Breast CSCs are enriched in the circulation and on arrival in the lung
Our fluorescent lentiviral CSC reporter (SORE6>FP) dynamically reports on stemness through activation of an artificial enhancer (‘SORE6’) by the stemness master transcription factors OCT4 and SOX2 or their paralogues14,15. A control construct lacking the SORE6 enhancer but retaining the minimal CMV promoter (minCMV) is used for gating or thresholding (Fig. 1a and Extended Data Fig. 1a). The destabilized mCherry has a short half-life of ~1.4 h (Extended Data Fig. 1b), which confers high selectivity and temporal resolution to the reporter. For this study, tumour cells were transduced with a constitutive nuclear GFP marker (all tumour cell nuclei green), and either SORE6>dsmCherry (CSCs marked red/yellow) or the matched control construct. Representative images and flow cytometry profiles for transduced tumour cells are given in Fig. 1b and Extended Data Fig. 1c. Fluorescence-activated cell sorting (FACS)-sorted SORE6+ cells were significantly enriched for tumour-initiating ability by in vivo limiting dilution analysis in all models tested (Extended Data Fig. 1d,e), and are referred to as CSCs in our subsequent analyses. In breast cancer models, our reporter enriches both tumour-initiating and metastasis-initiating activity in the OCT4hi SOX2hi SORE6+ population14, and we use the umbrella term CSC throughout.
Fig. 1: Breast CSCs are enriched in the circulation and on arrival in the lung.
a, Schematic of SORE6 CSC reporter and matched control reporter. minCMVp is a minimal CMV promoter; FP, fluorescent protein. b, Epifluorescent images of LM2 cells transduced with the SORE6>dsmCherry reporter or minCMV>dsmCherry control construct growing in vitro. Representative of >1,000 images from >100 independent experiments. All tumour cells have a constitutive nuclear GFP marker. Arrows indicate CSCs. Scale bar, 100 µm. c,d, Confocal images of fresh orthotopic primary tumours from LM2 model (c) and 4T1 model (d), representative of n ≥ 3 tumours per model and multiple fields per tumour. Arrows indicate CSCs; double arrow indicates rare CSC doublet. Scale bars, 50 µm. e, High-resolution image of LM2 cells arriving in the lung 10 weeks after orthotopic implantation of a primary tumour. BV, blood vessel. Scale bar, 50 µm This event was seen only once. f, Image of circulating tumour cells from a mouse with an LM2 primary tumour, representative of 35 images. Scale bar, 10 µm. g, Quantitation of the percentage of CSCs in primary tumour, circulating tumour cells (CTCs) and single disseminated tumour cells (DTCs) in lungs, from LM2 or 4T1 orthotopic primary tumours. Two-sided Fisher’s exact test. Number of cells (n) assessed is indicated for each condition, pooled from n = 3 mice (LM2 primary tumour), n = 4 mice (4T1 primary tumour), n = 4 mice (LM2 CTCs and DTCs) or n = 9 mice (4T1 CTCs and DTCs). NS, not significant.
Source data
To address CSC population characteristics during the early stages of triple-negative breast cancer (TNBC) metastasis to the lung, we used the human MDAMB231-LM2 (hereafter ‘LM2’)16 and the mouse 4T1 models17. Due to the immunogenicity of the dsmCherry protein, the 4T1 model was used in immunocompromised mice. The SORE6 reporter marked a minor population of tumour cells in orthotopic primary tumours from LM2 (Fig. 1c) or 4T1 cells (Fig. 1d). CSCs typically appeared as isolated single cells, with only occasional doublets, usually towards the tumour edge. In Fig. 1e, we captured several tumour cells from an LM2 orthotopic primary tumour simultaneously arriving in the vasculature at the lung metastatic site; 25 of 32 tumours (78%) were CSCs, suggesting a strong enrichment of CSCs among intravasated cells. Indeed, circulating tumour cells recovered from the blood of mice with LM2 or 4T1 orthotopic primary tumours were 75–80% CSCs, compared with 1–2% CSCs in the originating primary tumours (Fig. 1f,g), confirming enrichment of CSCs in the circulation as previously observed by us and others15,18. Similarly, single disseminated tumour cells detected in the lungs of mice with orthotopic primary tumours were 72–75% CSCs, suggesting that this enrichment persists after extravasation (Fig. 1g). Thus, our imaging tool is performing as expected and enables quantitative insights into CSC and nonCSC populations during tumour progression.
