Transient cell-in-cell formation underlies tumor resistance to immunotherapy

Despite the remarkable success of immunotherapy in cancer, most patients will develop resistant tumors. While the main conceptual paradigm suggests that relapsed clones emerge through a process of clonal selection and immunoediting, currently little evidence directly demonstrates this process in epithelial cancers. To study this process, we established several mouse models in which tumors drastically regress following immunotherapy, yet resistant tumors relapse within a few weeks of treatment cessation. Whole exome analyses indicated that relapsed tumors share hundreds of neo-antigens with the primary tumors and are comparably killed by reactive T cells. Examination of tumor cells that survive immunotherapies revealed that they structure a transient cell-in-cell formation, which is impenetrable to immune-derived cytotoxic compounds and to chemotherapies. This formation is mediated predominantly by a cell-membrane protein on activated T cells, which subsequently induces epidermal growth factor receptors and STAT3 phosphorylation in tumors cells. In contrast to previous reports on cell-in-cell formations, here both cells remain alive and can disseminate into single tumor cells once T cells are no longer present. Overall, this work highlights a powerful resistance mechanism which enable tumor cells to survive immune pressure and provides a new theoretical framework for combining chemotherapies and immunotherapies.


Abstract
Despite the remarkable success of immunotherapy in cancer, most patients will develop resistant tumors. While the main conceptual paradigm suggests that relapsed clones emerge through a process of clonal selection and immunoediting, currently little evidence directly demonstrates this process in epithelial cancers.
To study this process, we established several mouse models in which tumors drastically regress following immunotherapy, yet resistant tumors relapse within a few weeks of treatment cessation. Whole exome analyses indicated that relapsed tumors share hundreds of neo-antigens with the primary tumors and are comparably killed by reactive T cells. Examination of tumor cells that survive immunotherapies revealed that they structure a transient cell-in-cell formation, which is impenetrable to immune-derived cytotoxic compounds and to chemotherapies. This formation is mediated predominantly by a cell-membrane protein on activated T cells, which subsequently induces epidermal growth factor receptors and STAT3 phosphorylation in tumors cells. In contrast to previous reports on cell-in-cell formations, here both cells remain alive and can disseminate into single tumor cells once T cells are no longer present.
Overall, this work highlights a powerful resistance mechanism which enable tumor cells to survive immune pressure and provides a new theoretical framework for combining chemotherapies and immunotherapies.

Main
Given the direct correlation between the prevalence of cytotoxic T-cell immunity and improved clinical outcomes, and survival [1][2][3] , most therapeutic strategies are aimed at harnessing T-cell immunity to fight cancer. These attempts include de novo expansion of tumor-infiltrating cytotoxic T 4,5 , engineered T cells 6,7 , or blocking antibodies directed against suppressive receptors 8,9 . However, long-term follow-up indicates that while patients will experience tumor regression, it is frequently followed by recurrence of tumors that are largely resistant to subsequent treatments [10][11][12] . As tumor cells often express altered or neo-antigens 13,14 , it remains unclear why eliciting the full spectrum of T-cell immunity to these antigens is not sufficient to eradicate tumors.
Two theoretical frameworks have been widely employed to understand how tumors escape Tcell-based therapies. The main conceptual paradigm suggests that in order to avoid killing by T cells, tumor cells edit or lose their targeted antigens, or down-regulate HLA molecules [15][16][17] .
Alternately, tumor resistance to immunotherapy was shown to be acquired through tumor cellintrinsic mechanisms, including loss-of-function of JNK and PTEN, MYC overexpression, and constitutive WNT signalling [18][19][20] .
Nonetheless, most of our knowledge on this process comes from experiments comparing individuals who respond or do not respond to immunotherapy and not from experimental settings that measures the cell-intrinsic changes, neo-antigen landscape, and immune parameters within the same individual. Furthermore, whole exosome and genome sequencing comparisons of paired primary and relapsed tumors indicate that the majority of clones in relapsed tumors are shared with the primary tumors, suggesting relative low rates of clonal deletion [21][22][23] .
Therefore, to study this process in melanoma and breast cancer, we developed several mouse models in which resistant tumors relapse following immunotherapies. In both mouse and humans, we found that the tumor cells remaining after immunotherapy form unique cell-in-cell structures and generate membrane architecture that is impenetrable by immune-derived lytic granules, cytotoxic compounds, and chemotherapies. Both cells in this formation maintain their integrity and viability and can survive for weeks in culture containing reactive T cells. Once T cells are removed, tumor cells disseminate back into their parental single cells. This biological process has never been described before and is the main mechanism through which tumor cells escape T-cell immunity and give rise to relapsed tumors.

