Immune Escape After Adoptive T-cell Therapy for Malignant Gliomas

Immunotherapy has revolutionized cancer care (1, 2). However, tumor escape is common and poorly understood (2–4). Herein, we studied tumor escape variants after immunotherapy to draw meaningful insights about escape mechanisms. We then applied that information to study secondary immunotherapy based on escape variant total tumor RNA to treat tumor escape variants.

Gliomas are resistant to chemotherapy, radiation, surgical resection, and even recent developments in immunotherapy, yet glioma escape mechanisms remain poorly understood (5–14). One of the hypothesized methods of brain tumor escape is immunoediting, or the elimination of cells expressing targetable epitopes, the equilibration of remaining tumor, and the outgrowth of tumor escape variants. In peripheral tumors, immunoediting is amplified in the presence of IFNγ and Fas-mediated targeting of tumor, two primary components of T cell-mediated killing (15, 16). Furthermore, it was also recently demonstrated that programmed cell death protein-1 (PD-1) checkpoint blockade can promote T-cell immunoediting of tumors in the periphery (14, 17, 18). The expectation is that once the immunogenic antigens are deleted during immunoediting, the optimal opportunity to target immunogenic tumor antigens has largely passed (12, 17, 19).

Recent evidence in human trials has shown evidence of immunoediting including widespread loss of single antigen targets in gliomas and other cancers after monoclonal chimeric antigen receptor (CAR) T-cell therapy (4, 6–9). While some preclinical studies have indicated that this single antigen loss may not affect antitumor immunity (20), conclusions from these recent human studies recommend employing cell therapies with multiple antigen targets and the use of combinatorial therapies to activate host immunity and overcome the immunosuppressive tumor microenvironment (7, 21). Given these findings and similar evidence in the periphery, there is now an expectation that treatments focused on single or limited antigen pools may have limited long-term success and may even promote immunoediting and formation of tumor escape variants.

Additional tumor escape mechanisms implicated in peripheral tumors include loss of MHC class I (MHC-I) and upregulation of immune checkpoint molecules (3, 18). MHC-I is required for CD8⁺ cytotoxic T-cell targeting and killing of cells that present the T-cell's cognate antigen. Tumor cells can evade T-cell targeting by downregulation or deletion of MHC-I (18, 22). In this setting, natural killer (NK) cells possess cytotoxic capacity against MHC-I^lo tumors because MHC-I is a key inhibitory ligand for NK immunoglobulin-like receptors (23). While various regimens of lymphokine-activated killer cells have been investigated for the treatment of brain tumors, convincing demonstrations of NK-cell antitumor efficacy remain elusive (24–28). It also remains unclear if NK cells provide any role during adoptive cellular therapy (ACT) for brain tumors.

Our group developed an ACT platform that targets multiple tumor antigens with one infusion and is demonstrably efficacious in multiple murine models of brain malignancies (29–31). ACT employs bone marrow–derived dendritic cells (DC) pulsed with total tumor RNA to ex vivo activate a polyclonal population of tumor-specific T cells (29, 31). These cells are adoptively transferred into tumor-bearing hosts following host conditioning and hematopoietic stem and progenitor cell (HSPC) transplant and antitumor immunity is maintained with weekly tumor RNA-pulsed DC vaccines (Fig. 1A). This combinatorial strategy strongly modulates the tumor microenvironment and promotes continued intratumor T-cell activation (29–32).

ACT prolongs median survival and mediates approximately 30% long-term cures in malignant brain tumor-bearing hosts. However, approximately 70% of treated animals succumb to disease for unknown reasons (31). We hypothesized, based on our previous data highlighting the ACT-mediated upregulation of Ifng and Fasl, that brain tumors are primed for immunoediting during ACT (15, 16, 31). While investigation of viable human brain tumor tissue after failure of immunotherapy is limited because of a small number of biopsies postprogression, mouse brain tumor tissue is readily available after escape of immunotherapy. Therefore, we isolated treatment-resistant tumor escape variants from mice that succumbed to disease after initial suppression of tumor growth and eventual escape from ACT (termed TOGA). We used these immune-escaped models and in vitro T-cell functional assays, flow cytometry, and RNA analysis to investigate mechanisms of immune escape. These studies revealed that escape mechanisms include a shift in the immunogenic tumor antigens, downregulation of MHC-I on tumor, and upregulation of checkpoint molecules on tumors, NK cells, and T cells.

To evaluate the retreatment of escape variants with immunotherapy, we developed escape variant-specific T cells that were primed and expanded using DCs pulsed with tumor escape variant total tumor RNA. This polyclonal population of escape variant-specific T cells demonstrated heightened ability to target TOGA tumors compared with primary glioma-specific T cells. When administered with DC vaccines, host conditioning, and HSPC transplant to TOGA-bearing animals, TOGA-ACT prolonged median survival by 60% compared with untreated animals. When we introduced PD-1 blockade during ACT administration, TOGA-ACT+PD-1 blockade prolonged median survival by 110% compared with untreated animals. When CD8⁺ T cells or NK1.1⁺ NK cells were depleted during TOGA-ACT+PD-1 therapy, the therapeutic benefit was significantly ablated. With this flexible combinatorial approach, ACT+PD-1 blockade can be employed to immunologically reject primary and recurrent gliomas.

