another 2 days. Lastly, cells were briefly fixed with 10% formalin and rinsed with 60% isopropanol twice. After the plates were completely dried, cells were stained with fresh Oil Red O working solution (Alfa Aesar, cat. #A1298914) for 15 minutes and washed with distilled water to remove excess stain. For the ALP staining, 5E3 BMSCs pretreated with a 7-day osteogenic differentiation medium were seeded in each well of the 96-well plate. On the second day, the cells were briefly fixed and the staining was performed with the Alkaline Phosphatase Detection Kit (Millipore Sigma, cat. #SCR004). The images were acquired by a Leica MZ125 stereomicroscope.

To isolate the total mRNA, primary MSCs or cells after 7-day treatment of osteogenesis differentiation medium were trypsinized, collected, washed, and then dissolved in 0.5 mL TRIzol LS Reagent (Invitrogen, cat. #10296028). Total mRNA was extracted by the Direct-zol RNA miniPrep Kit (Zymo Research, cat. #R2052) with in-column DNase I treatment as instructed by the manufacturer's protocol. Total RNA (100 ng) was then converted into cDNA using the RevertAid First Strand cDNA Synthesis Kit (Thermo Scientific, cat. #K1622). qPCR was carried out with PowerUp SYBR Green Master Mix (Thermo Fisher, cat. #A25743) on a Bio-Rad CFX real-time system. The fold changes of mRNA levels were calculated by 2^−ΔΔCt with β-actin mRNA level as the internal control. The qPCR primer sequences used in this study were as follows: Actb, 5′-GGCTGTATTCCCCTCCATCG-3′, 5′-CCAGTTGGTAACAATGCCATGT-3′; Pou5f1, 5′-GGCTTCAGACTTCGCCTTCT-3′, 5′-TGGAAGCTTAGCCAGGTTCG-3′; Sox2, 5′-AGGAAAGGGTTCTTGCTGGG-3′, 5′-GACCACGAAAACGGTCTTGC-3′; Klf2, 5′-CACCTAAAGGCGCATCTGCGTA-3′, 5′-GTGACCTGTGTGCTTTCGGTAG-3′; Klf4, 5′-CTATGCAGGCTGTGGCAAAACC-3′, 5′-TTGCGGTAGTGCCTGGTCAGTT-3′; Pdgfra, 5′-AAAATGCGGGTTTTGAGCCC-3′, 5′-GACCAGAAAGACCTGGTGGG-3′; Sp7, 5′-AGCGACCACTTGAGCAAACAT-3′, 5′-GCGGCTGATTGGCTTCTTCT-3′; Cdh2, 5′-GGGAGGGGTAAAAGTTCTTAGCA-3′, 5′-TGGTACACAACACAGACGCA-3′; and Spp1, 5′-GAGGAAACCAGCCAAGGACTAA-3′, 5′-CTGAGATGGGTCAGGCACCA-3′.

For Western blotting, the primary MSCs were directly lysed with RIPA buffer and the total proteins were prepared accordingly. Total protein (20 μg) was loaded in the NuPAGE Novex Gel system (Invitrogen) and then transferred to the nitrocellulose membrane with the iBlot Transfer System (Invitrogen). After 1-hour blocking with 5% milk in TBST buffer, the membrane was incubated with rabbit anti-mouse N-cadherin (Abcam, cat. #ab12221, RRID: AB_298943, 1 mg/mL, 1:1,000) and mouse anti-mouse Gapdh (Santa Cruz, cat. #sc-32233, RRID: AB_627679, 100 μg/mL, 1:1,000) overnight at 4°C and stained with corresponding fluorescent secondary antibodies (LI-COR Bioscience, RRID: AB_621843 and RRID: AB_621842) on the second day for 2 hours. The membranes were then scanned with the Odyssey infrared imaging system.

The transfection was performed at about 50% cellular confluency using Lipofectamine RNAiMAX (Invitrogen, cat. #13778075) according to the manufacturer's instructions. Specifically, primary MSCs were transfected with a pooled validated siRNA mixture against mouse N-cadherin (Silencer Select siRNAs, Invitrogen, cat. #4390771-s63770 and s63771) in 15-cm dishes 3 days prior the coculture assay. For the E-cadherin knockdown in tumor cells, murine tumor cells were seeded in 6-well plates and transfected with the pooled siRNA mixture targeting mouse E-cadherin (Silencer Select siRNAs, Invitrogen, cat. #4390771-s63752 and s63753) 2 days before the coculture assays. A validated control siRNA (Silencer Select siRNAs, cat. #4390843) was used in parallel as a negative control. The sequence of siRNAs was as follows: siCDH1, s63752, sense, 5″-GAAGAUCACGUAUCGGAUUtt-3′, antisense, 5″-AAUCCGAUACGUGAUCUUCtg-3′; siCDH1, s63753, sense, 5″-GACCGGAAGUGACUCGAAAtt-3′, antisense, 5″-UUUCGAGUCACUUCCGGUCgg-3′; siCDH2, s63770, sense, 5″-GUGCAACAGUAUACGUUAAtt-3′, antisense, 5″-UUAACGUAUACUGUUGCACtt-3′; and siCDH2, s63771, sense, 5″-CCAGAACCCAACUCAAUUAtt-3′, antisense, 5″-UAAUUGAGUUGGGUUCUGGag-3′.