Relapsed tumor cells share neoantigens with primary tumors and are equally susceptible to killing by T cell
We have recently reported that the combination of dendritic cell adjuvant with tumor-binding antibodies elicits an extremely potent T-cell reaction 24 . Indeed, when palpable melanoma tumors were injected with a combination of TNFCD40L, and an antibody against the melanoma antigen TRP1, tumors were almost completely eradicated in all mice. Nonetheless, after about ten days, about half the mice developed reoccurred tumors that were resistant to subsequent treatments (Fig. 1a). Similar patterns of tumor relapse and resistance were also observed upon treating BALB/c mice bearing breast tumors treated with allogeneic antibodies and DC adjuvant (Fig. 1a). Intriguingly, immune composition analysis of tumors following the second immunotherapy indicated treated mice across all immunotherapies show massive immune-cell and T-cell infiltration (Fig. 1b, Extended Data Fig. 1a) with a dominant representation of T-cell clones against TRP2 and gp100 (Fig. 1c). Furthermore, analysis of melanoma cells showed they maintain expression of MHC-I and MHC-II (Fig. 1d, Extended Data Fig. 1a-1b), as well as TRP2 and gp100 (Fig.1e). To assess the overall changes in immunogenicity of resistant cell lines, beyond TRP2 and gp100 expression, we analysed their neoantigen landscape in comparison to that of B16F10 from untreated mice (Fig. 1f). Whole exome analysis (WES) indicated comparable neoantigen burden across all samples ( Fig. 1g) with the same 25 germline mutations in the highest score ones (Fig. 1h To assess whether tumors acquire resistance through inherent mechanisms, we next tested the susceptibility of these cell lines to immunotherapy. Consistent with their patterns of gene and neoantigens expression, these cell lines were equally susceptible to immunotherapy as their parental B16F10 cells (Fig. 1j). To test if this occurrence represents a broader phenomenon, we also treated melanoma-bearing mice with splenic CD8 + T cells expressing TCR against gp100, or TRP1 melanoma antigens. Both treatments induced significant tumor regression, followed by tumor reoccurrence in all treated mice (Fig. 1k). As observed with the previous immunotherapy, reoccurred tumors were resistant to subsequent treatment with gp100-or TRP2-reactive T cells (Fig. 1k). Consistently, in vitro analysis indicated that cell lines established from resistant tumors were susceptible to killing by gp100-reactive T cells (Fig. 1l). Moreover, mice bearing these cell lines were equally susceptible to treatment with gp100-reactive T cells as mice bearing the parental melanoma cells (Fig. 1m).
These results suggest that transient in vivo mechanisms rather than inherited ones govern resistance of relapsed tumors. In support of this hypothesis, RNAseq analysis indicated that of all established cell lines cluster within the same principle components, whereas the expression profile of freshly sorted tumor cells from immunotherapy-treated mice is markedly distinct (Fig. 1n, Extended Data Fig 1c, Supplementary Table-3).

