CD8-Targeted IL2 Unleashes Tumor-Specific Immunity in Human Cancer Tissue by Reviving the Dysfunctional T-cell Pool

Over the past decade, T-cell–based immunotherapies have emerged as a groundbreaking approach leveraging the potential of tumor-specific T cells to recognize and kill malignant cells (1–5). In particular, the PD-1/PD-L1 axis has been identified as a major pathway limiting effective antitumor T-cell immunity, and immune-checkpoint blockade (ICB) therapies targeting this pathway have demonstrated remarkable success in reactivating tumor-specific T cells to unleash their antitumor potential (6–9). Nevertheless, the durable benefit from this treatment is currently limited to a small fraction of patients. Interestingly, T-cell presence in the tumor only provides limited predictive value for response to PD-1 blockade (10, 11). This observation suggests that a considerable portion of tumors characterized by T-cell infiltration remains unresponsive to ICB, and that barriers other than PD-1 signaling preventing T-cell activation may exist. Hence, novel strategies are required to restore effector activity in tumor-specific T cells that are not susceptible to current ICB.

Interleukin-2 (IL2) is a powerful cytokine inducing the activation, differentiation, and proliferation of lymphocytes (12). Notably, IL2 appears critical for the effects of PD-1 blockade, as inhibition of IL2 signaling impeded T-cell activation and tumor control following PD-1 blockade in vitro and in vivo (13, 14). Furthermore, we and others previously demonstrated that the addition of IL2 can overcome ICB resistance both ex vivo in human tumors and in murine tumor models (15–17), indicating its potential to improve immunotherapeutic outcomes. Systemic administration of IL2 was the first immunotherapy to be approved for the treatment of metastatic melanoma and renal cell carcinoma (18); however, its clinical utilization has been challenging because of IL2’s pleiotropic functions. The IL2 receptor (IL2R) consists of three receptor subunits—IL2Rα, IL2Rβ, and IL2Rγ—that together form either the intermediate-affinity dimeric IL2Rβγ present on naïve and memory T cells, and NK cells, or the high-affinity trimeric IL2Rαβγ, which is strongly expressed on regulatory T cells (Treg) and transiently on activated T cells (19, 20). When given at low doses, IL2 thus preferentially induces Treg activation, which counteracts effector T-cell activity (21, 22). At higher doses, IL2 can elicit immune activation and tumor control (23) but also lead to severe systemic toxicities such as inflammation and vascular leak syndrome (24, 25). Although the engineering of IL2 variants (IL2v) with reduced or abolished IL2Rα binding with the goal of favoring Teff over Treg activation (26–30) has improved toxicity issues, these molecules did not show meaningful clinical efficacy (31). These observations suggest that nonspecific targeting of all IL2Rβγ-expressing cells does not provide the necessary therapeutic window to successfully exploit the clinical potential of IL2. To overcome these problems, cis-targeting of IL2 variants to T-cell-surface receptors has emerged as a powerful new strategy, making it possible to deliver the IL2 signal specifically to the T-cell subset of interest.

We here explored a novel IL2 molecule targeted to CD8β (CD8–IL2) that allows the selective stimulation of the CD8⁺ T cells critical for tumor control, while minimizing the activation and amplification of immunosuppressive Tregs and of other cell types such as NK cells that may contribute to systemic toxicity (32). Utilizing the patient-derived tumor fragment (PDTF) platform (10, 33) that provides a tumor microenvironment (TME) context that closely resembles the one present in patients with cancer, we assessed (i) whether CD8–IL2 can induce T-cell activation in human tumors and which specific cell types and cell states are responsive to this treatment; (ii) how CD8–IL2 reinvigorates antitumor T-cell immunity, and (iii) how T-cell reactivation upon CD8–IL2 differs from that achieved upon PD-1 blockade. Our data demonstrate that CD8–IL2 broadly arms tumor-infiltrating CD8⁺ T cells with effector capacity. Restoration of effector function, leading to potent immune reactivation, was most profound for the tumor-reactive dysfunctional T-cell pool and required simultaneous T-cell receptor (TCR) signaling. Furthermore, reinvigoration of dysfunctional T cells by CD8–IL2 was broader and qualitatively superior compared with anti–PD-1 in the same tumors. Finally, CD8–IL2 also induced functional T-cell reinvigoration in tumors that are resistant to PD-1 blockade, suggesting the presence of a tumor-specific T-cell pool that either cannot be reinvigorated by PD-1 blockade alone or is distinct from the one susceptible to reprogramming by PD-1 blockade.

CD8⁺ T cells are key effectors of antitumor immunity (3, 34, 35), and strategies that offer the possibility to specifically boost the function of antigen-activated CD8⁺ T cells are thus of value. We here explored a novel IL2 mutein cis-targeted to the CD8β chain (CD8–IL2, also called AB248) allowing the selective delivery of an immunostimulatory IL2 signal to CD8⁺ T cells. To avoid Treg expansion and antigen-independent T-cell activation, CD8–IL2 was designed to contain an IL2 mutein with no detectable IL2Rα binding and reduced binding to the signaling IL2Rβγ complex (32). To evaluate the specificity of the molecule, we stimulated single-cell digests of human tumors with different concentrations of either CD8–IL2, human recombinant IL2 (hrIL2), or an untargeted not-alpha IL2 variant (IL2v). Assessing the induction of the proliferation marker Ki-67 as a direct IL2 effect, we observed that, whereas both IL2v and hrIL2 induced Ki-67 expression in tumor-infiltrating Tregs and conventional CD4⁺ T (Tconv) cells, CD8–IL2 did not activate these cell types, even at the highest concentration tested (Supplementary Fig. S1A). Moreover, CD8–IL2 induced Ki-67 expression in CD8⁺ T cells at 1,000-fold lower concentrations as compared with hrIL2 and IL2v, demonstrating its potency. Analysis of T-cell expansion over 5 days confirmed that Tregs expanded significantly less following CD8–IL2 treatment as compared with IL2v treatment (Supplementary Fig. S1B). Interestingly, Tregs in human tumors expressed high levels of not only the IL2Rα subunit but also of the β-subunit relative to CD8⁺ T cells (Supplementary Fig. S1C), providing a further incentive for CD8 targeting to achieve CD8⁺ T-cell specificity.

Next, we aimed to address the potential of CD8–IL2 to reinvigorate tumor-infiltrating CD8⁺ T cells in their native tissue context, using a recently developed PDTF platform (10, 33). PDTFs are 3D fragments of human tumor tissue preserving the TME contexture outside of the patient and allowing short-term ex vivo culture in the absence or presence of immunotherapies. Of note, PDTFs display immune activation upon ex vivo PD-1 blockade that is highly correlated with the clinical response of the patient to the same treatment (10), underlining the relevance of such early immune activation in situ. To examine the effect of selective delivery of IL2 to CD8⁺ T cells on intratumoral T-cell activity, PDTFs from 23 human tumors were cultured in the absence or presence of CD8–IL2 for 48 hours (Fig. 1A). The PDTFs analyzed comprised five different cancer types, renal cell carcinoma (RE), ovarian carcinoma (OV), melanoma (MEL), non-small cell lung cancer (LU), and breast cancer (BR), and were enriched for immune-infiltrated tumors (Supplementary Fig. S1D and S1E; Supplementary Table S1). Ex vivo treatment with CD8–IL2 induced markers associated with cytotoxicity (granzyme B), proliferation (Ki-67), and activation (PD-1, CD137) in CD8⁺ T cells in the vast majority of tumors (Fig. 1B and C; Supplementary Fig. S2). In contrast, stimulation with anti-CD8β alone did not induce any of the observed changes (Supplementary Fig. S3A). Notably, CD8–IL2 also increased IFNγ production in CD8⁺ T cells (Fig. 1D), suggesting their functional reinvigoration.

The CD8-targeted IL2 mutein used here binds to the IL2Rβ/γ subunits, which also form a receptor for IL15. As similarities in IL2 and IL15 signaling have been observed (36), we compared the effects of CD8–IL2 and IL15 in PDTFs. As expected, we observed similar activation of CD8⁺ T cells upon both stimulations (Supplementary Fig. S3B). However, IL15 also strongly activated NK cells (Supplementary Fig. S3C), which can mediate toxicity but are dispensable for antitumor activity (refs. 32, 37, 38), highlighting the value of directing cytokines toward specific cell types.

Intratumoral T cells can acquire diverse phenotypes, and particularly dysfunctional T cells form a gradient of states, ranging from dysfunctional precursor cells with capacity for self-renewal to terminally differentiated late dysfunctional T cells (39–41). Recent reports in murine models described that IL2 molecules that are cis-targeted to PD-1⁺ T cells (16, 42) or that are provided in combination with PD-1 blockade (15) may mainly drive the differentiation of early dysfunctional T cells into functional effectors. To understand to what extent this observation would translate to human tumors, we analyzed the effects of CD8–IL2 on CD8⁺ T cells with distinct differentiation states. Using CD39 as a marker to differentiate CD8⁺ T cells with early (PD-1⁺CD39⁻) and late (PD-1⁺CD39⁺) dysfunction (Fig. 1E), we observed that CD8–IL2 increased granzyme B and Ki-67 expression in both subsets. Conversely, CD137 was mainly increased in PD-1⁺CD39⁺ T cells, suggesting differences in the reinvigoration of the early and late dysfunctional T-cell pools (Fig. 1F). Supporting this notion, only PD-1⁺CD39⁺ T cells showed increased degranulation after 6 hours of CD8–IL2 treatment (Supplementary Fig. S3D), suggesting that despite the increase in granzyme B expression in both subsets, only late dysfunctional T cells may acquire cytolytic immune function.

