Identifying and reprogramming softness-driven cancer stem-like cells overcomes CAR-T cell resistance in solid tumours

Abstract

Solid tumours show substantial mechanical heterogeneity, yet how such cues influence the susceptibility of cancer cells to T cell-based therapies remains unclear. Here we discover that cancer cells grown on soft matrices are less sensitive to chimeric antigen receptor T cell cytotoxicity and exhibit elevated extracellular adenosine triphosphate and sustained calcium activity. To understand the mechanisms underlying this reduced killing, we sought to identify the cells that respond to mechanical softness. We engineered a doxycycline-gated calcium-activated transcriptional mechano-recorder that integrates softness-induced calcium activity over a defined recording window and converts this prior signalling history into a stable fluorescent output that persists for days. Unlike real-time calcium indicators, which report instantaneous calcium activity only at the moment of imaging, this recorder preserves a sortable transcriptional mark, enabling selective labelling and profiling of cells according to their past mechanosensing activity. Transcriptomic analyses showed that recorder-positive cells adopt a stem-like programme, including epithelial–mesenchymal transition, hypoxia responses, oncogenic signalling and elevated stemness markers, across cancer cell lines and patient-derived samples. To render these resistant cells targetable, we rewired the mechano-recorder into a mechano-reprogrammer by replacing the fluorescent output with the clinically validated antigen CD19, enabling softness-responsive cells to be recognized by CD19-directed T cells. This rewired system improved elimination of stem-like cancer cells in culture and animal models, converting mechanobiological resistance into therapeutic vulnerability.

The tumour mechano-microenvironment is a complex and dynamic system where individual cancer cells adopt unique mechanosensing modalities and develop heterogeneous phenotypes^1,2. Cancer cell softening is a notable biomechanical feature³ that is closely linked to cancer malignancy and metastasis⁴. With the revolutionary advent of immunotherapies, there has been increased interest in understanding how these mechanical microenvironments impact the cancer cell phenotypes and consequently their interactions with immune cells^5-8.

Among immune cells, cytotoxic T lymphocytes play a pivotal role in immunosurveillance by forming immune synapses through the binding of T cell receptors (TCRs) to the major histocompatibility complexes (MHCs) on cancer cells and secreting granzyme proteases and pore-forming perforin to induce apoptosis in cancer cells^9,10. The mechanical strength at these synapses is essential for T lymphocyte activation and effective cancer cell killing^11,12. Cancer cells, including those from breast and bladder cancers, glioma and melanoma, can soften their cortical structure via cytoskeleton reorganization¹³ or cholesterol enrichment in the plasma membrane⁵. Such cell softening reduces the traction force exerted by the T cells^5,6,11 and impairs perforin pore formation^6,8,11, potentially leading to immunosurveillance evasion and cancer relapse. While bulk stiffening of tumours through collagen crosslinking and stromal remodelling is recognized¹⁴, tumours show striking intratumoural heterogeneity, where softer niches coexist alongside stiffer regions^4,15. At the single-cell level, such softness is closely associated with stem-like states, immune evasion and therapeutic resistance^3,6,16, implicating these subpopulations as key drivers of relapse and metastasis.

Cancer relapse is commonly observed in patients undergoing T cell immunotherapies, for example, chimeric antigen receptor T (CAR-T) cell therapy¹⁷, and has been reported in clinical trials treating glioblastoma with EGFRvIII or IL13Rα2 CAR-T cell therapies^18,19. In addition, the lack of definitive and specific surface markers across different solid tumour subtypes and patients, along with the dynamically evolving nature of cancer cells under different mechanical tumour microenvironments, poses a substantial challenge^20,21, resulting in suboptimal outcomes for monoclonal antibody and CAR-T cell therapies²². Consequently, there is currently no Food and Drug Administration-approved CAR-T cell product against solid tumours despite the remarkable success of CAR-T cell therapy for blood cancers²³. Thus, understanding the underlying regulatory mechanisms of the mechanical microenvironment impact on cancer cells and developing corresponding therapies would greatly enhance the therapeutic effectiveness of immunotherapy.

In this study, we observed that a soft extracellular matrix (ECM) two-dimensional environment correlates with a weakened cytotoxicity of CAR-T cell-based immunotherapy, accompanied by a markedly higher calcium signalling and extracellular ATP (exATP) levels. We designed and applied a genetic mechano-recorder to convert mechanically associated calcium activity into a durable transcriptional reporter. This system differs from conventional calcium indicators such as GCaMP, which report calcium activity at the time of imaging and lose signal as calcium returns to baseline. Instead, the mechano-recorder integrates calcium activity during a doxycycline-defined recording window and stores this prior signalling history as a stable fluorescent mark, enabling the barcoding of individual cancer cells for fluorescence-activated cell sorting (FACS) and RNA sequencing according to their past mechanosensing activity. We found that soft ECM promotes the transition of cancer cells towards a cancer stem-like cell (CSLC) state, termed ‘mechano-stemness’, in both cancer cell lines and patient-derived breast cancer cells, which correlates with high recorder expression. By changing the mechano-recorder to mechano-reprogrammer, we further rewired this mechano-stemness into the production of CD19, a clinically validated antigen²⁴, thus allowing CD19-CAR-T cells to target CSLCs more effectively in a heterogeneous population and leading to improved breast tumour control both in vitro and in vivo.

Reduced CAR-T cell efficacy against cancer cells in soft mechanical environments

We first examined the effect of mechanical environment on CAR-T cell killing of cancer cells. As stiffer ECM substrates have been shown to generate stiffer cells²⁵, we used a polyacrylamide hydrogel platform with varying Young’s modulus²⁶ to represent extreme cases of soft and stiff breast tumour mechanical environments¹⁵ (Supplementary Fig. 1). Triple-negative MDA-MB-231 breast cancer cells expressing endogenous antigen epidermal growth factor receptor (EGFR) or exogenous antigen CD19 (MDA-MB-231-CD19), as well as glioblastoma U87-MG cells expressing disialoganglioside GD2 antigen, could be efficiently eliminated by their corresponding CAR-T cells, demonstrating targeted antigen-specific killing (Supplementary Fig. 2a-c). Interestingly, regardless of the antigen or cancer types, tumour cells cultured on softer substrates exhibited greater resistance against CAR-T cell killing, although these tumours tended to grow more slowly on soft substrates (Fig. 1a,b and Supplementary Figs. 3 and 21). Surface marker staining showed that cancer antigen expression levels did not differ on varying ECM stiffnesses (Fig. 1c and Supplementary Fig. 4a-f), indicating that the increased immunotherapy resistance of cancer cells grown on softer ECM is not due to differences in antigen expression.

Fig. 1 ∣. Two-dimensional mechanical softness impairs CAR-T cell killing of cancer cells with distinctive calcium signals.

a, Schematic illustration of CAR-T cell killing preference of cancer cells cultured on stiff over soft ECM. b, Cytotoxicity (cell death percentage) of EGFR-CAR, CD19-CAR and GD2-CAR-T cells targeting MDA-MB-231, MDA-MB-231-CD19 and U87-MG cells, cultured on coverslip, stiff or soft ECM, at an E:T ratio of 1:1 for 24 h, respectively. One-way analysis of variance (ANOVA) with Tukey’s multiple comparisons test (two-sided). Error bars, mean ± s.d. EGFR-CAT T/MDA-MB-231, coverslip versus stiff or coverslip versus soft, ****P < 0.0001, stiff versus soft, **P = 0.0031; CD19-CAR-T/MDA-MB-231-CD19, coverslip versus stiff, *P = 0.048, coverslip versus soft, ***P = 0.0006, stiff versus soft, **P = 0.0088; GD2-CAR-T/U87-MG, coverslip versus stiff or coverslip versus soft, ****P < 0.0001, stiff versus soft, **P = 0.0019, respectively. n = 3 in each group, where n is independent replicates of cytotoxicity assay. c, EGFR, CD19 and GD2 surface expression in MDA-MB-231, MDA-MB-231-CD19 and U87-MG cells, cultured on coverslip, stiff or soft ECM for 24 h, respectively. Antigen levels were assessed with phycoerythrin-labelled anti-CD19, AF488-labelled anti-EGFR and APC-labelled anti-GD2 antibody, respectively. RFI, relative fluorescence intensity. AU, arbitrary unit. One-way ANOVA with Tukey’s multiple comparisons test (two-sided). Error bars, mean ± s.d. MDA-MB-231-EGFR, coverslip versus stiff, NS, not significant P = 0.8780, coverslip versus soft, P = 0.2808, stiff versus soft, P = 0.7256; MDA-MB-231-CD19, coverslip versus stiff, P = 0.7445, coverslip versus soft, P = 0.9727, stiff versus soft, P = 0.4738; U87-MG, coverslip versus stiff, P = 0.9982, coverslip versus soft, P = 0.4739, stiff versus soft, P = 0.3742, respectively. MDA-MB-231-EGFR or MDA-MB-, nonstain/coverslip/stiff/soft, n = 100; MDA-MB-231-CD19, nonstain/coverslip/stiff/soft, n = 100; U87-MG-GD2, nonstain, n = 84, coverslip, n = 87, stiff, n = 88, soft, n = 87. n depicted individual cells detected. d, Representative images of calcium levels, indicated by GCaMP, and constitutive mCherry levels as a normalizing reference in MDA-MB-231 cells transduced with PGK-jGCaMP8s-P2A-mCherry and cultured on coverslip, stiff or soft ECM. Scale bars, 100 μm. Images are representative of 3 independent experiments. e, Accumulative calcium levels in MDA-MB-231 cells cultured on coverslip, stiff or soft ECM. Accumulative calcium levels were assessed by taking the ratio of GCaMP signals over mCherry signals per cell every 10 s for 180 s in total to calculate the area under the curve. One-way ANOVA with Tukey’s multiple comparisons test (two-sided). Error bars, mean ± s.d. Coverslip versus stiff, ***P = 0.0006, coverslip versus soft, ****P < 0.0001, stiff versus soft, *P = 0.0260, respectively. Coverslip, n = 38; stiff, n = 38; soft, n = 33. n depicts cell number. f, exATP levels in MDA-MB-231 cells cultured on coverslip, stiff or soft ECM for 24 h or 48 h. One-way ANOVA with Tukey’s multiple comparisons test (two-sided). Error bars, mean ± s.d. 24 h, coverslip versus stiff, NS, not significant P = 0.1534, coverslip versus soft or stiff versus soft, ****P < 0.0001; 48 h, coverslip versus stiff, NS, not significant P = 0.5294, coverslip versus soft or stiff versus soft, ****P < 0.0001, respectively. n = 6 in each group, where n is independent replicates of exATP assay. g, The ECM stiffness effect on exATP and calcium levels in cancer cells, and the CAR-T cell cytotoxicity against these cancer cells.