Tumor cells remaining after T-cell killing organize in a cell-in-cell formation
To better characterize the tumor cells that survive in mice following immunotherapy, we enzymatically digested treated tumors and sorted the live melanoma cells. Confocal analyses indicated that most of the tumor cells organize in clusters of several nuclei surrounded by a single membrane and cortical actin, suggesting they structure in a cell-in-cell formation (Fig.   2a). Transmitting electron microscopy (TEM) of freshly sorted cells further corroborated a distinct segregation of the membranes and cytosols of the two cells. The inner cell was dense and seems to have been compact within another cell (Fig. 2b). To ensure that these structures do not result from the isolation procedure, we also analysed histological sections of tumors, whose nucleus and cell membranes are fluorescently labelled. While untreated tumors showed low levels of cell-in-cell clusters, about half the cells that survived in treated mice were organized as such ( Fig. 2c-2d).
To test if this formation was merely the result of incomplete cell division, we injected into mice tumor cells whose nuclei are labelled with GFP or with tdTomato, and treated the mice with immunotherapy. About one third of the tumor cells were a mixture of both GFP and tdTomato nuclei (Fig. 2e). Furthermore, monitoring the dynamics of single cells sorted from treated mice indicated that they spontaneously clustered in a cell-in-cell formation during their whether a specific immune cell type mediate these cell-in-cell formations, we isolated the main effector cells from tumor-bearing mice and incubated them overnight with B16F10 tumor cells.
We found that only T cells-mainly CD8 + , and with slower kinetics also CD4 + -induce a cellin-cell formation with the same special characteristics as the ones observed in vivo following immunotherapy ( Fig. 2f- Consistently, no cell-in-cell structures were observed in Nude SCID-gamma -/-(NSG) mice treated with immunotherapy (Fig. 2h).
We next tested whether this phenomenon occurs in other tumor cell types. Indeed, the same structures were also observed in 4T1 mammary carcinoma incubated with allogeneic T cell, but not in immortalized mammary epithelial cells (Extended Data Fig. 2c-2d). Most importantly, similar structures were also found in human breast carcinoma and melanoma cell lines incubated with allogeneic T cells from healthy donors (Extended Data Fig. 2e-2f).
Furthermore, incubation of cell lines established from two melanoma patients with their corresponding TIL indicated that the vast majority of cells that survive in culture form a cellin-cell formation, similar to the one found in mice (Extended Data Fig. 2g).
We next assessed how this structure protects the tumor cells. Initially, we incubated tumor cells with T cells expressing TCR against the known melanoma antigen gp100 and tested whether they escape T cells by reducing their MHCI or gp100 expression. We found that all cells in the cluster expressed similar levels of MHCI or gp100 compared to untreated tumor cells (Fig 2i-2j). Furthermore, scanning electron microscope (SEM) analysis indicated that multiple T cell were attached to the clustered tumor cells, and many pores were observed on the outer cell membrane, suggesting their ability to form an immunological synapse on their membrane (Fig. 2k). Next, we tested if tumor cell lines established following killing with TRP2and gp100-reactive T cells would be less sensitive to killing by the same reactive T cells. Thus, cell lines were then incubated again with gp100 and TRP2-reactive T cells, and their rates of killing were compared to that of the parental cell line (Fig 2l). Instead, tumor cells re-organized in cell-in-cell formation regardless of the nature of the tumor-reactive T cells (Fig. 2m), suggesting that this formation provides a transient protection, rather than an inherited one. Thus, we compared the sensitivity of single-tumor cells and those in a cell-in-cell formation to a variety of cytotoxic compounds and chemotherapies. In all the tested compounds, cell-in-cell formations survived longer and higher concentrations compared to their parental cell lines ( Fig.   2n and Extended Data Fig. 2h). To visualize the penetration rates of cytotoxic compound to these structures, we incubated CFSE-labelled lysosomes with single-tumor cells with cell-incell formed tumor cells. Indeed, lysosome penetration accumulated on the multiple membranes of the cell-in-cell formations, and penetration of the latter was much slower compared to singletumor cells (Extended Data Fig. 2i).

