Autocrine TGF-β1 drives tissue-specific differentiation and function of resident NK cells

Group 1 innate lymphoid cells (ILCs), comprising natural killer (NK) cells and ILC1s, mediate type 1 immune response through the production of IFN-γ and by directly eliminating infected and transformed cells (Spits et al., 2013). Both subsets share the expression of numerous surface receptors, including NK1.1 and NKp46, depend on the transcription factor T-bet, and employ a comparable array of effector functions (Vivier et al., 2018; Gordon et al., 2012). However, key differences among group 1 ILC subsets arise from their circulation and residency properties, which are defined by ontogeny and tissue niche (Meininger et al., 2020; Hernández-Torres and Stehle, 2022). For example, NK cells are continuously replenished by the bone marrow (BM) and patrol peripheral tissues via the bloodstream until they are recruited to a site of infection (Yokoyama et al., 2004). ILC1s, conversely, seed peripheral tissues early in life, where they persist and differentiate into resident sentinels with functional properties well-tailored to the tissue environment (Friedrich et al., 2021; Sparano et al., 2022). Together, circulating and resident group 1 ILCs provide broad and versatile protection against infections and cancer.

Early lineage decisions during hematopoiesis are thought to define the resident fate of ILC1s and the circulatory nature of NK cells (Klose et al., 2014; Constantinides et al., 2014; Gasteiger et al., 2015). ILC1s, however, are not the only tissue-resident subset among group 1 ILCs. Similar to T cells differentiating into resident memory T (T_RM) cells, NK cells can adapt to a state of tissue residency during inflammatory conditions such as cancer (Gao et al., 2017; Ducimetière et al., 2021; Dean et al., 2024) or infection (Park et al., 2019; Schuster et al., 2023; Torcellan et al., 2024). These tissue-resident NK (trNK) cells share a common set of transcriptional adaptations with ILC1s, yet display considerable heterogeneity depending on the tissue and pathological context (McFarland et al., 2021; Schuster et al., 2023; Torcellan et al., 2024). Thus, tissue residency among group 1 ILCs is also driven and regulated by local differentiation programs.

Interestingly, the salivary gland (SG) and uterus harbor an abundant subset of trNK cells even in the absence of inflammation (Tessmer et al., 2011; Sojka et al., 2014; Boulenouar et al., 2016; Cortez et al., 2016; Torcellan et al., 2024). Conceptually, these trNK cells are different from inflammation-induced trNK cells that require a pathological trigger to induce residency, like T_RM cells (Mueller and Mackay, 2016). Instead, trNK cells emerge during homeostasis, raising questions about the development and function of these cells. SG and uterine trNK cells are strictly dependent on TGF-β signaling (Cortez et al., 2016; McFarland et al., 2021), yet their origin is unknown as they do not derive from the Plzf-dependent ILC lineage (Constantinides et al., 2014; Erick et al., 2016) and only partially rely on common NK cell gene programs (Erick et al., 2016; Boulenouar et al., 2016; Cortez et al. 2014, 2016; Pikovskaya et al., 2016; Nixon et al., 2022). SG trNK cells further comprise cytotoxic subsets (McFarland et al., 2021; Yomogida et al., 2021; Nixon et al., 2022; Friedrich et al., 2021); however, whether they derive from local differentiation programs as seen in hepatic ILC1s (Friedrich et al., 2021) or represent a separate entity is currently unclear. While uterine trNK cells are essential in arterial remodeling during pregnancy (Ashkar et al., 2000; Boulenouar et al., 2016; Han et al., 2023), the function of SG trNK cells remains largely elusive. The SG is a major site for murine cytomegalovirus (MCMV) replication, and the limited capacity of group 1 ILCs to produce IFN-γ in this tissue has been suggested to contribute to MCMV persistence (Cortez et al., 2016). On the other hand, a subset of long-lived NK cells in the SG has been shown to prevent autoimmunity during chronic MCMV infection (Schuster et al., 2023) and may regulate LCMV-induced T_RM cells (Woyciechowski et al., 2020). Together, our limited knowledge of trNK cell development and their apparent context-dependent heterogeneity add further complexity to understanding their functional roles.

In the present study, we investigated programs driving NK cell residency during ontogeny, homeostasis, and infection. We uncovered that trNK cells in various glandular organs require continuous autocrine production of TGF-β1 for their generation and persistence. TGF-β1 imprints and maintains a major trNK cell population in the SG, which develops during early adolescence from immature NK cells. Later during development, TGF-β and IL-15 cooperatively induce a Hobit-dependent differentiation pathway toward a cytotoxic trNK cell population, which contributes to MCMV control.

Residency of group 1 ILCs is often attributed to early lineage decisions in BM progenitors, whereas factors that induce committed cells to establish residency are not well understood. Among those, TGF-β can profoundly influence group 1 ILC biology (Cortez et al., 2016, 2017; Gao et al., 2017; Viel et al., 2016), yet the sources and exact mechanisms are unclear. To identify the producers of TGF-β sensed by group 1 ILCs, we first analyzed publicly available single-cell RNA sequencing (scRNA-seq) datasets for Tgfb1 expression (Sparano et al., 2022; Kimmel et al., 2019; Yomogida et al., 2021). Surprisingly, we found that group 1 ILCs do not only express the receptor for TGF-β, Tgfbr2, but also the cytokine TGF-β1 itself in all tissues analyzed under homeostatic conditions (Fig. 1 A and Fig. S1 A). Since TGF-β production involves complex posttranslational regulation (Tzavlaki and Moustakas, 2020), we next measured the presentation of latency-associated peptide (LAP), which binds mature TGF-β until its release. Unstimulated splenic NK cells displayed little LAP on their surface; however, upon stimulation up to 10% of the cells were LAP⁺ (Fig. 1, B and C), suggesting that group 1 ILCs can produce their own TGF-β. Among NK cells, immature CD27⁺ NK cells displayed more LAP compared with mature CD11b⁺ cells and further upregulated its expression upon stimulation (Fig. 1 C). Interestingly, human peripheral blood NK cells, both CD56^Bright and CD56^Dim, also present LAP on their surface (Fig. 1 D and Fig. S1 B).