Elevated intracellular calcium and exATP levels in cancer cells on soft ECM

To better understand the mechanical effect on CAR-T cell killing, we assessed the calcium levels in MDA-MB-231 cells cultured on varying ECM stiffnesses, since calcium signalling plays a crucial role in cancer cell sensing of mechanical forces²⁷. While we anticipated that the dynamics of calcium signalling would increase with a higher ECM stiffness by potentially triggering more openings of ion channels, such as Piezo channels²⁸, our findings surprisingly showed higher intracellular calcium levels (GCaMP detection) in MDA-MB-231 cells grown on softer ECM, where cumulative calcium activity was consistently high in cells on soft ECM despite similar spike frequencies (Fig. 1d,e and Supplementary Fig. 5a,b). This unexpected finding led us to consider factors other than the pure mechanical tension effect on calcium channels in affecting intracellular calcium levels. The cell-released ATP (exATP) is known to induce rapid calcium spikes through the activation of P2Y receptors²⁹, and ECM stiffness has been shown to affect the ATP-induced calcium influx^30,31. Enhanced ATP release upon mechanical deformation has also been shown to promote metastasis in breast cancer cells³². We hence checked exATP levels in MDA-MB-231 cells cultured on varying ECM stiffnesses and found that exATP levels were indeed markedly higher in cells on softer ECM (Fig. 1f). This increase of exATP on soft ECM cannot be attributed to the variations in cell growth and densities, as the exATP levels are proportionally correlated with cell numbers (Supplementary Fig. 7) and ECM softness impaired cancer cell proliferation (Supplementary Fig. 3). Consistently, pharmacological softening with low-dose blebbistatin (2 μM) mimicked soft ECM, elevating exATP and calcium (Supplementary Fig. 7a-d). Blocking calcium with BAPTA-AM or 2-APB reduced ATP release and restored CAR-T cytotoxicity, confirming calcium as a key mediator of softness-triggered resistance (Supplementary Fig. 8a,b). These results led us to conclude that cancer cells on softer ECM increased their exATP and its associated calcium signalling, which may have subsequently led to resistance to CAR-T killing^33,34 (Fig. 1g).

Development and validation of a genetic mechano-recorder

We further investigated whether ECM softness-associated calcium signalling can be utilized to identify and label cancer cells resistant to CAR-T cell killing under different mechanical environments. We designed a mechano-recorder to integrate intracellular calcium activity during a doxycycline-defined recording window and convert this prior signalling history into fluorescent reporter expression, thereby barcoding individual cells. Unlike real-time calcium indicators reporting instantaneous calcium activity, the mechano-recorder stores prior calcium activity as a stable transcriptional output that can be read out after the initial calcium signal has subsided. The recorder utilizes an AND-gate design integrating Tet-On 3G and nuclear factor of activated T cells (NFAT), with a reverse tetracycline-controlled transactivator (rtTA) that connects to NFAT to serve as the regulator and the doxycycline (DOX)-gated inducible TRE promoter driving mCherry as the reporter (Fig. 2a). Elevated intracellular calcium levels can activate calcineurin to dephosphorylate NFAT in the regulator, which can then translocate with rtTA into the nucleus to initiate transcription of a target reporter gene (for example, mCherry) under TRE promoter in the presence of DOX³⁵. This system allows us to barcode cells with different mechanical experiences at any defined timeframe controllable by DOX into a stable fluorescence marker for sorting and transcriptome profiling (Fig. 2a and Supplementary Fig. 9).

Fig. 2 ∣. Engineer and characterize genetic mechano-recorder of calcium activity.

a, Schematic illustration of mechano-recorder design and mechanism to barcode cells based on distinct intracellular calcium signals owing to heterogeneous mechanosensing. Mechano-recorder detects distinct calcium signals from cells cultured on different ECM stiffnesses and barcodes the cells to be sorted and sequenced for transcriptome profiling. b,c, Representative images (b) and quantification (c) of recorder expression in MDA-MB-231 and MCF7 cells cultured on coverslip, stiff or soft ECM, with or without doxycycline treatment. Two-way ANOVA with Šidák’s multiple comparisons test (two-sided). Error bars, mean ± s.d. MDA-MB-231, coverslip/−DOX versus +DOX, *P = 0.0284, coverslip/+DOX versus stiff/+DOX, ***P = 0.0006, stiff or soft/−DOX versus +DOX, stiff/+DOX versus soft/+DOX, ****P < 0.0001; MCF-7, coverslip/−DOX versus +DOX, NS, not significant P = 0.0660, coverslip/+DOX versus stiff/+DOX, ***P = 0.0004, stiff/+DOX versus soft/+DOX, **P = 0.0077, stiff or soft/−DOX versus +DOX, ****P < 0.0001, respectively. MDA-MB-231, soft/−DOX, n = 69, soft/+DOX, n = 121, stiff/−DOX, n = 74, stiff/+DOX, n = 97, coverslip/−DOX, n = 117, coverslip/+DOX, n = 152; MCF-7, soft/−DOX, n = 98, soft/+DOX, n = 163, stiff/−DOX, n = 131, stiff/+DOX, n = 242, coverslip/−DOX, n = 246, coverslip/+DOX, n = 262. n depicts individual cells detected from ≥3 individual replicates. d,e, Representative images (d) and quantification (e) of recorder expression in MDA-MB-231 cells seeded with varied density on coverslip, with or without doxycycline treatment. Two-way ANOVA with Šidák’s multiple comparisons test (two-sided). Error bars, mean ± s.d. Low/−DOX versus +DOX, NS, not significant P = 0.5259, low/+DOX versus medium/+DOX, NS, not significant P = 0.2574, medium or high/−DOX versus +DOX, ****P < 0.0001, high/−DOX versus +DOX, **P = 0.0019, respectively. Low/−DOX, n = 138, low/+DOX, n = 66, medium/−DOX, n = 99, medium/+DOX, n = 136, high/−DOX, n = 406, high/+DOX, n = 218. n depicts individual cells detected from ≥3 individual replicates. f,g, Representative images (f) and quantification (g) of recorder expression in MDA-MB-231 cells seeded on coverslip, stiff or soft ECM, in the presence of doxycycline, and with or without apyrase treatment. Apyrase was added with doxycycline for 2 h. Two-way ANOVA with Šidák’s multiple comparisons test (two-sided). Error bars, mean ± s.d. Coverslip/−APY versus +APY, *P = 0.0128, stiff or soft/−APY versus +APY, ****P < 0.0001, coverslip/+APY versus stiff/+APY or stiff/+APY versus soft/+APY, NS, not significant P > 0.9999, respectively. Soft/−APY, n = 401, soft/+APY, n = 101, stiff/−APY, n = 410, stiff/+APY, n = 191, coverslip/−APY, n = 319, coverslip/+APY, n = 219. n depicts individual cells detected from ≥3 individual replicates. Cells were cultured overnight under indicated conditions and treated with doxycycline for 2 h and imaged 6 h after drug addiction. Images shown were mCherry expression and quantification were normalized against miRFP670 to normalize for protein expressions in individual cells. Scale bars, 100 μm. Colour scale indicates the pseudo-colour intensities of mCherry channel. Illustrations in a created in BioRender; Qu, Y. https://biorender.com/qm5u7j5 (2026).