Different population dynamics for CSCs and nonCSCs in early lung colonization
We next undertook a detailed analysis of CSC population dynamics following tumour cell arrival in the lung. To generate sufficient metastatic events, we used a pseudo-orthotopic format with tail vein injection for the LM2 model (Fig. 2a), and orthotopic implantation with primary tumour resection for the 4T1 model. Unless otherwise noted, unsorted tumour cell cultures containing both CSCs and nonCSCs were injected, and subpopulations in metastatic lesions were quantitated by ex vivo imaging of freshly excised lungs at different times. For the LM2 model, nonCSCs from unsorted cultures died off rapidly within 2 days of arrival in the lung, while CSCs mostly maintained their numbers (Fig. 2b). Injection of FACS-sorted LM2 CSCs and nonCSCs showed that only CSCs seeded metastases efficiently (Extended Data Fig. 2a). Thus CSCs have a selective early survival advantage in the lung and are uniquely capable of initiating progressively growing metastases.
Fig. 2: CSCs and nonCSCs have different population dynamics in early lung colonization.
a, Schematic for lung metastasis study formats for LM2 model. b, Number of surviving LM2 tumour cells in each subpopulation (CSC versus nonCSC) in the lung, normalized to time t = 6 h. Results are mean ± s.d. for n = 3 (6 h and 24 h) or 6 mice (48–72 h) per timepoint. Two-sided Mann–Whitney U test. NS, not significant (P > 0.99 for 6 h and P = 0.10 for 24 h). c, Metastatic lesion size (number of tumour cells per lesion) as a function of time after tail-vein injection of LM2 cells. Boxplots show median ± interquartile range, with whiskers extending to maximum and minimum values, for pooled lesions from three mice per timepoint. Two-sided Dunns multiple comparisons test versus day 3. NS, not significant (P = 0.86). d, Images of LM2 lung metastatic lesions of different sizes, representative of hundreds of images from >50 mice. Arrow indicates representative CSC. Scale bar, 50 µm. e, CSCs as the percentage of total tumour cells in LM2 metastatic lesions of different sizes. Data pooled from multiple timepoints for a total of 26–88 lesions analysed per size bin from 12 mice. Red lines, medians; dotted lines, quartiles. Kruskal–Wallis test. f, Different kinetics for expansion of LM2 CSC and nonCSC populations in the lung as a function of lesion size following tail-vein injection. n = 1–63 depending on lesion size (pooled from 12 mice and 390 lesions over multiple timepoints). Data for lesion sizes with n ≥ 3 datapoints are plotted as mean ± s.d. g, LM2 lung metastases immunostained for Ki67 (proliferation), GFP (tumour cells) and mCherry (CSCs). Representative of >20 images from 6 mice. Arrow indicates CSC. Scale bar, 10 µm. h, Percentage of cells within CSC or nonCSC subpopulations that are Ki67+ in LM2 lung lesions of different sizes. Results are mean ± se.m., n = 3–54 lesions per size bin, from 6 mice. Two-sided unpaired t-test. *P = 0.0143, **P = 0.00187, ***P = 0.000612, ****P = 0.000073. i, Percentage of cells within subpopulations that are Ki67+, for larger lesions from an independent experiment (n = 15 metastases from 2 mice). Solid lines indicate medians, and dotted lines indicate quartiles. Two-sided Mann–Whitney U test). j, Composition of two-cell lesions in LM2 model (n = 56 lesions from 9 mice, days 3–5). Representative of two independent experiments. k, Model for CSC and nonCSC kinetics in early metastatic colonization. Schematic in a created in BioRender; Wakefield, L. https://biorender.com/1tbn57h (2026).
Source data
A major advantage of our imaging approach is that the number of CSCs can be assessed for individual metastatic lesions. Median metastatic lesion size increased with time after tumour cell injection, but a wide range of metastasis sizes existed at each timepoint (Fig. 2c). The percentage of CSCs in a lesion of a given size was consistent over time (Extended Data Fig. 2b), allowing us to pool data across timepoints, with increasing lesion size generating a pseudotime trajectory. With this approach, we showed that the percentage of CSCs decreased as metastatic lesion size increased (Fig. 2d,e). Quantitating absolute CSC and nonCSC numbers in individual lesions revealed complex population dynamics. The CSC subpopulation reached an early plateau, while nonCSCs expanded continuously (Fig. 2f). Similar kinetics were observed for LM2 metastasis from the orthotopic primary site, though with fewer evaluable lesions, and for the 4T1 orthotopic model (Extended Data Fig. 2c–g).
To address the basis of these very different subpopulation dynamics, we assessed proliferation by co-immunostaining lungs for Ki67 (Fig. 2g). In the smallest lesions, most CSCs are proliferating, while the few nonCSCs are quiescent and only start to proliferate as lesions enlarge (Fig. 2h). Unexpectedly, CSCs were consistently more proliferative than nonCSCs, even in much larger metastases (Fig. 2i). Looking at the very earliest CSC division event by analysing the composition of two-cell lesions showed that nearly half were expansive self-renewing divisions (Fig. 2j), a third involved asymmetric divisions that jump-start nonCSC expansion, and the rest were fully differentiating divisions generating two nonCSCs. This latter is probably a metastatic dead-end, as we never observed lesions of >32 cells without a CSC (Extended Data Fig. 2h), suggesting that nonCSCs can undergo a maximum of 5 population doublings before exhausting their limited proliferative potential.