For the coculture assays, murine tumor cells and BMSCs were trypsinized, rinsed, and resuspended in 10⁵/mL and 10⁶/mL density with serum-free DMEM/F12 medium, respectively. For the 2D coculture assay, 200 tumor cells with or without 5,000 BMSCs were seeded into each well of a 96-well plate and cultured in serum-free DMEM/F12 medium. For the noncontact coculture assay, 2,000 tumor cells in 1 mL medium were seeded in the lower wells of a 24-well plate and then incubate for 30 minutes to allow the attachment of cells to the plates. Then, 3E4 BMSCs in 200 μL medium were seeded in the upper transwell inserts (3.0-μm pore size, Greiner Bio-One, cat. #662630). 0.5X B27 (Thermo Fisher Scientific, cat. #17504044) was added to sustain the survival of BMSCs in serum-free conditions. For the 3D coculture assay, 2E4 tumor cells with or without 2E4 BMSCs were seeded in each well of a 24-well low-attachment plate (Corning, cat. #3473) with 1 mL serum-free DMEM/F12 medium. The medium was changed every 3 days. For 2D and noncontact coculture assays, the cell growth was monitored by IncuCyte, and the total tumor cell area on day 5 for each well was quantified. For the 3D coculture assay, the fluorescent images of the sphere formed in each well were acquired by an Echo Revolve epi-fluorescent microscope (Figs. 5B and 6E) or IncuCyte (others), and the size of the sphere was determined by measuring the GFP⁺ area by ImageJ. The spheres from the same experimental groups were then pooled and transferred to a 1.5-mL Eppendorf tube. After 5-minute centrifugation at 500 × g, the cell pellets were fixed in 4% PFA for 2 hours and then sent to the pathologic core for paraffin embedding and sectioning.

For the comigration assay, the FluoroBlok inserts with 8-μm pore size (Corning, cat. #351152) were used. Basically, 1E4 tumor cells with or without 1E4 BMSCs were seeded in the top chambers. Serum-free DMEM/F12 medium (0.6 mL) was added in the bottom chambers. Then, the migration of tumor cells and BMSCs was monitored by IncuCyte. The colored membranes blocked the light transmission; therefore, the cells staying in the top chambers were invisible. Once the fluorescently labeled cells migrated through and were present under the membrane, cells could be detected.

GSE14020 (breast), GSE77930 (prostate), GSE101607 (mixed), and GSE143791 (prostate) were downloaded from the Gene Expression Omnibus database (https://www.ncbi.nlm.nih.gov/geo/). GSE143791 is an scRNA-seq dataset. Gene counts were read into the R Statistical Program. Genes with a count number > 1 in a cell were considered expressed by the cells. Coexpression between two genes was evaluated by Fisher exact tests. The P values of the tests were jointly determined by the number of cells expressing either genes or the number of cells expressing both genes. Therefore, we used the reciprocal of log-transformed P values as an abundance index of cells with multiple properties in Fig. 7C and D. Specifically, such index was used for cells expressing both NG2 and pericyte markers (perivascular NG2⁺ cells) and for cells expressing both CDH2 and osteoblast differentiation signature. When multiple genes were examined against a single gene, the median index of the multiple genes was used. Other datasets are microarray and RNA-seq datasets. Normalized data were read in the R Statistical Program. For individual genes, we used log₂-transformed values. For gene signatures, we used a single-sample gene set enrichment assay implemented by the “gsva” package (62). All figures were generated by the “ggplot2” package. The signatures used in the analyses and the corresponding link or reference are specified as follows: Osteoblast differentiation signature, https://www.gsea-msigdb.org/gsea/msigdb/cards/REACTOME_RUNX2_REGULATES_OSTEOBLAST_DIFFERENTIATION.html; Prostate cancer proliferation signature, Cuzick et al., Lancet Oncol (2011;12:245–55; ref. 47); Embryonic stem cell signature expressed in aggressive tumors, https://www.gsea-msigdb.org/gsea/msigdb/cards/BENPORATH_ES_1.html; Estrogen response signature (early), https://www.gsea-msigdb.org/gsea/msigdb/cards/HALLMARK_ESTROGEN_RESPONSE_EARLY.html; and Estrogen response signature (late), https://www.gsea-msigdb.org/gsea/msigdb/cards/HALLMARK_ESTROGEN_RESPONSE_LATE.html.

