CXCL9 recombinant adeno-associated virus (AAV) virotherapy sensitizes glioblastoma (GBM) to anti-PD-1 immune checkpoint blockade

The promise of immunotherapy to induce long-term durable responses in conventionally treatment resistant tumors like glioblastoma (GBM) has given hope for patients with a dismal prognosis. Yet, few patients have demonstrated a significant survival benefit despite multiple clinical trials designed to invigorate immune recognition and tumor eradication. Insights gathered over the last two decades have revealed numerous mechanisms by which glioma cells resist conventional therapy and evade immunological detection, underscoring the need for strategic combinatorial treatments as necessary to achieve appreciable therapeutic effects. However, new combination therapies are inherently difficult to develop as a result of dose-limiting toxicities, the constraints of the blood-brain barrier, and the suppressive nature of the GBM tumor microenvironment (TME). GBM is notoriously devoid of lymphocytes driven in part by a paucity of lymphocyte trafficking factors necessary to prompt their recruitment, infiltration, and activation. We have developed a novel recombinant adeno-associated virus (AAV) gene therapy strategy that enables focal and stable reconstitution of the GBM TME with C-X-C motif ligand 9 (CXCL9), a powerful call-and-receive chemokine for cytotoxic T lymphocytes (CTLs). By precisely manipulating local chemokine directional guidance, AAV-CXCL9 increases tumor infiltration by CD8-postive cytotoxic lymphocytes, sensitizing GBM to anti-PD-1 immune checkpoint blockade (ICB). These effects are accompanied by immunologic signatures evocative of an inflamed and responsive TME. These findings support targeted AAV gene therapy as a promising adjuvant strategy for reconditioning GBM immunogenicity given its excellent safety profile, TME-tropism, modularity, and off-the-shelf capability, where focal delivery bypasses the constrains of the blood-brain barrier, further mitigating risks observed with high-dose systemic therapy.


Introduction
Breakthroughs in immunotherapy including ICB, monoclonal and bispeci c antibodies, and CAR T cell therapy, have ignited the hope of achieving durable remission in even the most recalcitrant tumors.Although these strategies are capable of producing remarkable responses, therapeutic bene t is seen in only a small proportion of patients with many proposed reasons for lack of response 1 .Presently, T cell in ltration and abundance within the tumor microenvironment is one of the most predictive biomarkers for response to immunotherapy [2][3][4] .Unfortunately, diseases like GBM and many other solid tumors demonstrate low baseline in ltration of lymphocytes that is only marginally improved by treatment 5,6 .For GBM patients in particular, generalized lymphopenia as a result of impaired lymphocyte egress 7 alongside the lymphodepleting nature of conventional treatment further detracts from successful adaptive immune recognition of these tumors 8 .Preclinical studies on cellular therapy imaging and tra cking show that a comparably lower fraction of T cells can be found in these brain tumors as compared to lung, liver, and spleen tissues in comparative mouse models 9,10 .Similar hurdles with respect to CAR T cell penetration into brain tumors in the clinical setting have necessitated alternative locoregional delivery strategies 11 .Together these observations indicate that ineffective T cell migration and in ltration into GBM tumors may represent a principal barrier to immunotherapy.Solving this problem requires both an improved understanding of the chemical signals that govern T cell chemotaxis into tumors and identifying a method to amplify those signals.
CTLs are recruited via long-range signaling mediated by the diffusion of chemokines present in in ammatory environments.During glioma formation, tumors manufacture immune-suppressive chemokines and cytokines that co-opt resident cells, resulting in the preferential recruitment of immune suppressor cells from the periphery 12 .Additionally, we have found that primary patient glioma samples are de cient in lymphocyte-speci c chemokines.To overcome this problem, we propose the use of in situ recombinant AAV expressing a lymphocyte tra cking chemokine payload to restore expression of lymphocyte chemotaxis.AAV vectors are the leading platform for gene delivery due to their e cacy, ease of use, and safety pro le 13 .Additionally, multiple AAV gene therapies have reached FDA approval for a variety of diseases 13,14 .Thus, AAV vectors are an ideal platform for transforming targeted cell populations while minimizing potential patient risk.Following a comprehensive chemokine screen of clinical tumor specimens, we have identi ed CXCL9 as a candidate lymphocyte call-and-receive signal absent in GBM.We examined AAV delivered CXCL9 transgene tropism, durability, and impact on CTL tumor migration.Using syngeneic preclinical model systems of GBM, we further evaluated the therapeutic bene t of our AAV gene therapy alongside anti-PD-1 ICB.Single cell RNA sequencing (scRNAseq) of immune cells isolated from tumors during treatment reveal widespread immunological reconditioning of tumors, with improved effector lymphocyte recruitment that yields long-term survival outcomes in aggressive preclinical GBM models.This work suggests direct conditioning of the tumor microenvironment by AAV-CXCL9 constructs could potentially overcome a key resistance factor in GBM, sensitizing tumors to immune mediated therapies.

Chemokine analysis of Human GBM favors MDSC recruitment over lymphocyte in ltration
Upon examination of human glioma tumors for CD3 protein expression provided through the Human Protein Atlas, we found that > 80% of tumor specimens evaluated are negative for lymphocyte in ltrates (Extended Data Fig. 1a), corroborating prior literature describing the lymphocyte-replete nature of these tumors 15 .To identify if a paucity of lymphocyte chemotactic factors is a contributing aspect, human glioma samples were screened via chemokine proteome array.Of these, CXCL4, CXCL7, CXCL8 (IL-8), CXCL16, LCF (IL-16), TIG-2, and midkine (MDK) emerged as the most abundant secreted chemokines detected (Fig. 1a,b), where these ligands play a signi cant role in recruiting myeloid-derived suppressor cells in the context of gliomagenesis 16 .Notably, chemotactic factors that favor lymphocyte recruitment were poorly expressed, including CXCL9 (MIG) and MIP-1α/β (CCL3/CCL4) (Fig. 1a,b).CXCL9 is a powerful attractant known to induce the migration, differentiation, and activation of CTLs 17 .Expression of CXCL9 has been shown to correlate with anti-tumor immune activity and is predictive of response to ICB in several solid tumors 18 .We hypothesized that AAV delivery and restoration of CXCL9 as a "call-and-receive" signal for T lymphocytes within the TME would enhance their recruitment and in ltration in the tumor.
Historically, targeted transduction of cancer with AAV has proven challenging despite elegant efforts in capsid evolution studies and capsid engineering.To achieve su cient transgene expression, transduction of either tumor cells or tumor-associated stroma is likely necessary, where intra-tumoral delivery would minimize the potential for systemic toxicities and decrease off target homing of T cells.To identify an appropriate AAV capsid for targeting glioma tumors, we performed an in vitro capsid screen.Enhanced green uorescent protein (EGFP) was encoded into an AAV2 single stranded vector, utilizing the non-cellautonomous, constitutively active CBA promoter to drive transgene expression (AAVn-EGFP).These constructs were pseudotyped into 29 unique capsids as previously described 19 .Transduction in 15 unique glioma models, including primary human 20 and murine xenografts, was assessed via EGFP relative uorescence expression (Extended Data Fig. 1b).AAV6 was selected for further examination as it demonstrates moderate to high transduction in nearly all models tested and is further substantiated by excellent CNS transduction in other studies 19,21 .AAV6 capsids encoding each CXCL9, EGFP, blue uorescent protein (BFP), and empty vector control (Fig. 1c, Extended Data Fig. 1c,d,e) were designed for further testing.Quantitative ow cytometric evaluation of EGFP expression in three distinct syngeneic murine GBM models 72 hours following AAV6-EGFP transduction shows moderate transduction of GL261 and KR158, with > 25% of cells positive for the transgene at this time point, with lower transduction observed in CT-2A cells (< 20%) (Extended Data Fig. 1f).

