The transcription factor IRF8 drives tumor-specific exhaustion in CD8⁺ T cells

In chronic infection and cancer, persistent antigen exposure drives CD8⁺ T cells toward exhaustion, a dysfunctional differentiation state characterized by progressive loss of effector function, expression of multiple inhibitory receptor, and decreased proliferative capacity (Wherry, 2011). Recent advances have provided insights into the complex transcriptional networks regulating exhaustion in chronic infection and in cancer (Chen et al., 2026; Doering et al., 2012; Jansen et al., 2019; Philip et al., 2017; Sun and Dong, 2026; Wherry and Kurachi, 2015; Wu et al., 2026), with evidence for substantial differences between the two conditions due to their distinct microenvironments (Luxenburger et al., 2026; McLane et al., 2019; Philip and Schietinger, 2019; Tille et al., 2023). Among the central regulators of this state, several families of transcription factors (TFs) play a dominant function, including TOX (Alfei et al., 2019; Khan et al., 2019; Scott et al., 2019), NFAT (Martinez et al., 2015; Tille et al., 2023), and NR4A (Chen et al., 2019) families. Several members of the IFN regulatory factor (IRF) family have also been associated with CD8⁺ T cell exhaustion, including IRF2, IRF4, IRF5, and IRF7 (Kasmani et al., 2023; Lukhele et al., 2022; Mai et al., 2025; Man et al., 2017; Seo et al., 2021), raising interest in their role in CD8⁺ T cell differentiation fates. Through analysis of the inferred TF activity based on the analysis of regulatory programs of CD8⁺ T cells from multiple cancer and chronic infection studies, we identified IRF8 among the most active TFs in exhausted CD8⁺ T cells in a tumor-specific context. Despite studies showing its crucial role in myeloid cells (Aliberti et al., 2003; Grajales-Reyes et al., 2015; Kurotaki et al., 2014; Yanez et al., 2015), B cells (Carotta et al., 2014) and subsets of CD4⁺ T cells (Humblin et al., 2017; Ouyang et al., 2011), the role of IRF8 in CD8⁺ T cells is not fully understood, as different studies reached divergent conclusions. IRF8 was reported to drive effector differentiation of CD8⁺ T cells in a mouse model of graft-versus-host disease (Miyagawa et al., 2012) but also to restrain excessive CD8⁺ T cells proliferation during herpes simplex virus 1 infection (Sun et al., 2016). More recently, IRF8 was found to regulate the 3D chromatin architecture by linking promoters to distal enhancer regions, thereby improving gene transcription through interaction with the TF CTCF in exhausted CD8⁺ T cells. Based on the known function of IRF4 (Li et al., 2025), the close structural homology of IRF8 based on phylogenetic analysis of the IRF family (Antonczyk et al., 2019) raises the possibility that IRF8 plays a role in the regulation of differentiation of CD8⁺ T cells. Indeed, the two TFs share several binding partners, such as the AP-1 factors JUNB and BATF, PU.1, ETV6, and other IRFs, pointing toward similar and possibly redundant regulatory mechanisms for certain IRF family members.

In this study, we characterized the function of IRF8 in the regulation of CD8⁺ T cell exhaustion in tumors and chronic viral infection. We found that IRF8 expression is high in murine and human CD8⁺ tumor-infiltrating lymphocytes (TILs) and is driven by TCR in the tumor microenvironment (TME) but is repressed during chronic infection. The Irf8 locus was less accessible and less expressed in virus-specific CD8⁺ T cells from lymphocytic choriomeningitis virus clone 13 (LCMV cl13) infected animals compared with tumor-specific CD8⁺ T cells. Type I IFN, typically associated with chronic infections, can diminish IRF8 expression in CD8⁺ T cells. IRF8 KO in CD8⁺ T cells improved the anti-tumor response by dampening TOX expression and increasing the production of effector molecules, while IRF8 overexpression (OE) aggravated their exhausted phenotype. By comparing the phenotype and transcriptome of IRF8 KO with IRF2 and IRF4 KO CD8⁺ TILs, we showed convergence of the IRF family for the regulation of TOX-dependent T cell exhaustion. Altogether, we highlighted the role of IRF8 as a novel TF regulating tumor-specific CD8⁺ T cell exhaustion, which could unravel new therapeutic opportunities to better harness immunotherapies.

To identify TFs specifically active in exhausted CD8⁺ T cells, we inferred regulatory programs (regulons) from publicly available single-cell RNA sequencing (scRNA-seq) profiles from mouse CD8⁺ TILs using the SCENIC workflow (Aibar et al., 2017). We found that the regulon signature of IRF8 was among the top discriminant regulons for both progenitor exhausted (T_PEX) and terminally exhausted (T_EX) CD8⁺ T cells (Fig. 1, a and b). By analyzing mRNA expression levels from the same datasets, we observed a strong and specific expression of IRF8 in both T_PEX and T_EX compared with other IRF family members (Fig. 1 c), supporting a role for IRF8 in regulating CD8⁺ T cell exhaustion. To assess IRF8 protein expression in exhausted CD8⁺ T cells, we engrafted B16-gp33 melanoma cells by s.c. injection into C57/BL6 mice (B6), transferred with TCR transgenic CD8⁺ T cells that recognize the LCMV-derived gp_33–41 (gp33) epitope (P14 cells) and analyzed tumors after 14 days after tumor engraftment. TOX⁺ CD44^highPD-1⁺ CD8⁺ T cells showed significantly increased IRF8 levels compared with TOX⁻ CD44^highPD-1⁺ CD8⁺ T cells in the same tumor in both endogenous (host) CD8⁺ and tumor-specific P14 compartments, indicating that IRF8 is specifically expressed in exhausted CD8⁺ T cells (Fig. 1, d and e). We confirmed these results in B16F10 melanoma, murine EL-4 thymoma, LLC1, colon adenocarcinoma (MC38), and in an inducible orthotopic muscle invasive bladder cancer model (MIBC) (Leblond et al., 2020) (Fig. 1 f). We confirmed IRF8 OE in human exhausted CD8⁺ TILs, analyzing tumor samples from nine metastatic melanoma patients. IRF8 expression was significantly higher in TOX⁺PD-1^high compared with TOX⁻PD-1^high CD8⁺ T cells from the same tumor (Fig. 1, g and h), indicating that IRF8 expression is not restricted to murine CD8⁺ T cells. IRF8^high CD8⁺ T cells showed higher expression of exhaustion associated molecules, such as LAG3, PD-1, TIM3, and TOX, compared with IRF8^low CD8⁺ T cells, showing that IRF8 levels correlate with the exhaustion state in both murine and human TILs (Fig. 1, i–n). Altogether these results suggest that IRF8 is associated with CD8⁺ T cell exhaustion in cancer.

Compared with the expression of IRF8 in exhausted CD8⁺ TILs, scRNA-seq data showed that Irf8 mRNA levels were lower in exhausted CD8⁺ T cells responding to chronic LCMV cl13 infection in both T_EX and T_PEX (Fig. 2 a). Moreover, inferred IRF8 regulon activity was not specific to T_PEX and T_EX in this setting (Fig. S1 a). To compare IRF8 expression at the protein level between cancer and chronic infection, we adoptively transferred naive P14 cells into B6 mice 1 day prior to s.c. injection with B16-gp33 tumor cells. In parallel, we transferred naive P14 cells into TCRβ (Vβ5) transgenic mice 1 day prior to the infection with LCMV cl13, which causes chronic infection. Vβ5 mice display a reduced endogenous T cell response to LCMV and, similarly to CD4-depleted B6, have higher LCMV cl13 titers compared with B6 mice (Utzschneider et al., 2016; Utzschneider et al., 2013). 3 wk after tumor grafting or chronic infection, P14 cells (Fig. 2 b) displayed equivalent levels of TOX but expressed slightly higher PD-1 levels in the tumor compared with chronic infection, indicating the establishment of exhaustion across all conditions (Fig. 2, c and d). IRF8 was highly expressed in tumor-derived CD44^high PD-1⁺ P14 cells but was expressed at low levels in the corresponding chronic infection–derived P14 cells present in the spleen in both T_EX and T_PEX (Fig. 2, e–h; and Fig. S1, b and c). IRF8 was comparably low in spleens of tumor-bearing mice, confirming that IRF8 OE was restricted to tumor-associated CD8⁺ T cells (Fig. 2, e and f).

We also compared IRF8 expression in the naive P14 cells responding to tumors with that of in vitro preactivated P14 cells that were transferred into B6 mice 7 days after tumor engraftment (Fig. 2, b–h). The data confirmed our previous results, indicating that tumor-specific CD8⁺ TILs express IRF8 independently of whether priming occurred in vivo or in vitro. We further observed that IRF8 expression was significantly higher in tumor-derived SLAMF6⁻ TIM3⁺ T_EX compared with SLAMF6⁺ TIM3⁻ T_PEX cells, in line with Irf8 mRNA expression (Fig. 1 c). IRF8 was also higher in chronic infection–derived T_EX than T_PEX cells, although the overall expression was much lower than in tumors. Altogether, these observations suggest that IRF8 OE is associated with CD8⁺ T cell exhaustion specifically in the context of cancer.