Overall, our data lead to a model for early metastasis dynamics (Fig. 2k). After arrival in the lung, most nonCSCs and some CSCs die off. Metastatic lesion initiation is driven solely by CSC proliferation, combining symmetric expansive divisions with asymmetric divisions that generate CSCs and nonCSCs. In larger lesions, the CSC population stops expanding despite ongoing CSC proliferation, so CSC divisions in larger metastases must be asymmetric (either physically or statistically). This strategy maintains CSC numbers while refreshing the nonCSC pool, which has limited proliferative potential. Expansion of larger metastases is driven primarily by proliferation of the more numerous nonCSC progeny. The observation that the CSC population plateaus early despite ongoing CSC proliferation implies a strong negative feedback mechanism in larger metastases that regulates CSC fate to limit CSC population expansion.
Early inhibition of CSC population expansion is density dependent and driven by increased CSC differentiation
We reasoned that understanding this inhibitory feedback mechanism might help identify targetable pathways to selectively eliminate CSCs. Using Incucyte live cell imaging for simultaneous real-time monitoring of CSCs and nonCSCs (Fig. 3a), we found that the early shutdown of CSC expansion could be replicated in two-dimensional culture in vitro (Fig. 3b). The CSC population reached an early plateau while the nonCSCs in the same culture continued to expand, as seen in the metastatic lung. Consequently the percentage of CSCs in the culture is continually changing. Broadly similar kinetics were seen in other breast cancer models (Fig. 3c–g), although in some the CSC peak was followed by a loss phase.
Fig. 3: Early inhibition of CSC population expansion is driven by increased CSC differentiation.
a, CSCs and nonCSCs in unsorted LM2 cultures in vitro with increasing time in culture. Images are representative of >100 similar experiments. Scale bar, 200 µm. b, Expansion kinetics for CSCs and nonCSCs in unsorted LM2 cells in vitro as a function of time. The percentage of CSCs is shown on the right y axis. Mean ± s.d., n = 3. c–g, Expansion kinetics of CSCs and nonCSCs in unsorted cultures of the additional breast cancer cell lines 4T1 (mouse) (c), BT549 (d), SUM149 (e), SUM159 (f) and MCF10Ca1h (g). Mean ± s.d., n = 3. Data in b, c and e–g represent ≥2 independent experiments. h, Schematic for all possible CSC cell fates in fate-mapping experiments. i, Time-lapse images showing different types of CSC cell division, representative of ≥20 such events. Red: cytoplasmic mCherry (CSCs); white: nuclear GFP (all tumour cells). Yellow asterisks indicate parent CSC and mitotic daughters. Scale bar, 30 µm. j, Timeframe for fate mapping of CSCs in relation to their expansion and maintenance/loss phase in unsorted LM2 cultures. Mean ± s.d., n = 3. k, CSC cell fates during expansion (early) phase and maintenance/loss phase (late) for LM2 and MCF10Ca1h cells. Two-sided Fisher’s exact test for self-renewal versus other fates. l, NonCSC cell fates in LM2 cultures in early (expansion) phase. m, Origin of CSCs present at t = 60 h (end of expansion phase) in LM2 cultures. In l and m, n indicates number of cells mapped. n, Schematic for relative frequency of major cell fate changes in LM2 cultures in the 36–60 h timeframe (expansion phase). The graphic shows the percentage of mapped CSCs or nonCSCs undergoing the indicated fate change during the 24-h period. The remaining 10% of nonCSCs and 14% of CSCs are unchanged during the mapping period.
Source data
To address what happens at the CSC plateau, we performed single-cell fate mapping. Possible CSC fates are schematized in Fig. 3h, with representative images in Fig. 3i and mapping intervals in Fig. 3j. CSC differentiation frequency increased significantly in LM2 cultures at the plateau phase, with a concomitant reduction in expansive self-renewal (Fig. 3k). The shift was more pronounced in MCF10Ca1h, which showed a net loss of CSCs in the late phase, due to a bigger increase in direct differentiation (Fig. 3k). Thus at later times in culture, CSC cell fate decisions tip from expansive self-renewal to differentiation, with reduced CSC proliferation in some but not all models.