AAV6 transduces tumor-reactive astrocytes in vivo in preclinical models of GBM
While we demonstrate good targeting of glioma cells in an in vitro setting, AAV6 has also been reported capable of targeting other cell populations in the CNS 19,21 .These studies have largely been performed in the context of naïve mice or in models of neurodegenerative disease, which may or may not be directly applicable to CNS malignancy.To de ne AAV6 tropism in murine GBM in vivo, AAV6-EGFP (Extended Data Fig. 1c) was intratumorally injected into established intracranial KR158 and GL261 tumors.EGFP expression was detected 1 week following tumor transduction in both models (Extended Data Fig. 2a,b), however the morphological appearance and contiguous distribution of transduced cells suggest that AAV6 targeted cells are likely tumor-associated, and not cancer cells directly.Microglia and tumorassociated macrophages are reported to comprise a signi cant cellular proportion of glioma tumors 22 , and so we sought to identify if AAV6 was targeting either population.KR158 and GL261 were implanted into CCR2 RFP CX3CR1 GFP (B6.129(Cg)-Cx3cr1 tm1Litt Ccr2 tm2.1lfc /JernJ) dual reporter mice, where microglia can be identi ed via GFP expression, and bone marrow-derived in ammatory cells via RFP expression 23,24 .Tumors were evaluated by 3D IHC for viral transduction one week following intratumor injection with AAV6 encoding a BFP reporter (Extended Data Fig. 1d).In both tumor model systems, we found no co-localization between BFP and either RFP or GFP, indicating that neither microglia nor tumorassociated macrophages are the principal target of AAV6 transduction (Extended Data Fig. 3a,b).To assess the degree of AAV6 transduction speci cally in tumor cells we implanted mice with RFP-labeled KR158 or GL261 cells.Tumors were evaluated by 3D IHC for viral transduction one week following intratumor injection with AAV6-EGFP.Resected tumors were immuno-labeled against glial brillary acidic protein (GFAP) to detect astrocytes, another candidate tumor-associated cell population.Both tumor models reveal a high degree of overlap between GFAP (red pseudocolor) and EGFP (green pseudocolor), with minimal overlap between tumor cells (light blue pseudocolor) and EGFP (Fig. 1d, Extended Data Fig. 3c), indicating that EGFP-positive cells are likely astrocytes.Voxel-based co-localization algorithms to quantitate EGFP co-localization with each tumor or astrocytes con rm astrocytes as the principal cell target of AAV6 transduction, accounting for ~ 60-70% of EGFP-positive cells in both GL261 (Fig. 1e) and KR158 (Extended Data Fig. 3d) intracranial tumors.Because tumor presence can stimulate different activation states in astrocytes that may cause them to be more or less susceptible to viral transduction 25,26 , we also evaluated CNS tropism of AAV6 in age-matched naïve mice.AAV6 was equally e cient at transducing astrocytes in naïve animals as shown by co-localization between GFAP immunostaining and EGFP transgene expression (Fig. 1e,f).
One of the unique features of AAV gene transduction is that it rarely integrates into the host genome.Following uncoating in the host nucleus, single-stranded genomes are converted to double-stranded multimeric circular concatemeric episomes 27 .As such, AAV transgene expression can persist long-term in post-mitotic cells.Because tumor cells undergo rapid cell division, it may be possible that transgene expression is lost over time through sequential dilution of episomes passed down to daughter cells.To explore this, we evaluated AAV6-EGFP transgene expression longitudinally across early time points in mice harboring RFP labeled GL261 cells, which demonstrate the highest transduction e ciency in vitro (Extended Data Fig. 1f).Tumors were resected 3, 5, and 7 days following AAV6-EGFP intra-tumor injection as outlined in Fig. 1g, and EGFP transgene expression in tumors and astrocytes was measured by ow cytometry.Even at early time points, AAV6 predominantly transduces astrocytes identi ed as GFAP + RFP-(70-80% EGFP + cells), with limited EGFP expression observed in RFP + tumor cells (< 15%) (Fig. 1g, Extended Data Fig. 4a).By 7 days post intratumor viral injection, less than 5% of EGFP + cells on average were RFP + tumor cells, where astrocytes consistently comprised 70-80% of EGFP + cells at each time point.These data indicate that AAV6 more selectively transduces astrocytes in vivo, with limited and transient expression in tumor cells.