We next investigated the regulation of IRF8 expression in CD8⁺ T cells. Previous studies have shown that IRF8 is induced by TCR engagement in CD8⁺ T cells (Miyagawa et al., 2012; Nelson et al., 1996; Sun et al., 2016). We reproduced these findings by activating CD8⁺ T cells with αCD3 and αCD28 antibodies in vitro, observing strong IRF8 upregulation during the first 72 h of activation (Fig. 3 a). To understand whether IRF8 expression was maintained during prolonged TCR engagement, we took advantage of an in vitro chronic TCR stimulation model in which CD8⁺ T cells undergo multiple rounds of restimulation with αCD3 antibodies (Fig. S2 a). Chronically stimulated cells gradually lost SLAMF6, upregulated multiple inhibitory receptors, such as PD-1, TIM3, and LAG3, and acquired granzyme B as well as TOX expression, suggesting the establishment of an exhausted-like phenotype (Fig. 3 b). Compared with resting CD8⁺ T cells (activated for 3 days and subsequently cultured with IL-7 and IL-15), chronically stimulated CD8⁺ T cells showed higher levels of IRF8, indicating that IRF8 expression was maintained during prolonged TCR stimulation (Fig. 3 c).

To test whether IRF8 expression was driven by TCR triggering in vivo, we co-transferred in vitro preactivated P14 and OT-I TCR-transgenic CD8⁺ T cells, recognizing the gp33 and the chicken OVA-derived OVA_257–264 epitope, respectively, into B6 mice bearing B16-gp33 and B16-OVA tumors on opposite flanks (Fig. 3 d). At the time of adoptive transfer, both P14 and OT-I cells expressed similarly high levels of CD44, CD25, PD-1, and IRF8 (Fig. 3 e). 7 days after transfer, B16-gp33 tumor associated P14 TILs (tumor-specific) displayed an exhausted phenotype, with higher expression of PD-1 and TOX compared with the non-tumor–specific OT-I TILs (Fig. 3 f). Conversely, B16-OVA tumor-associated OT-I TILs (tumor-specific) showed higher PD-1 and TOX expression compared with P14 TILs (Fig. 3 g). Despite similar IRF8 expression levels in P14 and OT-I cells at the moment of co-transfer, IRF8 OE was exclusively observed in tumor-specific CD8⁺ T cells, while bystander cells expressed low IRF8 levels (Fig. 3, f and g), similar to cells that had been rested in vitro (Fig. 3 c). Together, these results show that TCR stimulation is required to induce and maintain IRF8 expression in tumor-associated CD8⁺ T cells.

Since persistent TCR engagement is a shared core feature of cancer and chronic infection, we speculated that IRF8 was repressed during chronic infection. To test this hypothesis, we transferred naive P14 cells into Vβ5 mice that were then infected with LCMV cl13 or into B6 mice infected with LCMV Armstrong (Arm), which results in acute resolved infection, and collected cells 28 days later (Fig. 4 a). We stimulated the recovered P14 cells for 24 or 72 h with gp33 peptide to assess their capacity to upregulate IRF8 expression ex vivo. P14 cells from the acute resolved LCMV Arm infection highly upregulated IRF8 at 24 h (8.9-fold) and 72 h (6.7-fold) after restimulation. In comparison, P14 cells from chronically infected mice displayed considerably lower upregulation (1.7-fold), despite increased expression of CD44, PD-1, and TOX (Fig. 4 b and Fig. S2 b).

We next determined the accessibility of the Irf8 locus based on our scATAC-seq (single-cell assay for transposase-accessible chromatin with sequencing) from tumor-associated P14 TILs (Tille et al., 2023) with published scATAC-seq data from chronic infection–derived P14 cells (Giles et al., 2022). While the accessibility of the Tox locus was comparable (Fig. S2 d), several regions close to the Irf8 locus, which overlap with TF–binding regions determined thanks to mouse CHIPseq CD8⁺ T cell datasets from CHIP-Atlas (see Materials and methods), displayed decreased accessibility in chronic infection-derived P14 cells (Fig. 4 c). The reduced accessibility correlated with reduced H3K27ac deposition, explaining the substantially reduced Irf8 transcription in chronic infection–derived P14 cells (Fig. 4, c and d) (Ma et al., 2025). Interestingly, the scATAC-seq profile of P14 cells displayed lower accessibility of the Irf8 locus during the chronic phase of infection compared with acute phase of the infection (day 8). These data suggested that IRF8 expression underwent epigenetic repression during chronic infection (Fig. S2 c).

Several factors differ between chronic infection and cancer and may regulate IRF8 accessibility. Among them, increased inflammation with increased IFN signaling has been described during chronic infection (Teijaro et al., 2013; Wilson et al., 2013). We tested the effect of IFNγ, described to induce IRF8 concomitant to TCR engagement (Nelson et al., 1996), and of IFNβ as hallmark for chronic type I IFN exposure during chronic infections on IRF8 expression during the first 72 h after CD8 T cell activation. While IFNγ and IL-12 did not alter IRF8 levels, IFNβ strongly repressed IRF8 induction (Fig. S2, e and f), suggesting a negative regulation of IRF8 expression by type I IFN. To confirm the effect of IFNβ during chronic TCR stimulation, we used our in vitro chronic TCR stimulation assay and added IFNβ at different time points. We observed increased levels of IRF7, an IFN-responsive TF (Fig. S2 g). In contrast, IRF8 protein levels did not change when IFNβ was present form from day 0 to 3 but were drastically decreased when IFNβ was present after day 3 (e.g., day 0 to day 9, day 3 to day 9, or day 6 to day 9), showing that type I IFN repressed IRF8 during chronic TCR stimulation (Fig. 4 e). To assess the effect of type I IFN in vivo, we transduced P14 cells from CD4-Cre STOP^fl/fl-Cas9 mice (hereafter P14-Cas9) with a retroviral vector containing a sgRNA-targeting Ifnar1 (IFNAR1 KO) or a nontargeting sgRNA (control) (Fig. S3 a). IFNAR1 KO and control P14 cells were then co-transferred in a 1:1 ratio into B16-gp33-tumor bearing Rosa26-Cas9 knock-in mice (hereafter B6-Cas9) (Fig. 4 f). The use of B6-Cas9 mice as host prevents the rejection of P14-Cas9 cells due to the tolerance of these mice to the Cas9 protein. Mice were intratumorally injected with CpG to induce the production of type I IFN (Wang et al., 2016). IFNAR1 KO P14 cells showed significantly higher IRF8 levels compared with control P14 cells, while IRF7 levels were decreased in IFNAR1 KO P14 cells, confirming the CD8⁺ T cell–intrinsic suppression of IRF8 by type I IFN signaling (Fig. 2 g).

To investigate whether type I IFN impacts IRF8 expression during chronic infection, we injected αIFNAR1-blocking antibodies at different time points during the course of infection. While late blockade (from day 22 to 27) significantly increased IRF8 expression compared with isotype control antibody-treated mice, blockade at mid time point (from day 10 to 20) showed no significant difference, while early blockade decreased IRF8 expression (Fig. 4, i and j). The gradual decrease in IRF8 expression in control mice over time is consistent with our observation that the Irf8 locus is less accessible on day 30 compared with day 8 after infection, suggesting progressive IRF8 repression over the course of infection (Fig. 4, c, i, and j; and Fig. S2 c).

Taken together, these data show that IRF8 expression in CD8⁺ T cells is driven and maintained by TCR engagement. While it is maintained in the TME, it is repressed during chronic infection through epigenetic remodeling of the Irf8 locus. Our results suggest that type I IFN participates at least partially in the downregulation of IRF8 at later stages during chronic infection. However, additional mechanisms likely contribute to IRF8 repression in CD8⁺ T cells.

Since IRF8 expression strongly correlated with exhaustion of tumor-associated CD8⁺ T cells where we inferred specific transcriptional activity, we tested whether IRF8 contributed to CD8⁺ T cell exhaustion. We performed adoptive cell transfers (ACTs) of P14-Cas9 control or P14-Cas9 IRF8 KO cells into B6-Cas9 mice 7 days after B16-gp33 or MC38-gp33 tumor inoculation (Fig. 5 a). Mice injected with P14 IRF8 KO cells displayed a significant delay in tumor growth compared with mice that received P14 control cells in both tumor models tested (Fig. 5, b and c). To investigate the effect of IRF8 ablation in CD8⁺ TILs, we co-transferred P14 IRF8 KO and P14 control cells in a ratio 1:1 into B16-gp33–bearing B6-Cas9 mice (Fig. 5 d). Based on differential expression (DE) of CD45.1 and CD45.2, we could discriminate between endogenous CD8⁺ T cells (CD45.2⁺) and transferred P14-Cas9 cells transduced with IRF8-targeting or control sgRNA (CD45.1⁺ and CD45.1⁺CD45.2⁺). This strategy allowed a direct and unbiased comparison of the cell-intrinsic effects of IRF8 ablation in CD8⁺ T cells exposed to the same TME, controlling as well for tumor size effect. At the time of the adoptive transfer, IRF8 KO and control P14 cells displayed a similar activation profile based on the expression of CD44, CD25, and PD-1 (Fig. S3 c). Transduced cells were followed thanks to the expression of CD90.1. 7 days after co-transfer, IRF8 KO cells accumulated slightly less efficiently in the tumor compared with control cells, but not in the draining LN (DLN) (Fig. 5 e). When analyzing T_PEX (SLAMF6⁺ TIM-3⁻) versus T_EX (SLAMF6⁻ TIM-3⁺), we observed a weak but significant increase in the proportion IRF8 KO T_EX (Fig. 5 f). Nevertheless, IRF8 KO cells displayed significantly lower levels of TOX with no effect on PD-1 expression (Fig. 5 h), together with an increase in IFNγ and GrzB production upon ex vivo peptide restimulation compared with control cells from the same tumor (Fig. 5 g). We observed similar effects in the DLN, suggesting that IRF8 ablation alleviates exhaustion features in tumor-associated CD8⁺ T cells (Fig. 5 i).