To understand the role of group 1 ILC–derived TGF-β1, we used Ncr1^CreTgfb1^fl mice, which lack TGF-β1 in NKp46⁺ group 1 ILCs, and Ncr1^wtTgfb1^fl littermate control mice, from now on referred to as “WT.” The resulting Ncr1^CreTgfb1^fl mice showed a disruption of the Tgfb1 open-reading frame specifically in group 1 ILCs (Fig. S1 C). In-depth analysis of the immune landscape in various organs revealed no major changes, indicating that group 1 ILC–derived TGF-β1 is dispensable for most immune cells in steady-state (Fig. S1 D). In contrast, we observed a significant reduction of group 1 ILCs in the SG (Fig. 1 E). Characterization of SG group 1 ILCs revealed that the subset of Eomes⁺CD49a⁺ trNK cells, which normally makes up around 65% of group 1 ILCs, was virtually absent in Ncr1^CreTgfb1^fl mice, whereas Eomes⁺CD49a⁻ conventional NK (cNK) cells were slightly increased (Fig. 1 F). This drastic shift among SG group 1 ILCs phenocopied the Ncr1^CreTgfbr2^fl model, in which group 1 ILCs lack the receptor for TGF-β (Cortez et al., 2016) (Fig. S1, E and F). To understand if the dependency on group 1 ILC–produced TGF-β1 was unique to the SG or a common feature of trNK cells, we screened additional organs. Strikingly, we discovered trNK cells in multiple tissues that contain or function as glands, including the uterus, pancreas, and choroid plexus (Fig. 1 G). Similar to the SG, these trNK cells depended on their own production of TGF-β1 (Fig. 1 H). Of note, ILC1s (Eomes⁻CD49a⁺) in the same tissues were also reduced in the absence of autocrine TGF-β1 production (Fig. 1 F and Fig. S1 G). In contrast, liver ILC1s of Ncr1^CreTgfb1^fl mice were unchanged, indicating no direct effect on the ILC1 lineage (Fig. S1 H). We observed no differences between WT and Ncr1^CreTgfb1^fl mice in terms of group 1 ILC abundance in the BM nor in the distribution of NK maturation stages (Chiossone et al., 2009) (Fig. S1, I and J). Thus, our results show that group 1 ILC–derived TGF-β1 is essential for the development of trNK cells in glandular tissues.

To understand if trNK cells continuously require autocrine TGF-β1 or only at a specific phase during development, we used Ncr1^CreERT2Tgfb1^fl mice, in which TGF-β1 can be ablated in group 1 ILCs upon tamoxifen administration (Fig. 1 I). 1 wk after a 3-day pulse of tamoxifen, SG trNK cells and ILC1s were significantly reduced, similar to the constitutive Ncr1^CreTgfb1^fl model, demonstrating a continuous requirement for TGF-β1 production (Fig. 1 J). As Ncr1-driven Cre systems target all group 1 ILC subsets, we next asked whether trNK cells have a cell-autonomous requirement for TGF-β1 or if more complex paracrine interactions between different group 1 ILC subset might be involved. To address this, we generated mixed BM chimeras using CD45.1⁺ WT BM combined with either CD45.2⁺ WT or CD45.2⁺Ncr1^CreTgfb1^fl BM (Fig. 1 K). Since group 1 ILCs from Ncr1^CreTgfb1^fl can still respond normally to TGF-β1 (Fig. S2 A), potential signals from TGF-β1–competent group 1 ILCs in the tissue might rescue TGF-β1–deficient cells. 6 wk after reconstitution, we detected donor-derived cNK cells in the SG, spleen, and liver with similar contributions from CD45.1 and CD45.2 donors regardless of their genotype (Fig. 1 L and Fig. S2 B). Similarly, rare donor-derived hepatic ILC1s were derived equally from CD45.2⁺ WT or CD45.2⁺Ncr1^CreTgfb1^fl BM, corroborating their independence on group 1 ILC–derived TGF-β1 (Fig. S2 B). In contrast, trNK cells and ILC1s in the SGs of mice receiving a mix of CD45.1⁺ WT and CD45.2⁺Ncr1^CreTgfb1^fl BM were almost exclusively of WT origin (Fig. 1, L and M), suggesting that TGF-β1 from paracrine sources is unable to rescue the development of these cells. In further support of primarily autocrine signaling, we did not observe a reduction of CD103⁺ T_RM cells in Ncr1^CreTgfb1^fl SGs, which share a similar niche and also rely on TGF-β (Fig. S2 C) (Mueller and Mackay, 2016). Altogether, these results define continuous autocrine TGF-β1 as the main mode of signaling required for the generation and maintenance of trNK cells.

ILC1s seed tissues in developmental windows whereas cNK cells are continuously replenished from the BM (Sparano et al., 2022). To understand the ontogeny of trNK cells, we analyzed SGs from WT and Ncr1^CreTgfb1^fl mice at different ages (Fig. 2, A–C). We found ILC1s and cNK cells as early as 1.5 wk after birth. Of note, early life cNK cells were immature and displayed low to intermediate CD49a expression as described previously (Sparano et al., 2022), which was however independent of group 1 ILC–derived TGF-β1 (Fig. 2 A and Fig. S2 D). The first TGF-β1–dependent trNK cells appeared between 1.5 and 3 wk after birth and expanded until adulthood (Fig. 2, A–C). Fate-mapping of SG group 1 ILCs using Id2^CreERT2R26R^EYFP mice (Sparano et al., 2022) further showed a drastic loss of EYFP⁺ trNK cells from 2 to 7 wk of age (Fig. 2 D). Id2 is expressed in ILCs, cNK, and trNK cells, as well as ILC progenitors (Klose et al., 2014) (Fig. S2 E). Thus, the loss of Id2-fate-label suggested that SG trNK cell expansion is unlikely to be driven by local proliferation of any of these populations. Instead, this indicated that recruitment of newly generated group 1 ILCs during the neonatal period establishes the SG trNK cell pool, which later remains resident and largely independent from the circulation (Cortez et al., 2016).

To better characterize the emergence of trNK cells, we performed scRNA-seq of SG group 1 ILCs at 3 wk of age (Fig. 2, E–J). Consistent with our flow cytometric analysis, we identified a cluster of ILC1s expressing Cd200r1, Il7r, Cxcr6, Tmem176b, a cluster of trNK cells marked by Itga1 (encoding CD49a) and Eomes co-expression, two clusters of cNK cells expressing Klra4, Sell, and Zeb2, and three clusters of proliferating cNK cells (Fig. 2, E and F; and Fig. S2 G). Regulatory network inference using SCENIC (Aibar et al., 2017) corroborated the distinct cluster identity, with ILC1s showing Maf, Ikzf2, Gata3, and Stat1/2 regulon activity; trNK cells showing Runx3 and Tbx21; and cNK cells Eomes, Irf8, Tcf4, and Egr1 (Fig. 2 G). To find potential developmental relationships between these clusters, we performed pseudotemporal trajectory analysis with RNAVelocity (Bergen et al., 2020), which revealed trNK cells as the main trajectory endpoint arising from proliferating and other cNK cells (Fig. 2 H). This endpoint also overlapped with the highest TGF-β pathway activity and imprinting score, corroborating the role of TGF-β1 in driving group 1 ILC differentiation in this organ (Fig. 2 I and Fig. S2 F). Interestingly, Tgfb1 and Tgfbr2 were upregulated in TGF-β1–imprinted cells, along with the α-integrin Itgav, which forms complexes with β-integrins Itgb1 and Itgb3 to activate TGF-β1 through cleavage of LAP (Tzavlaki and Moustakas, 2020) (Fig. 2, J and K; and Fig. S2 G). This suggests that autocrine TGF-β1 operates through a positive feedback loop, reinforcing and maintaining the differentiation of trNK cells. Our trajectory analysis further indicates that trNK cells emerge postnatally from immature cNK cells, which have a higher capacity to produce LAP (Fig. 1 C) and lack mature NK cell–associated genes such as S1pr5 (Nixon et al., 2022) (Fig. S2 G). Together, our data suggests that immature cNK cells contribute to the postnatal expansion of the trNK cell pool in the SG through a differentiation path that requires cell-intrinsic expression of TGF-β1.