AAV6 transgene signal distribution and durability in GBM
Next, we examined the distribution of transgene signal in both the GL261 and KR158 tumor models via 3D IHC to better understand the avidity of AAV6 for tumor-associated versus distal astrocytes following direct intra-tumor injection.BFP transgene expression was observed in a peritumoral pattern in and around the tumor body in both model systems (Fig. 2a,b), redolent of glial scar formation found in human brain malignancies 28 .As CXCL9 is a small, secreted chemokine, we wanted to determine if signal expression was still focal to the tumor or could be detected in contralateral brain and/or systemically.The whole brain was collected at 1 and 2 weeks following AAV6-CXCL9 or AAV6-EGFP intratumor injection as outlined in Fig. 2c.Cerebellar tissue was removed, and remaining tissue was dissected into the tumor containing and contralateral hemispheres.Serum was collected following peripheral blood draws taken from the posterior vena cava.Brain tissue and serum were also collected from non-transduced (sham) tumor controls, and naïve (non-tumor bearing) controls to establish CXCL9 baseline values.Serum levels of CXCL9 following intratumor delivery of AAV6-CXCL9 measured using high sensitivity ELISA assay did not exceed those observed in naïve controls (Fig. 2d,e).In the brain, elevated CXCL9 expression was selectively detected in the tumor bearing hemisphere transduced with AAV-CXCL9, with minimal signal observed in the contralateral hemisphere in both GL261 and KR158 model systems (Fig. 2f,g).Transgene CXCL9 expression appears to be stable, as signal intensity was consistent in AAV6-CXCL9 transduced tumors at both the 1-and 2-week time points in each tumor model (Fig. 2f,g).Of note, a small increase in CXCL9 expression was observed in AAV6-EGFP control transduced GL261 tumors and could be indicative of a mild in ammatory response to AAV6, however these values were not found to be statistically signi cant.Together these data demonstrate that AAV6 intratumor delivery of CXCL9 results in focal and durable expression of encoded transgene, where tumor-reactive astrocytes are the target of AAV6 transduction.

AAV6-CXCL9 enhances lymphocyte chemotaxis
To evaluate the biologic activity of AAV6-CXCL9 on lymphocyte recruitment, we performed competitive in vitro chemotaxis assays.Brie y, CTV-labeled splenic-derived T lymphocytes were anked by target cells transduced with AAV6 encoding either EGFP or CXCL9, and migration was monitored via uorescence microscopy at 1-and 24-hours following co-culture (Fig. 3a).Using GL261 tumor cells as the target population for AAV6 transduction, signi cantly more T lymphocytes co-localized in the CXCL9 transduced tumor eld as compared to EGFP at 24 hours (Fig. 3b).Given that astrocytes are the principal target of AAV6 transduction in vivo, chemotaxis was reassessed via competitive co-culture using astrocytes (C8-D1A) in lieu of GL261 glioma cells.Astrocytes transduced with AAV6-CXCL9 similarly showed enhanced recruitment of T lymphocytes (Fig. 3c), con rming that transgene encoding CXCL9 produces a biologically functional chemokine.To determine the effect of AAV6-CXCL9 on T lymphocyte recruitment in vivo, multiparametric ow cytometry was performed to quantitate the number of T cells present in dissociated tumors following intratumor delivery.These studies were done in combination with anti-PD-1 ICB, where tissue was collected one day following the nal dose of ICB to capture events within the therapeutic response window as outlined in Fig. 3d.In both GL261 and KR158 tumor models AAV6-CXCL9 alone had minimal impact on enhancing T cell recruitment to the tumor, however signi cant increases in T lymphocyte in ltration were observed in the context of combination treatment.AAV6-CXCL9 plus ICB increased CD8 T lymphocytes > 2.5-fold in the GL261 model and > 4.5-fold in the KR158 model (Fig. 3e,f and Extended Data Fig. 4b).While no signi cant changes in CD4 T lymphocyte recruitment in response to treatment was observed in the GL261 model (Fig. 3g), a > 3-fold increase was detected in the KR158 model (Fig. 3h).Anti-PD-1 ICB treatment in combination with control AAV6 (EGFP) modestly increased CD8 T lymphocyte recruitment in GL261 by 1.4-fold and in KR158 by 2.7-fold, indicating that CXCL9 markedly improves tumor in ltration by these cells.These data highlight a potential role for anti-PD-1 ICB in mobilizing T lymphocytes systemically, where sequestration of T lymphocytes was recently proposed as a novel mechanism of immune suppression in brain tumors 7 .

AAV6-CXCL9 sensitizes preclinical GBM to anti-PD-1 ICB
To assess if enhanced lymphocyte recruitment and immunological reprogramming through combination treatment could produce anti-tumor responses against GBM, we performed survival analyses in both the GL261 and KR158 syngeneic model systems.5 days following tumor implantation, AAV6 encoding CXCL9 or EGFP control transgene was injected intratumorally, with anti-PD-1 ICB (10mg/kg) administered intraperitoneally for a total of 4 doses given every 72 hours (Fig. 3d).In the GL261 model we found that anti-PD-1 ICB produced a small, but non-signi cant increase in overall survival as compared to sham treated control animals (p = 0.060), where AAV6-CXCL9 treatment yielded no survival bene t as a monotherapy (Fig. 4a).Combination treatment signi cantly improved overall survival, with 50% of animals exhibiting durable outcomes (Fig. 4a).We observed similar results in the KR158 model, which carries a low mutational burden and is recalcitrant to immunotherapy, with combination treatment signi cantly improving median survival, and long-term survival observed in 25% of this cohort (Fig. 4b).As an additional metric to validate the ability of combination AAV6-CXCL9 plus anti-PD-1 ICB to immunologically transform GBM tumors, GL261 tumors were implanted in GREAT transgenic mice to evaluate tumor-wide IFNγ expression following treatment as described in Fig. 3d.IFNγ was detected in combination treated tumors evidenced by EYFP signal detection via 3D IHC (Fig. 4c).Immunolabeling of tissues for CD45 con rms that EYFP (IFNγ) + cells are immune cells (Fig. 4d), indicative of proin ammatory immune activation.
To determine if CD8 lymphocytes contribute to the therapeutic survival effect, we repeated combinatorial treatment with concomitant CD8 depletion (Fig. 4e) in the GL261 model.We found that on study day 18 all animals treated with CD8 depleting antibodies had no detectable levels of circulating CD8 T lymphocytes, and no changes in the quantity of circulating CD4 T lymphocytes (Fig. 4f, Extended Data Fig. 4b).CD8 depletion reversed the survival bene t observed with combination AAV6-CXCL9 plus anti-PD-1 ICB, and this cohort progressed as quickly as control treated subjects (Fig. 4g).To determine if combination treatment could confer long-term immune memory formation, we performed a GL261 tumor rechallenge in long-term survivors (> 55days) that had received AAV6-CXCL9 plus anti-PD-1 ICB.No observable residual tumor was present from the initial tumor implantation during the second implantation.A second cohort of age-matched naïve animals was intracranially injected with GL261 as a control.Control animals all succumbed to tumor burden within 30 days of tumor implantation, whereas 100% of rechallenge animals remained disease free (Fig. 4h).These data con rm that therapeutic response to combination therapy is dependent on tumor in ltration by CD8 T lymphocytes as part of the adaptive immune cascade, and combination therapy can convey long-term immune memory protection against recurrence.
Top differentially expressed genes from each pooled population were identi ed, and cluster cell types were de ned using the expression of known marker genes resulting in the identi cation of 13 unique cell clusters 29,30 (Fig. 5f and Extended data Table 1).Analyses of lymphocyte tumor recruitment across treatment groups recapitulate our earlier observation, with combination therapy yielding a signi cant increase in total in ltrating CD8 T lymphocytes (Fig. 5g), identi ed using the gene expression markers Cd3d, CD8a, Cd8b1 as previously described 31 .T regulatory lymphocytes (Treg), de ned by Cd4, Foxp3, and Il2ra gene expression 31 , were also increased in response to combination therapy, although collectively these represent < 1% of the total tumor-associated immune population (Fig. 5h).Increased tumor in ltration by monocytes, classi ed by high Ly6c1 expression, was observed across all treatment groups as compared to sham control mice (Fig. 5i), with an enrichment of non-classical monocytes characterized by Spn, Cx3cr1, and Tnfrsf1b expression 32 in groups receiving anti-PD-1 treatment (Fig. 5j).Graphical summaries for all remaining cells clusters in response to each treatment are shown in Extended Fig. 5a-h.