Given that T cells upregulate PD-1 on their surface after antigen encounter, previous studies have explored PD-1 cis-targeting of IL2 for the selective activation of the antigen-activated T-cell pool (16, 42). In our data set, on average 80% of CD8⁺ T cells also expressed PD-1, suggesting that most of the cells would be targeted by both approaches (Supplementary Fig. S3E). In contrast, only half of the PD-1⁺ T cells were CD8⁺, suggesting that PD-1–targeted compounds may also exert effects on other cells including Tregs (Supplementary Fig. S3F). As a comparison, CD8β-expressing cells were almost exclusively CD3⁺ T cells (Supplementary Fig. S3G). To directly compare IL2 targeting to either CD8 or PD-1, we treated PDTFs from seven tumors with CD8–IL2 or an IL2v fused to PD-1 (PD1–IL2). Both CD8–IL2 and PD1–IL2 induced Ki-67 expression in CD8⁺ T cells; however, PD1–IL2 showed a lower effect in 5 of 8 tumors (Supplementary Fig. S3H). To understand whether similar CD8⁺ T-cell populations are activated, we compared the phenotype of Ki-67⁺ and Ki-67⁻ cells following both treatments. We found that CD8–IL2 and PD1–IL2 similarly induced the proliferation of CD8⁺ T cells expressing CD39 and CD103 (Supplementary Fig. S3I), indicating that both targeting strategies activate a subset that is likely to be tumor-reactive (43, 44). In line with the observed PD-1 expression by intratumoral Tregs, we saw an increase in this subset following ex vivo PD1–IL2 but not CD8–IL2 treatment (Supplementary Fig. S3H). Jointly, these data suggest that CD8–IL2 and PD1–IL2 treatments are comparable in their effects on CD8⁺ T cells in human tumor tissue, but PD1–IL2 treatment may result in a higher fraction of Tregs, potentially limiting antitumor immunity.

We previously found that immune reactivation, here defined by the induction of proinflammatory cytokines and chemokines, in PDTFs is a strong predictor of clinical response to anti–PD-1 (10). Therefore, we next assessed whether intratumoral T-cell activation by CD8–IL2 translated into such a functional cytokine response. Ex vivo CD8–IL2 treatment induced production of multiple soluble mediators including granzyme B, TNFα, IFNγ, or its downstream chemokines in part of the tumors (Fig. 2A). To distinguish tumors with or without such a CD8–IL2-induced immunologic response, we quantified the overall change in soluble mediator activity by cumulating the normalized delta values between the treated and untreated condition for each parameter. Based on this analysis, tumors were segregated into CD8–IL2 responders (CD8–IL2-R, 13/23) and CD8–IL2 nonresponders (CD8–IL2-NR, 10/23), distinguished by high and low cumulative scores, respectively (Fig. 2A–C). To delineate the mechanisms under-lying the heterogeneity in immunologic reinvigoration, we next explored whether CD8–IL2-R and -NR tumors differed in baseline tumor properties or the extent and quality of T-cell activation. Profiling the TME composition of each tumor by flow cytometry revealed that, with the exception of CD4⁺ T cells, the main immune populations were not substantially different between responding and nonresponding groups (Supplementary Fig. S4A). CD8–IL2-R tumors contained significantly fewer CD4⁺ Tconv cells in favor of more Tregs, i.e., of a CD39⁺ subset associated with high suppressive capacity (45, 46). We did not observe any difference in IL2Rβ/γ expression on CD8⁺ T cells, and PD-L1 expression on myeloid and tumor cells, respectively. MHC class I expression on cancer cells was higher in CD8–IL2-R, though without reaching statistical significance (Supplementary Fig. S4A). Based on our observation that CD8⁺ T cells proliferating upon CD8–IL2 expressed a dysfunctional profile, we assessed the baseline frequency of dysfunctional (PD-1⁺CD39⁺CD103⁺) and memory (PD-1⁺IL7R⁺) T-cell populations. The CD8⁺ compartment in responding tumors harbored significantly more dysfunctional cells as compared with nonresponding tumors, which conversely were enriched for memory T cells (Fig. 2D). CD8⁺ T cells in responding tumors also showed higher production of CXCL13 (Fig. 2D), another property associated with tumor reactivity (47–49).

Altogether, these data suggest distinct CD8⁺ T-cell landscapes in CD8–IL2-R and NR tumors. Thus, we next aimed to understand whether immunologic response to CD8–IL2 is associated with qualitative differences in T-cell activation. Surprisingly, Ki-67 and granzyme B expression were increased in both CD8–IL2-R and -NR upon treatment, suggesting comparable acquisition of markers associated with effector capacity and proliferation independent of cytokine production in the TME. By contrast, the induction of PD-1 and particularly CD137 expression was largely restricted to the CD8–IL2-R group (Fig. 2E). Notably, CD8–IL2-R tumors already displayed higher CD137 and Ki-67 expression at baseline (Supplementary Fig. S4B), compatible with ongoing tumor recognition in those tumors. This also corresponded with higher basal cytokine production in these samples (Supplementary Fig. S4C), suggesting that CD8–IL2 may reinvigorate a preexisting antitumor response. Based on these observations and since CD137 is upregulated shortly after TCR triggering (50), we hypothesized that simultaneous antigen recognition is required for cytokine responses to ex vivo CD8–IL2, but not for the induction of effector capacity. To test this, we pretreated eight CD8–IL2-R tumors with an Lck inhibitor (LCKi), which stalls signaling directly downstream of the TCR (Fig. 3A). Remarkably, Lck inhibition before CD8–IL2 treatment resulted in a significant decrease, and in some cases complete abolition, of cytokine and chemokine secretion, also evidenced by a marked reduction in the cumulative score (Fig. 3B–D; Supplementary Fig. S4D). Consistent with antigen recognition being essential for these responses, CD8⁺ T cells did not upregulate CD137 when TCR signaling was blocked before CD8–IL2 treatment. Strikingly, granzyme B, PD-1, and Ki-67 expression still increased despite the absence of TCR signaling (Fig. 3E). Similar observations were made when an MHC class I–blocking antibody was used instead of Lck inhibition (Supplementary Fig. S4E). These data imply that CD8–IL2 may induce effector capacity in T cells independent of them receiving a TCR signal, but that effector function occurs only in the presence of the latter. To test this hypothesis in a model in which the presence of a TCR stimulus can be experimentally controlled, we stimulated human PBMCs with CD8–IL2 and increasing concentrations of anti-CD3. In line with our observations in PDTFs, CD8–IL2 treatment alone was able to induce intracellular Granzyme B expression, implying that IL2 signaling is sufficient to promote T-cell effector capacity (Fig. 3F). Addition of CD8–IL2 to anti-CD3 stimulation further increased granzyme B expression compared with anti-CD3 alone in a dose-dependent manner. In contrast, the extracellular release of granzyme B was not induced by CD8–IL2 alone, but by anti-CD3 stimulation, and was further boosted when CD8–IL2 and anti-CD3 were combined (Fig. 3G). Similar effects were observed for IFNγ secretion, representing another important effector function of CD8⁺ T cells (Fig. 3G). Jointly, these results are compatible with a model in which CD8–IL2 broadly induces effector capacity in intratumoral CD8⁺ T cells but restores effect function only in the presence of simultaneous antigen recognition.