We co-transfected MDA-MB-231 cells and another luminal subtype breast cancer cell line, MCF7, with the recorder constructs for regulator and reporter expressions (Supplementary Fig. 9). We systematically examined the recording and reading periods and found that a 2 h recording (DOX incubation time) and a 6 h reading (reporter expression quantification) periods provided robust and stable reporter expressions (Supplementary Fig. 10a-d). As the reporter expression is doxycycline dosage dependent³⁵, 500 nM doxycycline was used to provide consistent outcomes. These conditions were used for all subsequent studies. The recorder’s sensitivity to calcium signalling was further calibrated and confirmed by using ionomycin-induced calcium signals, which increased the recorder expression in both cell lines (Supplementary Fig. 11a,b). When depleting calcium from the extracellular space with calcium-free medium and the chelator BAPTA (Ca²⁺ free/BAPTA), or blocking inositol 1,4,5-triphosphate (IP3) receptor (IP3R) mediating ER calcium release with 2-aminoethoxydiphenyl borate (2-APB), the recorder expression dropped substantially, verifying that the recorder is sensitive to intracellular calcium levels modulated by extracellular-to-cytosol influx or ER-release (Supplementary Fig. 11c,d). ER-released calcium was also confirmed to increase recorder expression by using thapsigargin to increase intracellular calcium, following the depletion of calcium stored in the ER by inhibiting the sarcoendoplasmic reticulum calcium ATPase (SERCA) (thapsigargin) (Supplementary Fig. 11c,d). Importantly, mechano-recorder output positively correlated with cumulative GCaMP-reported calcium activity measured over 6 h in cells cultured on soft ECM (Supplementary Fig. 12), supporting that the recorder captures integrated intracellular calcium activity. These two readouts are nevertheless temporally distinct: GCaMP provides a live readout of calcium dynamics during imaging, whereas the mechano-recorder converts calcium activity during the doxycycline-gated recording window into a stable transcriptional output that can be read after the initial calcium signal has resolved.

Tracking cellular mechanical events with mechano-recorder

To examine whether the recorder can robustly measure the mechanical experience and the ensuing exATP-calcium signalling of individual cells via the recorder expression, we seeded MDA-MB-231 or MCF7 cells engineered with the recorder gene cassettes on ECM with varying stiffnesses. Consistent with our calcium findings (Fig. 1e), the recorder expression was notably higher in cells cultured on soft ECM, in both cell lines (Fig. 2b,c). Notably, despite similar copy numbers of recorder gene cassette introduced into different groups of cells (Supplementary Fig. 9), MCF7 cells exhibited much lower overall expression levels of the recorder than MDA-MB-231 cells across all ECM stiffness conditions, possibly reflecting a more aggressive potential of MDA-MB-231 cells (Fig. 2b,c). To verify the consistence of this trend in three-dimensional (3D) environments by using a 3D collagen system with tunable stiffness³⁶, we found more cells expressing mCherry when MDA-MB-231 cells were embedded in softer (1 mg ml⁻¹) than stiffer (6 mg ml⁻¹) collagen gels (Extended Data Fig. 1a-d). This difference is due to the mechanical environments perceived by the cells rather than the difference in 3D DOX perfusion, as all cells expressed the same level of mCherry in response to the DOX induction utilizing a constitutively expressed rtTA, independent of the mechanosensitive calcium signalling and NFAT nuclear translocation (Extended Data Fig. 1e-h). We further examined and verified the role of mechanosensitive exATP in regulating the recorder function. Cells seeded at higher densities, which produce more exATP, exhibited higher recorder expression (Fig. 2d,e). Addition of ATP to the culture medium also increased recorder levels (Extended Data Fig. 2), while the treatment with apyrase, which hydrolyses exATP, markedly reduced recorder signals regardless of the stiffnesses of ECM on which the cells were seeded, thus confirming the role of exATP in recorder expression (Fig. 2f,g). In addition, low-level blebbistatin treatment enhanced recorder activation under doxycycline induction, confirming that the system reliably captures calcium signalling changes induced by mechanical softening (Supplementary Fig. 13a,b). These findings demonstrate that the mechano-recorder can effectively document the mechanical experience of individual cells under varying microenvironments, mediated by exATP and its associated calcium signalling.

Mechano-recorder revealed that mechanical softness enhanced cancer cell stemness

The recorder allowed us to barcode and sort individual cells under different mechanical environments, showing that softness induces more recorder-positive cells (Fig. 2a-c). The recorder-positive and -negative cells from different groups were selected for RNA sequencing, which revealed distinct transcriptional profiles between these two groups of cells (Fig. 3a). Gene Ontology enrichment analysis showed that the pathways related to cell adhesion, migration and response to hypoxia were enriched in recorder-positive cells on either soft or stiff ECM (Fig. 3b and Extended Data Fig. 3a). Gene set enrichment analysis (GSEA) further identified upregulated key pathways in recorder-positive cells such as EMT, hypoxia and KRAS upregulation, which are essential regulators in CSLCs or CSCs^20,21 (Fig. 3c and Extended Data Fig. 3b). The recorder-positive cells, particularly those from soft ECM, also expressed marker genes reflecting cancer stemness, for example, high in CD44, ALDH1, but low in CD24^3,37,38 (Fig. 3d). The sorted recorder-positive cells from either stiff or soft ECM in low-attachment conditions formed more mammospheres than the recorder-negative cells, demonstrating their higher stemness and tumorigenic potential (Fig. 3e,f and Extended Data Fig. 3c,d). Consistent with previous reports that tumour-repopulating cells are softer^3,6, we observed that recorder-positive cells exhibited lower cortical stiffness with atomic force microscopy (AFM) measurements (Fig. 3g and Extended Data Fig. 3e). Coupling recorder fluorescence with AFM at the single-cell level also revealed a strong inverse correlation with stiffness (Supplementary Fig. 14). In addition, MDA-MB-231 cells, known for their greater stemness compared with MCF7 cells³⁸, showed higher expression of the reporter (Fig. 2b,c and Supplementary Fig. 11a,b), further supporting the notion that the recorder expression can reflect the level of stemness in individual cancer cells. Our experiments further showed that these recorder-positive cells were more resistant to EGFR-CAR-T cell-mediated killing, confirming their CSLC phenotype (Fig. 3h and Extended Data Fig. 3f). Consistent to these findings, a previous study reported an increase in cancer stem cell markers in cells evading CAR-T killing³⁹. As such, our results indicate that soft environment promotes recorder-positive cells, which show CSLC characteristics, that is, ‘mechano-stemness’ (Fig. 3i), resulting in a higher resistance against CAR-T cell therapy.

Fig. 3 ∣. Mechano-recorder identified CSLCs coupled with mechanical softness.

a, Principal component analysis plot showing the clustering of RNA sequencing data from different groups: stiff recorder-positive (recorder+), stiff recorder-negative (recorder−), soft recorder-positive and soft recorder-negative cells. b, Gene Ontology enriched pathways from bulk RNA-seq analysis in soft recorder-positive cells with upregulated differentially expressed genes comparing to soft recorder-negative cells, P-adjusted <0.05 and log₂ fold change >0.5. Benjamini–Hochberg FDR test (two-sided). c, GSEA analysis using normalized gene counts in soft recorder-positive (soft positive) cells versus soft recorder-negative (soft negative) cells to show enriched pathways of EMT, hypoxia and KRAS signalling upregulation in soft positive cells. NES, normalized enrichment score. d, Heat map showing CSLC stemness marker gene expression in different groups: stiff recorder-negative, stiff recorder-positive, soft recorder-negative and soft recorder-positive cells. e,f, Representative images (e) and quantification (f) of mammosphere formation from soft recorder-negative and soft recorder-positive cells. Two-tailed t-test. Error bars, mean ± s.d. ****P < 0.0001. n = 42 in each group represents individual wells seeded. Scale bars, 100 μm. g, Cellular stiffness of soft recorder-negative and soft recorder-positive cells was measured by AFM. Two-tailed t-test. Error bars, mean ± s.d. ***P = 0.0008. Recorder−, maxima = 1.890, centre = 1.060, minima = 0.3247, n = 31; recorder−, maxima = 1.580, centre = 0.6517, minima = 0.07895, n = 38. n depicts individual measurements from ≥3 individual replicates. h, Cytotoxicity of soft recorder-negative and soft recorder-positive cells co-cultured with EGFR-CAR-T cells at an E:T ratio of 1:1 for 24 h. Two-tailed t-test. Error bars, mean ± s.d. **P = 0.0045. n = 3 in each group, where n is independent replicates of cytotoxicity assay. i, Schematic illustration of recorder-selected CSLCs, with high calcium and positive recorder expression promoted by ECM softness.

Enhanced cancer stemness associated with mechanical softness in patient-derived cells

Triple-negative breast cancer (TNBC) is notoriously known for its high recurrence rate⁴⁰. Metaplastic squamous cell carcinoma, a very rare case of TNBC, carries a worse prognosis and twice the risk of recurrence than non-metaplastic TNBC, with shorter disease-free period and lower overall survival^41,42. We accordingly examined three different types of breast cancer patient-derived cell with diverse subtypes related to TNBC: (1) 110-T, ductal adenocarcinoma; (2) 015-R, triple-negative ductal carcinoma; and (3) 296-R, triple-negative mixed ductal carcinoma and metaplastic squamous cell carcinoma. Transcriptome profiling clearly distinguished these three kinds of cell (Fig. 4a). We then introduced similar amounts of mechano-recorder constructs into these three cell types (Supplementary Fig. 15). Using ionomycin, we first confirmed the recorder functionality in reporting calcium signals in these patient-derived cells (Supplementary Fig. 16a-f).