Combination AAV6-CXCL9 and anti-PD-1 ICB treatment increases cellular crosstalk in lymphocytes
As shown in Fig. 6a, we identi ed 2,260 differentially expressed genes (DEGs) associated with AAV6-CXCL9 treatment, 2,607 DEGs associated with anti-PD-1 treatment, and 2,649 DEGs associated with these treatments combined.Of these, 70, 194, and 151 DEGs appear to be unique to each given treatment strategy, respectively, and may provide unique insight toward treatment impact on immune cell functional states.Through transcriptional expression of distinct ligands and receptors, cell-type-speci c interactions were inferred, providing additional insight towards the in ammatory pro le of tumors and how they change in response to treatment 33 .Using our prede ned cell clusters, a simpli ed DEG set was established for each.DEGs were then queried against public ligand-receptor databases (see Methods).Summary results are shown in Chord Plots, where line thickness represents the number of predicted interactions between two de ned cell clusters (Extended Data Fig. 6a-e).Next, we performed direct comparisons of interactome activity between treatment groups to elucidate heightened or decreased connectivity associated with AAV6-CXCL9 and anti-PD-1 ICB, where heatmap relative values in red indicate increases and blue decreases in prospective ligand-receptor interactions.In evaluating AAV6-CXCL9 in combination with either anti-PD-1 ICB or IgG2 control to resolve the contributions of ICB, notable increases in signals emanating from each macrophages (Mac), border-associated macrophages (BAM), microglia (Mg), and NK cells signaling to CD8 + and regulatory T cells were observed (Fig. 6b).Decreased incoming signals were noted in BAMs, CD4 + T cells, and dendritic cells (DCs) stemming from nearly all cell clusters (Fig. 6b).Cell-cell interactions associated with AAV6-CXCL9 shown in Fig. 6c reveal heightened communication directed toward both CD4 + and CD8 + T cell subsets, and NK cells prompted by all clusters excluding B cells and DCs.Signaling originating from all lymphocyte populations, and most innate immune cells including Macs, Mgs, Monocytes, and NK cells was increased, suggesting that AAV6-CXCL9 treatment broadly stimulates immune activity.
Given that combination treatment promotes CD8 T cell tumor in ltration, which is required for anti-tumor e cacy, we sought to resolve how treatment might impact CD8 T cell effector function via pathway analysis of DEGs speci cally within these cells.Comparative pathway analysis between CD8 T cell DEGs shows selective enrichment of thrombospondin (THBS), poliovirus receptor (PVR, CD155), CD137 (4-1BB), bronectin-1 (FN1), laminin, and major histocompatibility complex class I (MHC I), among others, as uniquely a liated with combination therapy when compared to AAV6-CXCL9 plus IgG2 control (Fig. 6d).
We next leveraged the NanoString nCounter® Immune Exhaustion Panel to further characterize immune status and in ammatory signatures associated with each respective treatment.A summary of pathway activation across all cell subsets in response to individual treatments is shown in Fig. 6f and Extended Data Fig. 6h, with CD8 T cell clusters outlined in black for each treatment group.Through deeper analysis of differential pathway activation, speci cally in the CD8 T cell subset, we found heightened antigen presentation (Fig. 6g), chemokine signaling (Fig. 6h), cytotoxicity (Fig. 6i), T cell exhaustion (Fig. 6j), TCR signaling (Fig. 6k), and PD-1 signaling (Fig. 6l).In particular, combination AAV6-CXCL9 plus anti-PD-1 ICB was associated with the highest increase in cytotoxicity, TCR signaling, and PD-1 signaling.Increased T cell exhaustion appears to be associated with AAV6-CXCL9 treatment.Given that monocyte tumor in ltration was additionally increased in response to treatment (Fig. 5i,j), we evaluated pathway activation in these cells to better understand their functional status.We found enhanced activation across 12 pathways, including antigen presentation, chemokine signaling, cytotoxicity, IL-10 signaling, JAK/STAT signaling, other interleukin signaling, T cell checkpoint, TGFβ signaling, TNF signaling, Type I interferon signaling, and Type II interferon signaling (Extended Data Fig. 7).Of these, AAV6-CXCL9 treatment appears to be associated with increased antigen presentation, cytotoxicity, JAK/STAT signaling, and Type I interferon signaling, where anti-PD-1 ICB induces IL-10 signaling, TLR signaling, and TNF signaling.Together these data are suggestive that treatment may augment the pro-in ammatory function of these cells.