To understand how CD8–IL2 treatment reactivates intratumoral CD8⁺ T cells, we performed single-cell RNA and TCR sequencing (scRNA- and TCR-seq) of PDTFs in the absence or presence of CD8–IL2. To benchmark the observed alterations, we cultured additional fragments from the same tumors with a number of other stimulations including CD8–IL2 + LCKi, IL2v, anti–PD-1, anti–PD-1 + CD8–IL2, anti-CD3, and anti-CD3 + CD8–IL2. Our cohort included five tumors (RE, N = 2; OV, N = 1; MEL, N = 2) with distinct immunologic responses to CD8–IL2 and anti–PD-1, respectively, based on previous experiments (Fig. 2A; Supplementary Fig. S5A). Comparable immune cell numbers were isolated and sequenced for each treatment condition per tumor (Supplementary Fig. S5B). Clustering analysis distinguished the main immune cell populations, including CD4⁺ T cells, CD8⁺ T cells, B cells, γδ T cells, and NK cells, which were consistently identified across all tumors and conditions (Supplementary Fig. S5C–S5F). As CD8–IL2 exclusively activates CD8⁺ T cells, we focused our further analysis on this subset. Treatment conditions containing anti-CD3 stimulation were not included in the subsequent analysis but were kept as an independent reference for TCR-induced activation (see below). Clustering of the intratumoral CD8⁺ T-cell pool from the remaining conditions yielded data on 8,712 cells. As patient-specific clusters were observed, data integration was performed leading to a more homogenous representation of clusters across patients without overcorrection of sample expression profiles (Supplementary Fig. S6A and S6B). Next, we performed a more detailed characterization of the distinct CD8⁺ T-cell clusters. To this end, clusters were annotated using differential gene-expression analysis, revealing 13 distinct CD8⁺ states including naïve-like (Naïve_c1_BNIP, Naïve_c2_RP), memory (Mem_c1_IL7R, Mem_c2_CAPG), transitional (Trans_KLRG1), dysfunctional (Dys_c1_TOX, Dys_c2_HLA-DRB1), cycling (Cycl_c1, Cycl_c2, Cycl_c3), effector (Eff_FCGR3A), NK-like (NK-like_KLRC1), and IFNγ-sensing (IFN_ISG15) clusters (Fig. 4A; Supplementary Fig. S6C). These annotations were further validated by overlaying previously described gene-expression signatures for effector (51), memory (51), exhausted (52), and proliferating (53) states on the UMAP (Fig. 4B). We next assessed canonical CD8⁺ T-cell markers previously described in scRNA-seq studies of human intratumoral T cells (54) to further characterize the distinct clusters (Fig. 4C). The Mem_c1_IL7R and Mem_c2_CAPG states showed similar expression of T-cell memory markers; however, the Mem_c2_CAPG cluster also expressed genes associated with cytotoxicity and effector differentiation (GZMA, GZMB, PRF1, and CXCR3), suggesting that it may represent a more activated state of Mem_c1. Similar patterns were observed in the two dysfunctional clusters, with Dys_c2_HLA-DRB1 displaying a highly activated state as compared with Dys_c1_TOX, with increased expression of granzyme gene family members, IFNG, TNFRSF9, TNFRSF18, and HLA class II genes. The Trans_KLRG1 cluster expressed low levels of inhibitory receptors and effector molecules and was positive for TCF7, in line with a more precursor-like state. Interestingly, the three cycling clusters showed distinct profiles, with Cycl_c2 displaying a more memory-like phenotype and lower expression of dysfunction markers as compared with Cycl_c1 and Cycl_c3. Moreover, Cycl_c3 contained mostly cells in the G₂–M phase of the cell cycle, in contrast to Cycl_c1 and Cycl_c2, which corresponded to the S phase (Supplementary Fig. S6D).

To understand how the identified cell states relate to treatment, we first compared cluster distribution between untreated and CD8–IL2-treated PDTF conditions (Fig. 4D; Supplementary Fig. S7A). Notably, five clusters newly formed or substantially increased upon CD8–IL2, comprising all three cycling clusters (Cycl_c1, Cycl_c2, Cycl_c3), as well as the activated Mem_c2_CAPG and Dys_c2_HLA-DRB1 clusters. Conversely, the Mem_c1_IL7R and Dys_c1_TOX clusters decreased, suggesting that CD8–IL2 treatment may trigger the transcriptional reconfiguration of CD8⁺ T cells shifting them toward more activated states. Underscoring this notion, we found that the newly formed states upon CD8–IL2, including the dysfunctional cluster, showed increased expression of a transcriptional signature derived from effector T cells generated during acute infection (ref. 53; Supplementary Fig. S7B and S7C). We next compared the CD8–IL2-induced changes to those elicited by an IL2v lacking the CD8-targeting moiety. At the same concentrations, IL2v induced minimal shifts in clusters compared with CD8–IL2 (Supplementary Fig. S7D), in line with our flow data (Supplementary Fig. S1A) and prior reports that untargeted IL2Rβγ agonists fail to induce better effector cells (42).

As a next question, we aimed to identify from which cell states the CD8–IL2-induced clusters originate. Therefore, we first conducted label transfer analysis of the proliferating Cycl_c1 and Cycl_c2 clusters (Cycl_c3 contained too limited cell numbers for this analysis; Fig. 4E). This revealed that cells from Cycl_c1 were derived mainly from dysfunctional and transitional states, with Dys_c2_HLA-DRB1 contributing most, whereas Cycl_c2 comprised multiple cell states including naïve, transitional, dysfunctional, memory, and NK-like states. As an alternative approach to reconstruct developmental relationships between the clusters, we performed diffusion pseudotime (DPT) analysis (56). Ordering the clusters along DPT revealed that naïve clusters were early in pseudotime, followed by memory, transitional, and dysfunctional states, consistent with the dysfunctional differentiation gradient of CD8⁺ T cells in the TME (Fig. 4F and G). The effector and NK-like states were situated between the naïve-like and dysfunctional states and appeared to form a separate trajectory. The IFN_ISG15 cluster, likely formed by cells responding to the secreted IFNγ, appeared later in pseudotime. In line with their treatment-induced differentiation, the Mem_c2_CAPG and Dys_c2_HLA-DRB1 clusters were later in DPT than their matched Mem_c1_IL7R and Dys_c1_TOX states that were predominant in untreated samples. Notably, the three cycling clusters exhibited differences in pseudotime, with Cycl_c2 appearing earlier and more spread over DPT compared with the other two cycling clusters, supportive of distinct trajectories as suggested by the label transfer analysis.

To further corroborate these findings, we parsed the TCR-seq data to connect clonal information with the individual cellular phenotypes. TCR clonal expansion per cluster was lowest in the naïve clusters, increased slightly in memory and transitional clusters, and reached the highest level in the effector and dysfunctional states (Fig. 4H and I), consistent with prior studies and potential tumor recognition by these clones (57–59). Again, we observed distinct expansion patterns among the cycling clusters, with a high fraction of expanded clones in Cycl_c1 and Cycl_c3, and lower TCR expansion in Cycl_c2, in line with the origin of the latter from multiple clonally diverse cell states. To further probe relationships between clusters, we next determined clonal overlap between the CD8⁺ states (Fig. 4J). Focusing on the five cell states emerging upon CD8-IL2, we observed low-level TCR sharing of Mem_c2_CAPG and Cycl_c2 with the Mem_c1_IL7R and Trans_KLRG1 clusters. By contrast, Dys_c2_HLA-DRB1, Cycl_c1, and Cycl_c3 states showed substantial clonal overlap among themselves, with Dys_c1_TOX, and to a lesser extent with the Trans_KLRG1 cluster, suggesting a separate differentiation trajectory induced within the dysfunctional axis.

Taken together, the above observations are compatible with the model suggested by our flow analyses, in which CD8–IL2 can broadly activate tumor-infiltrating CD8⁺ T cells, but may induce distinct rewiring of tumor-specific T cells that also receive a TCR stimulus. To connect the single-cell data to our earlier flow experiments, we overlayed transcript levels of TNFRSF9 (CD137), MKI67 (Ki-67), GZMB (granzyme B), and IFNG (IFNγ) on the UMAP (Fig. 5A). Expression of these transcripts enriched mostly in cluster areas induced upon CD8–IL2 treatment but not in those already present in unstimulated samples, in line with CD8–IL2-induced effector capacity. MKI67 expression was logically restricted to the cycling clusters. Of note, we found distinct expression patterns of GZMB, TNFRSF9, and IFNG between the activated clusters. In agreement with our earlier protein data, GZMB was highly expressed in all induced clusters, whereas IFNG and especially TNFRSF9 expression were more concentrated in cells of the dysfunctional axis, compatible with the latter also receiving TCR-dependent signals.

To visualize changes in the distribution of the most expanded TCRs upon treatment, we performed clonal overlay analysis on the cell states derived from untreated and CD8–IL2-treated PDTFs (Supplementary Fig. S7E). In the untreated condition, the most expanded TCRs were present in the Dys_c1_TOX cluster, whereas upon CD8–IL2 treatment they enriched in the Dys_c2_HLA-DRB1, Cycl_ c1, and Cycl_c3 clusters, supporting the notion that the activation observed in these clusters is dependent on IL2R stimulation in the context of TCR triggering. To test this hypothesis, we compared the effects of CD8–IL2 with a CD8–IL2 + LCKi condition tested in the same tumors (Supplementary Fig. S5A). DPT analysis of CD8⁺ T cells from the untreated, CD8–IL2, and CD8–IL2 + LCKi conditions revealed that TCR blockade decreased the overall progression of cells in pseudotime, implying that they were unable to undergo full activation and differentiation without receiving a TCR stimulus (Supplementary Fig. S7F). Plotting the cell densities from those conditions on the UMAP illustrated that CD8–IL2 still induced new cell states even when TCR signaling was blocked (Fig. 5B). Specifically, the Mem_c2_CAPG and Cycl_c2 clusters appeared to be induced largely independent of TCR signaling, whereas the formation of Dys_c2_HLA-DRB1, Cycl_c1, and Cycl_c3 showed varying dependence on TCR triggering (Fig. 5C). The remaining clusters were mostly unaffected by TCR blockade (Supplementary Fig. S7G). To further probe the influence of IL2R and TCR signals on these clusters, we performed label transfer of cell states from the CD8–IL2 condition on CD8⁺ T cells from anti-CD3-or anti-CD3 + CD8–IL2-treated PDTFs, a setting in which all T cells equally receive a TCR stimulus (note that these conditions were excluded from the clustering). Upon anti-CD3 stimulation, a small fraction of CD8⁺ T cells acquired the activated Dys_c2_HLA-DRB1 or Cycl_c1 states, whereas the Mem_c2_CAPG and Cycl_c2 clusters were only marginally induced, confirming the latter states’ dependency on IL2. In contrast, all five activated cell states emerged upon anti-CD3 + CD8–IL2 encompassing >70% of CD8⁺ T cells in this condition (Supplementary Fig. S7H), indicating that at least part of these states require both IL2 and TCR signals.