Fig. 4 ∣. Mechano-recorder detects stemness of patient-derived cells coupled with mechanical softness.

a, Principal component analysis plot showing the clustering of RNA sequencing data from 110-T, 015-R and 296-R cells. b,c, Quantification (b) and representative images (c) of recorder expression in 110-T, 015-R and 296-R cells cultured on coverslip, stiff or soft ECM, with and without doxycycline treatment. Two-way ANOVA with Šidák’s multiple comparisons test (two-sided). Error bars, mean ± s.d. 110-T, coverslip or stiff or soft/−DOX versus +DOX, ****P < 0.0001, coverslip/+DOX versus stiff/+DOX, NS, not significant P = 0.2746, stiff/+DOX versus soft/+DOX, **P = 0.0014; 015-R, all comparison, ****P < 0.0001; 296-R, coverslip/−DOX versus +DOX, ***P = 0.0004, all other comparisons, ****P < 0.0001. 110-T, soft/−DOX, n = 227, soft/+DOX, n = 327, stiff/−DOX, n = 353, stiff/+DOX, n = 395, coverslip/−DOX, n = 301, coverslip/+DOX, n = 224; 015-R, soft/−DOX, n = 356, soft/+DOX, n = 753, stiff/−DOX, n = 577, stiff/+DOX, n = 504, coverslip/−DOX, n = 476, coverslip/+DOX, n = 413. 296-R, soft/−DOX, n = 107, soft/+DOX, n = 164, stiff/−DOX, n = 198, stiff/+DOX, n = 323, coverslip/−DOX, n = 95, coverslip/+DOX, n = 121. n depicts individual cells detected. Scale bars, 100 μm. Colour scale indicates the pseudo-colour intensities of mCherry channel. d, GSEA analysis using normalized gene counts in 015-R cells versus 296-R cells showing enriched pathways of EMT, TNFα signalling via NF-κB and hypoxia in 296-R. e, Heat map showing breast cancer stem-like cell stemness marker gene expression in 110-T, 015-R and 296-R cells. Each column in the heat map corresponds to one biological replicate, with one data point depicting in vitro culture whereas other derived from mice xenograft.

Consistent with previous observations from breast cancer cell lines, all these patient-derived cells exhibited higher recorder expression on soft ECM, with the most aggressive and resistant 296-R showing the strongest signals (Fig. 4b,c). When different cell types were grouped together, recorder signals were clearly ranked from highest to lowest in the order of 296-R, 015-R and 110-T cells (Supplementary Fig. 17a,b). These patterns further support our hypothesis that the recorder can faithfully label the resistant CSLCs in response to soft mechanical environment. GSEA analysis of 296-R versus 015-R further revealed enrichment of similar sets of CSLC-associated pathways^17,18 and markers^3,37,38, including EMT, NF-κB–TNFα axis and hypoxia (Fig. 4d). These findings were supported by the lower expression of CD24, but higher expression of all other cancer stemness markers, in 296-R cells (Fig. 4e). Altogether, these observations confirmed the recorder’s capacity in detecting and reporting mechano-stemness of various types of patient-derived breast cancer type with different recurrence potentials. In addition, similar recorder responsiveness to ECM softness was observed in other aggressive tumour cell lines, including U87 glioblastoma, Panc-1 pancreatic and PC-3 prostate cancer cells (Extended Data Fig. 4a-f and Supplementary Figs. 18 and 19), underscoring the generalizability of this approach across multiple solid tumour types.

Rewiring mechano-stemness into synthetic ‘kill me’ signal for CAR-T cells

After verifying the recorder’s capacity to detect mechano-stemness and identify CSLCs, we examined whether we could rewire the recorder design to convert the mechano-stemness and the more resistant CSCs into a CAR-T cell ‘kill me’ signal. To this end, we rewired the recorder into a mechano-reprogrammer by replacing the fluorescent reporter with a truncated CD19 antigen for CAR-T cell recognition and killing of CSLCs, while retaining the same doxycycline-gated calcium-responsive input module (Fig. 5a). In this configuration, softness-associated calcium activity induces expression of a clinically validated antigen target rather than a fluorescent reporter, converting mechano-stemness into a synthetic ‘kill me’ signal for CAR-T cell recognition and killing. This truncated CD19 version is without intracellular domain⁴³, and hence its expression per se did not alter tumour mechanics (Supplementary Fig. 20). We cultured MDA-MB-231 cells on ECM with varying degrees of stiffness and observed higher CD19 expression in cells seeded on softer ECM (Fig. 5b), which was particularly prominent with the DOX incubation time at 2 h and the reading time at 6 h (Fig. 5b-d). This pattern was further validated using the 3D collagen model, where cells embedded in softer collagen gel exhibited higher CD19 expression (Extended Data Fig. 5a-c). This rewiring led to enhanced killing of MDA-MB-231 cells by CD19-CAR-T cells at different E:T ratios, overcoming the resistance of cancer cells on soft hydrogels and reversing the preferred CAR-T killing of cancer cells on stiffer ECM (Fig. 5e, Extended Data Fig. 6a,b and Supplementary Fig. 21). Consistently, calcium blockade abolished the cytotoxicity advantage conferred by rewiring, confirming that the effect depends on softness-triggered calcium signalling (Supplementary Fig. 22). RNA-seq further showed that rewired tumours enhanced CAR-T activation and effector programmes, including IL2–STAT5, NF-κB and IFNγ signalling (Extended Data Fig. 7a-c). These results suggest that our approach can rewire mechano-stemness into synthetic antigens for the CAR-T recognition and killing of CSLCs.

Fig. 5 ∣. Mechano-reprogrammer converts mechano-stemness into synthetic ‘kill me’ signal in vitro.

a, Schematic illustration of rewiring mechanosensing into CD19 signals in CSLCs for CD19-CAR-T cell recognition and killing. b, Histogram showing rewired CD19 expression in MDA-MB-231 cells cultured on coverslip, stiff or soft ECM, with indicated recording and reading time periods. c, RFI of CD19 in MDA-MB-231 cells cultured on coverslip, stiff or soft ECM with 2 h recording and 6 h reading time stained with or without phycoerythrin-CD19 antibodies. One-way ANOVA with Tukey’s multiple comparisons test (two-sided). Error bars, mean ± s.d. ****P < 0.0001. n = 100 in each group. d, Percentage of CD19 cells in MDA-MB-231 cells cultured on coverslip, stiff or soft ECM, with no recording or 2 h recording and 6 h reading. One-way ANOVA with Tukey’s multiple comparisons test (two-sided). Error bars, mean ± s.d. Recording time 0 h/reading time 0 h, coverslip versus stiff, NS, not significant P = 0.9895, stiff versus soft, NS, not significant P = 0.9868, coverslip versus soft, NS, not significant P = 0.9536; recording time 2 h/reading time 6 h, all comparisons, ****P < 0.0001. n = 3 in each group, where n is independent replicates of induction assay. e, Normalized cytotoxicity of MDA-MB-231 cells with and without rewiring was co-cultured with CD19-CAR-T cells at an E:T ratio of 1:1 for 24 h. One-way ANOVA with Tukey’s multiple comparisons test (two-sided). Error bars, mean ± s.d. No rewiring, coverslip versus stiff, NS, not significant P = 0.3145, stiff versus soft, NS, not significant P = 0.9867, coverslip versus soft, NS, not significant P = 0.2603; rewiring, coverslip versus stiff, NS, not significant P = 0.1883, coverslip versus soft, **P = 0.0014, stiff versus soft, **P = 0.0087. n = 3 in each group, where n is independent replicates of cytotoxicity assay.

Rewiring improves CAR-T cell killing of CSLCs in vivo

Building on the promising in vitro results, we extended our investigation to evaluate the efficacy of this rewiring approach in vivo. We selected the two most distinctive MDA-MB-231 cell populations with or without recorder expression, that is, stiff recorder-negative versus soft recorder-positive cells. We first checked recorder inducibility by DOX in NSG mice (Fig. 6a). Normalized by the constitutive renilla luciferase (Rluc) signals as a reference reflecting tumour size, recorder-positive cells clearly showed DOX-induced higher firefly luciferase (Fluc) signals than those of the recorder-negative cells (Fig. 6b,c). The consistently high expression of Fluc signal in recorder-positive cells comparing to the low expression in recorder-negative cells after induction in vivo may suggest a possible cancer cell mechanomemory⁴⁴. We observed that recorder expression dropped to basal level after 3 days of DOX injection, demonstrating the controllable and reversable expression of synthetic antigens. In addition, we used in vivo ultrasound elastography, which showed that recorder-positive tumours remained notably softer and exhibited elevated Nanog and Sox2 expression, confirming the persistence of a stem-like, soft phenotype in vivo (Supplementary Figs. 23 and 24).

Fig. 6 ∣. Rewiring improves CAR-T cell killing of CSLCs in vivo.

a, Timeline of in vivo recorder inducibility experiment using NSG mice. Recorder-negative MDA-MB-231 cells from stiff ECM and recorder-positive MDA-MB-231 cells from soft ECM were collected from FACS. Tumour cells were injected into the m.f.p. 5 days before experiments. Doxycycline was injected intraperitoneally 5 mg kg⁻¹ on day 0 and day 3. b,c, Normalized Fluc (recorder)/Rluc (tumour) signals (b) and representative BLI images (c) of tumours. Two-way ANOVA with Šidák’s multiple comparisons test (two-sided). Error bars, mean ± s.e.m. Day 1, **P = 0.0089, day 4, *P = 0.0465, respectively. n = 7 mice in each group. d, Timeline of in vivo CAR-T cell killing experiment using NSG mice. e,f, Normalized tumour sizes (e) and representative BLI images (f) of mice with (1) recorder− or (2) recorder+ tumours only, or (3) with recorder+ tumours treated by cocktail CAR-T cells. Two-way ANOVA with Šidák’s multiple comparisons test (two-sided). Error bars, mean ± s.e.m. Recorder− versus recorder+, *P = 0.0464, recorder+ versus recorder+/cocktail CAR-T, ***P = 0.0001, respectively. n = 5 in each group. BLI images units, radiance (p s⁻¹ cm⁻² sr⁻¹). g, Normalized tumour sizes in mice with recorder+ tumours and treated with cocktail or EGFR-CAR only T cells. Two-way ANOVA with Šidák’s multiple comparisons test (two-sided). Error bars, mean ± s.d. D4, NS, not significant P > 0.9999, D11, NS, not significant P = 0.9990, D17, *P = 0.0241, D20, **P = 0.0014, respectively. n = 5 mice in each group. i.p., intraperitoneal; i.v., intravenous.