Tumour cells show more phenotypic plasticity than their normal counterparts1,13. To assess whether nonCSC dedifferentiation also contributes to CSC population expansion, we fate-mapped LM2 nonCSCs. The spontaneous nonCSC dedifferentiation frequency was 2 in 100 cells per 24 h (Fig. 3l), compared with a CSC differentiation frequency of 16 in 100 cells (Fig. 3k). Thus, the forward conversion frequency of CSC to nonCSC is ~8× higher than the reverse conversion of nonCSC to CSC. However, because nonCSCs are more numerous, their low dedifferentiation frequency can still contribute substantially to CSC population expansion. Origin mapping of CSCs present at t = 60 h showed that 22% were generated by dedifferentiation of nonCSCs (Fig. 3m). Thus, in this aggressive TNBC model, nonCSC plasticity helps drive CSC population expansion. The relative frequency of the major CSC and nonCSC fate changes in the LM2 expansion phase is schematized in Fig. 3n.
CSCs are hyper-responsive to microenvironmental inputs
We next addressed possible mechanisms driving the negative feedback and fate decision changes at the CSC plateau. With higher initial seeding densities, the CSC population peaked earlier in time, but when plotted against total cell number, it always peaked at the same cell density (Fig. 4a,b), suggesting that some feature of a dense culture governs CSC fate decisions. Although our initial hypothesis was that the expanding nonCSC population might feedback inhibit the CSCs, CSC numbers plateaued at the same total cell density regardless of culture composition (CSC versus nonCSC) (Extended Data Fig. 3a,b).
Fig. 4: CSCs are hyper-responsive to microenvironmental inputs.
a, Effect of initial seeding density on LM2 CSC expansion kinetics as a function of time in unsorted cultures containing both CSCs and nonCSCs. Arrows indicate peak CSC numbers. Mean ± s.d., n = 3 per timepoint. Data are representative of independent experiments with two different cell lines. b, Data from a replotted as a function of culture density (total cells = CSCs + nonCSCs). c,d, Experimental design (c) and effect (d) of relief of cell crowding on CSC (orange/red) or nonCSC (grey/black) population dynamics in unsorted LM2 cultures. Arrow indicates time of reseeding. e, Effect of integrin signalling blockade with FAK inhibitor (defactinib) on LM2 CSC and nonCSC dynamics. Arrow indicates time of inhibitor addition. f, Effect of collagen or fibronectin coating of well on LM2 CSC and nonCSC dynamics. g, Effect of matrix stiffness on relative CSC numbers by flow cytometry. Mean ± s.d. for n = 4 (1 kPa and 12 kPa) or n = 3 (3 GPa). Adjusted P values from Tukey’s multiple comparisons test. h, Contact-independent effects of additional high-density LM2 cells cultured in a well insert on the expansion of LM2 CSCs and non-CSCs. i, Effect on LM2 cells of replacing spent medium with fresh medium at 48 h (arrow). j, Effect of LPA or glucose addition at 90 h (arrow) on LM2 cultures. k, Effect of LPAR inhibitor BMS986020 addition at 24 h (arrow) to LM2 cultures. In a–f and h–k, data for CSCs and nonCSCs in these mixed cultures are plotted separately for clarity. Results are mean ± s.d., n = 3 replicates. In a–k, results are representative of at least two independent replicate experiments. l, Model for role of CSC as a microenvironmental sensor for the tumour cell mass. Schematics created in BioRender: c, Wakefield, L. https://biorender.com/dr7tj94 (2026); h, Wakefield, L. https://biorender.com/l56dlm1 (2026); l, Wakefield, L. https://biorender.com/sj1jd9c (2026).
Source data
Increasing culture density entails many microenvironmental changes, including increased cell–cell contact and crowding, reduced cell spreading and matrix interaction, nutrient depletion and build-up of inhibitory factors. To address the contribution of cell crowding, we took LM2 cells in CSC plateau phase, removed the spent medium, trypsinized the cells and reseeded them into the same wells at the original density (1-in-1) or at a 1-in-5 split ratio, in the presence of the spent medium (Fig. 4c). The 1-in-5 split led to a rapid expansion of the CSC population, while reseeding at the original cell density had no effect (Fig. 4d). NonCSCs in the same culture responded far less to low-density reseeding (Fig. 4d), suggesting that relief of cell crowding can selectively promote CSC population expansion.
Cell crowding decreases extracellular matrix (ECM) contact and reduces integrin-driven activation of FAK/SRC signalling19,20 (Extended Data Fig. 3c). Consistent with a positive role for integrin signalling in CSC self-renewal, the FAK inhibitor defactinib slowed CSC population expansion earlier and more extensively than nonCSC expansion (Fig. 4e). Conversely, plating cells on collagen or fibronectin to stimulate integrin signalling selectively increased the CSC population (Fig. 4f), and the percentage of CSCs also increased with matrix stiffness (Fig. 4g), consistent with prior observations21. Thus CSCs show a heightened sensitivity to biochemical and biomechanical features of their microenvironment.