To further investigate the capacity of IRF8 to regulate T cell exhaustion, we overexpressed IRF8 by transducing P14 cells with a retroviral vector encoding Irf8 (hereafter P14 IRF8 OE) or an empty control vector (Fig. S3 b). We co-transferred the cells in a 1:1 ratio into B6 mice 7 days after B16-gp33 engraftment and analyzed the tumor 7 days later (Fig. 5 j). IRF8 OE and control cells did not differ in terms of activation at the time of transfer (Fig. S3 d). OE of IRF8 did not change the accumulation of P14 cells in the tumor and the DLN (Fig. 5 k) but significantly impacted the proportions of T_PEX and T_EX with an increase in the T_EX population, suggesting that IRF8 promotes terminal exhaustion (Fig. 5 l). We observed a significant increase in TOX and PD-1 expression together with a decrease in IL-2, GrzB, and IFNγ production after restimulation of P14 IRF8 OE compared with P14 control cells, mirroring the results found with IRF8 KO CD8⁺ T cells (Fig. 5, d–i). Collectively, these results showed that IRF8 promoted exhaustion and reduced the effector functions of tumor-associated CD8⁺ cells. Indeed, ACT using IRF8 KO CD8⁺ T cells ameliorated tumor control.

Our results indicated that IRF8 is specifically upregulated in the tumor, while CD8⁺ T cells responding to chronic viral infection expressed low levels of IRF8. To examine the role of IRF8 in chronic viral infection, we co-transferred P14 IRF8 KO together with control transduced cells into Vβ5 mice and infected them with LCMV cl13 (Fig. 6 a). At day 21 after infection, similar numbers of P14 IRF8 KO and P14 control cells were found in the spleen, LN, and liver, without significant changes in the proportions of T_PEX and T_EX cells (Fig. 6, b and c). IRF8 levels in IRF8 KO P14 cells were lower than in control P14 cells, showing that CD8⁺ T cells responding to chronic infection express low IRF8 levels. Consistent with this, P14 IRF8 KO cells showed a modest but significant decrease in TOX expression with no effect on PD-1 expression (Fig. 6 d). However, we did not observe changes in cytokine production upon ex vivo restimulation (Fig. 6 e). We also overexpressed IRF8 in P14 cells before transfer into LCMV cl13-infected mice (Fig. 6 f). P14 IRF8 OE cells showed a trend to increased accumulation in secondary LNs and the spleen but not the liver of LCMV cl13-infected mice. There were no changes in the proportions of T_PEX and T_EX cells (Fig. 6, g and h). Increased IRF8 expression resulted in a slight but significant increase in PD-1 and TOX expression, but did not further reduce the already low capacity to produce cytokine upon restimulation (Fig. 6, i and j). Alternatively, IRF8 might not be able to integrate into the regulatory network as other TFs might dominate the regulatory landscape.

In view of these findings, we next wondered whether IRF8 OE impacted the phenotype of CD8⁺ T cells when these cells expand in the absence of persistent antigen stimulation, but rather driven by homeostatic cytokines. To address this, we co-transferred P14 IRF8 OE and P14 control into Rag1^−/− mice and analyzed splenocytes 7 days later (Fig. 6 k). We recovered significantly higher numbers of IRF8 OE compared with control cells in both the spleen and LNs of Rag1^−/− mice (Fig. 6 l). Interestingly, P14 IRF8 OE cells showed increased PD-1 levels compared with control cells, while TOX level did not change in the spleen but were slightly increased in LN (Fig. 6 m and Fig. S3 e). Upon ex vivo restimulation with gp33 peptide, IRF8 OE cells upregulated PD-1 more strongly compared with the control cells (Fig. 6 n and Fig. S3 e). Moreover, IRF8 OE cells displayed strongly impaired effector functions with lower capacity to produce IL-2, TNFα, and IFNγ (Fig. 6, o and p; and Fig. S3 e), showing that OE of IRF8 is sufficient induce some features associated with exhaustion in CD8⁺ T cells even in absence of persistent TCR stimulation, similarly to what has been described for TOX (Khan et al., 2019). Altogether, these results show that IRF8 promotes TOX expression and hinders effector functions.

Recently, IRF2 in tumors (Lukhele et al., 2022) and IRF4 during chronic infection and in tumors (Man et al., 2017; Seo et al., 2021; Yan et al., 2023) have been shown to regulate CD8⁺ T cell exhaustion. We hypothesized that these IRF family members may play a convergent/redundant role in the regulation of exhaustion. To address this point, we directly compared KO of IRF2 and IRF4 with IRF8 KO in CD8⁺ TILs. We performed co-transfers of IRF2, IRF4, or IRF8 KO P14 cells with control transduced P14 cells into B16-gp33–bearing mice (Fig. 7 a and Fig. S4, b and c). The recovered transduced (CD90.1⁺) population was constant in all conditions in the case of the control transduction. In contrast, compared with the injected T cell product, the recovery of transduced (CD90.1⁺) population was decreased for the different KO, with minimal changes in the proportions of T_PEX versus T_EX populations (Fig. 7, b and c). IRF4 KO CD8⁺ T cells were drastically decreased in number, consistent with a critical role of IRF4 in the clonal expansion during infection (Man et al., 2017; Yao et al., 2013). IRF2 was also decreased, but to a lesser extent, similarly to IRF8 KO. TOX levels were reduced in IRF4 and IRF8 KO, and to a lesser extent in IRF2 KO (Fig. 7 d). Conversely, LAG3 or CD39 showed the opposite trend (Fig. S4 a) while PD-1 level remained elevated (Fig. 7 d). The functionality of IRF KO cells was partially improved, with an increase of IFNγ production in IRF2 and IRF8 KO cells, while GrzB and TNFα slightly increased in all IRF KOs upon restimulation (Fig. 7 e). Altogether, several IRF family members decrease TOX and increase effector functions in tumor-specific CD8⁺ T cells, with the strongest phenotype observed in IRF8 KO cells. These data suggest possible redundancy of IRF family members in anti-tumor–specific CD8⁺ T cell responses.

To further investigate the role of the three TFs, we performed scRNA-seq on ex vivo TILs using the same experimental setting. P14 IRF KO CD8⁺ T cells were co-transferred with P14 control cells into B16-gp33–bearing mice. 7 days after co-transfer, tumors were processed, individually labeled with a hashtagging antibody, and pooled before FACS sorting. This protocol allowed us to directly compare the transcriptome of P14 IRF KO cells with P14 control cells derived from the same tumor (Fig. 7 f and Fig. S4, d and e). After batch integration and correction, we defined four populations of P14 cells based on unsupervised clustering together with the expression of marker genes spanning progenitor, exhaustion, effector, and proliferative programs. IRF KOs did not result in major changes in the cell distribution to the four clusters (Fig. 7, f and g). However, IRF4 and IRF8 KO resulted in the emergence of a TOX-low population with effector-like features (C4), while the terminally differentiated T_EX population (C1) declined (Fig. 7 g). Gene set enrichment analysis (GSEA) confirmed the increased overlap with genes associated with effector CD8⁺ T cells and reduced overlap with genes associated with exhausted CD8⁺ T cells (Fig. 7 h and Fig. S4 f).

We then focused on the differentially expressed genes (DEGs) in the various KO (Table S1 and Table S2). In IRF8 KO, we confirmed our functional data with an increase in mRNA encoding in Gzmb and Ifng while Tox level was decreased (Fig. 8 a). More broadly, these DEG were common to IRF4/IRF8 KO and IRF2/IRF8 KO, while a smaller fraction was shared by the three KOs or by IRF4/IRF2 (Fig. 8, a and b). Among the genes upregulated in IRF4/IRF8 KO, we found Xcl1, Crtam, Gzma, Tnf, Ccl1, Ccl3, Ccl4, Ccl5, and Id3, while Prf1, Gzmf, Gzmc, Ramp1, Isg15, and Isg20 were upregulated in IRF2/IRF8 KO cells, suggesting a stronger activation status for the three KOs. The exhaustion-associated genes Tox, Arid1a, Arid1b, and Ptprj and the pro-survival gene Bcl2 were downregulated in all IRF KOs, while Themis and Ikzf2 were specifically downregulated in IRF4/IRF8 KOs. While the overlap of DEGs among the IRF2, IRF4, and IRF8 KO groups was relatively modest (Fig. 8 b), Principal component analysis (PCA) revealed a high degree of transcriptomic proximity between these samples (Fig. 8, c and d). This suggests that while each TF may regulate a distinct set of primary targets, their loss triggers a convergent global gene expression program. The PCA captures these broad, system-wide shifts—including subthreshold expression changes—that reflect a shared phenotypic trajectory toward cellular exhaustion, which is not fully captured by discrete threshold-based Venn analysis. Shared perturbed transcriptional programs analysis between IRF2, IRF4, and IRF8 KO revealed that the three TFs converge on dampening the expression of genes associated with pathways involved in activation or inflammation in tumor-infiltrating CD8⁺ T cells (Fig. 8 e and Fig. S5 a). IRF2 and IRF8 KO resulted in the upregulation of IFN-α/γ responses, IL2/STAT5 signaling, and the allograft rejection signature, pointing to enhanced inflammatory and cytotoxic activity consistent with a shift toward a more effector-like state. IRF4 and IRF8 KO similarly induced TNFα/NF-κB–driven inflammatory together with increased IL-2-STAT5 signaling and allograft rejection signature, again reflecting stronger functional activation. At the same time, both IRF2/IRF8 KO and IRF4/IRF8 KO pairs showed downregulation of pathways linked to cellular homeostasis and persistence, including protein secretion, complement, oxidative phosphorylation, and proliferative programs, suggesting a dual role in establishing exhaustion together with promoting survival (Fig. 8 f and Fig. S5, a–c). Pathways enriched based on each IRF KO-specific up- and downregulated genes were mostly associated to the ones already identified using common DEGs, strengthening our findings (Fig. S5 c).