To elucidate the physiological role of NK cell residency in the SG, we next characterized tissue localization, surface marker profile, and functional properties of group 1 ILCs in this organ. Immunofluorescence staining of NKp46⁺ SG group 1 ILCs showed most cells in close proximity to the EpCAM⁺ epithelium but distant from the CD31⁺ vasculature (Fig. S3 A). Consistently, all subsets of SG group 1 ILCs, including CD62L⁺Eomes⁺CD49a⁻ cNK cells, were unlabeled by intravenously injected anti-CD45 antibody already from 1.5 wk onwards, confirming their extravascular localization (Fig. S3, B and C). Interestingly, this was unchanged in Ncr1^CreTgfb1^fl mice (Fig. S3, A and B), suggesting that extravasation is not directly TGF-β dependent, and demonstrating that CD62L⁺Eomes⁺CD49a⁻ NK cells can enter the healthy gland tissue even in adult mice.

Next, we performed high-parametric flow cytometry immunophenotyping of SG group 1 ILCs (Fig. 3, A–C). Using dimensionality reduction and FlowSOM clustering, we identified six clusters, including the previously described clusters of CD200R⁺CD49a⁺IL18R⁺ helper-like ILC1s and CD200R⁺CD49a⁺GzmB⁺ cytotoxic ILC1s (Friedrich et al., 2021), and two clusters of mature (Eomes⁺CD49a⁻CD27⁺ CD11b⁻) and mature (Eomes⁺CD49a⁻CD27⁻CD11b⁺) cNK cells (Fig. 3, A–C). Consistent with their developmental relationship, Eomes⁺CD49a⁺ trNK cells appeared as a continuous gradient starting from immature cNK cells and were entirely missing in Ncr1^CreTgfb1^fl mice (Fig. 3, A–C). Of note, also cNK cell clusters were affected by the loss of TGF-β1, suggesting that the entire cNK–trNK spectrum in the SG is imprinted by TGF-β1 (Fig. 3 A and Fig. S3 D). Along the cNK–trNK gradient, we observed an early upregulation of CD49a, followed by other common TGF-β associated surface proteins such as CD103, and ATP-converting enzymes CD39 and CD73 (Mackay et al., 2013; Regateiro et al., 2011). Unexpectedly, we noticed that trNK cells with the highest expression of CD49a, CD103, CD39, and CD73 also displayed high levels of granzymes (Gzm) B and C (Fig. 3, A–E). We confirmed our flow cytometry findings using publicly available scRNA-seq data (Yomogida et al., 2021), which also showed a gradient from cNK cells toward CD103⁺ GzmB⁺ GzmC⁺ cytotoxic trNK cells (Fig. S3, E–G). Although reminiscent of cytotoxic ILC1s, which also lie at the extreme of a niche-specific differentiation spectrum (Friedrich et al., 2021), cytotoxic trNK cells were clearly distinguishable from ILC1s due to their expression of Eomes and lack of CD200R (Fig. 3 C; and Fig. S3, H and I). Finally, we could also detect GzmB⁺ cytotoxic trNK cells in the uterus, pancreas, and choroid plexus (Fig. 3, F–H), however not always co-expressing CD103⁺, indicating additional requirements beyond TGF-β1 for the expression of this integrin. Thus, autocrine TGF-β1 imprints the NK cell compartment and drives a differentiation spectrum which includes cytotoxic trNK cells.

To better understand the functional capabilities of SG group 1 ILCs, we next analyzed the expression of activating and inhibiting Ly49 receptors, which are stochastically expressed on NK cells and rarely found on ILC1s (Lassen et al., 2010). Intriguingly, the Ly49 receptor profile among cNK and trNK cells was almost identical and clearly different from ILC1s, consistent with the idea of an NK cell differentiation spectrum in the SG, which excludes ontogenically disparate ILC1s (Fig. 3 I). Further, the expression of the activating receptor NKG2D and GzmA was restricted to cNK cells and was downregulated or lost in trNK cells (Fig. S3 J), which conversely featured high amounts of granzymes B and C and perforin (Fig. 3, E and J). In addition to the generally restricted IFN-γ production reported for SG group 1 ILCs (Tessmer et al., 2011; Cortez et al., 2016), we found that the capacity to produce IFN-γ was further decreased from cNK to trNK cells with the lowest levels in cytotoxic trNK cells (Fig. 3 K). In contrast, the in vitro killing capacity of SG trNK cells against YAC-1 target cells was high and comparable with splenic cNK cells (Fig. 3 L). Therefore, TGF-β1–imprinted trNK cells have a diverse set of Ly49 receptors and are poor IFN-γ producers but retain a strong cytotoxic capacity.

We next sought to better understand how resident NK cells acquire the cytotoxicity program. High-parametric flow cytometry on SG group 1 ILCs from 1.5 wk of age until adulthood revealed that, in contrast to the adult, 1.5- and 3-wk-old SGs contained almost no cytotoxic trNK cells (Fig. 4, A–C and Fig. S4 A). The cytotoxic trNK cell population, clearly identified by high levels of CD49a, CD103, GzmB, and GzmC, emerged at around 5 wk of age and expanded until adulthood (Fig. 4, A–C and Fig. S4 A). GzmB⁺ group 1 ILCs had furthermore a flattened shape and were closely associated with EpCAM⁺ glandular epithelium, suggesting a specific niche for these cells (Fig. 4, D and E; and Fig. S4 B). Since a similar cytotoxic program in ILC1s is dependent on the transcription factor Hobit (Friedrich et al., 2021; Yomogida et al., 2021), we next used Hobit-TdTomato reporter mice (Behr et al., 2020) to measure Hobit expression in trNK cells. We found that cytotoxic trNK cells expressed Hobit, and the percentage of Hobit⁺ trNK cells in the SG increased in parallel to the emergence of the cytotoxic subset (Fig. 4, F and G). Furthermore, Hobit-deficient (Hobit^KO) mice had fully lost the cytotoxic population, demonstrating that Hobit controls the development of cytotoxic trNK cells (Friedrich et al., 2021) (Fig. 4 H).