Culture density effects that were independent of cell contact were also observed. Tumour cells in a Transwell insert decreased peak CSC numbers in the well below, with a much smaller effect on nonCSCs (Fig. 4h), implying a contact-independent inhibitory signal. Medium refreshment at the CSC plateau caused a transient increase in expansive CSC self-renewal that was faster and larger than any effect on nonCSCs (Fig. 4i), while conditioned medium from high-density cultures caused CSC loss (Extended Data Fig. 3d). The stimulatory effect of medium refreshment on LM2 CSCs could be mimicked by lysophosphatidic acid (LPA), a serum mitogen22, but not by additional glucose (Fig. 4j), and CSC expansion was blocked by the LPA receptor inhibitor BMS986020 (Fig. 4k).
In every perturbation tested, the response of CSCs was more extensive and/or faster than the nonCSC response (Fig. 4d–k), leading us to suggest that CSCs serve as sensitive sensors of microenvironmental quality for the tumour cell population as a whole (Fig. 4l). In this model, CSCs continuously evaluate their local microenvironment and calibrate their numbers in response to available resources and space by altering the balance between self-renewal and differentiation. Because nonCSCs derive from CSCs, these rapid CSC responses ultimately drive downstream changes in the nonCSC population, but this will occur with a temporal lag, as most nonCSCs have some residual proliferative capacity and nonCSC proliferation is less immediately affected by environmental conditions. Similar results were seen with 4T1 or SUM159 cells (shown for FAK inhibition in Extended Data Fig. 3e,f).
Multiple CSC sensor inputs converge on YAP/TAZ
To identify underlying molecular mechanisms, we performed RNA sequencing on CSCs and nonCSCs sorted from LM2 in vitro cultures or metastasis-bearing lungs (Extended Data Fig. 4a). The CSC transcriptome showed strong enrichment of stemness, drug resistance and metastasis gene signatures (Extended Data Fig. 4b). Gene set enrichment analysis identified the Hippo pathway as the most highly enriched Reactome Signaling pathway in the CSCs in vitro (Fig. 5a), and ingenuity pathway analysis identified YAP and TAZ (WWTR1 gene) as the most significant positive upstream regulators of the CSC transcriptome (Fig. 5b).
Fig. 5: Multiple microenvironmental inputs to the CSC sensor converge on YAP/TAZ.
a, Top Reactome Signaling pathways enriched in CSC versus nonCSC transcriptomes from sorted LM2 cells in vitro. b, Ingenuity Pathway Analysis showing predicted upstream regulators of the LM2 CSC transcriptome in vitro. P values for overlap were calculated using a one-sided Fisher’s exact test with Benjamini–Hochberg (BH) correction for multiple hypothesis testing. c,d, Multiplexed immunofluorescence (c) and quantitation (d) of nuclear YAP intensity in CSCs and nonCSCs in unsorted LM2 low-density culture. Image representative of ≥20 similar fields. Scale bar, 5 µm. n = 24 cells (CSCs) and n = 493 cells (nonCSCs); two-sided Mann–Whitney test. e, Effect of siRNA knockdown of YAP or TAZ on CSC dynamics in unsorted LM2 cultures. Note that results are plotted on the cell density axis (total number of cells). Mean ± s.d., n = 3. Results replicated in an independent experiment using shRNA knockdown. f, Effect of overexpression of constitutively active mutant YAP-S127A on CSC dynamics in unsorted LM2 cultures. Mean ± s.d., n = 3. g, Single-cell fate mapping of LM2 CSCs following YAP or TAZ knockdown. Two-sided Fisher’s exact test versus control for all self-renewal versus all differentiation. n indicates the number of cells assessed. h, Effect of YAP/TAZ knockdown on CSC self-renewal:differentiation ratio in LM2 cells (data from g). The horizontal dashed line indicates balanced self-renewal and differentiation. i, Schematic summarizing effect of YAP on CSC fates. Dashed line: event numbers too small to determine effect. j, Effect of stable YAP knockdown on metastatic efficiency of LM2 cells. Median ± interquartile range for n = 9 mice per group; two-sided Mann–Whitney U test. k, Schematic of upstream inputs to YAP/TAZ node. Red outlines indicate pathway components targeted by inhibitors in this study. l, Gene set enrichment plots for YAP signature and Rho-GTPase signalling in CSCs recovered from LM2 lung metastases in vivo. NES, normalized enrichment score; FDR, false discovery rate. m,n, Effect of LATS kinase inhibitor TRULI (m) or ROCK inhibitor Y27632 (n) on CSC dynamics in unsorted LM2 cultures. Mean ± s.d., n = 3. Western blots show effect of inhibitors on phospho-YAP-S127 (inactive form of YAP). o, Bubble plot summarizing relative weight of select input pathways across three TNBC models, as assessed by response to inhibitors (see Extended Data Fig. 5a for original plots). The plots in e, f, m and n are representative of at least two independent experiments. Schematic in k created in BioRender; Wakefield, L. https://biorender.com/gg1wwwlv (2026).