If not specified otherwise, all the quantitative data were generated and analyzed by GraphPad Prism 9. The number of animals or independent replicates is denoted in the figure panels or legends. In most in vivo experiments, the investigators were blinded to the genotype or allocation of subjects until the assessment of outcomes. All the in vitro experiments were repeated three times or involved 3 or more biological replicates. No repeated measurement on the same sample was applied in this study. Preliminary experiments with a small group size were performed to determine the group sizes in each in vivo experiment. The final results were pooled from multiple batches of in vivo experiments including the preliminary experiments. No animal that reached the experimental endpoint was excluded from the final analysis. The statistical methods are noted in the figure legends, and two-sided tests were used. Normally distributed data were analyzed with parametric statistics whereas nonparametric statistics were used for the comparison of metastatic burden, such as bioluminescent intensities and derived data. For the comparison of two normally distributed samples, an F-test was performed to assess the variance of two samples. If the F-test was rejected, Welch correction was applied to the Student t test. P values less than 0.05 were considered statistically significant.

All the datasets for bioinformatics analysis were downloaded from a public database, and the accession numbers are provided in the corresponding figure panels. The numeric values of all figures are provided in Supplementary Raw Data. Other raw data and codes are available upon request from the corresponding author.

S.T.C. Wong reports grants from Houston Methodist during the conduct of the study. No disclosures were reported by the other authors.

W. Zhang: Conceptualization, data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. Z. Xu: Formal analysis, validation, investigation, methodology. X. Hao: Formal analysis, validation, investigation, methodology. T. He: Data curation, software, formal analysis, validation, investigation, visualization, methodology. J. Li: Formal analysis, validation, investigation. Y. Shen: Data curation, software, formal analysis, validation, investigation, visualization. K. Liu: Formal analysis, validation, investigation. Y. Gao: Formal analysis, validation, investigation. J. Liu: Formal analysis, validation, investigation. D.G. Edwards: Formal analysis, validation, investigation. A.M. Muscarella: Formal analysis, validation, investigation. L. Wu: Formal analysis, validation, investigation. L. Yu: Formal analysis, validation, investigation. L. Xu: Formal analysis, validation, investigation. X. Chen: Resources. Y.-H. Wu: Formal analysis, validation, investigation. I.L. Bado: Formal analysis, validation, investigation. Y. Ding: Formal analysis, validation, investigation. S. Aguirre: Formal analysis, validation, investigation. H. Wang: Formal analysis, validation, investigation, methodology. Z. Gugala: Resources. R.L. Satcher: Resources. S.T.C. Wong: Conceptualization, resources, supervision, funding acquisition, methodology. X.H.-F. Zhang: Conceptualization, resources, data curation, software, supervision, funding acquisition, visualization, methodology, writing–original draft, project administration, writing–review and editing.

We are grateful to the discussion and suggestions from the Zhang laboratory members. We also thank Dr. Dongsu Park (Baylor College of Medicine) and Dr. Yangjin Bae (Baylor College of Medicine) for providing relevant animal strains. X.H.-F. Zhang is supported by the U.S. Department of Defense DAMD W81XWH-16-1-0073 (Era of Hope Scholarship), NCI CA183878, NCI CA251950, DAMD W81XWH-20–1-0375, the Breast Cancer Research Foundation, and the McNair Medical Institute. S.T.C. Wong and J. Li are supported in part by NCI U01CA252553, the John S. Dunn Research Foundation, and the T.T. and W.F. Chao Center for BRAIN. Y. Gao received training support from the Translational Breast Cancer Research Training Program (NIH T32 CA203690, principal investigator: Suzanne Fuqua). This project was supported by the RNA In Situ Hybridization Core with funding from the NIH (1S10OD016167), the Cytometry and Cell Sorting Core with funding from the Cancer Prevention & Research Institute of Texas (CPRIT) Core Facility Support Award (CPRIT-RP180672), the NIH (P30 CA125123, S10 RR024574, and S10 OD025251), the Pathology Core of Lester and Sue Smith Breast Center, the Optical Imaging and Vital Microscopy Core and the Integrated Microscopy Core at Baylor College of Medicine, and the MD Anderson Cancer Center Research Histology Core and Bone Histomorphometry Core Laboratory. We also acknowledge Dr. Cecilia Ljungberg, Joy Guo, Rena Mao, and Joel M. Sederstrom at the Baylor College of Medicine and Leah Guerra at MD Anderson Cancer Center for their expert assistance.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.