To better understand which factors trigger the cytotoxic program in trNK cells, we sorted immature (CD27⁺) and mature (CD11b⁺) subsets of splenic cNK cells and stimulated them in vitro with TGF-β1 or IL-15 for either 24 h or 7 days. Of note, all conditions were supplemented with low levels of IL-15 (10 ng/ml) to ensure NK cell survival. Compared to stimulation with high levels of IL-15 (50 ng/ml), the stimulation of cNK cells with TGF-β1 (5 ng/ml) led to an upregulation of CD49a, CD73, and CD103 (Fig. 4 I and Fig. S4 C). Interestingly, induction of CD49a and CD73 by TGF-β1 after 24 h was more prominent on CD27⁺ cNK cells in comparison with the CD11b⁺ subset (Fig. 4 I and Fig. S4 C). Furthermore, while TGF-β1 induced the upregulation of CD103 in CD27⁺ cNK cells after 7 days, CD11b⁺ cNK cells showed almost no expression of this integrin (Fig. 4 I and Fig. S4 C), suggesting that immature cNK cells might be more prone to TGF-β1 imprinting compared with the mature subset (Viel et al., 2016). While this recapitulated most of our in vivo findings, TGF-β1 effectively suppressed GzmB and GzmC (Fig. 4 J) (Viel et al., 2016; Thomas and Massagué, 2005). Consistently, next to the loss of trNK cell numbers in Ncr1^CreERT2Tgfb1^fl mice upon tamoxifen administration (Fig. 1 K), absence of Tgfb1 led to decreased CD49a and CD103 expression but an upregulation of GzmB, which suggests that TGF-β1 is actively suppressing cytotoxicity also in vivo (Fig. 4, K and L; and Fig. S4 D).

To understand how a GzmB⁺GzmC⁺ cytotoxic population can emerge despite the continuous imprinting by TGF-β1, we searched for factors that might counterbalance the suppressive activity of TGF-β1. In addition to its essential role for NK cell survival, IL-15 is also a potent inducer of granzyme expression and cytotoxicity (Fehniger et al., 2007). Furthermore, IL-15 can induce Runx3 and T-bet (encoded by Tbx21) (Levanon et al., 2014), both of which are predicted to be active in trNK cells (Fig. 2 G). Indeed, after a 24-h pulse of TGF-β1, culturing NK cells with IL-15 could partly overcome the suppressive effects of TGF-β1 and induce a CD49a⁺GzmB⁺GzmC⁺ population, reminiscent of cytotoxic trNK cells found in vivo (Fig. 4, M and N; and Fig. S4, E and F). Without prior TGF-β stimulation, IL-15 did not induce the formation of this trNK-like cytotoxic subset (Fig. 4, M and N; and Fig. S4, E and F). Consistent with IL-15 serving as a counterbalance for the suppressive activity of TGF-β, in vivo treatment with IL-15/IL-15Rα complex led to an expansion of cytotoxic trNK cells in the SG (Fig. 4 O). Finally, to test if IL-15 and TGF-β drive the expression of Hobit, we isolated cNK cells from SGs of Hobit-TdTomato reporter mice and cultured them in vitro. While IL-15 alone only slightly influenced Hobit expression, the combination of IL-15 and TGF-β1 led to much higher Hobit expression (Fig. 4 P) (Yomogida et al., 2021). Overall, these data suggest that early imprinting by TGF-β1 in cooperation with IL-15 drives NK cells toward a Hobit-dependent program of residency and cytotoxicity.

After uncovering a crucial role of autocrine TGF-β1 in the steady-state development of trNK cells, we next asked whether a similar mechanism would be responsible for the emergence of trNK cells in a pathological setting. In the tumor microenvironment, TGF-β signaling is required for the conversion of NK cells into a subset of CD49⁺CD49b^intEomes⁺ trNK-like cells with reduced antitumor activity (Gao et al., 2017; Cortez et al., 2017; Ducimetière et al., 2021). Using a subcutaneous MC38 tumor model, we confirmed that trNK cells were only present in WT mice and not in Ncr1^CreTgfbr2^fl mice, consistent with TGF-β sensing being essential for the emergence of this population (Fig. S4, G–I). In contrast, Ncr1^CreTgfb1^fl mice contained similar amounts of trNK cells in the tumor compared with WT controls, concomitant with unchanged tumor growth and immune cell infiltration (Fig. S4, J and K). These results indicate that the intratumoral formation of trNK cells is not driven by the same intricate autocrine TGF-β1 program but is instead enforced by external TGF-β sources, highlighting context-dependent differences in trNK cell development.

The SG represents a main target for multiple viruses including CMV (Atyeo et al., 2024). To understand the relevance of TGF-β1–dependent resident group 1 ILCs in the context of this viral infection, we challenged WT mice with MCMV and analyzed SG group 1 ILCs at different days post infection (dpi), corresponding to MCMV appearance in the SG (8 dpi), infection peak (16 to 22 dpi), and latency establishment (50 dpi) (Fig. 5, A and B). To better simulate a natural viral infection, we used MCMV deficient for glycoprotein m157 (m157⁻), which is otherwise directly and very strongly recognized by Ly49H on cNK and trNK cells of C57BL/6 mice (Lisnić et al., 2015) (Fig. 3 I). At 8 dpi, we observed a strong immune response in the SG, as seen by the 10- to 15-fold increase of leukocytes which consisted primarily of recruited T cells (Zangger et al., 2021) (Fig. S5, A and B). Group 1 ILC numbers also increased by three to sixfold, with the largest expansion seen among trNK cells (Fig. 5, A and B). Already representing the largest subset in steady state, trNK cells at 8 dpi comprised 90% of group 1 ILCs (Fig. 5 B). At 16–22 dpi, trNK cell and ILC1 numbers were slowly decreasing, whereas cNK cells reached their peak expansion (Fig. 5, A and B). Finally, at 50 dpi, we found that all subsets except ILC1s returned to baseline. Strikingly, GzmB⁺ cytotoxic trNK cells accounted for the most expanded subset among group 1 ILCs (Fig. 5, C and D) and were found throughout the tissue (Fig. 5 E), suggesting a specific role for these cells in MCMV control.