To validate the efficiency of CAR-T cell therapy targeting rewired CSLCs, we injected recorder-negative or recorder-positive MDA-MB-231 cells into mice mammary fat pad (m.f.p.) (Fig. 6d). In the tumour-only groups, recorder-positive tumours grew more aggressively than the recorder-negative tumours, with one mouse bearing recorder-positive tumour died at day 25, highlighting the tumorigenic potential and aggressiveness of the recorder-selected CSLCs. Ultrasound elastography revealed that CAR-T treatment drove tumour softening, verifying the evasion of immune cell killing engendered from softness (Extended Data Fig. 8a-j). Flow cytometry showed greater CD3⁺ T cell infiltration in rewired recorder-positive tumours, revealing the enhanced immune recruitment resulted from rewiring (Extended Data Fig. 9a,b). We then used a cocktail strategy, with EGFR-CAR-T cells targeting the bulk cancer cell population, and CD19-CAR-T cells targeting the rewired CSLCs expressing high CD19 induced by DOX every 3 days. This cocktail approach ultimately achieved robust killing efficacies (Fig. 6e,f). Notably, the cocktail treatment also had a notably stronger efficacy of tumour suppression than that with EGFR-CAR-T cells only (Fig. 6g). These results suggest that our genetic engineering strategy can rewire mechano-stemness of CSLCs into a ‘kill me’ signal to improve the effectiveness of CAR-T cell therapies in solid tumours in vivo.

Discussion

The mechanical properties of tumour microenvironment strongly influence the interactions between cancer and immune cells. Softer cancer cells, associated with higher malignancy and stemness, are more adept at evading immune cell killing than the stiffer cells. This raises the question of whether CAR-T cell efficacy, which relies on antigen–receptor interactions rather than MHC–TCR recognition, is similarly affected by cellular softness. In this study, we observed that cells cultured on softer ECMs are more resistant to CAR-T cell killing, despite similar antigen densities. Further investigation revealed that calcium and exATP levels were elevated in these cells. We therefore developed a genetic mechano-recorder that integrates mechanically associated calcium activity over a doxycycline-defined time window and converts this prior signalling history into a stable fluorescent reporter. This design enables individual cells to be barcoded according to their past mechanical experience, rather than merely reporting calcium levels instantaneously. These recorders lead to the finding that softer ECM promotes a higher calcium signalling and CSLC formation, a phenomenon we term ‘mechano-stemness’. We then demonstrated that this mechano-stemness can be rewired into a mechano-reprogrammer with synthetic ‘kill me’ signal by replacing the fluorescent reporter in the recorder system with the CD19 antigen, making these resistant CSLCs more susceptible to CAR-T targeting and killing. Our in vivo results further reinforced this concept: recorder-positive tumours remained softer, expressed higher Nanog and Sox2, and showed enhanced CD3⁺ T cell infiltration after rewiring. These findings indicate that soft environments correlate with stemness in vivo, consistent with reports that soft tumour regions are enriched for aggressive, therapy-resistant cells and can reprogram tumour-repopulating cells through H3K9 demethylation and Sox2 expression¹⁶. Overall, our study highlights the potential of recording and reprogramming cancer cell mechano-stemness to enhance the efficacy of CAR-T cell therapy against CSLCs associated with soft microenvironment, offering a novel strategy to improve therapeutic outcomes in challenging cancer subtypes.

Different calcium recorders have been developed with major applications in neuronal monitoring, including a calcium-dependent BioID (Cal-ID) recorder with three elements: (1) the calcium sensor calmodulin, (2) the calmodulin-binding peptide RS20 and (3) circularly permuted, split BioID enzyme⁴⁵; and another recorder consisting of a split tobacco etch virus (TEV) protease separately tethered to a calmodulin and a calmodulin-binding peptide M13⁴⁶. Since these calcium recorders are irreversible and lack controllable recording windows, the background noise may accumulate to overwhelm the recording specificity. With a DOX-gated and controllable recording window, our mechano-recorder should be suited for the precise recording and barcoding of individual cancer cells as a function of their mechanical experiences at specific periods. This mechano-recorder also allows the detection of softness-associated CSLCs for sorting and characterization.

CSLCs are key players in tumour recurrence and relapse⁴⁷. However, CSLCs are notoriously difficult to identify and target as they not only evade traditional chemotherapy^20,21 but also are softer, thus limiting the efficacy of T cell immunotherapies by reducing the traction force of T cells’ grip to form perforin pores^6,8,11. Methods have been proposed to stiffen the cancer cells to strengthen the synapse between the immune cells and cancer cells^5-7. One such method uses methyl-β-cyclodextrin to deplete cholesterol from cell membranes, which improves T cell interaction but poses the risks of non-specific toxicity. This method may not have sufficient suppression of CSLCs, leading to potential tumour recurrence^20,21,47. Furthermore, direct and specific modulation of cell mechanics in vivo remains a great challenge. Our approach offers an alternative, more feasible approach by rewiring the mechanostemness of CSLCs into the clinically validated CD19 expression for CAR-T cell targeting and killing. Of note, doxycycline, a Food and Drug Administration-approved drug, has shown effects in reducing CSLC formation in breast cancers⁴⁸. As such, DOX and the CAR-T targeting of rewired CSLCs in our system may lead to a synergistic effect to enhance the cancer immunotherapy against the softness-associated CSLCs. While our study focuses on engineered CSLCs as a mechanistic tool, the underlying principle, rewiring immune-evasive, soft tumour cells to become recognizable by CAR-T cells, is compatible with clinically viable approaches, including both targeted gene delivery platforms and advanced CAR-T cell engineering strategies^49-51. The recent advancements in tumour vaccination strategies using adeno-associated viruses to deliver DNA directly to target cells in vivo should further expand the potential translational impact of our system, offering new opportunities for therapeutic applications. As such, our engineered circuits showcase the potentials of (1) using the mechano-recorder to select the softness-associated CSLCs from a heterogeneous cancer population and (2) rewiring mechano-stemness into CD19 for CAR-T cell targeting and killing. This mechano-recorder system can be readily extended to screen CSLCs in patient samples. Furthermore, as calcium signalling plays an essential role in cancer metastasis⁵², the mechano-recorder could be used to identify and label these metastatically aggressive cells. By using barcoding and sorting-sequencing strategies, we can gain insights into the features of metastatic cells and use rewiring to specifically target and prevent cancer metastasis.

Methods

Cloning, lentiviral production and engineered cells

The plasmids used in this study are listed in Supplementary Table 1. PCR was performed using synthesized primers (Integrated DNA Technologies) and Q5 DNA polymerase (NEB, M0491). Cloning was achieved by using Gibson Assembly (NEB, E2611L), and the sequences of the constructed plasmids were verified by Sanger sequencing (Genewiz).

For lentiviral production, the plasmid of interest, the 2nd generation lentiviral packaging plasmid pMD2.G and PsPax2 were cotransfected into Lenti-X 293T cells (Clontech Laboratories, #632180) using the ProFection Mammalian Transfection System (Promega, E1200) according to the manufacturer’s protocol. After 48 h, the supernatant was collected and filtered through 0.45 μm PES filters (Genesee Scientific, #25-246), and the virus was further concentrated using the Lenti-X Concentrator (Takara Bio, 631231). Functional titres (transduction units) were determined by infecting Lenti-X 293T cells with titrated lentivirus and calculated based on the infection-positive rate measured by flow cytometry 3 days post-infection.

The engineered cells used in this work are listed in Supplementary Table 2. Per 0.2 × 10⁶ tumour cells, 5 μl 100× concentrated lentiviruses were added to deliver transgenes, which were assessed 72 h post-transduction with flow cytometry (BD Accuri C6), and then FACS (SONY SH800) was performed following standard protocols to achieve >90% positive rates. Plasmid and cell lines made in this study are available upon request.

General cancer cell lines, patient-derived cells and cell culture

Human breast epithelial adenocarcinoma cells, ER positive (MCF7, ATCC, HTB-22) or ER negative (MDA-MB-231, ATCC, HTB-26), human glioblastoma cell (U87-MG, ATCC, HTB-14), human prostatic adenocarcinoma cell (PC-3 CRL-1435) and pancreatic epithelioid carcinoma (Panc-1, CRL-1469) were cultured in Dulbecco’s modified Eagle medium (DMEM) (Gibco, 11995115) supplemented with 10% fetal bovine serum (FBS) (Gibco, 10438026) and 1% penicillin–streptomycin (P/S) (Gibco, 15140122). Patient-derived cells, 581153-110-T-J1-PDC, 879694-015-R-J1-PDC and 885512-296-R-J2-PDC, were obtained from NIH Patient-Derived Models Repository (PDMR). Patient-derived cells were cultured in advanced DMEM/F12 1X (Invitrogen, 12634-010), supplemented with 10% FBS, 1% P/S, 0.4 μg ml⁻¹ hydrocortisone (Sigma, H4001), 0.01 μg ml⁻¹ epidermal growth factor (EGF) recombinant human protein (Invitrogen, PHG031), 24 μg ml⁻¹ adenine (Sigma, A2786), 2 mM L-glutamine (Invitrogen, 25030-081) and 10 μM Y-27632 dihydrochloride (Tocris Bioscience, 1254). Primary human T cells were cultured in X-VIVO 15 (LONZA, 04-418Q) supplemented with 5% FBS, 1% P/S, 50 μM 2-mercaptoethanol (Gibco, 31350010) and 100 U ml⁻¹ recombinant human IL-2 (PeproTech, 200-02). Cells were cultured at 37 °C in a humidified 5% CO₂ incubator.