Cytokine pro ling of combination AAV6-CXCL9 plus anti-PD-1 ICB
As combination therapy increases DEGs of both the CCL and CXC superfamily of secreted chemokines and cytokines, we sought to parse out transcriptional changes within CD8 T cells as an additional means to evaluate the activation state of these cells given the central role of these ligands in directing migration and activation of immune cells during in ammation 53 .A summary of all CCL and CXC family ligand and receptor transcripts expressed by CD8 T cells is presented in the heatmap in Fig. 7a.CD8 T cell mediated stimulation of monocytes/macrophages is evidenced by increased transcription of CCL2, CCL3, and CCL12 across all treatment groups as compared to sham control (Fig. 7b-d).CCL4 transcription was also increased (Fig. 7e), indicative of NK stimulation by CD8 T cells.CCL5 was found to be the most differentially upregulated soluble ligand in combination treated CD8 T cells as compared to all other treatment groups (Fig. 7f), and is strongly indicative of CD8 T cell effector function 54,55 .While we show that AAV6 delivered CXCL9 transgene expression predominantly emanates from tumor-reactive astrocytes, our scRNAseq data shows that each anti-PD-1 and AAV gene therapy induces CXCL9 transcription within CD8 T cells (Fig. 7g), additionally demonstrating immune activation as a result of treatment 17,18,56 .CXCL10 was also found to be transcriptionally upregulated in response to anti-PD-1 and AAV gene therapy (Fig. 7h), which prompts further CD4, CD8, and NKT lymphocyte recruitment 56 .Altogether, these data support that combination AAV6-CXCL9 and anti-PD-1 ICB both increases lymphocyte tra cking to intracranial GBM tumors and potently stimulates effector lymphocyte cellular communication and activation.
As described above, secreted cytokines can in uence the trajectory of tumors in a multitude of waysreprogramming tumor-associated cells and suppressing in ltrating in ammatory subsets which allows for tumor tolerance, progression, metastasis, and even therapeutic resistance or, alternatively, creating an environment favorable for innate and adaptive immune activation to facilitate tumor rejection 57 .
Moreover, the cytokine pro le of a tumor may serve as predictive and/or therapeutic biomarkers allowing for the detection of tumor presence, forecasting therapeutic response, and can also be used to guide therapeutic choices 57 .We performed a large-scale cytokine proteomic assessment of single agent and combination treated tumors to identify candidate biomarkers of response to therapy.Tumors were resected ten days after the onset of treatment as shown in Fig. 3d.Of the 111 soluble murine proteins on the array, relative expression of 65 was detected in treated and/or control GL261 tumor samples as summarized Extended Data Fig. 8a, with representative cytokine immunoblots shown in Fig. 7j. 10 secreted factors were identi ed as differentially expressed as compared to sham control tumors following either single or combination treatment with AAV6-CXCL9 and anti-PD-1 ICB: ADIPOQ, C1QR1 (CD93), CCL5, CCL12, CD40, CXCL9, CXCL10, CXCL16, LCN2, and MPO (Fig. 7k, Extended Data Fig. 8b-k).Of these, CCL5, CD40, and CXCL16 were most potently induced by combination treatment.These markers are highly indicative of lymphocyte presence and activation, where CCL5 is a potent pro-in ammatory ligand manufactured principally by CD8 T lymphocytes, and CD40 is a co-stimulatory ligand that triggers lymphocyte proliferation and cytokine production 54,55 .Of note, elevated CCL5 ligand expression demonstrates concordance with scRNAseq data (Fig. 7f).In addition, CXCL9, CXCL10 and CXCL16 are strong chemotactic signals for lymphocyte recruitment.Both CXCL10 and CXCL16 are induced by interferon gamma (IFNγ) and tumor necrosis alpha (TNFα), powerful catalysts of innate and adaptive in ammation 58,59 .A summary of treatment-induced secreted ligands and known receptor interactions are depicted via circular interactome analysis performed using Circos® visualization software 60 , revealing insight towards immune reprogramming that occurs in response to each respective treatment (Fig. 7l).These data combined validate that AAV6 delivery of CXCL9 to the tumor microenvironment in tandem with anti-PD-1 ICB not only facilitate lymphocyte recruitment to GBM tumors, but also reprograms the immunological landscape towards a pro-in ammatory phenotype.
In summation (Fig. 8), intra-tumor delivery of AAV6 encoded CXCL9 results in the production of a prolymphocyte chemotactic gradient by transduced tumor-reactive astrocytes.This, in concert with anti-PD-1 ICB, signi cantly increases tumor in ltration by lymphocytes likely through CXCL9 engagement with its cognate receptor expressed by these cells-CXC motif chemokine receptor 3 (CXCR3).In particular, CD8 T lymphocytes are the premier arbiters of anti-tumor response, where depletion of this lymphocyte subset negates therapeutic e cacy.Moreover, CD8 T cell effector activation and function is evidenced by heightened expression of co-stimulatory molecules, such as 4-1BB and ICOS, and production of proin ammatory chemokines and cytokines.Beyond CD8 T lymphocyte activation, AAV6-CXCL9 and anti-PD-1 ICB appear to contribute widespread immunological activation, demonstrated by heightened cellular cross-talk across numerous immune clusters, and protein detection of pro-in ammatory molecules.
Notably, CCL5, CXCL9, CXCL10, and CD40 are detected in response to combination therapy within resected tumors, and may serve as biomarkers of therapeutic response.