To further examine the role of IRFs in regulating exhaustion-associated gene expression, we performed CUT&RUN-seq assays for IRF2, IRF4, and IRF8 using CD8⁺ T cells chronically stimulated in vitro, as described above. Integration of scRNA-seq and CUT&RUN-seq data identified a list of eight genes (Fig. 8 g) consistently regulated across IRF2, IRF4, and IRF8 perturbations and supported by IRF binding. Genome browser visualization confirmed IRF binding to multiple regions in the Tox gene, which coincided with accessible chromatin regions in TILs, supporting the direct regulation of Tox by IRF (Fig. 8 h). IRF-bound genes included both upregulated and downregulated genes, indicating that IRFs exert context-dependent regulatory effects rather than uniform activation or repression (Fig. 8 g and Fig. S5 d).

Taken together, these patterns indicate that IRF2, IRF4, and IRF8 converge by promoting exhaustion at least in part through direct regulation of TOX regulation in TILs. KO of the different IRFs similarly promoted a more activated and less exhausted phenotype, underscoring their convergent roles in shaping the exhausted state in TILs.

Despite recent advances, the molecular network governing CD8⁺ T cell differentiation in the context of persistent antigen exposure, i.e., in cancer and chronic infection, is not fully elucidated yet. The IRF family is raising interest for its emerging role in regulating CD8⁺ T cell differentiation and more specifically exhaustion. We established the role of IRF8 as a positive regulator of CD8⁺ T cell exhaustion specifically in a tumor context. Although its expression is TCR-dependent, it was expressed in CD8⁺ TILs but not in virus-specific CD8⁺ T cell during chronic infection. Accessibility of the Irf8 locus was decreased during chronic infection, with a negative regulatory role for type I IFN, hallmark of chronic infections. CRISPR/Cas9-mediated deletion of IRF8 in tumor-specific CD8⁺ T cells resulted in better tumor control capacity in the setting of ACT in two different tumor models. Overexpressing IRF8 in CD8⁺ T cells increased exhaustion features by increasing TOX expression and limiting cytokine production. Deletion of IRF2, IRF4, or IRF8 in CD8⁺ TILs ameliorated their effector functions in a similar fashion by acting on redundant pathways controlling CD8⁺ T cell differentiation, providing a broader picture of the role of this TF family in exhaustion.

While the function of IRF8 in many immune compartments is well described, its role in CD8⁺ T cells is less clear. Miyagawa et al. (2012) showed that IRF8 is induced by integration of TCR and γc-cytokines in settings of suboptimal antigen stimulation. In their experimental model, IRF8 KO CD8⁺ T cells issued form global Irf8^−/− mice resulted in impaired maturation of CD8⁺ T cells into effector phenotype. Conversely, in Sun et al. (2016), the authors show using CD4-CRE Irf8^fl/fl that targeted IRF8 deletion resulted in more robust immune responses against herpes simplex virus infection. IRF8 KO CD8⁺ cells showed stronger proliferation and higher effector function capacity, measured by IFNγ level assessment, suggesting a role for IRF8 in limiting activation and expansion of CD8⁺ T cells. These seemingly contrasting results underscore the importance of selecting appropriate experimental models to address research questions, as specificity of gene ablation might heavily impact functional outcomes. A more recent publication identified IRF8 as an important factor in reorganization of the spatial proximity between enhancer and promoter regions of exhaustion-associated genes (Li et al., 2025). Li et al. show using a CD8-CRE Irf8^fl/fl experimental mouse model that targeted IRF8 KO in CD8⁺ T cells disrupted differentiation on T_PEX into T_EX, leading to heavily impaired cytokine production, resulting in impaired tumor control capacity. We believe that a major difference explaining the opposing findings in Li et al. (2025) and our work relies on the timing of IRF8 deletion. IRF8 is strongly upregulated shortly after activation, which might indicate that it could exert essential functions during early phases of activation. Using a CRISPR/Cas9-mediated approach on preactivated cells, we believe that we could circumvent these potential confounding effects due to early depletion and better assess IRF8 function in the setting of persistent antigen stimulation, such as in the tumor and during chronic infection. This underscores the need for further investigation of the role of IRF8 in CD8⁺ T cells during early steps of activation.

We found that IRF8 is highly expressed in exhausted CD8⁺ T cells in the tumor but not during chronic infection. As previously reported (Miyagawa et al., 2012; Nelson et al., 1996), we confirmed that IRF8 expression in CD8⁺ T cells is driven by TCR engagement, similarly to its homolog IRF4 (Man et al., 2013). The low expression of IRF8 during chronic infection is interesting and underscores substantial differences between the two conditions, due to inherently distinct microenvironments. We observed clear differences between tumor-specific and virus-specific CD8⁺ T cells at the epigenetic level, providing an explanation to the DE of IRF8 in the two conditions. Chromatin was less accessible at the Irf8 locus during chronic infection and the H3K27ac modification was lower compared with CD8⁺ TILs, confirming that it is transcriptionally less active. We propose here a role for type I IFN, hallmark of chronic infections (Teijaro et al., 2013; Wilson et al., 2013), as repressor of IRF8 expression during chronic infection. In our experimental models, addition of IFNβ in vitro or induction of type I IFN signaling in vivo in tumor were able to dampen IRF8 levels in exhausted CD8⁺ T cells. Blocking IFNAR1 in LCMV cl13-infected mice showed opposite effects depending on the timing of treatment. While late blockade during the chronic phase resulted in enhanced IRF8 expression, early blockade during the acute phase of infection decreased IRF8 levels. We think that although our results support the notion that type I IFN is playing a role in altering IRF8 expression in CD8⁺ T cells during chronic infection, additional mechanisms are in place for the DE observed in cancer and chronic infection, and that the biology behind type I IFN signaling in CD8⁺ T cells during chronic infection requires further investigation.

We show here that IRF8 plays a central role in regulating CD8⁺ T cell exhaustion by promoting TOX expression and restricting effector molecules production in tumor-specific CD8⁺ T cells. The comparison of IRF8 KO with IRF2 and IRF4 KO in CD8⁺ T cells allowed us to demonstrate the convergence of these central IRF family members in the regulation of T cell differentiation in the tumor. IRF2 was reported to induce exhaustion in CD8⁺ TILs as a negative feedback on persistent IFN signaling (Lukhele et al., 2022). IRF4, closer to IRF8, has first been described as a TCR-responsive TF driving CD8⁺ T cell exhaustion in LCMV cl13 setting (Man et al., 2017) and tumors (Yan et al., 2023). A subsequent study showed instead that IRF4 was cooperating with its transcriptional partner BATF, when this last was overexpressed, to counter exhaustion in CAR T cells, rewiring the transcriptional outcome as result of redistribution of IRF4 among its target sites in chromatin (Seo et al., 2021). These different reports are issued from different experimental approaches, rendering it difficult to comprehensively integrate the roles of the different TFs. With our approach, we were able to directly compare the cell-intrinsic effect of IRF8 deletion in an unbiased manner by directly comparing IRF8 KO with control TILs issued from the same TME. Moreover, using the same experimental setting, we could directly compare the effects of IRF8 KO with those of IRF2 and IRF4 KO, putting our findings in context of the currently published literature. The three IRFs studied here converge, though by different means, on dampening activation modules and upregulating exhaustion programs in CD8⁺ TILs, unraveling a previously unappreciated redundancy of the IRF family in orchestrating exhaustion. A further point supporting this is the upregulation of the IL2/STAT5 signaling pathway in the IRF2/IRF8 and IRF4/IRF8 KO pairs, as STAT5 has been described to induce a program opposing TOX-driven exhaustion (Beltra et al., 2023). The closer distance between IRF4 and IRF8 KO observed in the PCA analysis and in the shared DEGs suggest a higher degree of redundancy, as already described in B lymphocytes (Shukla and Lu, 2014). Interestingly, a recent publication screening the impact of gene KO in CAR T cells found no advantage to KO IRF2 or IRF4 in their system (Knudsen et al., 2025). It would be interesting to study whether a KO of IRF8, that displayed a better effect on both effector functions and survival/accumulation in our system, has a better effect in CAR T cells engineering too.

The lack of effect of IRF8 KO during chronic infection could also be attributed to the functional redundancy of the various IRF, including IRF2 and IRF4, that may take over IRF8 function. Interestingly, IRF7 has been shown to play an important role in the regulation of CD8⁺ T cell exhaustion during LCMV cl13 infection (Kasmani et al., 2023). IRF7 is barely expressed in exhausted TILs, while its level is much higher in exhausted T cells obtained after chronic infection (Fig. 1 c and Fig. 2 a). Additionally, IRF7 is positively regulated by type I IFN (Fig. 4, g and j), explaining potentially the mutual expression of IRF7 and IRF8 and making of IRF7 a good candidate to take over IRF8 functions during chronic infection.

Altogether, our results established a pivotal role for IRF8 in the context of tumor-induced CD8⁺ T cell exhaustion, but not during chronic infection. The observed functional redundancy of various IRF members during T cell exhaustion underscores the complexity of regulatory networks defining CD8⁺ T cell differentiation trajectories. We confirmed that IRF8 is also expressed human CD8⁺ TILs. Although we did not address the function of IRF8 in human CD8⁺ T cells, our data provide evidence for an application in a human setting. Indeed, a recent study investigating the phenotype of CAR T cells in chronic lymphocytic leukemia found that IRF8 levels were highly enriched in the pre-infusion CAR T cells product from nonresponders compared with complete responders (Fraietta et al., 2018). Moreover, another study found highly reduced TF-binding sites of several IRF members, including IRF8, in open chromatin of CAR T cells with long persistence compared with short persistence, correlated with clinical response, in acute lymphocytic leukemia patients (Bai et al., 2024). The elucidation of mechanistic processes during exhaustion in cancer could unravel new therapeutical opportunities to better harness immunotherapies, for instance, through specific inhibitors or via genetic engineering of IRF8 in tumor-specific CD8⁺ T cells.