Isolation, culture and transduction of primary human T cells

Human peripheral blood mononuclear cells were isolated from buffy coats (San Diego Blood Bank) using lymphocyte separation medium (Corning, 25-072-CV). Primary human T cells were isolated using MACS with the Pan T Cell Isolation Kit (Miltenyi, 130-096-535) and activated with Dynabeads Human T-Expander CD3/CD28 (Gibco, 11141D) at a 2:1 beads-to-cells ratio. After 48–52 h, the T cells were mixed with lentivirus at a multiplicity of infection of 5 and seeded in plates pre-coated with 3 μg ml⁻¹ RetroNectin (Takara, T100B) at 4 °C overnight, followed by centrifugation at 1,800 × g at 32 °C for 1 h. Dynabeads were removed 5 days post-activation, followed by AF647 anti-F(ab′)₂ antibody staining and FACS. T cells were expanded every 15 days with a subsequent round of Dynabeads activation at a 1:1 beads-to-cells ratio. For reproducibility, T cells from two additional healthy donors (donor 2 and donor 3, San Diego Blood Bank) were isolated, activated and engineered as above, and used in cytotoxicity assays.

Fabrication of polyacrylamide hydrogel substrates

Polyacrylamide hydrogel substrates were prepared following a protocol adapted from literature²⁶. In brief, glass coverslips were chemically activated and polyacrylamide hydrogels of defined stiffness were prepared using standard protocols. Gels were polymerized using acrylamide and bis-acrylamide with ammonium persulfate and TEMED, functionalized with sulfo-SANPAH under ultraviolet irradiation and coated with fibronectin (50 μg ml⁻¹ in 1× PBS) 4 °C overnight, followed by 3 washes with 1× PBS before cell seeding.

Mechanical testing

Young’s moduli of the hydrogel samples were determined using a nanoindenter (Piuma, Optics11 Life). Samples were fixed to the bottom of a 3.5 cm Petri dish and immersed in PBS solution before testing. Nanoindentation tests were performed at room temperature using a spherical probe with a 50 μm radius and a cantilever stiffness of 0.5 N m⁻¹, selected based on the range of potential stiffness. During the measurement, the cantilever applied a pressure depth of 14,000 nm at a rate of 10 μm s⁻¹. The acquired data were fitted with the Hertz contact model and a Poisson’s ratio of 0.5 to obtain the stiffness measurement.

Staining and flow cytometry

Antibodies used in this paper are listed in Supplementary Table 3. Cells were washed and resuspended in 100 μl staining solution (1× PBS containing 1% FBS) with indicated dilutions for 30 min on ice. Following incubation, the samples were washed 3 times with 1× PBS before flow cytometry or FACS analysis. The FCS data were analysed using the FlowJo software (FlowJo v10).

Luciferase-based cytotoxicity assay

Fluc⁺ cancer cells were seeded at 2 × 10⁴ cells per well in a 96-well plate or 1 × 10⁵ cells per well in a 48-well plate with indicated ECM condition 3 h before adding T cells with the indicated E:T ratios. To test the role of Ca²⁺ signalling, tumour cells were cultured in Ca²⁺-free DMEM with BAPTA (100 μM) or 2-APB (50 μM) for 30 min before CAR-T addition and maintained during 24 h co-culture. After 24 h of co-culture, a ONE-Glo Luciferase Assay System (Promega, E6110) was used following the manufacturer’s instructions to quantify luminescence. Reference groups for calculating growth rates are the same amount of tumour cells used for seeding at 0 h. The growth (%) of sample X was calculated as (luminescence of X/luminescence of reference) × 100%. For cytotoxicity assay, the reference groups are cancer cells seeded with the same density and cultured on the same ECM conditions without T cell addition. The cytotoxicity (%) of sample X was calculated as (1 – luminescence of X/luminescence of reference) × 100%. Normalized cytotoxicity was that normalized to killing of tumour cells with constitutive antigen expression seeded on respective ECM stiffness with respective CAR-T cells at an E:T ratio of 1:1 for 24 h.

Immunofluorescent staining and fluorescence imaging

MDA-MB-231 and U87-MG cells were cultured on indicated ECM overnight, fixed with 4% paraformaldehyde for 1 h at room temperature and blocked with 5% bovine serum albumin (BSA; Gemini Bio-Products) in DPBS overnight at 4 °C. The primary antibodies and secondary antibodies used in this study are listed in Supplementary Table 3. The samples were incubated in primary antibody mixture in staining buffer (BioLegend) overnight at 4 °C. After washing 3 times with 1× PBS, the samples were stained in secondary antibodies and DAPI (1:1,000, Santa Cruz Biotechnology sc-3598) for 1 h in the dark at room temperature. The samples were rinsed 3 times again with 1× PBS and mounted for imaging with a Nikon Eclipse Ti inverted microscope. Quantifications of the mean fluorescence intensity were performed using ImageJ 1.54p, and relative fluorescence intensity was calculated over the mean fluorescence intensity of individual cells cultured on coverslip.

exATP measurements

MDA-MB-231 cells were seeded at 2 × 10⁴ cells per well in a 96-well plate with indicated ECM condition, and exATP levels were measured as luminescence at 24 h or 48 h after seeding with RealTime-Glo Extracellular ATP Assay (Promega, GA5010). References are measurements of empty wells with the same reagent and medium as sample groups (no cell control), and exATP levels are calculated as percentages of no cell control. For mechanistic studies, cells were treated with blebbistatin, BAPTA or 2-APB under conditions described above, and exATP was quantified at 24 h.

Cell seeding and drug treatments in mechano-recorder testing

Cancer cells (MDA-MB-231, MCF7, 110-T, 015-R and 296-R cells) with mechano-recorder were seeded on indicated ECM for overnight and treated with doxycycline (500 nM, Sigma, D1822-500MG) for indicated treatment windows. Ionomycin (1 μg ml⁻¹, Sigma, I0634-1MG), thapsigargin (2 μM, Selleck, S7895), apyrase (15 U ml⁻¹, Sigma, A6410-100UN) and ATP (100 μM, Sigma, A1852-1VL) were added with doxycycline and treated for 2 h. 2-APB (50 μM, Sigma, 100065), complete calcium free DMEM (Gibco, 21068028) and BAPTA (100 μM, Thermo Scientific, B22258) were added 30 min before the addition of doxycycline and washed out with doxycycline after 2 h. Cells with different densities (0.005 × 10⁶ (low), 0.01 × 10⁶ (medium) and 0.05 × 10⁶ (high)) were seeded on the stiffness gel and cultured for 24 h before DOX addition and imaging. For pharmacological softening, cells were treated with blebbistatin (2 μM, Sigma, B0560) for 24 h before DOX addition and 2 h with doxycycline for mechano-recorder induction imaging. For all other trials, 0.01 × 10⁶ cells were seeded on the indicated stiffness gel or with indicated drug treatments for consistency purposes. Images were taken with a Nikon Eclipse Ti inverted microscope. Mechano-recorder activity was measured from single-endpoint imaging where cell segmentation was done using miRFP670 images, and the ratio of mCherry/miRFP was calculated by diving mCherry MFI over miRFP MFI for individual cells.

3D culture in collagen I matrix

3D cell culture in collagen I were prepared following a protocol adapted from literature³⁶ by embedding cells in rat tail collagen I (1 mg ml⁻¹ or 6 mg ml⁻¹), followed by polymerization at 37 °C. The cells in collagen were allowed to grow for >24 h, up to 30 h, washed once using PBS and retrieved from dissolving using collagenase (3 mg ml⁻¹, Sigma, 11213857001) for 30 min under 37 °C and subsequently pelleted. The cells were then stained, and fluorescence signals were measured using flow cytometry as described.

Measurement of cellular stiffness by AFM

Cellular stiffness was measured using a Park FX200 AFM (Park Systems). The cantilevers used had a nominal spring constant of 0.06 N m⁻¹, manufactured by Nano World. The force/distance (F/D) spectroscopy of each cell was measured using the contact mode, with a tip velocity of 0.5 μm s⁻¹. The peak force setpoint for feedback was set to 1 V. Young’s modulus was calculated using the Park Systems XIE data analysis software.

To correlate cellular stiffness with recorder activity on a single-cell level, AFM was coupled with fluorescence microscopy. An Asylum 3D-BIO AFM (Asylum Research) mounted on a Ti-U fluorescent inverted microscope (Nikon Instruments) was operated in force contact mode. InvOLS calibration was performed on glass, and a silicon nitride pyramid-tipped cantilever (NanoWorld; nominal spring constant ~40 pN nm⁻¹ by thermal calibration) was used for indentations. Cells were maintained in culture medium at 37 °C. Force–indentation curves (1–2 nN, ≥10 per cell) were analysed with a Hertz model (conical tip, Poisson ratio = 05) fitted to the first 30% of the approach curve. Fluorescence images were simultaneously captured using a C-HGFI Intensilight (Nikon) and QuantEM camera (Photometrics). Regions of interest were manually drawn around AFM-measured cells in Fiji (ImageJ 1.54p) to quantify mean fluorescence intensities.