DISCUSSION
Perhaps one of the most consequential advantages of AAV gene therapy for the treatment of GBM, and possibly other solid tumors, is that fundamentally AAV is a modality for distributing encoded transgene into the tumor microenvironment.Vectors carrying unique biotherapeutic transgenes capable of targeting particular cells can be pooled, offering a straightforward method for delivering personalized anti-cancer combination treatment targeting one or multiple aspects of tumorigenicity.Examples include transgenes encoding anti-angiogenics, anti-migrastics, direct cytotoxic agents (e.g.suicide genes), immune stimulating elements, immune checkpoint inhibitor decoys, and even gene-editing elements such as CRISPR/Cas9 or shRNA.Herein, we demonstrate that AAV encoding for the lymphotactic chemokine CXCL9 can be leveraged to engage the immune system to recognize and attack tumor cells by modulating anti-PD-1 immunotherapy in the combinatorial setting.
We found that GBM tumors possess a chemokine signature that favors the recruitment of myeloid and other 'suppressive' immune cells, notably lacking lymphocyte call-and-receive signals.This is unsurprising given that the immune contexture of human GBM is largely comprised of myeloid cells 61 .These data suggest that de ciencies in lymphocyte tra cking and tumor in ltration likely contribute to this problem.
To test if GBM reconstitution with a lymphocyte selective chemokine could improve tra cking, we leveraged AAV gene therapy to generate durable production of CXCL9.With the initial intent of transducing glioma cells to generate tumor-tropic transgene expression, we found that in vivo transduction was not redolent of in vitro screening methods, instead revealing potent transduction of tumor-reactive astrocytes with our lead serotype, AAV6.These ndings advise caution in extrapolating AAV transduction e cacy from in vitro or even ex vivo screening methods to anticipated cell/tissue tropism in the in vivo setting.Despite this discrepancy, we found that AAV6-transduced astrocytes confer a high degree of tumor tropism, where transgene expression was limited to the immediate tumor area.
This data suggests that tumor-associated astrocytes may be selectively susceptible to AAV6 transduction.Heparin sulfate proteoglycan (HSPG), one of the principal host cell receptors for AAV6 transduction 62 is reported to be upregulated by astrocytes in response to brain injury 25,26 , offering a possible explanation for the tumor-tropic nature of astrocyte transduction by AAV6.
AAV targeting of tumor-associated cells may carry several distinct advantages over directly targeting cancer cells.Given the genetic diversity of cancer cells, identifying a serotype that can reliably and consistently transduce tumor cells poses a challenge.Tumor-associated cells such as astrocytes, as demonstrated herein, or alternatively endothelial cells, microglia, etc. are more likely to have a homogenous genotype across patient tumors allowing for targeted off-the-shelf therapeutic development, drastically accelerating treatment timelines and reducing cost as compared to personalized medicine approaches necessary to guide serotype selection for cancer cell targeting.Our data also indicates that targeting of tumor-associated astrocytes is likely to produce a more durable response as these cells are less susceptible to genetic alterations.Direct tumor transduction, on the other hand, is short-lived, likely as a result of vector genome dilution due to the high proliferative capacity of these cells 63 and may require repeat treatments to sustain transgene expression.
Although AAV6 produced CXCL9 clearly improves lymphocyte chemotaxis in vitro, we found that AAV6-CXCL9 monotherapy did not induce robust lymphotaxis into intracranial GBM tumors.The addition of anti-PD-1 ICB dramatically improved lymphocyte tra cking in the combinatorial setting with AAV6-CXCL9, in particular CD8 T cells.These ndings speak toward the prospect of T cell sequestration as an auxiliary barrier to tra cking that jointly contributes to lymphopenia in GBM as rst reported by Chongsathidkiet and colleagues 7 .Therefore, treatment strategies to boost peripheral lymphocyte counts may be necessary to realize the potential of AAV-based chemotactic therapy.Anti-PD-1 ICB offers one such modality, as peripheral expansion of T cells has been validated as a clinical correlate of response to this immune checkpoint inhibitor in certain cancers 64,65 .GBM is largely refractory to anti-PD-1 ICB including in somatically hypermutated GBM, 8,66 reciprocally underscoring the need for multimodal treatment strategies.Our ndings echo this sentiment, with anti-PD-1 ICB showing minimal improvements in overall survival in the preclinical setting.Another consideration is to combine AAV6-CXCL9 gene therapy alongside adoptive cellular transfer of ex vivo modi ed T cells, such as CAR T cell therapy.This strategy would bypass host lymphocyte sequestration altogether by direct intravenous delivery of antigen-speci c T cells, where tumor-tropic CXCL9 expression would facilitate directed tra cking.
Likewise, tumor-speci c homing of CAR T remains an unresolved issue for solid tumors [9][10][11] and may bene t from this particular combinatorial strategy.Both AAV6-CXCL9 and anti-PD-1 ICB also confer secondary mechanisms of action that protect against immune tolerance.CXCL9 is reported to promote lymphocyte differentiation and maturation towards an effector phenotype 67 and anti-PD-1 ICB protects against TME PD-L1 induced lymphocyte exhaustion, thus contributing toward adaptive immune activation and prolongation of cytotoxicity in these cells 68 .Our ndings corroborate this, con rming enhanced CD8 T cell cytotoxicity and TCR signaling pathway activation alongside detection of proin ammatory chemokines and cytokines associated with adaptive immunity, particularly in the context of combination treatment.Total abrogation of therapeutic e cacy with CD8 lymphocyte depletion validates adaptive immune activation as the principal mechanism of anti-tumor response.Combination virotherapy with pooled AAV vectors targeting different aspects of immunogenicity should also be explored further as a viable multimodal strategy for overcoming GBM immune evasion.
We found that direct intratumor delivery of AAV6 was su cient to establish tumor-selective production of CXCL9.While intratumor delivery is considered an invasive method of treatment, generating focal expression of chemokine is critical for e cacy as the mechanism by which these secreted molecules work to facilitate immune cell tra cking is by establishing concentration gradients that immune cells expressing the cognate receptor follow, prompting selective in ltration into in amed tissues 53 .While BBBcrossing serotypes have received signi cant attention for their novel ability to transduce CNS tissue following intravenous delivery, higher doses are required to maintain su cient transduction e ciency, and off-target transduction of peripheral tissues remains a consequence of this delivery method [69][70][71][72] .
Peripheral expression of CXCL9 or other lymphotactic chemokines could counterproductively deter homing to the CNS, reducing the e cacy of this treatment strategy.Furthermore, systemically delivered AAV could also encourage immunological responses resulting in host complement activation and antibody-mediated neutralization, or could prompt adverse toxicity such cytopenias, hepatoxicity, and even neurotoxicity [71][72][73] .To generate focal AAV transgene expression, virus could be delivered via stereotactic injection into unresectable gliomas, for example in tumors that arise in vital structures of the brain or in the event of recurrence, or into the resection cavity during surgery.
In summary, the use of AAV gene therapy has the potential to disrupt the existing treatment paradigm for GBM which relies on radiation, surgery, and cytotoxic chemotherapy.Systemic administration of immunotherapeutics and single-target chemotherapy agents have shown limited clinical e cacy due to dose-limiting toxicities, the constraints of the blood brain barrier (BBB), and the suppressive nature of the TME.This study combines the excellent safety pro le of AAV 13 with focal delivery directly to the TME, bypassing the restrictions and limitations of systemic delivery.AAV biotherapy is minimally invasive, tunable, and enables simultaneous delivery of multiple anti-cancer agents that can be customized to targets unique to each brain tumor.This platform has further application across multiple metastatic tumors where the TME limits the e cacy of immunotherapy.