Human tumor-infiltrating T cells were obtained from patients treated at the Department of Oncology, University Hospital (CHUV), Lausanne, Switzerland, based on the patient’s informed consent, in the framework of clinical studies (NCT01308294, NCT00112242, NCT00112229, NCT00112216, NCT00002669, LUD 96-010, and PB_2016-02178 [188/12] approved by the Commission cantonale d’éthique de la recherche sur l’être humain). Frozen human samples were analyzed by flow cytometry analysis to assess IRF8 protein levels on a single-cell resolution.

The experiments conducted in this research project comply with Swiss ethical regulations. Experiments with animals were performed in compliance with the University of Lausanne’s internal regulations and approved by the “Service de la consummation et des affaires vétérinaires” (authorization VD3593). All mice used for the experiments were housed in a specific-pathogen free facility of the University of Lausanne. To avoid sex bias, male and female mice were used in the whole project.

C57BL/6 (B6) (CD45.2⁺) mice were obtained from Charles River. CD45.1 congenic B6 mice were bred locally. B6 P14 TCR-transgenic mice (line 237), expressing a TCR specific for the LCMV gp33–41 epitope (gp33) in the context of H-2D^b (P14 T cells), were initially provided by H.P. Pircher (Institute for Immunology, Medical Center, University of Freiburg, Freiburg, Germany) (CD45.2⁺) (Pircher et al., 1989). CD45.1 and CD45.1.2 P14 CD4-CRE STOP^fl/fl-Cas9-eGFP (P14-Cas9) were used to induce targeted KO of selected genes. B6 Rosa26-Cas9-eGFP knock-in mice (B6-Cas9) mice were used as host mice for transfer experiments of Cas9-expressing T cells to avoid rejection. Cas9 strains were kindly provided by P.C. Ho (University of Lausanne, Lausanne, Switzerland [Franco et al., 2023]). OT-I CD45.1 and Rag1⁻ mice were bred in-house. CD45.1 Vβ5 mice were used as hosts for experiments with infection and were kindly provided by W. Held.

For cell culture we used RPMI 1640 GlutaMAX-I (Gibco) and DMEM GlutaMAX-I (Gibco) media. Complete RPMI was supplemented with 10% heat-inactivated FBS (Gibco), 100 U ml⁻¹ penicillin-streptomycin (Gibco), 10 mM Hepes (Gibco), 1 mM sodium pyruvate (Gibco), and 50 μM 2-mercaptoethanol (Gibco). Complete DMEM was supplemented with 10% heat-inactivated FBS and 100 U ml⁻¹ penicillin-streptomycin. Primary murine CD8 T cells were cultured in complete RPMI. Frozen human melanoma metastasis samples were thawed and cultured ON with complete RPMI. B16F10, LLC1, EL-4, and MC38 cell lines were cultured in complete DMEM. B16-gp33 and B16-OVA were cultured in complete DMEM with 100 μg ml⁻¹ G418 (Calbiochem). PlatinumE (PlatE) cells were cultured in complete DMEM with 10 μg ml⁻¹ blasticidin (Invivogen) and 1 μg ml⁻¹ puromycin (Invivogen).

The various sgRNA or scrambled controls were inserted into a PQG-Thy1.1 retroviral vector (Table S4), kindly shared by P. Reichenbach (Melita Irving’s lab, University of Lausanne, Lausanne, Switzerland). The IRF8-coding sequence used for OE experiments was purchased from Addgene (plasmid number 102872) and was inserted into pMSGV retroviral vector. The sequence was cloned using In-Fusion Cloning kit (Takara Bio) after the Thy1.1 transduction marker, separated by a self-cleaving P2A sequence. As control for OE experiments, a pMSGV vector containing only Thy1.1 was used.

For retroviral production, 7 × 10⁶ PlatE cells were seeded into a T75 flask ON before transfection. PlatE cells were then transfected using a mix of Lipofectamine 2000 (Invitrogen) with a total of 30 μg of DNA (7.5 μg pCL-Eco and 22.5 μg of PQG-sgRNA-Thy1.1 or pMSGV-Thy1.1) in pure DMEM. The culture medium was replaced after 24 h with complete RPMI. Retroviral supernatants were recovered after 48 h, filtered, snap-frozen, and stored at −80°C.

As efficient retroviral vector integration requires active cell proliferation, CD8⁺ T cells were transduced 24 h after activation with 1 μg ml⁻¹ GP33 peptide (KAVYNFATC) and 20 U ml⁻¹ recombinant human IL-2 (Proleukin aldesleukin) in 1.5 ml complete RPMI. For the transduction, 1 ml of culture medium was removed and replaced with 1 ml of retroviral supernatant supplemented with 10 μg ml⁻¹ polybrene infection/transfection reagent (Sigma-Aldrich). Cells were spun for 60 min at 1,800 r.p.m. at 30°C and incubated 2 h at 37°C. 1 ml of medium was then removed and replaced by the initially removed medium.

For validation of IRF2, IRF4, and IRF8 KO, transduced KO and control CD90.1⁺ P14 CD8 Cas9 T cells were FACS sorted 48 h after retroviral transduction. RNA was extracted with the RNeasy Plus Micro Kit (Qiagen) following the manufacturer’s recommendations. Reverse transcription was done with the High-capacity cDNA Reverse Transcription kit (Applied Biosystems), and qPCR was done with the KAPA SYBR Fast qPCR Master Mix (2x) Kit (Sigma-Aldrich). Primers are listed in Table S4. Amplification was done in 96-well plates on a QuantStudio 6 Pro Real-Time PCR systems machine (Applied Biosystems), and data were analyzed with the Design & Analysis 2 software.

Tumor cells were resuspended in PBS before s.c. injections into the flank of mice. 0.5 × 10⁶ cells for B16F10, B16-gp33, and B16-OVA, and 1 × 10⁶ cells for MC38, LLC1, and EL-4 were injected in 50 μl PBS volume. Orthotopic MIBC, used to assess IRF8 expression in exhausted CD8⁺ T cells, was induced though injection of a CRE-encoding adenovirus in the bladder of p53^fl/fl/pten^fl/fl mice, as described elsewhere (Leblond et al., 2020). Analysis was performed 8 wk after tumor initiation. For naive CD8⁺ T cell transfer experiments, CD8⁺ T cells were enriched by negative selection following the manufacturer’s instruction (STEMCELL) and intravenously transferred at 2 × 10⁶ cells/mouse. For preactivated CD8⁺ T cell transfer experiments, LN cell suspensions were cultured in complete RPMI + 1 μg ml⁻¹ GP33 peptide (KAVYNFATC) or 1 μg ml⁻¹ OVA peptide (SIINFEKL) with 20 U ml⁻¹ recombinant human IL-2 (Proleukin aldesleukin) for 48 h before intravenous injection in mice. For KO and OE experiments, retroviral transduction was performed at 24 h after activation. Cells were counted and mixed in a 1:1 ratio, with a total of 2 × 10⁶ CD8⁺ T cells injected per mouse (e.g., 1 × 10⁶ P14 control + 1 × 10⁶ P14 IRF8 KO CD8⁺ T cells). 7 days after T cell transfer mice were sacrificed and tumors, draining and non-draining LN were collected and passed through 70-μm cell strainer. Lymphocytes were enriched from tumors using a 40/70 Percoll gradient (VWR 17-0891). For tumor growth experiments, 4 × 10⁶ CD8⁺ T cells after 48 h in vitro activation were injected intravenously 7 days after tumor engraftment. Tumor volumes were assessed with a caliper and calculated using the following formula: volume (mm³) = length (mm) × width (mm) × height (mm). Endpoint for tumor growth experiments was 1 cm³ of tumor volume in compliance with the animal experimentation license.

Vβ5 CD45.1 mice were infected by intravenous injection with 2 × 10⁶ PFUs of LCMV cl13 1 day after ACT of 1 × 10³ naive P14 CD45.2 CD8⁺ T cells for chronic infection. CD45.1⁺CD45.2⁺ B6 mice were injected with 2 × 10⁵ PFUs of LCMV 53b Arm strain 1 day after ACT of 1 × 10⁴ naive P14 CD45.2 CD8⁺ T cells for acute infection. For IRF8 KO and IRF8 OE co-transfer experiments, transduced CD8⁺ T cells were FACS sorted based on CD90.1 expression and injected in a ratio KO/OE: Control of 1:1 at 4 × 10⁴ CD8⁺ T cells per mouse on the same day of infection. On the readout day, spleens were recovered, passed through cell strainers, and incubated 5 min with RBC lysis buffer before further processing. Livers were passed through cell strainers, and lymphocytes were enriched with a 40/70 Percoll gradient. For IFNAR1 blockade, mice were injected i.p. with 1 mg of anti-IFNAR1 (MAR1-5A3; BioXCell) (1 mg) or 1 mg of isotype control Ab (mouse IgG1, MOPC-21, BioXCell) at day 3, days 10 and 15, or day 22 after infection, as indicated in the Fig. 4.

CD8⁺ T cells were enriched by negative selection following the manufacturer’s protocol (STEMCELL). 0.4 × 10⁶ CD8⁺ T cells per well were plated in complete RPMI in a 24-well plate, previously coated with anti-mouse CD3 antibodies (clone 145-2C11; BioLegend) at 5 μg ml⁻¹, and activated for 72 h with addition of soluble anti-mouse CD28 (clone 37.51; BioLegend) at 1 μg ml⁻¹ and 20 U ml⁻¹ IL-2. Every 48 h, cells were plated to new anti-CD3–coated 24-well plate with 10 U ml⁻¹ IL-2 at 1 × 10⁶ cells per well. This step was repeated two additional times for a total of three restimulations after initial activation. For the resting condition, after the initial, activation cells were plated in complete RPMI supplemented with 5 ng ml⁻¹ IL-7 (PeproTech) and 5 ng ml⁻¹ IL-15 (PeproTech), and cytokines were refreshed every 2 days.