Source data
YAP and TAZ constitute a key nodal point in the transduction of biomechanical and other extracellular cues that regulate organ size and cell fate decisions20,23,24. Previously implicated in regulating stemness in breast and other cancers20,23, they are plausible integrators of inputs to the CSC in its role as a microenvironment sensor. Activation of YAP and TAZ involves their stabilization and translocation to the nucleus24, where they interact with TEADs to modulate transcription25. We showed preferential localization of YAP/TAZ in the nuclei of LM2 cells in low-density versus high-density cultures in vitro (Extended Data Fig. 4c), and in small versus large metastases in vivo (Extended Data Fig. 4d), as expected from known inhibitory effects of cell crowding on YAP/TAZ23. Western blots confirmed YAP inactivation in high-density LM2 cultures, and in low-density cultures treated with high-density culture conditioned medium (Extended Data Fig. 4e). Importantly, CSCs had more nuclear YAP than nonCSCs (Fig. 5c,d), confirming preferential YAP activation in CSCs.
Addressing their functional role, small interfering RNA (siRNA) knockdown of YAP or TAZ decreased CSC population expansion (Fig. 5e), while overexpression of constitutively active YAP-S127A increased CSCs and overrode the density inhibition (Fig. 5f). Because YAP and TAZ can affect cell proliferation independently from cell fate decisions23, the data are plotted against total cell number to deconvolute direct effects on CSCs from effects secondary to changes in culture density. Fate mapping showed that YAP or TAZ knockdown reduced CSCs mainly by decreasing expansive self-renewal and increasing differentiation (Fig. 5g,h), through more differentiating divisions (Extended Data Fig. 4f), and a big increase in direct differentiation (Fig. 5g). YAP/TAZ effects on CSC fates are summarized in Fig. 5i. Stable short hairpin RNA (shRNA) knockdown of YAP in LM2 cells selectively decreased CSCs two- to threefold in vitro (Extended Data Fig. 4g), and significantly decreased metastatic efficiency (Fig. 5j), confirming a key role for YAP in maintaining and expanding the CSC population during metastasis.
Multiple stimulatory and inhibitory microenvironmental inputs converge on YAP/TAZ20,23,24 (Fig. 5k). The Hippo pathway is the canonical negative regulator with LATS kinases as key mediators26, suppressing YAP/TAZ at high cell density24, and in response to cellular energy stresses27. On the stimulatory side, Rho-GTPases are particularly important positive upstream regulators28, and we found Rho-GTPase activity to be preferentially enriched along with YAP activity in the transcriptome of LM2 CSCs in vitro (Fig. 5a) and in vivo (Fig. 5l). Key metabolites such as serum lipids, fatty acids, glutamate and purines can activate Rho-GTPases and the effector kinase ROCK to regulate YAP/TAZ27. ROCK also plays a key role in mechanotransduction, mediating signalling from integrins to activate YAP/TAZ26. Consistent with this regulatory architecture, the LATS inhibitor TRULI blocked YAP inactivation and enabled continued expansion of CSCs at high density (Fig. 5m), while the ROCK inhibitor Y27632 inactivated YAP and reduced CSC expansion, by increasing CSC differentiation and reducing self-renewal (Fig. 5n and Extended Data Fig. 4h,i). Thus, both Hippo and Rho GTPase signalling were confirmed as upstream regulators of CSC dynamics through modulation of YAP/TAZ.
Testing inhibitors of key upstream nodes (Fig. 5k, red boxes) across three models of metastatic TNBC showed that all inputs influenced CSC self-renewal but the dominant input varied between models (Fig. 5o, Extended Data Fig. 5a and Supplementary Table 1). LM2 cells were particularly sensitive to exogenous bioactive lipids, probably because the LPA synthesizing enzymes autotaxin (ENPP2) and phospholipase A1 (PLA1A)22 are not expressed (Extended Data Fig. 5b). All three lines were sensitive to modulation of mechanosensing by FAK inhibition, while 4T1s were particularly sensitive to energy stress induced by AMPK activation. Thus, the YAP/TAZ node in CSCs integrates information from multiple microenvironmental inputs, with relative weight varying between models.
CSC hyper-responsiveness involves amplified input signals and a more TEAD-accessible chromatin organization
The greater responsiveness of CSCs to microenvironmental signals could involve more sensitive signal detection leading to more active YAP/TAZ (input arm), or more efficient conversion of YAP/TAZ activation into phenotypic output (effector arm). Using LPA as a representative stimulus in LM2 cells, we showed that CSCs express more LPA receptor1 (LPAR1) on the cell surface (Fig. 6a,b), and LPA treatment increased nuclear YAP/TAZ to a greater extent in CSCs than nonCSCs (Fig. 6c), indicating that LPA generates an amplified signal in CSCs.