We next investigated the importance of SG-resident group 1 ILCs during MCMV infection by comparing WT to Ncr1^CreTgfb1^fl mice. Despite elevated cNK cell numbers, trNK cells and ILC1s were almost absent in infected Ncr1^CreTgfb1^fl mice during both acute infection and later stages (Fig. 5, F and G; and Fig. S5 C). Importantly, expression of characteristic group 1 ILC surface markers remained stable throughout the infection, allowing consistent identification of the subsets (Fig. S5 D). Furthermore, we observed no major changes in NK cells or adaptive immune cells in central immune hubs such as the spleen, consistent with the notion that our Ncr1^CreTgfb1^fl model specifically targets resident group 1 ILCs (Fig. S5 E). Thus, during MCMV infection, the TGF-β1–dependent subsets of group 1 ILCs cannot be rescued by recruited cells nor expanded with TGF-β1 released by other cells. To understand which aspect of antiviral immunity might be affected by the absence of trNK cells and ILC1s, we next looked at IFN-γ producers and cytotoxic effectors in the infected SG. IFN-γ was mainly produced by infiltrating T cells and unchanged in Ncr1^CreTgfb1^fl mice, consistent with a limited capacity of SG group 1 ILCs to produce this cytokine (Fig. S5, F and G) (Cortez et al., 2016). Among group 1 ILCs, expansion of highly IFN-γ⁺ cNK cells fully compensated for the loss of trNK cells and ILC1s resulting in comparable numbers of IFN-γ⁺ group 1 ILCs (Fig. 5 H and Fig. S5 H). In contrast, the expansion of GzmB⁺ cytotoxic group 1 ILCs in the WT SG was severely impaired in Ncr1^CreTgfb1^fl mice, while other GzmB⁺ cells such as CD8 T cells were unchanged (Fig. 5, I and J; and Fig. S5 I). To investigate if this reduction had an impact on viral control in the SG, we measured MCMV genomes using quantitative PCR (qPCR). Central control of MCMV in the blood and spleen was unaffected (Fig. S5, J and K). Moreover, we found similar MCMV levels in SGs of WT and Ncr1^CreTgfb1^fl mice during the acute infection phase at 7 dpi (Fig. 5 K). However, at 16 and 28 dpi, when the virus had been cleared from most other organs, we observed significantly higher MCMV titers in the SG, demonstrating a critical role for SG-resident group 1 ILCs in local virus clearance (Fig. 5 K). Importantly, infection of TGF-β receptor 2–deficient Ncr1^CreTgfbr2^fl mice resulted in equally increased viral titers, indicating that the absence of resident group 1 ILCs, and not the production of TGF-β1 itself, is responsible for this increase (Fig. S5 L). In conclusion, these results show that cell-autonomous production of TGF-β1 is required for the differentiation and maintenance of cytotoxic trNK cells in the SG and for optimal local control of viral infection.

NK cell tissue residency can be induced both during homeostasis (Tessmer et al., 2011; Sojka et al., 2014; Boulenouar et al., 2016; Cortez et al., 2016; Erick et al., 2016; Flommersfeld et al., 2021) and in response to disease (Park et al., 2019; Schuster et al., 2023; Torcellan et al., 2024). Here, we show that trNK cells are present during steady-state in various glandular tissues, require continuous autocrine TGF-β1 signaling for their maintenance, and undergo a Hobit-dependent differentiation program towards cytotoxic effector cells, which are poised to protect the tissues from viral infections. Our study sheds light on the poorly understood development and function of trNK cells.

The term “tissue-resident NK cells” was historically used to describe what is now recognized as ILC1s, leading to inconsistent terminology in the field. Given the phenotypic and functional overlap between these subsets, particularly their shared expression of tissue residency markers and Hobit (Friedrich et al., 2021), a common term might seem justified. However, the discovery of ILC1s as a transcriptionally and ontogenetically distinct lineage (Klose et al., 2014; Constantinides et al., 2014; Sparano et al., 2022) and the recognition that circulating NK cells can establish tissue residency (Gao et al., 2017; Park et al., 2019; Ducimetière et al., 2021; Flommersfeld et al., 2021; Torcellan et al., 2024; Dean et al., 2024; Cortez et al., 2017) support classifying tissue-resident group 1 ILCs into two types: ILC1s, arising from distinct progenitors during ontogeny, and trNK cells, which comprise “ex-circulating” NK cells that acquired residency traits. While it is unclear whether trNK cells can convert into bonafide ILC1s, Ly49 receptor profiles, and gene signatures clearly distinguish these cells. Runx3, for example, appears to be a trNK cell-specific regulator that is not active in ILC1s (Levanon et al., 2014). Moreover, the conversion from ILC1s to cNK cells has not been observed, reinforcing their identity as a stable lineage (Friedrich et al., 2021; Yomogida et al., 2021; Nixon et al., 2022). Interestingly, CD27⁺ cNK cells with an “ILC1-like” phenotype can arise from ILCPs (Liang et al., 2024; Ding et al., 2024), raising the question of whether these cells are precursors of trNK cells. While this aligns with higher TGF-β production and responsiveness in CD27⁺ cNK cells, limited expression history of the ILCP signature gene Plzf in SG group 1 ILCs (Erick et al., 2016) and a diverse Ly49 repertoire present on trNK cells but not ILCP-derived cNK cells argue against this. Thus, a distinction between ILC1s and trNK cells acknowledges recent work showing cNK–trNK cell conversion in various contexts (Gao et al., 2017; Torcellan et al., 2024; Dean et al., 2024), while allowing future studies to clarify origin and interconvertibility of these two resident group 1 ILC subsets.

SGs act as reservoirs for viruses like CMV, mumps, and to a lesser extent EBV and influenza, facilitating both viral persistence and transmission (Atyeo et al., 2024). As part of the oral-pharyngeal immune system, SG immune cells must balance pathogen defense with tolerance to harmless food antigens, which may contribute to viral persistence (Wu et al., 2014; Atyeo et al., 2024). SG trNK cells, which localize near the epithelium, emerge around weaning age, coinciding with increased exposure to environmental pathogens due to dietary changes (Al Nabhani et al., 2019), supporting their protective role in this environment. Unlike resident group 1 ILCs in other tissues, which heavily rely on IFN-γ for protection (Klose et al., 2014; Weizman et al., 2017; Shannon et al., 2021; Torcellan et al., 2024), SG trNK cells display a stronger cytotoxic profile, reflecting the specific demands of the SG tissue. Interestingly, SG group 1 ILCs were also shown to limit overt CD4 T cell responses during MCMV infection, preventing tissue-damage and autoimmunity (Schuster et al., 2023). While it remains unclear if this tolerogenic function is dependent on TGF-β, our findings demonstrate that autocrine TGF-β1 directly suppresses IFN-γ and GzmB production in trNK cells. Furthermore, SG trNK cells express CD39 and CD73 (Cortez et al., 2016), which catalyze the formation of immunosuppressive adenosine (Regateiro et al., 2011). Hence, trNK cells seem to effectively maintain a balance between antiviral and tolerogenic activity.