Fluorescent antibody staining was performed for all experiments with the same protocol. Cells were incubated with anti-mouse CD16/32 (101320; BioLegend) in FACS buffer (PBS with 2% FBS, 5 mM EDTA, and 0.2% Na azide) for 15 min. Extracellular staining was then performed for 30 min at 4°C with an antibody mix in FACS buffer. Cells were washed twice with PBS and subsequently stained for viability with Aqua Vivid (L34966; Invitrogen) or Zombie NIR (423105; BioLegend) for 15 min at 4°C. Cells were then analyzed directly on a cytometer or fixed and permeabilized for intracellular/intranuclear staining for 20 min at 4°C using the FoxP3 staining kit (00-5523-00; Invitrogen). After one wash in permeabilization buffer, cells were incubated for 1 h at 4°C for intracellular/intranuclear staining. For cytokine staining, P14 CD8⁺ T cells were restimulated for 5 h at 37°C in presence of GP33 peptide in addition to GolgiStop (554724; BD). The complete list of antibodies used can be found in Table S3.

The expression profile figures of IRF genes from Fig. 1, c and g were created based on the publicly available SPICA atlas for tumor-infiltrating T cells and virus-specific CD8 T cells atlases (Andreatta et al., 2022). This collection comprises a highly curated and annotated collection of cancer and infection transcriptomic studies describing immune cell states at high resolution.

The input consisted of a publicly available, carefully curated compendium of scRNA-seq profiles from mouse CD8⁺ TILs (Andreatta et al., 2021). The collection covers five well-defined, naturally occurring CD8⁺ states with distinct transcriptional signatures: CD8_NaiveLike, CD8_EffectorMemory, CD8_EarlyActiv, CD8_Tpex, and CD8_Tex. Regulatory programs (regulons) were inferred with the SCENIC workflow (Aibar et al., 2017) (https://scenic.aertslab.org). To capture regulatory events specific to individual subpopulations, the dataset was partitioned by T cell state, and each subset was processed independently. SCENIC proceeds in three sequential stages—grnBoost2, RcisTarget, and AUCell—which we used as follows: Step 1: Co-expression modules (grnBoost2). Co-expression modules were derived using grnBoost2—a fast implementation related to GENIE3 (Huynh-Thu et al., 2010)—via the SCENIC R package. We invoked runSCENIC_1_coexNetwork2modules with nTopTfs = 50 and nTopTargets = 5 to aggregate targets into preliminary (raw) putative regulons. Step 2: Regulon/GRN refinement (RcisTarget). Modules were pruned to remove indirect targets by motif enrichment analysis with RcisTarget (Herrmann et al., 2012; Imrichová et al., 2015) and cis-regulatory motif resources. Specifically, we used the cisTarget ranking databases mm9-500bp-upstream-7species.mc9nr.feather and mm9-tss-centered-10kb-7species.mc9nr.feather, together with the motif table motifs-v9-nr.mgi-m0.001-o0.0.tbl. The motif database provides scores for motif–gene pairs, yielding per-TF motif–gene rankings. For each candidate TF target set, motif enrichment was quantified as the area under the recovery curve on these rankings (Aibar et al., 2017). When a TF’s motif was significantly enriched, a regulon was called by retaining target genes with high motif–gene scores; these curated TF-target edges formed the GRN.

We applied Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA), a multivariate statistical approach that leverages predefined class labels to reduce variability within each class and enhance separation between classes (Thévenot et al., 2015). The method produces a predictive component (p1), which represents a data projection that optimally distinguishes elements (cells) belonging to the two predefined groups (T cell states). Each feature (such as a gene or regulon) contributes to this predictive component, quantified by a value referred to as weightStarMN. To facilitate comparison, these values are rescaled to a range between 0 and 1 while retaining their sign, yielding the discriminant score for each feature. Within a given OPLS-DA model, these scores allow features to be ranked according to their relative importance in distinguishing between the classes. The OPLS-DA was performed using the ropls R packageREF. The input data for OPLS-DA were mean centered and divided by the SD to ensure proper normalization.

The scRNA-seq experiment was performed on ex vivo FACS-sorted CD8⁺ TILs issued from a co-transfer experiment. Mice were engrafted with B16-gp33 tumors, followed by a co-transfer of retrovirally transduced P14 control together with P14 IRF(2, 4, or 8) KO CD8⁺ T cells at day 7. 7 days after adoptive CD8⁺ T cell transfer, mice were sacrificed, and tumors were passed through a 70-μm cell strainer. CD8⁺ T cells were enriched using a positive CD8⁺ TILs selection (Miltenyi), following the manufacturer’s instructions. After CD8⁺ TILs enrichment, cells were stained for FACS sorting as described above (viability marker, CD8-biotin, CD45.1, CD45.2, and CD90.1). In a second step, cells were incubated with TotalSeq hashtagging streptavidin-APC antibodies (BioLegend) for multiplexing. Control and IRF KO cells issued from the same tumor were thus labeled with the same hashtagging sequence, allowing for a paired analysis of cells exposed to the same TME. The samples were pooled and FACS sorted based on viability and the expression of CD8a, CD90.1, and CD45.1/CD45.1.2 to separate P14 control from P14 IRF KO cells. Sequencing was performed on two lanes of an Element Aviti System device at the Lausanne Genomic Technologies Facility (https://wp.unil.ch/gtf/).

Raw scRNA-seq data were obtained from multiple experimental batches of IRF2, IRF4, and IRF8 control and KO samples. CellRanger-processed gene expression and multiplexing count matrices were imported into R (version 4.x) using Seurat (version 5) and related packages (SeuratWrappers, STACAS, scRepertoire, scGate, UCell, ProjecTILs, scIntegrationMetrics, and batchelor). For each dataset, Seurat objects were generated from the gene expression matrices and multiplexing capture assays, with sample- and condition-specific metadata annotations (condition, sample, batch, and Ctrl/KO). Gene expression counts were normalized per dataset prior to merging. To identify and exclude confounding cellular states, transcriptional signatures corresponding to cell cycle (G1/S, G2/M), IFN-stimulated genes (ISGs), and heat shock proteins (HSP) were derived from the SignatuR database and custom gene sets. UCell scores were computed per cell for these signatures across all datasets. All Seurat objects were merged into a single dataset and subjected to quality control filtering. Quality metrics included the number of detected genes (nFeature_RNA), UMI counts (nCount_RNA), and the proportion of reads mapping to ribosomal (Rp[sl]) and mitochondrial (mt-) genes. Cells were retained if they satisfied the following thresholds: nFeature_RNA: 800–4,900, nCount_RNA: 1,300–24,000, percent.ribo <50%, and percent.mito <10%. After filtering, the merged dataset was saved as a Seurat object for downstream analysis. Analyses were performed in R using Seurat (object handling, normalization, and HVG selection), PCAtools (dimensionality reduction, scree diagnostics, and biplots), ggplot2 (visualization), tidyverse/dplyr (data wrangling), Matrix, and ProjecTILs (heatmap utility). The SignatuR resource provided gene signatures. The in-house package ScPlusPlus was installed from GitHub and loaded for utilities. Default settings were used unless otherwise specified. A previously generated Seurat object containing merged scRNA-seq profiles from IRF2/IRF4/IRF8 control and KO conditions was read from disk. Sample-level metadata included condition, sample, batch, and previously computed signature scores (e.g., cycling). Cells were binarized into Prolif versus NoProlif based on the cycling signature score threshold (cycling ≥ 0.1 → Prolif; otherwise NoProlif). Counts per condition and per sample were tabulated to quantify group sizes. For both condition and sample strata, cell counts were aggregated by Prolif class, and two normalizations were computed: (1) normSize.Condition = cells in class/total cells within each condition (or sample) and (2) normSize.Cluster = cells in class/total cells within each Prolif category.

Class composition within each Prolif group was further adjusted by dividing normSize.Condition by the sum across classes for that group (clustComposition). Bar plots of normalized frequencies were generated and exported to PDF. For each batch (analyzed separately for batch1 and batch2), cells were subset to the given batch label and processed as follows: (1) Normalization: Seurat LogNormalize with scale.factor = 10,000. (2) Feature selection: FindVariableFeatures with selection.method = “dispersion” and nfeatures = 5,000. (3) Feature blacklist and pruning: High-variance features were filtered to exclude genes associated with confounding programs and technical artifacts. Gene sets included: Cell cycle G1/S and G2/M (from SignatuR), ISG genes (custom list: Isg15, Ifit3, Ifit1, Rsad2, Usp18, Isg20, Bst2, Phf11b, Zbp1, Stat1, Slfn5, Samhd1, Iigp1, and Ifit3b), heat shock genes (from SignatuR HeatShock), mitochondrial (prefix mt-), ribosomal (regex ^Rp[ls]), TCR (regex Tr[abgd][vjd]), and a custom blacklist (external resource). The final HVG set was defined as the initial 5,000 HVGs minus all blacklisted genes. Using ProjecTILs::celltype.heatmap with assay = “RNA” and the pruned HVG set, a sample-by-gene matrix (clustered by genes and samples) was computed and returned. Zero-sum columns (genes) were removed. Sample row names were parsed to extract state and batch identifiers; a metadata table was built with fields subset (=State), source (=Condition parsed from sample names), and a binary Ct_KO label (ctrl versus KOs). For each batch, the top 2,000 HVGs present in the summary matrix were retained, and the matrix was transposed to genes × samples for PCA. PCA was performed with PCAtools::pca using removeVar = 0.1 (removing the lowest 10% variance features). The number of informative components was assessed with parallelPCA (Horn’s test) and elbow detection from the variance profile. PCA biplots were generated with PCAtools::biplot, coloring samples by source with a predefined color key and encircling groups. For batch2, a single outlier sample (I4.KO_r3) was removed prior to PCA. Normalized composition bar plots and PCA diagnostics (scree plots with Horn’s and elbow references, PCA biplots) were produced with ggplot2 and PCAtools.