IFN induces membrane-bound protein on T cells that mediate tumor cell-in-cell formations
To understand why only T cells induce a cell-in-cell formation, we initially analysed their interactions with tumor cells using SEM. Interestingly, tumor-reactive T cells, but not PMA plus ionomycin-activated NK cells or LPS-activated macrophages, express the intact form of lysosomes on their cell surface ( Fig. 3a-3c, Extended Data Fig. 3a). Next, we tested if this formation can be induced by other extracellular vesicles. Hence, B16F10 cells were cultured overnight with extracellular vesicles from activated NK, macrophages, immortalized melanocytes (Melan A), B16F10-derived lysosomes, or commercial 1M latex beads. Confocal analysis of these cultures indicated that only T cell-derived lysosomes induced a cell-in-cell formation, suggesting this it is mediated by T cell-restricted molecules (Fig 3d-3f).
To shed light on the nature of these molecules, we isolated the lysosomes from the supernatants of activated T cells and subjected them to biochemical manipulations before adding them to a B16F10 culture. Boiling the lysosomes at 95 o C for 5 min completely abrogated their capacity to induce a cell-in-cell formation, but neither incubation at 60 o C for an hour nor freeze-thaw cycles abrogated this capacity, suggesting that these are globular proteins and not enzymes (Fig. 3g, Extended Data Fig. 3b). Size exclusion assays indicated that these proteins are larger than 30kDa, smaller than 100kDa, and are not digested by trypsin (Fig. 3g, Extended Data Fig. 3b). Nonetheless, incubation of tumor cells with gp100-reactive T cells in the presence of blocking antibodies against lysosome-associate proteins such as LAMP1, CD95, or CD63 did not alter their capacity to induce a cell-in-cell formation, suggesting that these molecules are also expressed on T-cell membranes (Fig. 3h). To elucidate the type of activations that induce these molecules on T cells, we incubated naïve splenic T cells overnight with various stimulators, washed them, and incubated them with tumor cells. We found that T cells activated with IFN, and to a lesser extent also anti-CD3 plus high-dose IL-2, induce the most cell-in-cell formation (  Table-4). Of these genes, only about 30 proteins are expressed on the cell membrane, and most of them are shared with lysosome membranes (Fig. 3l). We next tested if other IFN-activated cells can induce a cell-in-cell formation, or whether it is restricted to activated T cells. We found that IFN-stimulated NK cells, but not macrophages or B cells, could induce a tumor cell-in-cell formation, albeit at a lower percentage compared to activated T cells (Fig. 3m).

Cell-in-cell tumor formation is governed by STAT3 signalling
We then set out to understand how tumor cells form cell-in-cell structures and what mechanisms govern this structure. Initially, we tested whether it is cell fusion. For that, we incubated tumor cells whose cytosols are labelled with either GFP or mCherry with tumor-reactive T cells.
Confocal analysis indicated that each cell type in the cell-in-cell formation maintained its cytoplasm, and no mix between colours was detected (Fig. 4a). Furthermore, long-term followup of cell dissemination from this structure indicated that each cell maintains its initial single labelling colour (Extended Data Movie-4). To corroborate this, we also incubated tumor cells whose membrane, nucleus, and F-actin were labelled with different fluorophores. Similarly, each cell in this formation was separated and maintained the integrity of its original cell components (Fig 4b). We next tested the possibility that this formation is entosis by staining for molecules reported to mediate that structure. However, we observed no increase or changes in the cellular localization of phosphorylated  catenin, E-cadherin, and phosphorylated integrin , which are all important mediators of entosis 25 (Fig. 4c-4e and Extended data Fig. 4a-4d).
Furthermore, we observed no reduction in T cell-mediated cell-in-cell formation upon blocking of E-and N-cadherins, actin filaments disruption, or inhibition of Wnt signalling ( Fig. 4f-4g).
In sharp contrast, the blocking of mRNA synthesis or protein production completely abrogated tumor cells' capacity to form a cell-in-cell formation, suggesting that the structure requires de

Discussion
Cancer immunoediting has been the primary paradigm explaining how tumors escape T-cell immunity and the framework for most recent immunotherapies 16,26 . In support of this view, fibrosarcoma clones emerged in T cell-deficient mice and were deleted upon injection to immunocompetent mice 27,28 . Immunoediting is further supported by clinical observations in which tumors relapse by omitting their expressed antigen [29][30][31] . While it is clear that immunogenic clones can be eliminated by T cells, the presence of immunogenic clones expressing neoantigens despite having infiltrated reactive T cells has been well documented 4,32,33 . Consistently, the persistence of TAA-reactive T cells correlates with excellent prognosis in responsive individuals 34,35 , indicating that immunogenic clones do not change so frequently. Furthermore, several seminal studies demonstrated that TCR-MHC is promiscuous, allowing each T cell to recognize several up to hundreds of different peptides [36][37][38] . As a result, editing a given antigen may not be sufficient to prevent T-cell recognition; or alternately, the antigen may be recognised by other T cells.
Whereas a suppressive microenvironment provides some explanation regarding how immunogenic clones can survive T-cell pressure [39][40][41] , why tumors reoccur in patients undergoing complete response remains unknown. Here, we found that tumor cells escape Tcell-based immunotherapies by forming a transient cell-in-cell structure, and not by immunoediting. These structures generate multiple membrane layers, thus reducing the probability of compounds' internalization, hence leading to increased tumor-cell survival.
Several cell-in-cell formations have been described before including phagocytosis, cell cannibalism, and entosis 42 . Recently, these structures have been shown to enable tumor-cell survival after chemotherapy treatment by engulfing and digesting neighbouring cells 43 . Despite structural similarity to entosis, the cell-in-cell structure described in this study differs markedly in several key elements: First, it is triggered predominantly by reactive T cells, rather than glucose and nutrient starvation 44 , mitosis 45 , or loss of adherence 25 . Additionally, it requires de novo synthesis of genes regulated by STAT3, and not through Rho-Rock and -cathenin signalling 25 . Most importantly, the structure described here is reversible and only rarely leads to cell death and apoptosis.
Overall, the ability of tumor cells to transiently enter and disseminate from each other in response to T-cell killing is a biological process that has never been described heretofore. It better explains how immunogenic tumors can survive in the host and provides a novel framework for immunotherapies.