Mammosphere formation assay

MDA-MB-231 cells with recorder were seeded on stiff or soft ECM overnight and treated with doxycycline for 2 h. Six hours after doxycycline addition, soft recorder-negative and soft recorder-positive cells were sorted through FACS. Each group of cells was seeded 100 cells per well into 96-well low-attachment plates in DMEM/F12 with supplement B27 (1X, Life Technologies, 17504044) and 10 ng ml⁻¹ EGF. Medium (50 μl) was added per well every 5 days. After 23 days of culture, mammosphere growth was monitored using IncuCyte. Mammospheres with a diameter of >100 μm were counted.

Animals

Animal studies were approved in protocol 21479 by the USC Institutional Animal Care and Use Committee (IACUC). All researchers complied with animal-use guidelines and ethical regulations during the animal studies. All experiments used female NOD.Cg-Prkdcscid Il2rgtm1Wjl/SzJ (NSG) mice of 6–8 weeks old were purchased from Jackson Laboratory. The maximum allowable tumour size under both institutional protocols was 1.5 cm in diameter or approximately 10% of body weight. Tumour growth was closely monitored, and no animals exceeded these limits throughout the study.

In vivo bioluminescence imaging

Fluc substrate D-luciferin (75 mg kg⁻¹, GoldBio, LUCK-1G) or Rluc substrate coelenterazine (5 mg kg⁻¹, GoldBio, CZ25) was administered intraperitoneally into NSG mice, followed by in vivo bioluminescence imaging (BLI) approximately 8–12 min later. The signal is captured by Lumina LT Series III (PerkinElmer). BLI of Fluc and Rluc in the same mouse was performed >6 h apart. Images with peak signals were analysed with Living Image software (PerkinElmer). Induction signal or tumour size is analysed based on the integrated reading within a fixed region of interest over the tumour.

Inducible gene activation in vivo

MDA-MB-231 cells with recorder were seeded on stiff or soft ECM overnight and treated with doxycycline for 2 h. Six hours after doxycycline addition, the cells were sorted through FACS. NSG mice were injected with 0.5 × 10⁶ soft recorder-positive or stiff recorder-negative MDA-MB-231 cells (with constitutive Rluc and inducible Fluc) into the mammary fat pad. Five and eight days later (days 0 and 3), doxycycline (5 mg kg⁻¹) was administered intraperitoneally, and Fluc luminescence showing recorder expression was measured 6 h before Rluc luminescence showing tumour size. Induction of the mechano-recorder was quantified using the integrated Fluc luminescence intensity over Rluc luminescence intensity.

In vivo cytotoxicity of cocktail CAR-T cells

NSG mice were injected with 0.5 × 10⁶ soft recorder-positive or stiff recorder-negative MDA-MB-231 cells at their mammary fat pad. Five days later, 1.5 × 10⁶ EGFR-CAR-T cells or cocktail CAR-T cells (1.5 × 10⁶ EGFR-CAR-T cells + 1.5 × 10⁶ CD19-CAR-T cells) were administered through tail vein injection. Beginning on day 5, doxycycline (5 mg kg⁻¹) was administered intraperitoneally every 3 days. Tumour progression was monitored by BLI on the indicated days. Tumour size was quantified using the integrated Rluc luminescence intensity of the tumour region and normalized to that of the same tumour measured on day 4.

Elastography

Tumour elasticity was assessed using ultrasound elastography (HS2 system, KONICA MINOLTA, L18-4 probe) on mice with tumours with or without EGFR-CAR-T treatment at day 21. Regions of interest were manually drawn, and strain values were divided into 256 levels (0 = stiffest to 255 = softest). Pixel distributions were plotted as histograms. Different tumours were imaged and mean strain values were compared to assess softness.

Immunohistochemistry staining

Frozen tumour tissues were embedded in OCT and cryosectioned at 8–10 μm thickness. Sections were fixed in 4% PFA in PBS (10 min, room temperature), washed in PBS and incubated in 0.3% H₂O₂ for 10 min to quench endogenous peroxidase. After blocking with 3% BSA in PBS (1 h, room temperature), sections were incubated overnight at 4 °C with primary antibodies against Nanog and Sox2 (Supplementary Table 3). From this step onwards, all washes and antibody dilutions were performed in PBST (PBS + 0.1% Tween-20). Sections were washed 3 times in PBST, incubated with HRP-conjugated secondary antibodies for 30 min at room temperature, developed with 3,3′-diaminobenzidine (DAB) substrate, counterstained with haematoxylin, dehydrated and mounted. Images were acquired on an EVOS M7000 fluorescence microscope, and staining intensity was quantified using ImageJ 1.54p with DAB colour deconvolution.

T cell infiltration

Tumours were collected from NSG mice at day 15, minced into 1–2 mm fragments in cold RPMI medium and digested with Collagenase P (1 mg ml⁻¹; Millipore Sigma, catalogue number 11213857001) and DNase I (100 μg ml⁻¹; Roche, catalogue number 10104159001) at 37 °C for 45 min with gentle agitation. After digestion, samples were triturated, filtered through a 70 μm strainer and washed with PBS, followed by centrifugation at 400 × g for 5 min. Red blood cells were lysed with eBioscience 1X RBC Lysis Buffer (Thermo Fisher, catalogue number 00-4333-57) for 3 min at room temperature. Single-cell suspensions were resuspended in flow staining buffer (1× PBS + 1% FBS), stained with CD3 antibody for 30 min at 4 °C in the dark, washed and analysed using flow cytometry. Data were processed with FlowJo v10.

Bulk RNA sequencing and data analysis

MDA-MB-231 cells expressing the recorder were seeded on stiff or soft ECM overnight and treated with doxycycline for 2 h. Six hours after doxycycline treatment, the cells were sorted through FACS with top 20% recorder-positive and bottom 20% of recorder-negative cells for each ECM condition and processed without freezing. For T cell transcriptomic profiling, donor 2 CD19-CAR-T cells were mono-cultured or co-cultured with rewired or non-rewired tumour cells for 24 h (E:T = 1:1), after which T cells were isolated by FACS. Total RNA was extracted using the RNeasy mini kit (Qiagen). RNA quality check with BioAnalyzer (Agilent), sequencing libraries preparation and paired-end sequencing with Illumina NovaSeq 6000 were performed by Novogene to generate 20 million paired-end reads per sample (2 × 150 bp). Patient-derived cells RNA-seq raw counts were downloaded from the NIH PDMR public site (https://pdmr.cancer.gov) as individual biological replicates. Notably, 296-R and 015-R samples contain one in vitro culture with all others being xenograft samples, and 110-T only contains an in vitro culture sample as it is non-tumorigenic in mice.

For RNA-seq analysis, Trim Galore (v0.6.5, https://github.com/FelixKrueger/TrimGalore) was used to trim and filter low-quality reads. After quality control using fastqc, each fastq file was mapped to the human hg38 genome with gene annotation from GENCODE version 33, with >90% reads successfully aligned, and transcript numbers were quantified using Salmon (https://combine-lab.github.io/salmon/). Differentially expressed gene analysis was performed using the edgeR (version 4.3) package. Significant differentially expressed genes were determined by false discovery rate (FDR) < 0.05 and log₂ fold change greater than 0.5. Gene Ontology terms were identified using DAVID Bioinformatics Resources, and GSEA was performed with the GSEA desktop application (Broad Institute) and the molecular signatures database after normalization with DESeq2 (version 3.19) package.

Study design and statistical analyses

Statistical analyses were performed using Prism software (Graph-Pad). Detailed statistical information for each figure is provided in the corresponding figure captions. No statistical methods were used to pre-determine sample sizes, but our sample sizes are similar to those reported in previous publications^35,53. Mice were randomly assigned to treatment groups. Data collection and analysis were not performed blind to the conditions of the experiments. Data distribution was assumed to be normal, but this was not formally tested. No animals or data points were excluded from analysis.

Extended Data

Extended Data Fig. 1 ∣. The Calcium recorder has a higher expression in cells cultured in softer 3D ECM.

a, Constructs of regulator and reporter transduced in MDA-MB-231 cells. b-d, Percentage (b), RFI (c), and histogram (d) showing recorder expression in MDA-MB-231 cells transduced with constructs in (a) cultured in 6 mg/ml or 1 mg/ml collagen for 24 hr. e, Constructs of regulator with constitutive rtTA and reporter transduced in MDA-MB-231 cells to serve as a control. e-h, Percentage (f), RFI (g), and histogram (h) showing recorder expression in cells transduced with constructs in (e) cultured in 6 mg/ml or mg/ml collagen for 24 hr. RFI, relative fluorescence intensity. b,c, Two-way ANOVA with Šidák’s multiple comparisons test (two-sided). Error bars, mean ± SD. 6 mg/mL or 1 mg/mL −DOX/ + DOX, ****p < 0.0001, +DOX 6 mg/mL vs. 1 mg/mL, *p = 0.0237, respectively. f,g, Two-tailed t-test. Error bars, mean ± SD. ns, not significant p = 0.0952, ns, not significant p = 0.0954, respectively. n = 3 (b,f) representive independent replicates of induction assay, and n = 100 (c,g) represents individual cells detected.