To understand which of these CD8⁺ states have tumor recognition capacity, we sorted dysfunctional, memory-like, effector-like, and transitional subsets from four tumor samples based on key transcriptional markers (Fig. 5D and E). The respective cell states of the sorted subsets were confirmed by another scRNA-seq experiment (Supplementary Fig. S8A). Moreover, we validated that the subsets identified by these markers recapitulate the distinct patterns of CD137 and granzyme B expression in dysfunctional compared with other cell states upon CD8–IL2 (Supplementary Fig. S8B). We next expanded the different subsets and cocultered them with autologous tumor digest with or without CD8–IL2 (Fig. 5F; Supplementary Fig. S8C and S8D). In line with clonal expansion patterns, tumor recognition, as measured by CD137 expression, degranulation, and IFNγ/TNFα production, was largely restricted to the dysfunctional subsets in all four tumors with the exception of one tumor in which also T cells expanded from transitional cells showed reactivity (OV054-1). The addition of CD8–IL2 to the cocultures did not increase degranulation or cytokine production, as these cells probably already acquired maximal effector capacity during expansion in high-dose IL2. Of note, CD8–IL2 still increased the expression of CD137 in dysfunctional and transitional T cells upon tumor recognition (Fig. 5G), suggesting an additional effect via enhancing TCR signaling. Altogether, these data confirm that tumor reactivity is mostly restricted to T cells with dysfunctional phenotypes, and suggest a mechanism in which CD8–IL2 may promote reinvigoration both via TCR-dependent and -independent signals.

To further uncouple the effects induced by either IL2R or TCR signaling, we devised two specific gene programs (Supplementary Fig. S9A; Supplementary Data S1). Reasoning that T cells exposed to CD8–IL2 in the presence of Lck inhibition should only receive IL2 and no TCR signals, we constructed an IL2 program based on differential gene-expression (DGE) analysis between the untreated and CD8–IL2 + LCKi conditions (Supplementary Fig. S9B). Similarly, a TCR program was developed based on DGE between cells from PDTFs that were either untreated or stimulated by anti-CD3 (Supplementary Fig. S9C). First, we assessed the expression of the two programs across conditions. Notably, the IL2 program was increased in both the CD8–IL2 and CD8–IL2 + LCKi conditions, whereas the TCR program was specifically enriched following CD8–IL2 and not when the LCKi was coadministered (Fig. 6A). Vice versa, the increase in the IL2 program upon anti-CD3 stimulation was only marginal, implying that TCR signaling alone is not sufficient for its induction (Supplementary Fig. S9D). Interestingly, adding CD8–IL2 to anti-CD3 further increased the expression of the TCR program (Supplementary Fig. S9D), supporting the notion that IL2 may enhance TCR-mediated signals.

Overlaying both signatures on the UMAP confirmed that the IL2 program was broadly induced across cell states upon CD8–IL2 treatment, whereas the induction of the TCR program was restricted to the dysfunctional axis (Fig. 6B), aligning with the previously observed tumor reactivity patterns. DPT analysis of the two gene programs indicated that the IL2 program continuously increased with pseudotime, compatible with the CD8–IL2-induced clusters arising later in pseudotime (Fig. 6C). Notably, the increase of the IL2 program over pseudotime was unaffected by TCR blockade. In contrast, the TCR program was only acquired by fully differentiated cells at the end of pseudotime and was abolished upon blockade of TCR signaling. Assessing the gene programs across clusters, we found—in line with tumor reactivity patterns—that the IL2 program was active in all CD8–IL2-induced clusters, whereas the TCR program was mostly confined to the dysfunctional, Cycl_c1 and Cycl_c3 clusters (Fig. 6D and E). Interestingly, corresponding with the protein expression data, the TCR program was to some extent already detectable in untreated samples particularly in the Dys_c1_TOX cluster (Fig. 6C–E), implying ongoing TCR recognition in dysfunctional T cells that likely translates into functional immune reactivation upon reinvigoration by CD8–IL2.

In recent work, antigen-specific CD8⁺ T cells in murine cancer and chronic infection models have been found to acquire a “better effector” state after PD-1–targeted IL2 treatment (42). Comparison of this “better effector” state to the IL2 and TCR programs showed that the “better effector” signature slightly increased in the activated memory and dysfunctional states that emerged upon CD8–IL2 treatment, but not in the cycling clusters (Fig. 6E; Supplementary Fig. S9E and S9F). Additionally, effector and NK-like T-cell subsets are strongly enriched for the “better effector” signature, likely due to the high expression of effector genes characterizing these states. Thus, in human tumors, the “better effector” signature may either reflect a more general effector program that can be acquired by CD8⁺ T cells or a cell state that emerges later upon IL2 treatment.

As a final experiment to understand whether the global transcriptional reprogramming upon CD8–IL2 was independent of immunologic response, we performed a separate analysis of the one tumor (RE027) in our cohort that did not show an immunologic response to CD8–IL2 (Figs. 2A and 3A), reasoning that the T cells in this tumor should display IL2-induced changes that occur independent of concurrent TCR signaling. Analysis of IL2 and TCR program dynamics revealed that the IL2 signature increased over pseudotime independently of TCR signaling in all tumors (Supplementary Fig. S10A). In contrast, the TCR program was not detectable in RE027, both in untreated and treated conditions, suggesting the absence of tumor recognition (Supplementary Fig. S10B). Accordingly, a comparison of individual clusters from each tumor revealed consistent IL2 program scores across induced clusters for all tumors, whereas TCR program scores were noticeably lower in the dysfunctional and Cycl_c1 and Cycl_c3 clusters in RE027 (Supplementary Fig. S10C). Supporting the lack of a tumor-reactive repertoire in RE027, this tumor displayed substantially distinct clonal expansion patterns with most expanded TCRs found in the memory and effector, rather than dysfunctional cell compartments (Supplementary Fig. S10D). Additionally, by directly comparing individual expanded T-cell clones present in both the untreated and CD8–IL2-treated conditions, we observed that TCRs in tumors with immune reactivation upon CD8-IL2 transitioned into the Dys_c2_HLA-DRB1 or cycling clusters, whereas this transition could not be observed in the nonresponding RE027 tumor (Supplementary Fig. S10E). Collectively, these results support a mechanism in which CD8–IL2 profoundly rewires tumor-infiltrating T cells, endowing them with effector capacity, and thereby facilitating the reinvigoration of tumor-specific dysfunctional T cells to unleash potent antitumor immunity.

To understand how CD8–IL2 treatment may alter the TME beyond the CD8⁺ T-cell pool, we performed PDTF cultures of six responsive tumors for 24 hours instead of 48 hours, allowing better preservation of the innate immune compartment. After confirming that CD8–IL2 already activates CD8⁺ T cells at this time point (Fig. 6F), we assessed the expression of activation and maturation markers on Tconv, Treg, NK cell, B cell, and myeloid cell compartments (Fig. 6G). Aside from a small increase in CD86 expression on B cells in 4 of 6 tumors, no consistent phenotypic changes were observed across any of these subsets. To evaluate potential transcriptional changes, we performed a DGE analysis of all non-T immune cells between CD8–IL2-treated and untreated PDTFs. A limited number of genes was significantly upregulated upon CD8–IL2, including the IFN-inducible genes IFITM1, IFIT3, and CCL5 (Fig. 6H). We thus hypothesized that the IFNγ produced by CD8⁺ T cells upon reinvigoration may be sensed by other immune cells. To test this, we assessed the expression of IFN-responsive genes (60), which confirmed increased IFN-sensing upon CD8–IL2, both within the whole non-T immune cell compartment and specifically within B cells that formed the largest subpopulation (Fig. 6I). To directly test whether the alterations in the non–T-cell immune compartment are dependent on IFNγ, we pretreated PDTFs with an anti-IFNGR antibody before CD8–IL2 treatment. Although CD8⁺ T-cell activation remained unchanged by IFNGR blockade, a number of soluble mediators were no longer produced when IFNγ signaling was inhibited, particularly affecting IFN-downstream chemokines such as CXCL9, CXCL10, and CXCL11 (Fig. 6J and K). Collectively, these data suggest that, through IFNγ, TME-resident immune cells may act as amplifiers of CD8–IL2-mediated T-cell reinvigoration.