Neoantigen predication
Genomic DNA was extracted from 5x10 6 B16F10 tumor cell lines using NucleoSpin® Tissue

Statistics
Significance was calculated using the nonparametric two-way ANOVA with Tukey's correction for multiple hypotheses. In some cases, a Bonferroni-Sidak post-test was performed after two-way ANOVA. For two-group analysis, one-way ANOVA with Dunn's test was performed. The results were analyzed by Prism (GraphPad Software, Inc.). All statistical analyses were performed in Prism (GraphPad Software, Inc.). All experiments were performed at least three times with at least three replications. Fig. 1: Relapsed tumor cells share neoantigens with primary tumors and are equally susceptible to killing by T cells. a. Tumor size following treatments with anti-CD40, TNF and tumor-binding antibodies (n=4). b. Mean percentages of T cells out of CD45 + cells in relapsed tumors (n=4). c. Mean percentage of TRP2-and gp100-reactive T-cell clones in relapsed tumors (n=4). d. Representative Flow cytometry of MHC-I and MHC-II expression in relapsed B16F10 cells. e. Representative staining of TRP2 and gp100 in relapsed B16F10 cells. f. Illustration of neoantigen discovery pipeline. g. Neoantigen burden in B16F10 cells isolated from untreated tumors, or following immunotherapy treated tumors (IT). h. Allele frequency comparison of 25 neoantigens with the highest MHC-I affinity. i. RNAseq expression level of B16F10-known neoantigens. j. Tumor growth following treatments with anti-CD40, TNF and anti-TRP1 tumor-binding antibodies (n=4). k. Tumor size following adoptive transfer of gp100-reactive CD8 + or TRP1-reactive CD4 + T cell (n=4). l. Mean percentages of apoptotic tumor cells after incubation overnight with gp100-reactive CD8 + T cells (n=5). m. Tumor size following adoptive transfer of gp100-reactive T cell (n=4). n. Principal Component Analysis (PCA) of B16F10 tumor cell lines freshly isolated from treated mice. Experiments were repeated independently at least three times. Statistical significance was calculated using ANOVA with Tukey's correction for multiple comparisons (*** denotes p<0.001, **** denotes p<0.0001). Error bars represent standard error. Scale bars = 20 m. treated with anti-CD40, TNF and anti-TRP1 (Ab-IT), or treated with gp100-reactive (gp100 ACT) (n=4). e. Representative images of B16F10 labeled with H2B-tdTomato and H2B-GFP freshly isolated from tumor-bearing mice treated anti-CD40, TNF and anti-TRP1. f. Mean percentage of cell-in-cell formation following incubation of B16F10 cells with immune cells (n=3). g. Representative 3D projection and horizontal sections (Z-stack) of B16F10 cells incubated with gp100-reactive CD8 + T cells. h. Mean percentage of cell-in-cell and single cells in tumor-bearing NSG -/mice treated with anti-CD40, TNF and anti-TRP1 antibodies (n=3). i. Representative staining of gp100 expression in B16F10 tumor cells following incubation with gp100-reactive CD8 + T cells. j. Representative images of MHC-I expression in B16F10 tumor cells following incubation with gp100-reactive CD8 + T cells. k. SEM analysis of B16F10 incubated with gp100-reactive CD8 + T cells. l. Percentages of B16F10confluence tumor cells following incubation with gp100-and Trp2-reactive T cells (n=3). m. Mean percentage of cellin-cell in B16F10 culture following one (1 st ) or two (2 nd ) cycles of incubation with gp100reactive CD8 + T cells (n=3). n. Mean percentage of confluence change over time of B16F10 control (WT) or cell-in-cell tumors (n=3). Experiments were repeated independently at least three times. Statistical significance was calculated using ANOVA with Tukey's correction for multiple comparisons (*** denotes p<0.001, **** denotes p<0.0001). Error bars represent standard error. Error bars represent standard error. Scale bars = 20 m (a, c, I, j), 5 m (b left, k left), 500 nm (b right, k right). e. Representative images of B16F10 incubated overnight with latex beads, B16F10 and Melan A melanosomes, NK-and T cell-derived lysosomes. f. SEM analysis of B16F10 incubated with T cells-derived lysosomes (arrowhead). g. Mean percentage of cell-in-cell formation after incubation with modified T-cell lysosomes (n=3). h. Mean percentage of cell-in-cell formation after incubation with T cells and lysosomes-blocking antibodies (n=3). i. Mean percentage of cell-in-cell formation following incubation with pre-activated T cells (n=3). j. PCA of expression profiles of naïve and activated CD8 + T cells k. Heatmap of gene expression patterns of naïve and activated CD8 + T cells. l. Venn diagram of membrane-and lysosome-associated genes up-regulated in CD8 + T cells stimulated with IFN. m. Mean percentage of cell-in-cell formation following overnight incubation with IFN-activated immune cells. Experiments were repeated independently at least three times. Statistical significance was calculated using ANOVA with Tukey's correction for multiple comparisons (*** denotes p<0.001, **** denotes p<0.0001). Error bars represent standard. Scale bars = 5 m (a-c top, f top), 500 nm (a-c bottom, f bottom), 20 m (e).