The homeostatic emergence of trNK cells in the SG resembles that of T_RM cells, both of which rely on TGF-β (Mackay et al., 2013). Like trNK cells, skin T_RM cells also produce TGF-β in an autocrine fashion (Hirai et al., 2021), which promotes tissue retention via KLF2 inhibition and concomitant S1PR1 down- and CD103 upregulation (Skon et al., 2013; Mackay et al., 2013). TGF-β also represses Eomes and T-bet, reducing IFN-γ and granzyme levels (Thomas and Massagué, 2005; Mackay et al., 2015; Cortez et al., 2016). On the other hand, IL-15 counterbalances this imprinting in T_RM and trNK cells, preserving cytotoxic effector function through rescue of T-bet and induction of Hobit (Mackay et al., 2013, 2016; Kragten et al., 2018). Since Eomes actively suppresses Hobit (Parga-Vidal et al., 2021), TGF-β–mediated downregulation of Eomes may be a prerequisite of a Hobit-dependent cytotoxicity program in both T_RM and trNK cells (Kragten et al., 2018). These findings suggest that remarkably similar networks involving Eomes, T-bet, Hobit, and IL-15 coordinate tissue residency of both T and NK cells.

Given the difficulty of experimentally confirming tissue residency in humans, except through organ transplantation studies (Cuff et al., 2016; Strunz et al., 2021), much of the research on human trNK cells is based on studies in mice. Human trNK cells are CD56^bright, typically express CD69, CD103, and/or CXCR6 (Le et al., 2022), and are found in tissues like the lung, gut, liver, and uterus (Marquardt et al., 2019; Sagebiel et al., 2019; Dogra et al., 2020; Stegmann et al., 2016; Strunz et al., 2021). Importantly, human trNK cells not only show similar gene signatures with their murine counterparts (Sparano et al., 2022; Torcellan et al., 2024; Lopes et al., 2022) but, as we show here, are also capable of producing TGF-β. Thus, further experimental research will likely deepen our understanding of human NK cell residency and provide valuable insights for strategies aimed at enhancing their presence, particularly in the tumor microenvironment.

Female and male 6–10-wk-old C57BL/6 mice were purchased from Janvier Labs. R26R^Ai14 (B6;129S6-Gt(ROSA)26Sor^{tm14(CAG-tdTomato)Hze}/J, stock# 007908), Id2^CreERT2 (B6.129S(Cg)-Id2^{tm1.1(cre/ERT2)Blh}/ZhuJ, stock #016222), R26R^YFP (B6.129X1-Gt(ROSA)26Sor^tm1(EYFP)Cos/J, stock #006148), CD45.1 (B6;SJL-Ptprc^aPepc^b/BoyJ, stock# 002014), Tgfbr2^fl (B6;129-Tgfbr2^tm1Karl/J, stock# 012603), and Tgfb1^fl (C57BL/6J-Tgfb1^em2Lutzy/Mmjax, stock# 065809) were purchased from the Jackson Laboratory. Ncr1^Cre (B6.Cg-Ncr1^{tm1.1(icre)Viv}/Orl) mice (Narni-Mancinelli et al., 2011) were kindly provided by E. Vivier (Center of Immunology Marseille-Luminy), Ncr1^CreERT2 (B6;Ncr1^{tm1(icre/ERT2)Tfen}) (Wagner et al., 2020) were kindly provided by T. Fehniger (Washington University School of Medicine in St. Louis), and HobitKO (B6;129P-Znf683^tm3Gin/J) (Mackay et al., 2016) and Hobit-TdTomato (B6;Zfp683^{tdTomato-P2A-Cre-P2A-DTR}) mice (Behr et al., 2020) were kindly provided by K. van Gisbergen (University of Amsterdam). All mice were kept on a C57BL/6 background and housed in individually ventilated cages under specific-pathogen-free conditions according to institutional guidelines in the Laboratory Animal Services Center of the University of Zürich or at the ETH Phenomics center. All animal experiments were approved by the Swiss Cantonal Veterinary office.

Adult and neonatal mice were treated with tamoxifen (Sigma-Aldrich) reconstituted in corn oil (Sigma-Aldrich) with a final concentration of 25 mg/ml. 3 mg was administered via oral gavage to 8–10-wk-old mice on three consecutive days. Neonatal mice were injected intraperitoneally at postnatal day (P) 12 with 1.5 mg of tamoxifen.

6–8-wk-old female CD45.1.2 mice (B6.SJL-Ptprc^aPepc^b/BoyJ) were irradiated twice with 5.5 Gy. 6–8 h after irradiation, Tgfb1^fl (CD45.2) or Ncr1^CreTgfb1^fl (CD45.2) BM was combined with Tgfb1^WT CD45.1 BM (all female donors) and total 5 × 10⁶ cells were injected intravenously. For 2 wk after irradiation, mice were supplied with antibiotics in the drinking water. Chimerism was monitored in the blood by flow cytometry 6 wk after reconstitution.

MC38 colorectal carcinoma cells were grown in DMEM containing 1 mM GlutaMAX, 10% FBS, and 100 µg/ml penicillin/streptomycin (Thermo Fisher Scientific) at 37°C and 5% CO₂. For tumor inoculation, cells were detached using trypsin (Thermo Fisher Scientific) and resuspended in PBS/Matrigel (Corning) immediately before administration. Primary tumors were induced by subcutaneous injection of 500,000 MC38 tumor cells and harvested after 2 wk.

Recombinant MCMV lacking m157 (MCMV^Δm157) (Bubić et al., 2004) was propagated on primary mouse embryonic fibroblasts. 10–12-wk-old female mice were infected intravenously with 2 × 10⁵ PFU of MCMV. At the desired timepoint after infection, organs were collected and viral titers were determined by standard plaque assay or using qRT-PCR measuring M86 gDNA levels (Lemmermann et al., 2010).