PD-1 signaling is a major inhibitory brake that prevents the effector function of tumor-specific T cells. However, not all T-cell–infiltrated tumors respond to PD-1 blockade. Therefore, we next compared CD8–IL2 and PD-1 blockade for their capacity to reactivate antitumor immunity. To this end, PDTFs from 13 tumors were treated with either CD8–IL2, anti–PD-1, or CD8–IL2 + anti–PD-1. Hierarchical clustering of soluble mediators and cumulative score calculations discriminated samples with the absence or presence of a downstream immune response (Fig. 7A). Of note, immunologic responses were observed in all three treatment groups. We next evaluated the number of tumors exhibiting ex vivo responses to the different treatments (Fig. 7B). Among the 13 tumors, 4 (31%) responded to anti–PD-1, and interestingly, all these tumors also showed a response to CD8–IL2. Notably, two tumors that did not respond to anti–PD-1 treatment displayed a response to CD8–IL2 (2/13; 15%). Upon combination treatment, additional two tumors that were resistant to either anti–PD-1 and CD8–IL2 monotherapy became immunologically responsive (2/13; 15%), indicating that overall CD8–IL2 converted 44% (4/9) of the anti–PD-1 nonresponders.

As all anti–PD-1 responders also responded to CD8–IL2, we next investigated whether the two treatments resulted in distinct or similar response patterns. To this end, we compared the changes in individual soluble mediators in ex vivo responders (cumulative score >0) to each treatment. Although a large fraction of mediators were induced similarly by both treatments, some differences were noticeable (Fig. 7C). Specifically, the induction of IFNγ and its downstream mediators seemed more prominent upon CD8–IL2 than anti–PD-1 treatment, suggesting potential differences in T-cell reactivation. Therefore, we next questioned how transcriptional reprogramming by anti–PD-1 compared with that by CD8–IL2. Among the five tumors included for scRNA-seq, three were also ex vivo responders to PD-1 blockade, evidenced by a positive anti–PD-1 response score which was previously found to be highly predictive for clinical response to PD-1 blockade (Supplementary Figs. S5A and S11A; ref. 10). Comparing the UMAP visualizations of the anti–PD-1 and CD8–IL2 conditions from the three dual responding tumors, we observed that the remodeling of cell states induced by anti–PD-1 was minor compared with that caused by CD8–IL2 (Fig. 7D), compatible with previous observations in mouse models (15, 42). By contrast, the anti–PD-1 + CD8–IL2 combination led to similar effects as CD8–IL2 alone. Comparing the prevalence of individual clusters between treatments, we found that anti–PD-1 promoted a slight expansion of the Dys_c2_HLA-DRB1, and minimally of the Cycl_c1 clusters, which were substantially increased by CD8–IL2 alone or in combination with anti–PD-1 (Fig. 7E). As a control, we assessed the two anti–PD-1 nonresponders, which did not show expansion of these clusters, suggesting a relation to the immunologic responses observed (Supplementary Fig. S11B).

Next, we aimed to assess whether the different treatments led to distinct transcriptional changes in the induced T-cell clusters. As no apparent changes in cluster distribution were observed between the CD8–IL2 and CD8–IL2 + anti–PD-1 conditions, we performed a DGE analysis between the two treatments. This revealed only two genes (GZMB and NSD2) to be significantly different (Supplementary Fig. S11C), suggesting that the prominent transcriptional effects of CD8–IL2 may make it difficult to capture additional anti–PD-1 effects. Similar observations have been made in mouse models, where the combination of PD-1 and IL2 led to equal effects as IL2 alone (15). Consequently, we focused on comparing transcriptional changes elicited by the two monotherapies. To understand whether the treatments reinvigorate T cells originating from different cell states, we evaluated TCR sharing between cells in the untreated with those in either the anti–PD-1 or CD8–IL2 conditions (Fig. 7F). In this analysis, the anti–PD-1-induced Dys_c2_HLA_DRB1 and Cycl_c1 clusters displayed the largest clonal overlap with the Dys_c1_TOX cluster, and to some extent with the Trans_KLRG1 cluster, suggesting reinvigoration of these cell states. Of note, the three clusters that were formed by CD8–IL2 in the presence of TCR signaling showed similar sharing patterns, with all exhibiting a strong overlap with the Dys_c1_TOX cluster. In addition, the Cycl_c1 cluster also displayed TCR sharing with the Trans_KLRG1 state, in line with the induction of proliferation of precursor-like T cells. Of note, the fraction of reinvigorated cells was substantially lower upon anti–PD-1 than upon CD8–IL2 (Fig. 7D and F). This was further corroborated when examining individual TCRs present in both the untreated and anti–PD-1 conditions, where only minor shifts in the cell state were observed (Supplementary Fig. S11D). Intriguingly, DPT analysis showed that dysfunctional T cells arising late in pseudotime displayed a clear increase in the TCR program upon anti–PD-1 treatment, suggesting that they may undergo reactivation despite largely maintaining their original dysfunctional state (Supplementary Fig. S11E).

To characterize the differential effects of PD-1 blockade and CD8–IL2 in more detail, we focused on the tumor-reactive CD8⁺ compartment. Therefore, we inferred tumor reactivity based on the expression of two published gene signatures (57, 59). Of note, the expression of these signatures strongly correlated with clonal expansion and tumor recognition capacity observed in the sorted subsets (Fig. 7G; Supplementary Fig. S11F). In untreated tumors, the highly scored tumor-reactive cells (top 20%) were predominantly found in the dysfunctional cluster Dys_c1_TOX (Fig. 7H). Conversely, the cells scoring either intermediate (21%–40%) or low (41%–100%) for these tumor reactivity signatures comprised a mix of various cell states with fewer or no dysfunctional T cells, respectively. Upon anti–PD-1 treatment, approximately one third of highly scored tumor-reactive T cells acquired the activated Dys_c2_HLA-DRB1 state, with a minor fraction moving to the proliferative Cycl_c1 cluster. In contrast, upon CD8–IL2 the vast majority of these cells underwent activation, shifting to the Dys_c2_HLA-DRB1, Cycl_c1, and Cycl_c3 phenotypes. A similar but slightly less prominent effect of CD8–IL2 was observed on cells with intermediate tumor reactivity scores, whereas lowly scored cells mainly acquired the TCR-independent Mem_c2-CAPG and Cycl_c2 states. In contrast, PD-1 blockade did not induce the activation or proliferation of cells with intermediate and low tumor reactivity scores. CD8–IL2 in contrast to anti–PD-1 also strongly reduced TOX expression in dysfunctional, particularly in cells with high tumor reactivity scores (Fig. 7I; Supplementary Fig. S11G), suggesting qualitative differences in reinvigoration. DGE analysis between highly scored tumor-reactive cells in the untreated and either treated condition further demonstrated transcriptional activation, including an increase in genes related to HLA class II expression, cytotoxicity, and proliferation. Of note, the induction of proliferation was largely limited to CD8–IL2, which also transcriptionally activated a larger part of the tumor-reactive T-cell pool as compared with anti–PD-1 (Supplementary Fig. S11H). Altogether, our findings suggest that anti–PD-1 and CD8–IL2 may induce a similar trajectory of activation in tumor-reactive T cells present within the human TME. However, CD8–IL2 may reinvigorate a broader tumor-specific T-cell pool and generate more potent effector cells, establishing CD8–IL2 as a promising strategy to unleash antitumor immunity in human cancers.

In this study, we performed a comprehensive analysis of T-cell reinvigoration by a novel CD8-targeted IL2Rβγ agonist in human cancer tissue. Thereby, we observed that CD8–IL2 broadly reactivates the intratumoral T-cell landscape and specifically enables the revival of dysfunctional T cells to mediate immune reactivation upon antigen encounter. Reinvigoration by CD8–IL2 was superior to that by PD-1 blockade and could even mobilize antigen-specific T cells in anti–PD-1–resistant tumors.

IL2 therapy has been identified as a promising strategy to enhance T-cell responses in ICB-resistant tumors (17, 61–63). However, the clinical value of systemic IL2 administration is limited by toxicity and immunosuppressive effects following Treg activation (64–66). We here demonstrate that targeting of IL2 to the CD8β-chain allows for selective and potent reactivation of CD8⁺ T cells, requiring 1,000-fold lower concentrations than recombinant IL2 or an untargeted IL2Rβγ agonist, thus making it an attractive strategy for systemic cancer therapy. Similar cis-targeting approaches have recently been explored in a number of murine studies (15, 16, 42). In these studies, IL2 was directed toward PD-1⁺ T cells reasoning that PD-1 expression may mark the tumor-reactive T-cell pool. Interestingly, when comparing CD8-and PD-1-targeting of IL2, we observed no substantial differences in CD8⁺ T-cell reactivation, which may be explained by the fact that the vast majority of CD8⁺ T cells in human tumors express PD-1. Of note, PD-1 is also present on Tregs, and treatment with PD1-IL2 accordingly increased their proportion in contrast to CD8–IL2, highlighting CD8-targeting as a valid strategy to selectively enhance antitumor effector responses.