Fig. 4: Cell-in-cell tumor formation is governed by STAT3 signalling and EGF receptors a.
Representative images of cytosol-labeled B16F10 cell lines incubated overnight with gp100reactive T cells. b. Representative images of B16F10 cell lines labeled with H2B-GFP and MyrPalm-TdTomato, and B16F10 cell lines labeled with H2B-TdTomato and lifeactin-GFP, following overnight incubation with gp100-reactive T cells. c. Representative images of -catenin expression in B16F10 cells incubated with gp100-reactive T cells. d. Representative images of E-cadherin expression in B16F10 incubated with gp100-reactive T cells. e.

Representative images of phosphorylated integrin-1 expression on B16F10 cells incubated with activated T cells. f. Mean percentage of cell-in-cell of B16F10 cells treated with different inhibitors and incubated with T cells (n=3). g. Representative images of B16F10 cells treated with inhibitors and incubated overnight with gp100-reactive T cells.
h. Relative expression of highly enriched genes in B16F10 cells isolated directly from relapsed tumors (Tumor) and after incubation with T-cell lysosomes (Lyso), compared to B16F10 control cells (WT) (n=3). i. Representative images of phospho-STAT3 in B16F10 incubated with gp100-reactive T cells. j. Representative images of phosphorylated EGFR of B16F10 incubated with gp100-reactive T cells k. Mean percentages of cell-in-cell formation following overnight incubation with T cellderived STAT3 or EGFR inhibitors (n=3). l. Representative confocal images of B16F10 cells incubated with STAT3 or EGFR inhibitors and with T cell-derived lysosomes. Experiments were repeated independently at least three times. Statistical significance was calculated using ANOVA with Tukey's correction for multiple comparisons (*** denotes p<0.001, **** denotes p<0.0001). Error bars represent standard error. Scale bars = 20 m.

Extended Figures
Extended Data Fig. 1

Tumor models
For melanoma tumor studies, 2.5 × 10 5 B16F10 cells suspended in 50 L DMEM were injected s.c. into C57BL/6 mice above the right flank. For 4T1 triple-negative breast cancer model, 2 × 10 5 4T1 cells in 30 L DMEM were injected into mammary fat pad number five fat pad number five. Tumor size was measured twice a week using calipers. Treatment was applied at different days post-injection. When tumors reached 120 mm 2 , the mice were sacrificed due to ethical considerations.
Macrophages were isolated from peritoneal cavity of euthanized mice by washing and recollecting 5ml of HBSS twice, followed by centrifugation at 600 rcf.
Serum was acquired by pulling blood from venae cava of euthanized mice. Blood was incubated on ice for 30 minutes and centrifugation of at 400 rcf for 10 minutes. Serum was collected and recentrifuged at 21,000 rfc for 10 additional minutes.