Methods
Cell Culture KR158B-luc (Kluc) glioma line (provided by Dr. Karlyne M. Reilly, NCI Rare Tumor Initiative, NIH) and GL261 cells have been veri ed histologically as high-grade glioma, and gene expression analysis con rmed appropriate haplotype background and expression of astrocytoma-associated genes 74 .CT-2A were purchased from Millipore Sigma.Primary human glioma cells including L0, L1, L2, CA1, CA2, CA4, CA6, CA7, L23, L26, L31, L34, L38, L47, and HA2 were a kind gift from Dr. Brent A. Reynolds 20 .C8-D1A primary astrocytes were purchased from ATCC.All cells were cultured in DMEM (Fisher-Scienti c) supplemented with 10% FBS (VWR) and 1% Penn-Strep (Life Technologies), and maintained at 37 o C in humidi ed conditions with 5% CO 2 .At the beginning of the study, cells were expanded, stocks made, and thawed vials were maintained in culture for no more than 3 weeks.
In vivo studies C57BL/6J (Strain# 000664), CCR2 RFP CX3CR1 GFP (Strain# 032127), and GREAT (Strain# 017581) mice were purchased from Jackson Laboratory.On day 0, 5x10 4 tumor cells suspended in 50% methylcellulose and 50% saline (Fisher-Scienti c) were stereotaxically injected into murine brain at a depth of 3mm, 2mm lateral to bregma, at a volume of 2µl in 8-16-week-old animals.On day 5, AAV6 vectors were intratumorally injected in the same coordinates as tumor implantation.Where indicated, monoclonal antibody treatment (PD-1 ICB, IgG control, CD8a depletion) was administered beginning Day 5 via intraperitoneal injection and given every 72 hours.Protocols were reviewed and approved by the University of Florida Institutional Animal Care and Use Committee.

Clinical Specimens
De-identi ed patient tissues were procured by the Florida Center for Brain Tumor Research (FCBTR) under the University of Florida Institutional Review Board protocols 201300482.

Drug
InVivoMAb anti-mouse PD-1 and InVivoMAb rat IgG2a isotype control monoclonal antibodies were purchased from BioXcell, diluted in Sterile Saline 0.9% solution (Patterson Veterinary Supply, Inc.), and administered via intraperitoneal injection at a dose of 10mg/kg given every 72 hours for a total of 4 doses.InVivoMAb anti-mouse CD8α and InVivoMAb rat IgG2b isotype control monoclonal antibodies were purchased from BioXcell, diluted in Sterile Saline 0.9% solution (Patterson Veterinary Supply, Inc.), and administered via intraperitoneal injection at a dose of 300µg/mouse given every 72 hours for a total of 6 doses.
AAV Protocol HEK 293T cells (ATCC cat# CRL3216) were cultured to ~70% con uency in two Cellstacks (Corning cat# 3269) per construct and transfected using PEI 25k MW (Polysciences cat# 23966-1) for 3 days.The cells were then harvested via shaking and centrifugation until cell pellet was formed.The pellet was then digested with a nal concentration of 50U/mL of Benzonase (Sigma cat# E8263) and 0.5% sodium deoxycholate in a lysis buffer (150mM NaCl, 50mM Tris-HCl pH 8.4) for 30 minutes at 37 o C. Following incubation, the supernatant was supplemented with 5M NaCl until a 1M nal concentration was achieved.Afterwards, the supernatant was lysed via 3 freeze thaw cycles of -80 o C and 50 o C. The lysate was spun down and supernatant transferred to an ultracentrifuge tube (Beckman cat# 342414), where it is layered with discontinuous layers of iodixanol (Accurate Chemical cat# AN1114542) to separate out viral particles from the supernatant.This was spun for 1 hour at 18 o C at 69,000 rpm.The viral particles were isolated and removed, then washed four times in a dialysis column (Millipore cat# UFC910024) with PBS before being nally puri ed in a sterile ltration column (Millipore cat# UFC30DV00).

AAV Quanti cation
The viruses were titrated by quantitative PCR (Bio-Rad CFX384) using custom probes designed to target the ITR sequences.First, 1 uL of the virus was treated with DNAseI (Thermo Fisher cat# 18068015) for 15 minutes at room temperature, inactivated by heat and EDTA, protein coat of virus digested with Proteinase K (Thermo Fisher cat# 25530049) and nished with a second heat-inactivation step.Following incubations, the sample was diluted and mixed with a Taqman PCR Master Mix (Thermo Fisher cat# 4352042) and the custom designed probes (Thermo Fisher cat# 4332078).The probe sequences were as follows: Forward -GGAACCCCTAGTGATGGAGTT, Reverse -CGGCCTCAGTGAGCGA, Probe -CACTCCCTCTCTGCGCGCTCG.The samples were then compared to a standard curve consisting of a linearized plasmid with ITRs from a range of 1e4 to 1e8 genomic copies per mL.The samples were then run on a standard program of 10 minute denature at 95 o C, then cycled 39 times at 95 o C at 1 minute and 60 o C at 30 seconds.
Lentiviral transduction of tumor cells RFP labeled GL261 and KR158 tumor cells were transduced with a LentiBrite RFP Control Lentiviral Biosensor (Millipore-Sigma), MOI 50.Following cell expansion, RFP-positive cells were FACS sorted using a BD FACSAria-II cell sorter, yielding RFP-stable tumor cells.