One intriguing observation we made relates to the “duality” of the response induced by CD8–IL2. Although CD8–IL2 could induce T-cell activation in the majority of tumors across five cancer types, functional immune responses leading to the production of proinflammatory cytokines and chemokines were elicited in about 60% of tumors. Perturbation experiments revealed that these immunologic responses depended on simultaneous antigen recognition by the T cells as IL2 could not restore functional activity when TCR signaling was inhibited. Importantly, these observations also highlight that expression of effector markers such as granzyme B or IFNγ does not necessarily allow to infer functionality. A number of recent studies using engineered IL2 molecules to enhance antitumor or antiviral immunity in mouse models reported the induction of “better effector” cells similar to those generated in acute infections (15, 16, 42). Interestingly, this “better effector” state was different from the IL2-driven activation program that we identified early during reinvigoration and showed enrichment in both IL2-induced states and effector subsets already present in untreated tumors. Thus, the “better effector” program may either reflect a more general effector program acquired by T cells in human cancers or show a substantial increase only later during treatment. Supporting the latter notion, Moynihan and colleagues observed in the accompanying article that a cluster compatible with “better effectors” was not detectable after 2 days of CD8–IL2 treatment, but emerged on day 4 (32). The IL2 activation program in contrast was acquired by multiple CD8⁺ cell states early after CD8–IL2 treatment, suggesting a rapid and broad remodeling of the intratumoral T-cell landscape toward enhanced effector capacity. The induction of this program was, however, not sufficient to promote immunologic responses as tumors displaying cytokine production upon CD8–IL2 exhibited an additional activation trajectory prompted by antigen recognition and restricted to the dysfunctional axis. Increased clonal expansion and TCR sharing among the dysfunctional T-cell clusters as well as their enriched capacity for tumor recognition further supported the role of the dysfunctional pool in mediating CD8–IL2-induced immune responses. Although this is in sharp contrast to most studies in mouse models that observed the main effect of IL2 on stem-like T cells, a recent murine study investigating a PD-1–targeted low-affinity IL2 similarly reported preferential reinvigoration of PD-1⁺TIM3⁺ dysfunctional T cells (13). A possible explanation for these differences may be the distinct time points of analysis for these studies. Whereas IL2-induced expansion of stem-like cells was observed after 10 to 14 days of treatment, the reinvigoration of dysfunctional T cells observed by us and Ren and colleagues (13) was captured after 2 and 3 days, respectively. These data can be reconciled by a model in which the early response to IL2 is mainly mediated by tumor-specific dysfunctional T cells already located at the tumor site, followed by an influx of stem-like “resource” cells at later time points that sustain the response.

The observation that a large fraction of dysfunctional T cells can be reinvigorated is intriguing as they are often considered terminally exhausted and resistant to therapeutic reinvigoration. Of note, treatment of the same tumors with PD-1 blockade induced highly limited revival of dysfunctional T cells, in line with prior reports describing the accumulation of exhausted T cells upon anti–PD-1 likely due to an epigenetically fixed state that prevents major transcriptional rewiring (15, 42, 67, 68). These findings are also compatible with the recent observation in chronic LCMV models that Stat5, the major IL2R downstream signal, acts as an antagonist of Tox, a key transcription factor driving the epigenetic imprinting of exhaustion (69–72). Notably, induction of Stat5 activity using an orthogonal IL2:IL2Rβ pair led to increased functionality and reduced exhaustion, specifically decreasing Tox expression in CD8⁺ T cells with established dysfunction (73). This was accompanied by partial reprogramming of the epigenetic landscape, leading to increased accessibility of effector-related (Ifng, Gzma, Gzmb, Prf1, and Fasl) and proliferation-related (Mki67) peaks. Compatible with these observations, we observed that tumor-specific dysfunctional T cells strongly reduced TOX expression and acquired effector biology upon CD8–IL2. It will therefore be of interest to assess in future studies whether IL2-mediated signals can promote epigenetic reprogramming of dysfunctional T cells also in human cancers. Interestingly, despite the limited change in cell state, we observed that dysfunctional T cells acquired a TCR-driven activation program upon PD-1 blockade, which is also related to immune reactivation of the same tumors, suggesting that these cells can regain some degree of functionality. CD8–IL2 activated dysfunctional T cells along a similar transcriptional trajectory as anti–PD-1, however with more and qualitatively superior effectors as evidenced by a higher IFNγ response in the PDTFs.

The broad efficacy of CD8–IL2 across cancer types, the observed immunologic responses in ex vivo anti–PD-1 resistant tumors, and the potential synergistic effects of combining CD8–IL2 with anti–PD-1 posit this therapy as an attractive new treatment strategy for patients with cancer. However, these results also raise new questions about the interplay between the two therapies and how CD8–IL2 exerts activity in anti–PD-1–resistant tumors. Conceivable explanations are that resistant tumors contain a tumor-reactive T-cell pool that is either not susceptible to anti–PD-1 or for which PD-1 signaling is not the sole barrier preventing activation. However, further research is required to elucidate how PD-1 blockade and CD8–IL2 may synergize and be optimally combined for clinical application.

In conclusion, our study provides comprehensive insights into the therapeutic potential of CD8–IL2 and its ability to reinvigorate the intratumoral CD8⁺ T-cell landscape. Our findings are compatible with a model in which CD8–IL2 can broadly “arm” T cells with effector capacity, enabling particularly the dysfunctional T-cell pool to undergo functional reactivation upon target antigen encounter. Finally, this study also highlights the precision targeting of immunomodulatory signals to specific immune cell populations as an attractive strategy to induce effective antitumor immunity.

Human tumor tissue was collected from patients with cancer undergoing surgery for non–small cell lung cancer, ovarian cancer, melanoma, renal cell carcinoma, and breast cancer at the Netherlands Cancer Institute/Antoni van Leeuwenhoek Hospital (NKI-AVL) between April 2017 and December 2022 (Supplementary Table S1). The study was approved by NKI-AVL’s institutional review board (CFMPB484) and performed in compliance with all ethical regulations. Patients of whom material was collected for the study consented in writing either by opt-out procedure or via prior written informed consent (after May 2018) to their tissue being available for research after diagnostic procedures have been completed.

Resected human tumor tissue was collected on ice in RPMI-1640 (Thermo Fisher Scientific) supplemented with 2.5% fetal bovine serum (FBS; Sigma-Aldrich) and 1% penicillin–streptomycin (Roche). Directly after collection, tumor tissues were dissected into fragments (PDTF) of 1 to 2 mm³ on ice. Processed tumor fragments were mixed to obtain collections of 8 to 12 fragments that represent different areas of the tumor. Each collection of fragments was frozen in 1 mL freezing medium (FBS with 10% dimethyl sulfoxide (Sigma-Aldrich)). Vials were cryopreserved in liquid nitrogen until further use.

PDTF cultures were performed as described previously (10, 33). In brief, cryopreserved PDTFs were thawed and washed with wash medium [Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% FCS and 1% penicillin–streptomycin]. Subsequently, PDTFs were embedded in artificial extracellular matrix {sodium bicarbonate (1.1%; Sigma-Aldrich), rat-tail collagen I (1 mg/mL; Corning), Matrigel (2 mg/mL; Matrix High Concentration, Phenol Red-Free, BD Biosciences or Cultrex UltiMatrix, reduced growth factor basement membrane extract, R&D Systems), tumor medium [DMEM supplemented with 1 mmol/L sodium pyruvate (Sigma-Aldrich), 1 × MEM nonessential AA (Sigma-Aldrich), 2 mmol/L l-glutaminutese (Thermo Fisher Scientific), 10% FBS, and 1% penicillin–streptomycin]} as follows; 40 μL extracellular matrix was added to a flat-bottom 96-well plate and incubated at 37°C for 15 to 30 minutes. PDTFs were placed on top, and an additional 40 μL of the matrix was added and incubated at 37°C for 15 to 30 minutes. To treat the PDTFs ex vivo, tumor medium was supplemented with either CD8–IL2 (CD8-specific cis-targeting not-alpha attenuated-beta IL2, Asher Biotherapeutics, 10 nmol/L for comparison with PD1–IL2, 1 nmol/L for all other experiments,), PD1–IL2 (PD-1–specific cis-targeting not-alpha attenuated-beta IL2, Asher Biotherapeutics, 10 nmol/L), IL2v (not-alpha IL2, Asher Biotherapeutics, 1 nmol/L), anti-CD8β (clone 97/47, Asher Biotherapeutics, 1 nmol/L), recombinant human IL2 (PeproTech, see titration), recombinant human IL15 (Preprotech, 1 mmol/L), anti–PD-1 (Nivolumab, Bristol-Myers Squibb, 10 μg/mL), or anti-CD3 (OKT3, BioLegend, 0.5 μg/mL). PDTF cultures were incubated for 48 hours at 37°C. A shorter culture time of 24 hours was used for a subset of experiments as indicated. For response perturbations, PDTFs were preincubated for 2 hours at 37°C with either a LCKi (8 μmol/L, CAS no. 213743-31-8, Merck Millipore), anti-IFNGR1 (50 μg/mL, catalog no. 92101, R&D Systems), or anti-MHC I (100 μg/mL, clone W6/32, BioLegend). To exclude toxicity induced by the LCKi, different gene signatures associated with cell death and stress were assessed (74), which did not show any increase in LCKi-treated compared with untreated PDTFs (Supplementary Fig. S12).