All analyses were performed in R using Seurat (object handling, normalization, QC, and clustering), STACAS (batch integration), UCell (signature scoring), UMAP via Seurat, ProjecTILs (utilities), ggplot2/patchwork (visualization), dplyr/tidyr/tidyverse (wrangling), Matrix, gridExtra, and scIntegrationMetrics. Gene programs were retrieved from SignatuR. Unless otherwise noted, default parameters were used, and Seurat objects were handled in assay version “v3.” A previously curated Seurat object containing merged scRNA-seq profiles from IRF2/IRF4/IRF8 control and KO conditions was loaded and split by the batch metadata. Mouse gene annotations (EnsemblGeneTable.Mm) were used to standardize gene symbols, and each cell’s barcode was stored in metadata. Objects were slimmed to the RNA assay and log-normalized with Seurat’s NormalizeData (scale.factor = 10,000). Three confounding programs were defined: Cycling = union of G1/S and G2/M (SignatuR), ISG = curated ISG set (Isg15, Ifit3, Rsad2, Stat1, Samhd1, etc.), HSP = heat-shock program (SignatuR). Per-cell UCell scores were computed within each batch. Cells with elevated stress or IFN signatures were excluded (ISG < 0.25 and HSP < 0.6). Across batches, quality control thresholds were derived from the 2nd–98th percentiles of nCount_RNA (UMIs) and nFeature_RNA (genes), and only cells within these ranges were retained. Batch-specific HVGs were identified with STACAS::FindVariableFeatures.STACAS (requesting twice the target number, with nfeatures = 1,000 downstream). HVGs were pruned against comprehensive blocklists consisting of cell cycle, ISG, and HSP programs (from SignatuR), mitochondrial genes (^mt-), ribosomal genes (^Rp[ls]), and TCR segments (Tr[abgd][vjd]), and an external blacklist plus specific entries (Trbv12-1, Trbv13-2, Trgv2, Trbc1, and Seq1). For each batch, pruned HVGs were defined as setdiff(HVG, blocklists); pre- and post-filter HVG counts were recorded. A final cross-batch HVG panel was selected with SelectIntegrationFeatures (nfeatures = 1,000). Objects were integrated using STACAS::Run.STACAS (dims = 1:30, anchor.features = HVG panel), followed by UMAP embedding (dims = 1:30, seed = 123). Marker expression (e.g., Tox, Pdcd1, Gzmb, Cd69, Tcf7, Ifng, and Cx3cr1) was inspected on the RNA assay. Cells were clustered in the integrated space via shared nearest neighbor graphs (FindNeighbors, k.param = 20) and Louvain community detection (FindClusters, algorithm = 1) across multiple resolutions (0.2–0.4). Resolution 0.2 was selected. Cluster expression profiles were visualized by violin plots across markers spanning progenitor, exhaustion, effector, and proliferative programs (Lef1, Tcf7, Pdcd1, Tox, Gzmb, Ifng, Zeb2, Klrg1, Mki67, etc.). Clusters were then manually annotated and stored in the Seurat. Annotation field (e.g., C1_Tex, C2_Tex_Cycling, C3_Tpex, and C5_ToxLow_EffLike). Cluster abundances were tabulated per condition and sample. For each cluster, relative frequencies were computed as: normSize.Condition = cells in cluster ÷ total cells in that condition/sample, normSize.Cluster = cells in cluster ÷ total cells of that cluster. Cluster composition was further normalized within clusters to give clustComposition. Bar plots of relative cluster frequencies.

Analyses were performed in R using Seurat (object handling), edgeR (pseudobulk DE), fgsea, ScPlusPlus (GSEA helper utilities), tidyverse/dplyr/tidyr (wrangling), Matrix, ggplot2/patchwork (visualization), and ProjecTILs utilities. Seurat objects were handled in assay version v3 unless noted. An integrated Seurat object (object_integrated_ss_full,for.DEA.rds) was loaded and the default assay set to RNA. Cells were annotated by cluster/state in Seurat.Annotation. The genotype comparison was defined per IRFx target as KO versus control (e.g., baseline_target = “IRFxc,” contrast_target = “IRFxko”). Sample strings were parsed from the sample metadata to extract target (IRFx), genotype (C/KO), and replicate_id (e.g., T1, T2). Only balanced replicates (those containing both genotypes for the same IRFx target) were retained for analysis. Gene-level raw counts were aggregated per (replicate_id × genotype) by summing cell counts (sparse matrix row sums), producing a pseudobulk counts matrix. A corresponding sample metadata table recorded replicate and genotype for each pseudobulk column, with genotype factors ordered so that control served as the baseline. Pseudobulk counts were filtered using edgeR’s filterByExpr (blocking on replicate), library sizes normalized with TMM, and gene-wise/BCV dispersions estimated. A paired GLM with design replicate_id + genotype\texttt{∼ replicate\_id + genotype} replicate_id + genotype was fit using glmQLFit, and the genotype coefficient tested with glmQLFTest to obtain moderated quasi-likelihood statistics. Positive logFC indicates higher expression in IRFxKO relative to IRFxc within replicate. Full DE tables were exported as CSV (e.g., deg.IRF4KO_vsCtril_considering_pairing.csv). Genes were ranked by logFC from the edgeR contrast (IRFxKO versus IRFxc). Ranked lists were analyzed with ScPlusPlus::run_msigdbr_gsea_for_ranked_lists (wrapper around fgsea) against MSigDB C7: IMMUNESIGDB gene sets, filtered by the keyword “EFFECTOR_VS_EXHAUSTED.” GSEA parameters included top_n = 20, padj_cutoff = 0.1, and dot plots colored by NES and sized by leading edge. Enrichment plots were generated and saved as PDFs for the all-cells analysis and for the Tpex and Tex scopes; GSEA result tables were also written to disk.

All analyses were performed in R using Seurat (object handling), EnhancedVolcano (volcano plots), VennDiagram (set overlaps), ggplot2/patchwork (visualization), tidyverse/dplyr/tidyr (data wrangling), matrix, and ScPlusPlus (MSigDB enrichment wrappers). Seurat objects were handled in assay version v3 unless otherwise noted. An integrated Seurat object (object_integrated_ss_full, for.DEA.rds) was loaded for enrichment analyses. DE results from a paired pseudobulk design (KO versus control) were imported from CSV files for IRF2, IRF4, and IRF8. Each DE table contained at least gene, logFC, and FDR and was available at multiple scopes: all cells, Tpex (C3_Tpex), and Tex (C1_Tex). A curated gene list (MarcoList.csv) was used to highlight specific markers in plots. DE tables were visualized with EnhancedVolcano. Plots emphasized canonical CD8 T cell markers (e.g., Tox, Tcf7, Pdcd1, Gzma/b/c/d/e/f, Havcr2, Klrg1, and Cx3cr1), the curated list from MarcoList.csv, and the most extreme DE genes (top/bottom by logFC). Nominal thresholds of FDR <0.05 and |logFC| ≥0.3 guided labeling/interpretation, with axis ranges tuned per comparison. Figures were exported as PDFs (e.g., VolcanoPlot_deg.IRF8KO_vsCtril.pdf, plus Tpex and Tex variants). Within Tex comparisons, DE genes were classified as upregulated (logFC > 0.3, FDR < 0.05) or downregulated (logFC less than −0.3, FDR <0.05). From these, multiple contrasts were derived: (1) Target-specific sets: unique to each IRF after subtracting overlap with the other two (e.g., IRF2.Spec.Up, IRF4.Spec.Down). (2) Common sets: genes shared across all three IRFs (Common_All_Up, Common_All_Down). (3) Pair-specific overlaps: shared between two IRFs but absent in the third (e.g., IRF2.IRF4_Specific.Up, IRF8.IRF4_Specific.Down). Venn diagrams (PNG format) visualized overlaps among up or down gene sets, using custom colors and transparency. Pathway enrichment was performed with ScPlusPlus::run_msigdbr_ora_for_regulons, using the integrated object as background. Analyses targeted MSigDB collections C2: REACTOME and H: Hallmark (species = mouse). For C7, terms were filtered with keywords “CD8” and “IFN” in “all” mode. Parameters included top_n = 5, padj_cutoff = 0.05–0.10, and plot_style = “bar.” Enrichment bar plots were color-mapped to gene set categories and exported as PDFs (e.g., Immune_BAR.Up_Specific.pdf, Reactome_BAR.Up_SpecificPairs.pdf). All volcano plots, Venn diagrams, and ORA bar plots were saved in the plots/directory. Intermediate and combined gene lists were retained for plotting. Package versions and environment details were recorded via sessionInfo.