Fig. 6: CSC hyper-responsiveness to LPA involves amplified signalling and more TEAD-accessible chromatin.
a, Flow cytometry for LPAR1 expression in LM2 CSCs (mCherrypos) versus nonCSCs (mCherryneg). b, Quantitation of data in a. Mean ± s.d., n = 3. Two-sided unpaired t-test. Representative of two independent experiments. MFI, median fluorescence intensity. c, Immunofluorescent quantitation of YAP localization in LM2 CSCs versus nonCSCs in unsorted cultures with and without treatment with LPA. n = 524 and 400 for nonCSCs ± LPA; n = 55 and 80 for CSCs ± LPA. Violin plots show medians (solid lines) with interquartile range (dotted lines). Two-sided Kruskal–Wallis test for CSCs versus nonCSCs within treatment group, with Dunn’s multiple comparisons correction. Data pooled from two biological replicates. d, Schematic for the ATAC-seq strategy. e, Top ten enriched transcription factor (TF) binding motifs in DARs in sorted LM2 CSCs versus nonCSCs around TSSs or enhancer regions for two independent ATAC-seq experiments (run 2 had only nine significantly enriched motifs). The two-tailed P-value cut-off for motif enrichment from cumulative binomial distribution test was 0.01. The exact P values and Benjamini–Hochberg q values are given in the Source Data. O-S-T-N, OCT4–SOX2–TCF–NANOG. f, Genes with TEAD binding sites in enhancers that are preferentially accessible in CSCs in both of two independent experiments (see the Source Data for lists of all annotated genes with CSC enhancer DARs with and without TEAD sites and corresponding P values). The number of DARs in each major category is indicated. g, Effect of TEAD inhibitor GNE7883 treatment on expression of KLF5 and ID1 in LM2 cells. CTGF is a canonical YAP target as positive control. Mean ± s.d. for n = 3, normalized to vehicle control. Two-sided Student’s t-test. h, Effect of GNE7883 on CSC dynamics in unsorted LM2 cultures. Mean ± s.d., n = 3, representative of two independent experiments. i, Flow cytometry for effect of 1D1 or KLF5 knockdown on the percentage of CSCs in LM2 cultures. Mean ± s.d. for n = 3, Dunnett’s multiple comparisons test versus control (CON). j, Schematic for mechanisms underlying the hypersensitivity of CSCs to microenvironmental signals. Schematic in j created in BioRender; Wakefield, L. https://biorender.com/uv47yib (2026).
Source data
To address the effector arm, we performed assays for transposase-accessible chromatin using sequencing (ATAC-seq) on sorted LM2 CSCs and nonCSCs (Fig. 6d). There were no consistent global chromatin accessibility differences, but each subpopulation had some local differentially accessible regions (DARs). Focusing on CSC DARs in promoter proximal regions (transcriptional start site (TSS) ± 4 kb) and in previously identified MDAMB231 enhancers29, we searched for enriched transcription factor binding motifs. TEAD motifs were consistently enriched in CSC promoter and enhancer DARs in each of two independent experiments, while other motifs were more variable (Fig. 6e). Because >90% of YAP/TAZ/TEAD binding occurs at distal active enhancer regions (marked by H3K27Ac)25,30, we focused on CSC enhancer DARs with TEAD binding sites. The two genes consistently associated with such enhancers (Fig. 6f) were ID1, a well-characterized inhibitor of differentiation31,32, and KLF5, a key transcription factor in the maintenance of the stem state33. TEAD inhibition with GNE788334 reduced ID1 and KLF5 mRNAs (Fig. 6g) and decreased CSCs in LM2 cultures (Fig. 6h). Knockdown of ID1 or KLF5 with siRNA also reduced CSCs (Fig. 6i), consistent with TEAD-dependent expression of both proteins maintaining stemness.
Important interrelationships exist between YAP/TAZ/TEAD, stemness and the epithelial-to-mesenchymal transition (EMT). YAP/TAZ cooperate with oncogenic KRas to promote EMT35, and inducers of EMT or forced overexpression of core EMT transcription factors (EMT-TFs) increase the frequency of stem-like cells in cancer cell populations36. Furthermore, cells that have undergone a partial EMT have increased tumour and metastasis initiating ability37,38. While the stemness transcription factors SOX2, OCT4 (POU5F1 gene) and NANOG were strongly enriched in our CSC population, classic EMT transcription factors were not (Extended Data Fig. 6a). However, the CSC enhancer DAR of the ID1 gene has a Snail (SNAI1) binding site near the TEAD site (Extended Data Fig. 6b). SNAI1 knockdown reduced ID1 mRNA by ~30%, and suppressed CSC more than nonCSC expansion (Extended Data Fig. 6c–e), suggesting that EMT-TFs may functionally contribute to TEAD-mediated regulation of stemness.
Our data indicate that CSC hyper-responsiveness to microenvironmental change reflects enhanced reception and propagation of microenvironmental signals, as well as enhanced signal interpretation through an altered chromatin architecture that increases TEAD binding site accessibility at regulatory elements of genes that inhibit differentiation and sustain stemness (Fig. 6j).