After 48 hours of ex vivo tumor culture, supernatants were collected and frozen at −80°C, and PDTFs were pooled per condition in 2 mL ice-cold digestion medium [RPMI-1640 with 1% penicillin–streptomycin, pulmozyme (12.6 μg/mL, Roche), and collagenase type IV (1 mg/mL, Sigma-Aldrich)]. PDTFs were digested at 37°C under continuous rotation for 45 to 60 minutes, washed with phosphate-buffered saline (PBS), filtered and subsequently transferred to a 96-well plate for the flow cytometry staining procedure. First, the cells were incubated with human Fc-Receptor Blocking Reagent (eBioscience) and simultaneously stained with Live/Dead IR Dye (Thermo Fisher Scientific) or Zombie NIR (BioLegend) for 20 minutes on ice. Next, cells were washed and incubated with the surface antibody mix in a staining buffer (eBioscience) for 20 minutes on ice. After washing, cells were fixed and permeabilized using Fix/Perm solution (eBioscience) for 30 minutes at RT. Cells were subsequently washed twice with 1× permeabilization buffer (eBioscience) and incubated with the intracellular antibody mix in permeabilization buffer for 40 minutes at RT. Next, cells were washed twice and resuspended in FACS buffer [PBS supplemented with 2 mmol/L Ethylenediamine-tetraacetic acid (EDTA)] for data acquisition on a Symphony A5 (BD Biosciences) or Aurora (Cytek Bio). All antibodies used are listed in Supplementary Table S2. Data analysis was performed with FlowJo software (version 10.7).

Supernatants collected from PDTF cultures were thawed on ice and pooled for each experimental condition (8–10 blocks per condition). The presence of indicated cytokines, chemokines, and cytotoxic mediators was detected using the LEGENDplex Human CD8/NK panel and Human Proinflammatory Chemokine panels (both BioLegend). For all assays, 17 μL of volume was used; otherwise, the assays were performed according to the manufacturer’s instructions and measured on a BD LSRFortessa X-20 Cell Analyzer (BD Biosciences).

To assess the degranulation of intratumoral T cells upon CD8–IL2, cryopreserved PDTFs were thawed and digested as described above. Tumor digests were plated in a round-bottom 96-well plate and rested for 1 hour. Cells were subsequently stimulated with CD8–IL2 (1 nmol/L) in the presence of anti-CD107a–FITC (1:50, BioLegend). After 1 hour, GolgiPlug (1:1,000) and GolgiStop (1:1,500; both from BD Biosciences) were added, and the digest cultures were incubated for 6 hours to assess degranulation and overnight for IFNγ production. Afterward, the samples were subjected to the flow cytometry staining procedure as described above and measured on an Aurora (Cytek Bio) analyzer.

To isolate different populations of CD8⁺ tumor-infiltrating lymphocytes (TIL), cryopreserved PDTFs were thawed and digested as described above. Cells were washed with Cell Staining Buffer (BioLegend), resuspended in 300 μL PBS with Human TruStain FcX (1:10, BioLegend), and incubated for 10 minutes on ice. Next, cells were washed and incubated with 300 μL staining buffer containing anti-CD8–FITC (1:50, BioLegend), anti-CD127-PE-Cy7 (1:100, BioLegend), anti-CD39-BV711 (1:50, BD Biosciences), anti-CD62L-PE (1:50, BioLegend), anti-KLRG1-BV421 (1:20, BioLegend), and anti–PD-1-AF647 (1:20, BD Biosciences) for 20 minutes on ice. DAPI was added shortly before the sort, and cells were subsequently sorted on the BD FACSAria Fusion SORP cell sorter (v.8.0.1; BD Biosciences) into four populations: (i) dysfunctional T cells (CD8⁺ CD39⁺ PD1⁺), memory-like T cells (CD8⁺ CD127⁺), transitional T cells (CD8⁺ PD1int KLRG1⁺), and effector-like T cells (KLRG1⁺). Sorted cells were collected in 1 mL cold TIL medium [80% RPMI/20% AIM-V (Thermo Fisher Scientific), 1% penicillin–streptomycin, 1% L-glutamine, 10% human serum (Sigma-Aldrich)]. Cells were rested for 3 hours in a TIL medium supplemented with 30 IU/mL in IL2 (Clinigen Healthcare Ltd). Subsequently, TILs were expanded in vitro in the presence of 1:200 feeder cells (PBMCs irradiated at 40 Gy) and TIL medium supplemented with 30 ng/mL anti-CD3 (OKT3, BioLegend) and 3,000 IU/mL IL2 at 37°C. From day 7 onward, the cells were split 1:2 when necessary and the medium supplemented with 3,000 IU/mL IL2 was refreshed. On day 14, cells were frozen and stored in liquid nitrogen.

One day prior to coculture, cryopreserved TILs were thawed and rested in TIL medium with a low dose of IL2 (30 IU/mL) overnight. Matched PDTFs were thawed, digested, and rested in coculture media [RPMI supplemented with 1% sodium pyruvate, 1% MEM nonessential AA, 0.2% penicillin/streptomycin, 10% human serum] for 1 hour at 37°C. Expanded TILs were stained with CellTrace Violet (CTV) according to the manufacturer’s protocol (Thermo Fisher Scientific). The percentage of tumor cells from total live cells in digests was assessed by previous flow cytometry gating on CD45−, forward and side scatter-high cells. CTV-labeled TILs (50–100k) were cocultured with autologous tumor digest at an effector:tumor ratio of 1:1 in a coculture medium, in the presence or absence of CD8–IL2 (10 nmol/L). As a positive control, TILs were stimulated with anti-CD3 (OKT3, BioLegend, 10 μg/mL) and anti-CD28 (CD28.2, BioLegend, 5 μg/mL). To detect degranulation, anti-CD107a-FITC (1:50, BioLegend) was added to the coculture. For analysis of IFNγ, TNFα, and CD107a expression, GolgiPlug (1:1,000) and GolgiStop (1:1,500; both from BD Biosciences) were added after 1 hour, and the cocultures were incubated for 12 to 16 hours at 37°C. Afterward, the samples were subjected to the flow cytometry staining procedure as described above and measured on an Aurora (Cytek Bio) analyzer.

For scRNA-seq, PDTF cultures were processed as described above until a single-cell suspension was obtained. Subsequently, cells were transferred to 1.5 mL Eppendorf tubes and resuspended in 25 μL cold Cell Staining Buffer (BioLegend) with Human TruStain FcX (1:10, BioLegend) and incubated for 10 minutes on ice. In order to pool samples for sequencing, TotalSeq-C anti-human hashtag antibodies (numbers 1, 2, 4, 6, 7, 8, 9, and 10, 1 μg/mL final concentration, BioLegend) were added to the individual samples. Without washing, 25 μL of Cell Staining buffer containing anti-CD45-PerCP-Cy5.5 (1:50, Invitrogen) was added, accompanied by TotalSeq-C antibodies against PD-1 (EH12.1, 1:1,000), CD8 (SK1, 1:5,000), and CD4 (RPA-T4, 1:2,500). Subsequently, cells were incubated for 25 minutes on ice and washed three times with 1 mL staining buffer. Then, cells were resuspended in 500 μL MACS buffer [PBS with 0.5% bovine serum albumin (BSA, Sigma) and 2 mmol/L Ethylenediamine-tetraacetic acid (EDTA, Life Technologies)]. To pool equal cell numbers of each sample labeled with the different hashtag antibodies, aliquots of 5 μL from each sample were counted using AccuCount Blank Particles 13.0 to 17.9 μm (Spherotech). Dead cells were stained with propidium iodide (Sigma-Aldrich, 0.5 μg/mL) right before acquisition. Using flow cytometry, live immune cells were counted per 10,000 counting beads to pool equal numbers of each sample. CD45⁺ live cells from this mixture were subsequently sorted using a FACSAria Fusion Flow Cytometer (BD Biosciences) and collected in RPMI-1640 medium supplemented with 1% penicillin–streptomycin and 10% human serum. Lastly, cells were washed once with cold 1% BSA in PBS and once with 0.04% BSA in PBS, and resuspended in 0.04% BSA at a concentration between 800 and 1,200 cells/μL for 10× Genomics scRNA- and TCR-seq.

For scRNA- and TCR-seq, sorted CD45⁺ immune cells were loaded on each lane of the 10× Chromium instrument at a target capture rate of 1,000 to 10,000 individual cells per sample. Libraries of RNA, TCRs, and antibody barcodes were constructed according to the manufacturer’s protocol using the Chromium Next GEM Single-Cell V(D)J Reagent Kits (10× Genomics). These libraries were sequenced on a Novaseq instrument (Illumina) with read lengths of 26–28/58–130 for RNA and HTO libraries, and 26–28/92–130 for TCR libraries, aimed at recovery of 30,000 read pairs per cell for RNA libraries, and 5,000 reads for both antibody and TCR libraries.

Sequenced gene-expression reads were mapped to the human GRCh38-2020-A reference genome and quantified using Cell Ranger software (10× Genomics, v7.1.0). The resulting filtered gene-expression matrix and CITE-seq antibody count matrix were imported into Seurat (v4.9.9). Seurat objects were constructed on a per-patient basis and included all conditions, which were individually labeled using barcoded hashtag oligos (HTOs). Cells without dominant HTO signal or more than 1 dominant HTO were removed. The resulting data sets were merged. To ensure high data quality, cells were considered low quality and removed when they showed a mitochondrial RNA content of >15%, a ribosomal protein content of <7.5%, a unique molecular identifier (UMI) count of <800, or displayed a UMI count lower than 500 or exceeding 6,000.