Mice were euthanized via CO₂ inhalation and transcardially perfused with PBS. SG, liver, lung, spleen, uterus, and pancreas were collected in ice-cold PBS, cut into small pieces, and digested with collagenase IV (0.4–1 mg/ml; Sigma-Aldrich) and DNase (2 mg/ml) in HBSS (with Ca₂⁺⁺/Mg₂⁺⁺) supplemented with 5% FBS (Sigma-Aldrich). Choroid plexus was removed using a dissection microscope. All organs were digested under gentle rocking at 37°C for 20–40 min. After digestion, cells were passed through a 100-μm filter and washed, followed by 3-min incubation in ACK (ammonium–chloride–potassium, Sigma-Aldrich) buffer to lyse erythrocytes. For harvesting of BM cells, tibiae and femora were carefully separated from surrounding muscle and connective tissue, crushed in a mortar, and passed through a 100 μm filter, followed by lysis of erythrocytes. For measurement of specific surface markers (e.g., Ly49 receptors) tissues were only mechanically dissociated and then directly filtered to avoid antigen cleavage by collagenases. Human PBMCs were collected from healthy volunteers.

Single cells were incubated for 15 min in FcReceptor-blocking buffer (anti-CD16/32 in PBS; BioLegend) prior to staining. Cells were washed with PBS and incubated in antibody staining mix for 45 min at 4°C. In some cases, biotinylated antibodies for lineage (CD3ε, CD5, TCRβ, CD19, Ly6G, and Ter119) (Table S1) exclusion were added to the staining mix. Detection of the lineage with a streptavidin-fluorophore conjugate was performed in a separate step after washing with PBS. For intracellular and intranuclear staining, cells were fixed and permeabilized using the Foxp3 transcription factor staining buffer kit (Thermo Fisher Scientific) or using the Cytofix/Cytoperm (BD) kit, washed two times with PermWash (0.01% sodium azide, 0.5% saponin and 2% BSA in PBS), and then incubated with desired antibodies in PermWash overnight. Antibodies were purchased from BioLegend, BD or Thermo Fisher Scientific (Table S1). For intracellular cytokine staining, cells were stimulated with 500 ng/ml ionomycin (Sigma-Aldrich) and 50 ng/ml PMA (Sigma-Aldrich), or IL-12 (10 ng/ml; Preprotech) and IL-18 (100 ng/ml; Preprotech) for 4 h. GolgiStop (BD) and GolgiPlug (BD) were added 1 h after the start. For extracellular LAP staining, cells were incubated overnight (mouse) or for 48 h (human PBMCs; with 100 u/ml IL-2) without GolgiStop (BD) and GolgiPlug (BD). Stained cells were acquired on a Cytek Aurora spectral analyzer (Cytek Biosciences) and analyzed with FlowJo (BD) and R. Dead cells and doublets were excluded for analysis using SSC-A/H, FSC-A/H, and a Fixable Viability Kit (LIVE/DEAD Blue, Thermo Fisher Scientific).

Single-cell suspensions were prepared as described above but under sterile conditions. Single cells were stained with biotinylated antibodies against lineage (CD3ε, CD5, CD19, Ly6G, Ter119, and MHC-II; Table S1) for 20 min, followed by a washing step and negative selection of Lin⁺ cells using Streptavidin Nanobeads (BioLegend). Cells were stained with desired antibodies and then kept in HBSS without Mg₂ and Cl₂ (Gibco) supplemented with 30% FBS at 4°C before sorting. Sorting was performed using a FACSymphony S6 cell sorter (BD) or FACSAria III (BD). Purity of sorted cells was assessed regularly and always exceeded 98%.

Sorted splenic NK cells (Lin⁻NK1.1⁺CD49a⁻CD49b⁺) or SG NK cells (Lin⁻NK1.1⁺CD49a⁻CD62L⁺) were cultured in 96-well plates at a density of 50,000 cells per well. Cells were kept in RPMI, containing 1 mM GlutaMAX, 25 mM HEPES, 1 mM sodium pyruvate, 1% non-essential amino acids, 10% FBS and 100 µg/ml penicillin/streptomycin (Thermo Fisher Scientific) at 37°C and 5% CO₂. Cytokine concentrations were for IL-15 (recombinant murine; Preprotech) 10 ng/ml (minimal) and 50 ng/ml (high) and for TGF-β (recombinant human; Preprotech) 5 ng/ml. For long cultures, medium with cytokines was replaced every 48 h.

Sorted splenic NK cells (Lin⁻NK1.1⁺CD49a⁻CD49b⁺) were mixed with 5,000 YAC-1 cells at 1:1–10:1 ratio and kept in V-bottom 96-well plates for 6 h at 37°C and 5% CO₂, followed by flow cytometric analysis. Specific killing was calculated as $Experimental lysis-spontaneous lysis/maximum lysis-spontaneous lysis×100%$⁠.

As previously described (Dean et al., 2024), IL-15 (#210-15; Peprotech) and IL-15Rα (#551-MR-100; R&D) were reconstituted in PBS at a ratio of 1:5 and incubated at 37°C for 30 min to create IL-15/IL-15Rα complexes. IL-15/IL-15Rα complexes (2.5 μg IL-15:12.5 μg IL-15Rα) were administered twice intraperitoneally (48 h apart) followed by tissue harvest after another 48 h.

Tissues were harvested, immediately fixed with 4% paraformaldehyde for 4 h, subsequently dehydrated by incubation in 30% sucrose for 48–72 h, and snap-frozen in OCT for long term preservation. Samples were sectioned using a cryostat (Leica) to generate 10–16 µm thick slices, which were washed with PBS and incubated overnight at 4°C with primary antibodies (Table S2) in PBS with 10% FBS and 0.5% Triton-X (Sigma-Aldrich). The next day, slides were washed with PBS and stained with secondary antibodies (Table S2) and DAPI for 2 h at 4°C, then washed again and mounted. Images were taken on a Stellaris 5 (Leica) or a Z1 Slidescanner (Zeiss) and processed with Imaris (version 10.1) or ZEN (blue edition, version 3.1). Quantification of cell shape and GzmB levels was performed blinded, counting per mouse at least 50 cells from five different sections.

Analysis of high-parametric flow cytometry data was performed as described previously (Ingelfinger et al., 2021). In brief, raw data was preprocessed using FlowJo. Compensated live CD45⁺ Lin⁻ NK1.1⁺ NKp46⁺ cells were exported from FlowJo and imported into R. After arcsinh transformation and percentile normalization to values between 0 and 1, dimensionality reduction and visualization was achieved with the Uniform Manifold Approximation and Projection (UMAP) algorithm (McInnes et al., 2018). Unsupervised clustering was performed using the FlowSOM algorithm (Van Gassen et al., 2015) with a k-value resulting in a biologically meaningful number of clusters. Plots and visualizations were created using the ggplot2 package.