All analyses were performed in R version 4.3. Package dependencies included ArchR (core processing), Seurat/Signac (object conversion/visualization), ggplot2 (version 3.4.4; plotting), pheatmap (heatmaps), BSgenome.Mmusculus.UCSC.mm10 (genome), EnsDb.Mmusculus.v79 (gene annotation), and ArchRtoSignac (conversion utilities). Unless otherwise noted, the mouse reference genome was mm10. Computations used 16 threads (addArchRThreads(threads = 16)), and stochastic steps used a fixed seed = 1,234 for reproducibility. Accessible fragment files (three samples: Cl13d8, Cl13d30, naive (GSE199565 [Giles et al., 2022]) and TIL_NFAT5_wt (GSE237711 [Tille et al., 2023]) were provided as fragments.tsv.gz. For each sample, we created Arrow files with: Cell QC thresholds: minTSS = 15, minFrags = 5,000, maxFrags = 50,000, Matrices: genome-wide TileMatrix (500-bp bins) and GeneScoreMatrix (from the mm10 gene annotation defined via addArchRGenome(“mm10”)), Command: createArrowFiles(..., addTileMat = TRUE, addGeneScoreMat = TRUE, force = TRUE). Arrow files were used to instantiate an ArchRProject (copyArrows = TRUE), which was then filtered for doublets. Putative doublets were scored per sample using addDoubletScores with k = 10, knnMethod = “UMAP,” and LSIMethod = 1. Cells flagged as doublets were removed via filterDoublets. We performed iterative LSI on the TileMatrix in two rounds: (1) IterativeLSI (v1), addIterativeLSI(name = “IterativeLSI,” iterations = 4, varFeatures = 10,000, dimsToUse = 1:20, clusterParams = list(resolution = c(0.1,0.2,0.4), sampleCells = 10,000, n.start = 10), seed = 1,234, force = TRUE). (2) IterativeLSI (v2), addIterativeLSI(name = “IterativeLSI_2,” iterations = 4, varFeatures = 20,000, dimsToUse = 1:20, clusterParams = list(resolution = c(0.1,0.2,0.4), sampleCells = 10,000, n.start = 10), seed = 1,234, force = TRUE). ArchR’s depth-correlated LSI components were automatically handled (default corCutOff). SDs per component were inspected to confirm dimensionality choice. Graph clustering used Seurat’s Louvain/Leiden implementation via ArchR: Round 1: addClusters(reducedDims = “IterativeLSI,” name = “Clusters_1,” resolution = 0.2, seed = 1,234, force = TRUE). Round 2: addClusters(reducedDims = “IterativeLSI_2,” name = “Clusters_2,” dimsToUse = 1:20, resolution = 0.3, seed = 1,234, force = TRUE). Cluster/sample relationships were quantified with confusion matrices and visualized as heatmaps (pheatmap) after row-normalization. UMAP embeddings were computed from the LSI spaces: UMAP_1: addUMAP(reducedDims = “IterativeLSI,” nNeighbors = 30, minDist = 0.5, metric = “cosine,” name = “UMAP_1,” force = TRUE). UMAP_2: addUMAP(reducedDims = “IterativeLSI_2,” dimsToUse = 1:20, nNeighbors = 50, minDist = 0.5, metric = “cosine,” name = “UMAP_2,” force = TRUE). Cells were visualized with plotEmbedding, colored by sample, Clusters_1, Clusters_2, or user-defined meta-annotations (see below). Custom theming was applied for publication figures. Samples were grouped into experimental categories for downstream summaries: Chronic LCMV (Cl13d8, Cl13d30) naive and TILs. A derived factor Infection_vs_TIL (levels: naive, Arm, Cl13, and TILs) and batch labels were added (Cl13/Naive → “Giles_Wherry_NI_2022”; TILs → “Greg_NI_2024”). Pseudobulk replicates were created per Clusters_2 with addGroupCoverages(groupBy = Clusters_2, force = TRUE). Peaks were called using MACS2 discovered by findMacs2(), and a reproducible peak set was generated with: addReproduciblePeakSet(groupBy = Clusters_2, pathToMacs2 = ..., force = TRUE). The resulting peak set (as GRanges) was retrieved via getPeakSet() and saved to disk. A PeakMatrix was then added (addPeakMatrix(force = TRUE)), and available matrices were enumerated for QC. Browser tracks around marker genes were produced with plotBrowserTrack: Tox: ±500 kb window; grouped by Sample and restricted to selected groups (“naive,” “Cl13d8,” “Cl13d30,” and “TILs”), Irf8: ±30 kb window (same grouping approach). Plots were rendered to PDF (plotPDF). Group-level bigWigs were exported with: getGroupBW(groupBy = “Sample,” normMethod = “ReadsInTSS,” tileSize = 100, maxCells = 1,000, and ceiling = 4).

The aforementioned ATAC-sequencing bigWig files and H3K27ac Cut&Tag bigWig files were utilized for epigenetic analysis in selected loci. The H3K27ac data were sourced from the Gene Expression Omnibus (GEO) under accession number GSE235007 (Ma et al., 2025). Additionally, TF-binding regions for mouse CD8⁺ T cells were obtained from the ChIP-Atlas public database (Zou et al., 2022). Genomic tracks were visualized using the Integrative Genomics Viewer (IGV) version 2.16.0 (Kent et al., 2002). H3K27ac signal intensity within specific genomic regions, including the Irf8 and Tox loci, was quantified using the UCSC command-line utility bigWigAverageOverBed.

CD8⁺ T cells from WT B6 mice were enriched by negative selection following the manufacturer’s protocol (STEMCELL) and cultured following the in vitro chronic TCR stimulation protocol described above to induce exhaustion. CD8⁺ T cells were then FACS sorted and processed for CUT&RUN-seq following the manufacturer’s instructions (EpiCypher) with the modification of increasing the bead concentration to ensure smaller fragments were not lost during the purification process. Libraries were pooled and subjected to three further rounds of bead purification to exclude the adapter dimers present in the libraries. Libraries were sequenced on a NovaSeq instrument in 150-bp paired-end mode. CUT&RUN sequencing data were aligned to the mouse reference genome (mm10) using Bowtie2 (version 2.4.5), and BAM files were processed using SAMtools (version 1.19). Genome-wide signal tracks were generated using deepTools (version 3.5.5; bamCoverage) with CPM normalization to account for differences in sequencing depth across samples. Peak calling was performed using SEACR (version 1.3) in stringent mode with IgG as control. Reproducible peaks across replicates were identified using BEDTools (version 2.31.1), and peaks were annotated to genes using HOMER (version 4.11). BigWig files were generated using UCSC utilities, including bigWigMerge and bedGraphToBigWig (UCSC tools version 445), and used for downstream visualization. Direct targets were defined by intersecting DEGs with genes associated with CUT&RUN peaks. Shared targets across IRF2, IRF4, and IRF8 were defined as genes differentially expressed in all three perturbations and bound by at least one IRF. Lollipop plots were generated in R (version 4.4.1), where the x-axis represents log2 fold change and point size indicates –log10(FDR). For shared targets, log2 fold change values from the IRF8 KO condition were used. Genome browser tracks were visualized using the IGV (version 2.19.2).

Statistical testing was done using GraphPad Prism version 10. Use tests are indicated in each figure legend. For co-transfer experiments, control and KO/OE cells were exposed to the same tumor or infection environment, thus paired Student’s t tests were used. One-way and two-way ANOVA were used to compare more than two groups. P value threshold was set at 0.05, with recommended corrections when multiple testing was performed. For all tests, *P < 0.05, **P < 0.01, and ***P < 0.001.

Fig. S1 compares the irf family members’ expression level in TILs and chronic infection and shows gating strategies. Fig. S2 completes Fig. 4 with additional data for IRF8 level in the liver, additional ATAC-seq and ChIPseq data, and the effect of additional cytokines on IRF8 level. Fig. S3 shows a map of the vector used for KO and OE of IRF8, with the effect 48 h after activation on the indicated activation markers and completes Fig. 6, k–p, showing similar data in the LN. Fig. S4 shows the efficacy of the various KO and the strategy used for scRNA-seq. Fig. S5 complements Fig. 8 with additional GEO analysis. Table S1 shows the top 100 upregulated genes in IRF2, IRF4, and IRF8 KO versus control transduced TILs. Table S2 shows the top 100 downregulated genes in IRF2, IRF4, and IRF8 KO versus control transduced TILs. Table S3 shows antibodies used for flow cytometry and scRNA-seq multiplexing. Table S4 shows primer and sgRNA sequences.

All omics data are deposited on the GEO platform (GSE333230). All data are available from the corresponding author upon reasonable request.

This work was funded by the University of Lausanne and grants from the Max Cloëtta Foundation (G. Verdeil), the Swiss National Science Foundation (G. Verdeil: 310030_182680; W. Held: 310030_200898), ISREC Institute (D. Cropp), and the Swiss Cancer Research Foundation (G. Verdeil; KFS-5770-02-2023; W. Held: KFS-5386-08-2021).

Author contributions: Marco Ongaro: conceptualization, data curation, formal analysis, investigation, methodology, project administration, supervision, validation, visualization, and writing—original draft, review, and editing. Romane Thouenon: investigation and writing—review and editing. Isaac Crespo: formal analysis, visualization, and writing—original draft. Alexandre Dumez: investigation. Mélanie Charmoy: investigation and methodology. Eduardo Roman-Camacho: data curation, formal analysis, and visualization. Victoria J. Morel: investigation. Daniela Cropp: conceptualization, project administration, supervision, validation, and writing—review and editing. Emma Desponds: investigation. Giulio Zanette: investigation. Yi-Chuan Wang: data curation. Letizia Vergili: investigation. Laure Tillé: conceptualization and investigation. Inés Di Resta: formal analysis and investigation. Marine M. Leblond: investigation and Jésus Corria Osorio: data curation, formal analysis, methodology, resources, supervision, and visualization. Ping-Chih Ho: data curation. Werner Held: conceptualization, funding acquisition, supervision, and writing—review and editing. Grégory Verdeil: conceptualization, funding acquisition, methodology, project administration, resources, supervision, validation, visualization, and writing—review and editing.