Cell Ranger software (10× Genomics, v7.1.0) was used to assemble TCR reads into consensus sequences. Cells for which multiple TCRβ chains were captured were considered doublets and were removed from downstream analysis. Cells with a matching TCRα and TCRβ sequence were considered T-cell clones.

The transcript counts of cells that passed quality control were log₁₀-normalized using Seurat’s NormalizeData() to correct for differences in library size. CITE-seq counts were normalized using a center-log ratio, likewise using NormalizeData(). Highly variably expressed genes were selected using FindVariableFeatures(), with the selection.method set to “vst” and the number of genes set to 2,000. From the list of 2,000 genes, we excluded genes that contained the sequence “MT-,” “RP-,” “LINC-,” and “AC-” in their name, as well as long-noncoding RNAs “MALAT1” and “XIST.” IG genes (“IG[HJKL]”) and TCR V genes (“TRAV-” and TRBV-”) were likewise excluded to avoid formation of clusters based on clonal populations. The genes that remained after filtering were used for principal component analysis using runPCA() after gene expression was scaled and the effects of mitochondrial genes and dissociation induced genes were regressed out using ScaleData(). The first 20 principal components were selected as input for UMAP (RunUMAP()) and clustering. Clusters of transcriptionally similar cells were identified by constructing a k-nearest neighbors graph and applying the Louvain algorithm with a resolution set at 1.5, using the FindNeighbors() and Findclusters() functions, respectively.

The resulting clusters were annotated as either “B cell,” “CD4 T cell,” “CD8 T cell,” or “γδ T cell & NK cell.” We noticed that the CD4/CD8 T-cell annotation was imperfect and observed the signal of CD4⁺ T cells within the CD8 clusters and vice versa. To maximize the correct separation of CD4⁺ T cells and CD8⁺ T cells and minimize false-positive events for either compartment, we applied a series of filters. First, among the cells previously annotated as “CD4 T cell,” “CD8 T cell,” and “γδ T cell and NK cell,” we selected all cells that showed a normalized expression >0.5 of either CD3D, CD3E, or CD3G. Among these CD3⁺ cells, cells with normalized expression of CD4 CITE-seq antibody >0.7 or expression of CD4 >0.5 were annotated as “CD4⁺ T cell.” Likewise, cells that showed normalized expression of the CD8 CITE-seq antibody >1.2 or expression of CD8A >0.5 or CD8B >0.5 were annotated as CD8⁺ T cells. Cells that passed both the CD4 and CD8 filtering steps were annotated as “Double positive” and were not included in further downstream analysis.

The resulting cells that classified as “CD8⁺ T cells” were reclustered in order to derive more detailed CD8⁺ T cell subsets. The 4,000 most variable features were identified as outlined previously. IG genes and TCR genes were again excluded from this list. Data were scaled and the effects of mitochondrial genes were regressed out. To remove any batch effects that might arise, CD8⁺ T cells from the five patients were integrated using the RunHarmony() function from Seurat, with the max.iter.harmony parameter set to 30. UMAP projections were created using the first 16 dimensions (reduction set to “harmony”) and clusters were derived similarly as described earlier (resolution Louvain algorithm set to 0.8).

Seurat’s FindAllMarkers() function was used to obtain genes that were differentially expressed between clusters and to identify canonical markers that defined each cluster. Furthermore, to better identify various CD8⁺ T-cell subtypes, we applied gene signatures for terminally exhausted CD8⁺ T cells (52), effector T cells (51), memory T cells (51), and proliferating T cells (53). Scores for each signature were computed using the Seurat function AddModuleScore() using the gene signature of interest and setting the number of control genes from the same bin of expression at 5. The same strategy was used to compute the better effector signature from Codarri Deak and colleagues (42) in this data set, using the genes with FDR < 0.01.

To identify a TCR-specific gene signature, independent of CD8–IL2 exposure, we derived genes that were upregulated in T cells upon treatment with anti-CD3 compared with T cells that were untreated. For this, we compared the gene expression of all CD8⁺ T cells in the untreated condition to all CD8⁺ T cells in the anti-CD3-treated condition using Seurat’s FindMarkers() function. All genes that were enriched with an average log₂ fold change >0.3 and P value < 0.05 in the anti-CD3 condition were included in the signature. A similar approach was used to compile a signature of genes induced upon CD8–IL2 treatment independent of TCR signaling. To this end, transcriptomic differences between cells that were untreated or treated with CD8–IL2 + LCKi were identified. The TCR program score and IL2 program score were calculated for each cell using AddModuleScore().

Pseudotime Analysis and Trajectory Inference To model transitions between CD8⁺ T-cell subsets, DPT analysis was performed to construct differentiation trajectories. The first 20 harmony-corrected principal components were used as input to derive a diffusion map with 30 principal components, using the DiffusionMap() function (default parameters) from Destiny (V3.14). DPT was calculated by inputting the diffusion map into the DPT() function to rank the cells.

TCR annotation was performed using the 10× Cell Ranger vdj pipeline. If more than one β chain was detected in a cell, then the cell was considered as doublet and excluded from further analysis. scRepertoire package (v.1.10.0) was used for clonotype assignment and analysis of clonotype dynamics. To compare differences in CD8⁺ T-cell state between cells from different conditions, we performed clone-matched analysis of the cell state composition of clones derived from untreated and CD8–IL2 conditions, or untreated and anti–PD-1-treated conditions. Only clones that were present at least once in each condition were included.

Data are reported as the mean ± SEM. Statistical significance was determined using the Mann–Whitney U test, two-tailed Wilcoxon test, or Friedman test, as indicated. Differences were considered statistically significant if ****, P < 0.0001; ***, P < 0.001; **, P < 0.01; *, P < 0.05. All analyses were performed in either GraphPad (v.9.5.1) or R (v.4.2.3). Unless otherwise specified, experiments were performed without duplicates because of material restrictions.

Source data are provided in this paper. The scRNA- and TCR-seq data generated in this study are deposited in the European Genome–Phenome Archive (EGA) under accession number EGAS00001007712 and are accessible upon request. For inquiries regarding data access or any related information, please contact [email protected]. Other data supporting this report are available from the corresponding author upon reasonable request.

T.N. Schumacher reports personal fees from Asher Bio during the conduct of the study; personal fees from Neogene Therapeutics, Merus, Allogene Therapeutics, and Scenic Biotech outside the submitted work. Y.A. Yeung reports a patent for US20220251202A1 pending and a patent for US20220162314A1 pending. K.D. Moynihan reports other support from Asher Biotherapeutics during the conduct of the study; in addition, K.D. Moynihan has a patent for CD8-targeted IL2 pending. I.M. Djuretic reports other support from Asher Biotherapeutics during the conduct of the study; other support from Asher Biotherapeutics outside the submitted work; in addition, I.M. Djuretic has a patent for CD8-targeted IL2 molecules pending to Asher Biotherapeutics. D.S. Thommen reports grants from Asher Biotherapeutics during the conduct of the study; grants from Bristol-Myers Squibb outside the submitted work. No disclosures were reported by the other authors.

P. Kaptein: Data curation, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. N. Slingerland: Data curation, formal analysis, validation, visualization, methodology, writing–review and editing. C. Metoikidou: Data curation, formal analysis, validation, visualization, methodology, writing–review and editing. F. Prinz: Data curation, funding acquisition, investigation, writing–review and editing. S. Brokamp: Data curation. M. Machuca-Ostos: Data curation, investigation, writing–review and editing. G. de Roo: Data curation. T.N. Schumacher: Conceptualization, formal analysis, investigation, writing–review and editing. Y.A. Yeung: Formal analysis, investigation, methodology, writing–review and editing. K.D. Moynihan: Formal analysis, investigation, methodology, writing–review and editing. I.M. Djuretic: Formal analysis, investigation, methodology, writing–review and editing. D.S. Thommen: Conceptualization, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology,writing–original draft, project administration, writing–review and editing.

This work was supported by a Melanoma Research Alliance (MRA) Team Science Award (681127, to D.S. Thommen together with Christian Blank and Daniel Peeper), by research funding from Asher Biotherapeutics (to D.S. Thommen), by a Austrian Science Fund fellowship (FWF, DK-MCD W1226 to F. Prinz), and an institutional grant to the NKI of the Dutch Cancer Society (KWF) and the Dutch Ministry of Health, Welfare and Sport. We thank the NKI-AVL core facilities for their excellent technical support, the Flow Cytometry Facility and the Genomics Core Facility, i.e., Iris de Rink and Arno Velds for the technical support and input with single-cell analyses, and the Molecular Pathology and Biobanking Core Facility for assistance with collecting and processing of human tissue material. We thank Luca Braccioli and Elzo de Wit for assistance with the revision experiments. We thank Timm Reissig for the collection of clinical data, Roos Wagensveld for help with experiments, and all members of the Thommen laboratory for helpful discussions. Illustrations in Figs. 1A and 3A and Supplementary Fig. S5A were created with BioRender.com.