Live CD45⁺ cells from SGs of male and female 3-wk-old mice were sorted into PBS containing 0.05% BSA for the 10X Genomics protocol. Viability of the sorted cells was evaluated with trypan blue and was over 95% for all samples. Sorted cells were loaded into 10X Genomics Chromium in parallel. Libraries were prepared following manufacturer’s protocol (Chromium Next GEM Single Cell 3ʹ Reagent Kits v3.1 protocol) and sequenced on an Illumina NovaSeq sequencer according to 10X Genomics recommendations to a depth of around 50,000 reads per cell. Initial processing was done using Cell Ranger (v7.0.1) mkfastq and count (reads were aligned to GENCODE reference build GRCm38.p6 Release M23). Starting from the filtered gene-cell count matrices produced by CellRranger’s built-in cell calling algorithms, we proceeded with Seurat v4 workflow (Hao et al., 2021).

Analysis, including quality control (QC), processing, graph-based clustering, visualizations, and differential gene expression analyses of the scRNA-seq data were performed in R using the Seurat pipeline (version 4.0) (Hao et al., 2021). Potential low-quality cells with number of features below 500, above 5,000, or with >5% mitochondrial genes were filtered out using the QC parameters of nFeature_RNA and percent.mt. Following QC, global count normalization was performed, followed by selection of variable features (n = 2,000), linear transformation using ScaleData, and dimensionality reduction. Graph-based clustering and dimensionality reduction were performed using the first 12 principal components and a resolution value of 0.6. Cell type assignment and selection of group 1 ILCs were based on cluster-specific genes identified with the FindAllMarkers function (only.pos = TRUE, min.pct = 0.25, logfc.threshold = 0.25). For adult SG group 1 ILCs, the dataset of Yomogida et al. (2021) was used. Following normal processing, two clusters which likely represented doublets and stressed cells were excluded. TGF-β imprinting was measured using AddModuleScore with a list of TGF-β response genes (Plasari et al., 2009) and TGF-β pathway activity using PROGENy (Schubert et al., 2018). Single-cell regulatory network inference and clustering (SCENIC) on P20 group 1 ILCs was performed using the pySCENIC workflow with standard parameters (Aibar et al., 2017).

Annotations of unspliced/spliced reads were obtained using velocyto with default parameters (Bergen et al., 2020). Reads were aligned to GENCODE reference build GRCm38.p6 Release M23. Counts of unspliced reads were merged with the spliced count matrix of our fully processed Seurat object followed by RNAVelocity analysis employing scVelo workflow (Bergen et al., 2020). Briefly, before running the dynamical model, moments for velocity estimation were computed with n_pcs = 20 and n_neighbors = 30. The dynamical model was employed to learn the full transcriptional dynamics of splicing kinetics, transcriptional state, and cell-internal latent time across the complete dataset.

All experiments were performed using randomly assigned, sex-matched mice without blinding of investigators. All data points and n reflect biological replicates. No data were excluded. Statistical significance was calculated with one-way ANOVA with Holm–Šídák multiple comparison correction or unpaired two-tailed t tests: P values <0.05 were considered significant. *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001; ns, not statistically significant. Statistical analysis was performed using GraphPad statistical software (GraphPad Software Inc.). Error bars show mean ± SD unless otherwise indicated.

Fig. S1 shows additional data relevant to Fig. 1, including targeting of Tgfb1 in Ncr1^CreTgfb1^fl mice, its effects on the general immune composition and specifically on group 1 ILCs. Fig. S2 provides additional data for Fig. 1, showing the effects of stimulating NK cells from Ncr1^CreTgfb1^fl and Ncr1^CreTgfbr2^fl mice with TGF-β in vitro. Moreover, Fig. S2 shows Id2^CreERTR26R^EYFP fate-mapping data of SG group 1 ILCs and scRNA-seq data relevant for Fig. 2. Fig. S3 contains immunofluorescence images, flow cytometry data, and scRNA-seq data of adult SG group 1 ILCs showing their extracellular localization and distinct expression profile. Fig. S4 shows additional protein expression, IF images, and quantifications for Fig. 4, together with results from MC38 tumor challenge in Ncr1^CreTgfb1^fl and Ncr1^CreTgfbr2^fl mice. Fig. S5 contains supporting data for MCMV experiments in Fig. 5, including characterization of SG group 1 ILC during infection, IFN-γ producers, and viral loads in spleen and blood. Table S1 contains all antibodies used for flow cytometry and Table S2 contains all antibodies used for histology.

The sequencing data that were generated for this study have been deposited in Gene Expression Omnibus under the accession number GSE282148. The remaining data are presented here or available from the corresponding author upon reasonable request.

We thank T. Fehniger (Washington University School of Medicine in St. Louis, MO, USA), K.P.J.M. van Gisbergen (University of Amsterdam, Netherlands), and E. Vivier (Aix Marseille University, CNRS, INSERM, CIML, France) for providing mice. We further thank the Functional Genomics Center Zurich (University of Zurich) for technical support, and M. Kopf (ETH Zurich), M. Mayoux (University of Zurich), and A. Fonseca for intellectual input and technical support.

This work was supported by grants from the Swiss National Science Foundation (PR00P3_179775) to S. Tugues, the Swiss Cancer Research Foundation (KFS-5420-02-2021) to S. Tugues, the Universität Zürich Candoc (K-41311-01-01) to C. Sparano, the Sassella Foundation to C. Sparano and S. Tugues, the Vontobel Foundation to S. Tugues, the Novartis Foundation to S. Tugues, the Olga Mayenfisch Foundation to S. Tugues, and the Wilhelm Sander Foundation to S. Tugues.

Author contributions: C. Sparano: Conceptualization, Data curation, Formal analysis, Investigation, Project administration, Visualization, Writing - original draft, Writing - review & editing, D. Solís-Sayago: Conceptualization, Data curation, Formal analysis, Investigation, Project administration, Visualization, Writing - original draft, Writing - review & editing, N.S. Zangger: Data curation, Formal analysis, Investigation, Resources, L. Rindlisbacher: Investigation, Writing - review & editing, H. Van Hove: Investigation, M. Vermeer: Investigation, Writing - review & editing, F. Westermann: Investigation, Resources, C. Mussak: Investigation, E. Rallo: Investigation, S. Dergun: Investigation, G. Litscher: Investigation, Writing - review & editing, Y. Xu: Investigation, M. Bijnen: Investigation, Methodology, Resources, Writing - review & editing, C. Friedrich: Writing - review & editing, M. Greter: Conceptualization, V. Juranić Lisnić: Methodology, B. Becher: Funding acquisition, Resources, Supervision, Writing - review & editing, G. Gasteiger: Conceptualization, Resources, Writing - review & editing, A. Oxenius: Resources, Supervision, S. Tugues: Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Writing - original draft, Writing - review & editing.