Targeting CSC sensor nodes reduces chemotherapy-driven enrichment of CSCs in early metastasis
Finally, we addressed therapeutic implications. CSCs are relatively resistant to therapy9,39, and breast tumour tissue remaining after endocrine therapy or chemotherapy is enriched for stem features40. Paclitaxel treatment of the LM2 model significantly reduced metastatic lesion size (Fig. 7a,b) but increased the proportion of CSCs among surviving cells across all lesion sizes (Fig. 7c,d). Fate mapping in vitro showed that paclitaxel greatly reduced all division-dependent events in both subpopulations, although CSCs were less affected than nonCSCs (Fig. 7e). Strikingly, paclitaxel induced nearly fivefold more cell death in nonCSCs than CSCs (Fig. 7f). CSC differentiation and nonCSC dedifferentiation frequencies were relatively unaffected, although CSC differentiation became mostly division independent (Extended Data Fig. 7a,b). Overall, CSCs maintain some self-renewal capacity under paclitaxel treatment and show enhanced survival relative to nonCSCs.
Fig. 7: Targeting CSC sensor nodes reduces chemotherapy-driven enrichment of CSCs in metastasis.
a, Schematic for paclitaxel treatment. b,c, Effect of paclitaxel on LM2 lung metastasis size (b) and percentage of CSCs among tumour cells in the lung (c). Median ± interquartile range, n = 6–7; two-sided Mann–Whitney test. d, Number of CSCs per lesion as a function of LM2 lung lesion size in response to paclitaxel. n = 1–10 lesions per size bin. Data for lesion sizes with n ≥ 3 datapoints are plotted as mean ± s.d. Two-sided analysis of covariance test for difference in slope. e,f, Effect of paclitaxel on division-dependent fates (e) and cell death (f) of LM2 cells in culture from single-cell fate mapping starting after 32 h of paclitaxel treatment. Two-sided Fisher’s exact test. g, CSC sensor inputs converging on ROCK. h, Effect of combined paclitaxel + ROCK inhibitor treatment for 3 days on CSCs and nonCSCs in unsorted LM2 cultures in vitro assessed by flow cytometry. Mean ± s.d. for n = 3, two-sided Dunnett’s multiple comparison test versus CON. i, Schematic for paclitaxel + Y27632 treatment in LM2 model. j,k, Effect of therapeutic interventions on lung metastasis size (j) or percentage of CSCs per metastasis (k) at endpoint. Median ± interquartile range for 96–118 metastases per group pooled from n = 5–6 mice per group; two-sided Dunn’s multiple comparisons test for all pairwise combinations. l, Treatment effect on the percentage of lesions with no CSCs. Two-sided Fisher’s exact test for indicated treatment pairs. n indicates the total lesion number as indicated. NS, not significant. Schematics created in BioRender: a, Wakefield, L. https://biorender.com/2pcq816 (2026); g, Wakefield, L. https://biorender.com/urhbrdb (2026); i, Wakefield, L. https://biorender.com/z10b5so (2026).
Source data
We next asked whether blocking CSC self-renewal could reverse chemotherapy-induced CSC enrichment. In principle, the YAP/TAZ/TEAD node would be the optimal target, and promising TEAD inhibitors are being developed34,41. The TEAD inhibitor GNE788334 combined with paclitaxel significantly decreased metastasis number in the LM2 model (Extended Data Fig. 7c,d), although the recommended excipient precluded reliable CSC imaging. Instead, for proof of concept, we used Y27632 to inhibit ROCK, an upstream node for both bioactive lipid and mechanosignalling to YAP/TAZ (Fig. 7g). Flow cytometry confirmed that treatment of LM2 cells with paclitaxel alone selectively reduced nonCSCs, while Y27632 alone selectively reduced CSCs, and the combination reduced both in vitro (Fig. 7h). In vivo, ROCK inhibition did not compromise the ability of paclitaxel to reduce metastatic lesion size (Fig. 7i,j and Extended Data Fig. 7e) but did reverse the paclitaxel-induced enrichment of CSCs across all lesion sizes (Fig. 7k and Extended Data Fig. 7f,g) and increased the proportion of metastases that had no CSCs (Fig. 7l), suggesting that some metastases were being ‘sterilized’ of their CSCs. Similar results were seen with the LPAR inhibitor BMS5986020 in the LM2 model (Extended Data Fig. 7h–l). Y27632 with paclitaxel also significantly reduced metastatic burden in immunocompetent mice, using the 4T1 model without imaging markers to avoid immune rejection (Extended Data Fig. 7m,n). Thus targeting regulatory nodes in the CSC sensor mechanism combines with conventional chemotherapy in vivo to improve antimetastatic responses.