Proteome Arrays
Following resection, the right hemisphere (cerebellum removed) of murine brain (tumor-containing) were transferred to 1.5mL microtubes, snap frozen in LN2, and stored at -80 o C until lysis.De-identi ed ash frozen patient GBM tissue was procured from the FCBTR.Tissue shavings were collected on dry ice and transferred to 1.5mL microtubes.300-500µl PBS containing 1x Halt TM Protease/Phosphatase inhibitor (Thermo Fisher) and 1% Triton X-100 (Sigma) was added to samples and transferred to wet ice.Tissue was lysed manually using a 20-gauge needle attached to a 1mL syringe followed by vortexing every 5 minutes for 30-60 minutes.Supernatant was collected following centrifugation at 10,000G at 4 o C, and assayed for protein concentration using a NanoDrop Spectrophotometer.0.75mg of each human sample was used for the Human Chemokine Array Pro ler (R&D Systems, ARY017), and 1mg of each murine sample was used for the Mouse XL Cytokine Array (R&D Systems, ARY028) following manufacturer's instruction.Images were captured using BioRad ChemiDoc MP Imaging System with ImageLab 6.1 software over a series of exposure times.Mean voxel intensity per capture antibody was calculated using Imaris x64 v9.7.0, and protein signal was normalized against internal reference controls.Detected protein and predicted receptor interactions were analyzed and visualized using Circos® 60 .ELISA Tissue specimens were collected at 1 and 2 weeks post-AAV6 injection.Peripheral blood was taken from the anterior vena cava, centrifuged at 12,000rpm x 10 minutes @ RT, and serum collected and stored at -80C.Whole brain was resected, cerebellum removed, and divided into the tumor-bearing (AAV6 injected) and contralateral hemispheres.Naive brain and serum were collected and used to set the baseline for both week 1 and week 2 datasets.Tissue was snap frozen and stored at -80 o C until lysis.Tissue was lysed using RIPA buffer containing 2x Halt protease/phosphatase inhibitor cocktail (Thermo Fisher) with manual dissociation performed using a 20-gauge needle attached to a 1mL syringe followed by vortexing every 5 minutes for 30-60 minutes, and maintained on ice.Following lysis, tissue samples were centrifuged at 12,000 rpm @ 4 o C x 10 minutes.Supernatant was collected, and assayed for protein concentration using a NanoDrop.Protein concentrations were adjusted using RIPA buffer.MIG/CXCL9 ELISA (Thermo Fisher) performed according to manufacturer protocol.Serum diluted 1:2 using Assay Diluent B. Tissue sample concentration: 2mg.All samples run in duplicate.
In vitro Chemotaxis or C8-D1A cells were plated in 24-well dishes at 1x10 5 /well in pre-warmed complete media.AAV6-EGFP or AAV6-CXCL9 (RFP+) was added at a nal concentration of 10 5 VGS.24 hours following transduction, cells were transferred into the outer chambers of µ-Dishes with 3-well culture inserts (Ibidi), 10 4 , suspended in 15µl of growth-factor reduced Matrigel® (Corning).40µl of complete media was added following polymerization for 10 mins at 37 o C in humidi ed conditions with 5% CO 2 .CD3+ T cells were isolated from naïve C57BL/6 mouse spleen (8-12 weeks) using MojoSort CD3 T cell isolation kit (Biolegend) per manufacturer's protocol.T cells were labeled with Cell Trace Violet dye (CTV) (Thermo Fisher) per manufacturer protocol.1x10 4 labeled T cells were suspended in 15 µl cold growth-factor reduced Matrigel® (Corning), and added to the µ-Dish center well.Following polymerization as described above, media was removed from all wells, and 3-well insert was carefully removed.The gap between wells was lled with additional Matrigel to form a continuous substrate and allow for cell migration, and allowed to polymerize for 20 minutes.Complete media was added to cover cells, and incubated at 37 o C in humidi ed conditions with 5% CO 2 .IF images were acquired using a Nikon A1RMP confocal microscope at 1-and 24-hours following co-culture, and T cell chemotaxis was quanti ed as the number of migratory cells visible in either the EGFP or CXCL9 (RFP+) transduced tumor/astrocyte eld.
Single Cell RNA Sequencing, Quality Control, and Data Analysis Following brain resection, cerebellar tissue was removed and the right hemisphere collected for processing.Tissue dissociation and CD45 TIL isolation was performed as described under Tissue Dissociation above.The cells directly after depletion were washed with PBS and 0.04% bovine serum albumin two times and ltered with 40-µn cell drainer.Cells were collected by centrifugation at 500g for 5 min and subsequently counted with hemocytometer.Cells were diluted in ice-cold PBS containing 0.04% BSA at a density of 1000 cells/µL.The nal cell suspension volume equivalent to 8000 target cells was used for further processing.Cells were loaded into a Chromium NextGEM Chip G (10x Genomics, Pleasanton, California) and processed in Chromium X following the manufacturer's instructions.
Preparation of gel beads in emulsion and libraries were performed with Chromium Next GEM Single Cell 3' Kit v.3.1 (Dual Index) according to User Guide provided by the manufacturer.Libraries quality and quantity were veri ed with 2200 TapeStation (Agilent technologies, USA).Libraries were pooled based on their molar concentrations.Pooled library was sequencing on the NovaSeq 6000 instrument (Illumina, San Diego, California).For sequencing 3' gene expression libraries we used following read length: Read 1-28 cycles; i7Index-10 cycles; i5Index-10 cycles; Read 2-90 cycles.Raw base call (BCL) les generated by NovaSeq 6000 sequencer were processed using Cell Ranger software (10X Genomics, version 7) for demultiplexing, barcode processing, and single-cell 3'-gene counting.Mouse genome reference GRCm38 was used for sequence alignment using STAR aligner.A read was considered exonic, if at least 50% of it mapped to an exon, intronic (if it was non-exonic and intersected an intron), or intergenic otherwise.For reads that aligned to a single exonic locus but also aligned to 1 or more non-exonic loci, the exonic locus was prioritized and the read was considered to be con dently mapped to the exonic locus.Cell Ranger also aligned exonic reads to annotated transcripts.An annotated transcript that aligned to the same strand was considered to be con dently mapped to the transcriptome.These con dently mapped reads were used for unique molecular identi er (UMI) counting and subsequent analysis to generate h5 les.
The h5 le of each sample was then processed with Partek Flow analysis software (version 10).Cells meeting the following quality control (QC) parameters were included in the analysis: total reads between 1000 to 33,649; expressed genes between 187 to 5464; mitochondrial reads percentage <20%.Following this selection, we obtained 48159 cells that passed QC lters.Next, features were ltered in order to include only genes expressed in more than 0.01% of cells and 20,785 genes were retained.UMI counts were then normalized following Partek® Flow® recommendations: for each UMI in each sample the number of raw reads was divided by the number of total mapped reads in that sample and multiplied by 1,000,000, obtaining a count per million value (CPM), the normalized expression value was logtransformed.Starting from the normalized data node, we performed clustering analysis for each sample separately by means of graph-based clustering task in Partek® Flow® software which employs the Louvain algorithm.Clustering analysis was done based on the rst 100 principal components.To visualize single cells in a two-dimensional space, we applied Uniform Manifold Approximation and Projection (UMAP) plot using the rst 50 principal components for each sample separately and for the entire data set.Cell types were determined by the expression of marker genes that de ne speci c cell types (Supplemental table 1).Pathway enrichment analysis for tumor cells and immune cells was performed with AUCell algorithm using the NanoString nCounter Immune Exhaustion panel.Interactions between immune populations were analyzed and visualized using the CellChat algorithm 33 .The pheatmap package was used for unsupervised hierarchical clustering to create heatmaps 75 .

Statistical analysis
Statistical analyses performed using GraphPad Prism 9 as described in gure legends.Signi cance determined as p<0.05.Voxel-based co-localization was established using Imaris x64 v9.7.0 using the Coloc module with automatic threshold selection.For survival studies, animals were randomized prior to treatment.

Figure 7 In
Figure 7