Extended Data Fig. 2 ∣. ATP increased the calcium recorder expression.

a,b, Representative images (a) and quantification (b) of the recorder expression in MDA-MB-231 cells seeded on coverslips in the presence of doxycycline, with and without ATP treatment. ATP was added with doxycycline for 2 hr. One-way ANOVA. Error bars, mean ± SD. ****p < 0.0001. +DOX/−ATP, n = 172, +DOX/ + ATP, n = 177, where n is individual cells detected. Scale bar = 100 um. Color scale indicates the pseudo-color intensities of mCherry channel.

Extended Data Fig. 3 ∣. Recorder selects cancer stem-like cells on stiff ECM.

a, GO enriched pathways from bulk RNA-Seq analysis with upregulated differentially expressed genes in stiff recorder+ cells comparing to stiff recorder− cells, p-adjusted < 0.05 and log2 fold change > 0.5. b, GSEA analysis using normalized gene counts in stiff recorder+ (Stiff Positive) cells versus stiff recorder− (Stiff Negative) cells show enriched pathways of EMT, hypoxia, IL6-JAK-STAT3 signaling, and KRAS signaling upregulation in Stiff Positive cells. NES, normalized enrichment score. FDR, false discovery rate. c,d, Representative images (c) and quantification (d) of mammosphere formation from stiff recorder− and stiff recorder+ cells. Stiff recorder− and stiff recorder+ cells were collected from FACS and seeded 100 cells per well. After 23 days, mammosphere with > 100 um diameter are counted. Two-tailed t-test. Error bars, mean ± SD. ****p < 0.0001. n = 42 in each group represents individual wells seeded. e, Cellular stiffness of stiff recorder− and stiff recorder+ cells was detected by AFM. Two-tailed t-test. Error bars, mean ± SD. *p = 0.0181. Recorder−, n = 35, Recorder + , n = 39, where n is individual measurements. f, Cytotoxicity (cell death percentage) of stiff recorder− and stiff recorder+ cells co-cultured with EGFR-CAR T cells at an E:T ratio of 1:1 for 24 h. Two-tailed t-test. Error bars, mean ± SD. *p = 0.0272. n = 3 in each group where n is independent replicates of cytotoxicity assay.

Extended Data Fig. 4 ∣. Recorder expression is higher in U87-MG, Panc-1 and PC-3 cells when ECM is softer.

a-f, Representative images (a,c,e) and quantification (b,d,f) of recorder expression in U87-MG, Panc-1 and PC-3 cells cultured on coverslip, stiff or soft ECM, with and without doxycycline treatment. Two-way ANOVA with Šidák’s multiple comparisons test (two-sided). Error bars, mean ± SD, U87-MG, ****p < 0.0001; Panc-1: Coverslip/−DOX vs. Coverslip/+DOX, ns, not significant p > 0.9999, Stiff/−DOX vs. Stiff/+DOX, ***p = 0.0003, Soft/−DOX vs Soft/+DOX, ns, not significant p = 0.0522, all others, ****p < 0.0001; PC-3: Coverslip/+DOX vs. Stiff/+DOX, ns, not significant p = 0.4570, all others, ****p < 0.0001. U87-MG: Soft/−DOX, n = 149, Soft/+DOX, n = 549, Stiff/−DOX, n = 486, Stiff/+DOX, n = 686, Coverslip/−DOX, n = 257, Coverslip/+DOX, n = 768; Panc-1: Soft/−DOX, n = 41, Soft/+DOX, n = 191, Stiff/−DOX, n = 32, Stiff/+DOX, n = 224, Coverslip/−DOX, n = 78, Coverslip/+DOX, n = 137; PC-3: Soft/−DOX, n = 214, Soft/+DOX, n = 188, Stiff/−DOX, n = 573, Stiff/+DOX, n = 281, Coverslip/−DOX, n = 209, Coverslip/+DOX, n = 164. n depicts individual cells detected. Scale bar = 100 μm. Color scale indicates the pseudo-color intensities of mCherry channel.

Extended Data Fig. 5 ∣. Rewiring to induce higher CD19 expression in cells cultured in softer 3D ECM.

a-c, Percentage (a), RFI (b), and histogram (c) showing rewired CD19 expression in MDA-MB-231 cells cultured in 6 mg/ml or 1 mg/ml collagen. RFI, relative fluorescence intensity. Two-way ANOVA with Šidák’s multiple comparisons test (two-sided). Error bars, mean ± SD. ****p < 0.0001. n = 1000 (b) represents individual cells detected, and n = 3 (a) representive independent replicates of induction assay in each group.

Extended Data Fig. 6 ∣. CD19 rewiring restores CAR-T cytotoxicity against soft-ECM tumor cells across multiple donors.

a,b, Cytotoxicity (cell death percentage) of EGFR-CAR or CD19-CAR T cells from (a) donor 2 and (b) donor 3 targeting MDA-MB-231, MDA-MB-231-CD19 cells or MDA-MB-231 rewired to express CD19, cultured on coverslip, stiff or soft ECM, at an E:T ratio of 1:1 for 24 hr, respectively. One-way ANOVA with Tukey’s multiple comparisons test (two-sided). Error bars, mean ± SD. (a) EGFR-CAR T/MDA-MB-231: Coverslip vs. Soft, *p = 0.0474, ns, Coverslip vs. Stiff, not significant p = 0.1908, Stiff vs. Soft, ns, not significant p = 0.5473; CD19-CAR T/MDA-MB-231-CD19: Coverslip vs. Soft, ***p = 0.0005, Coverslip vs. Stiff, *p = 0.0471, Stiff vs. Soft, **p = 0.0073; CD19-CAR T/MDA-MB-231-rewired, Coverslip vs. Stiff, ns, not significant p = 0.3216, Coverslip vs. Soft, **p = 0.0047, Stiff vs. Soft, *p = 0.0249. (b) EGFR-CAR T/MDA-MB-231: Coverslip vs. Stiff, ns, not significant p = 0.1665, Coverslip vs. Soft, *p = 0.0133, Stiff vs. Soft, ns, not significant p = 0.1710; CD19-CAR T/MDA-MB-231-CD19: Coverslip vs. Stiff, *p = 0.0169, ns, Stiff vs. Soft, not significant p = 0.1142, Coverslip vs. Soft, **p = 0.0017; CD19-CAR T/MDA-MB-231-rewired, Coverslip vs. Stiff, ns, not significant p = 0.3472, Coverslip vs. Soft, *p = 0.0431, Stiff vs. Soft, ns, not significant p = 0.2910, respectively. n = 3 in each group where n is independent replicates of cytotoxicity assay.

Extended Data Fig. 7 ∣. Transcriptomic profiling reveals enhanced activation and effector programs in CAR-T cells co-cultured with CD19-rewired tumor cells.

a, Principal component analysis (PCA) of RNA-seq data from CAR-T cells cultured alone (mono-culture, black), co-cultured with non-rewired MDA-MB-231 tumor cells (blue), or co-cultured with CD19-rewired MDA-MB-231 tumor cells (red). CD19-rewired co-culture produces a distinct transcriptional profile separated along PC1. b, Heatmaps showing differential expression (Z-score) of genes related to functional states (left) and activation signaling pathways (right) in CAR-T cells from each culture condition. c, GSEA comparing rewired versus non-rewired co-cultures (NES = normalized enrichment score; FDR = false discovery rate).

Extended Data Fig. 8 ∣. In vivo ultrasound elastography reveals that both stiff and soft tumors become softer after EGFR-CAR-T treatment.

a,c, Representative B-mode ultrasound image with superimposed elastography map of a (a) stiff recorder-negative or (c) soft recorder-positive MDA-MB-231 tumor and xenograft without CAR-T treatment. b,d, Pixel-wise histogram of elasticity indices from the tumor in a,c, respectively. 1 = stiff, 255 = soft. e,g, Representative B-mode ultrasound image with superimposed elastography map of a (e) stiff recorder-negative or (g) soft recorder-positive MDA-MB-231 tumor and xenograft with CAR-T treatment. f,h, Pixel-wise histogram of elasticity indices from the tumor in e,g, respectively. i, Quantification of pixel-weighted mean elasticity index for stiff recorder-negative or soft recorder-positive tumors with and without CAR-T treatments. Two-tailed t-test. Error bars, mean ± SD. Stiff Recorder− vs. Stiff Recorder−/EGFR-CAR T, *p = 0.0409, Soft Recorder+ vs. Soft Recorder + /EGFR-CAR T, ns not significant p = 0.3566. n = 3 in each group where n is individual mice. j, Schematic showing CAR-T cells target stiffer tumor cells and leaving mechanically soft tumors.

Extended Data Fig. 9 ∣. Enhanced CAR-T infiltration in soft recorder-positive tumors.

a, Representative flow cytometry plots showing the percentage of CD3⁺ T cells within tumor samples from stiff recorder-negative (left) and soft recorder-positive (right) MDA-MB-231 xenografts following CD19-CAR T therapy. Gated CD3⁺ populations are indicated as a percentage of total events. b, Quantification of CD3⁺ T cell infiltration across tumors. Two-tailed t-test. Error bars, mean ± SD. *p = 0.0396. n = 4 in each group where n is individual mice.

Supplementary Material

Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41551-026-01722-7.

Data availability

The main data that support the results of this study are available within the paper and its Supplementary Information. Raw RNASeq datasets are available in NCBI SRA (BioProject PRJNA1416904). Source data are provided with this paper.

Code availability

The code for image analysis and quantification used in this study is available via GitHub at https://github.com/phuongho43/fluoressential.