Tumor cell lentiviral transduction
Afterwards, 80% of medium was replaced and T cells were cultured for an additional three days in T cell media containing 1,000 IU IL-2. Transduction efficacy was assessed by FACS as the percentages of GFP-expressing cells.
Adoptive T-cells Transfer T cells were isolated from spleens of naïve mice as specified above and infected with pMIGII encoding TCR recognizing MHCI-gp10025-33 46

mRNA extraction and RNAseq analyses
To extract mRNA from tumor cells, 5x10 6 B16F10 sorted from in vivo relapsed tumors, or from cell B16F10 cell lines established from relapsed tumors were collected. For preparation cell-incell culture induced by T cells-derived lysosomes, 1x10 6 B16F10 cells were seeded on a lowadherence 10 cm plastic plate pre-coated with 200 μg/mL poly-L-lysine and were let to adhere for three hours. Concentrated T cell-derived lysosomes in DMEM were added to the culture, For T cells RNAseq, reads were aligned using Kallisto 56 to mouse genome version mm10, followed by further processing using the Bioconductor package DESeq2 1.24.0 54 . The data was normalized using TMM normalization, and differentially expressed genes were defined using the differential expression pipeline on the raw counts with a single call to the function DESeq (FDR-adjusted P value <0.05). Heatmap figures were generated using pheatmap package 57 and clustered using Euclidian distance. Boxplots were generated using BoxPlotR.

Cell-derived lysosomes and melanosomes isolation
For isolation of T-cell-derived lysosomes, 2×10 7 CD8 + T cells were isolated from spleen and plated in 10 cm plate precoated with anti-CD3 antibodies and high-dose IL-2 (1000 IU/mL) for 48 hours in T cell media containing 10% FBS pre-centrifuged at 140,000 rpm for one h to deplete bovine-derived exosomes. Supernatants were collected and concentrated using Amicon® Ultra (Merk) with 100kD filter. For isolation of NK cells-derived lysosomes, splenic For isolation of secreted melanosomes from B16F10, Melan A, 8x10 6 cell were cultured in 10cm tissue culture plates and let to adhere overnight. Cells were incubated 48 h in DMEM supplemented with 10% exosome-free FBS. Supernatants was collected and centrifuged for 1 hour on glucose gradient, as previously described 58 .
Cell-in-cell inhibition assays

Cell-in-cell counts
For cell-in-cell count, 3,000 tumor cells per one cm 2 were plated in a 4 Chamber glass-bottom confocal plates (Cellvis) and let adhere for several hours. Cells were then incubated overnight to 48 hours with immune cells, or exosomes and counted under fluorescent microscope.
Percentages of cell-in-cells structures were calculated by counting 200 cells in three different chambers, three fields in each chamber.

Study approval
All animal protocols were approved by the Stanford University Institutional Animal Care and Use Committee under protocol: 01-16-095. The Tel Aviv University Institutional Review Board approved the human subject protocols, and informed consent was obtained from all subjects prior to participation in the study.

Acknowledgments
The authors would like to deeply thanks Rabin Medical Center Institutional Tissue Bank and especially Dr. Adva Levi-Barda, for their invaluable support of this research. We would also like to thank Dr. Vered Holdengreber from the Faculty of Life Sciences, Tel Aviv University and Dr. Gal Radovsky from Tel Aviv University Center for Nano Science and Nano Technology for their help with electron microscopy imaging.

Statistics
For time course experiments, significance was calculated using the nonparametric two-way ANOVA with Tukey's correction for multiple hypotheses. In some cases, Bonferroni-Sidak post-test was performed after two-way ANOVA. For two groups analysis, one-way ANOVA with Dunn's test was performed. The results were analyzed by Prism (GraphPad Software, Inc.). All statistical analyses were performed in Prism (GraphPad Software, Inc.). All experiments were performed at least 3 times with at least three replications.