Female six- to eight-week-old C57BL/6 mice (Jackson Laboratories, 000664), transgenic DsRed mice (Jackson Laboratories, 006051), transgenic GFP mice (Jackson Laboratories, 004353), and GREAT mice (Jackson Laboratories, 017580) were used. The facilities at the University of Florida Animal Care Services are fully accredited by the American Association for Accreditation of Laboratory Animal Care, and all studies were approved by the University of Florida Institutional Animal Care and Use Committee.

Imaging was performed as previously described using the IVIS Spectrum Imaging System (29).

Total tumor RNA was isolated from two sources: either in vitro cell lines or directly postexcision. Qiagen RNAeasy Kit (Qiagen, 74104) was utilized for all extractions. Manufacturer's guidelines were followed.

Untreated KR158B and GL261 tumors were harvested 3 weeks postimplantation and 6-week-old C57BL/6 mouse brains were harvested for transcriptome analysis. cDNA preparation and sequencing for these nine samples were described previously (33). TOGA1.1 tumor was harvested at humane endpoint at 63 days posttumor implantation. For this sample, RNA sequencing (RNA-Seq) libraries were generated using the SMARTerTM Ultra Low input RNA Kit and KAPA LTP Library Preparation Kit Illumina platforms following the manufacturers recommended protocols (Clontech, catalog No. 634935 and KAPA Biosystems, catalog No. KK8230). Analysis for all RNASeq samples was preformed on University of Florida High Performance Cluster (HiPerGator). Briefly, low-quality reads and adaptors of fastq data were trimmed by trim_galore (34) then aligned to Ensembl 91 mouse genome by RSEM to extract sample gene expression (35, 36). This algorithm allows us to align reads from different library preparation on transcript-level and normalized gene expression by transcripts per kilobase million (TPM) which makes samples more comparable among different groups. TPM were compared among the groups. CancerSubtypes (37) and pheatmap (38) were used for data normalization and visualization. The top 1,000 most variable genes were extracted by CancerSubtypes and clustered with pheatmap. Then, genes were extracted from each of six clusters (Fig. 1F). In addition, gene networks were generated with stringdb's confidence mode. Principal component analysis was performed with pca3d and rgl (39) (Supplementary Fig. S5).

Tumor-reactive T cells were generated as described previously (29–31, 40). For TOGA T cells used in TOGA-ACT, the same protocol was used with a different RNA species that was isolated from immune-escaped TOGA lines. Briefly, total tumor RNA was electroporated into DCs and tumor-specific DCs were then used to prime naïve hosts. One week later, primed splenocytes were harvested and cocultured with additional tumor RNA-specific DCs and IL2. After 5–7 days of coculture and periodic splitting, polyclonal, tumor-specific T cells were harvested and utilized. Treatment of tumor-bearing mice began with 5-Gy lymphodepletion or 9-Gy myeloablation on day 5 postintracranial injection with X-ray irradiation (X-RAD 320). On day 6 postintracranial tumor injection, mice received a single intravenous injection with 10⁷ autologous ex vivo expanded TTRNA T cells with either 5 × 10⁴ lineage-depleted (lin⁻) hematopoietic stem and progenitor cells (Miltenyi Biotec, catalog No. 130-090-858). Beginning day 7 posttumor injection, 2.5 × 10⁵ TTRNA-pulsed DCs were injected intradermally weekly for 3 weeks.

In vitro experiments utilized IFNγ release from T cells in functional assays as a measure of T-cell activity. Functional assays included effector T cells and targets (pulsed DCs or tumor cell lines) that are cocultured in a 10:1 ratio in 96-well U-bottom plates in triplicate. IFNγ Platinum ELISAs (Affymetrix, catalog No. BMS606) were performed on a cellular media that was harvested and frozen from the supernatants after 48 hours. The supernatant transfer system utilized the 10:1 ratio of T cells and DCs to generate supernatants.

KR158B-luc was murine glioma was used courtesy of Dr. Karlyne M. Reilly (29, 41) as described previously (31). TOGA cell lines were isolated from excision of brain tumors from mice after succumbing to KR158B-luc tumors post-ACT treatment (including 9-Gy irradiation, HSPC transplant, tumor-reactive T cells, and three DC vaccines). TOGA tumors were utilized identically to KR158B-luc tumors. Cell lines tested negative for Mycoplasma contamination (IDEXX, September 26, 2017).

In vivo antibodies included 10 mg/kg anti–PD-1 mAb (BioXcell, BE0146), 250-ug anti-CD8 depletion antibody (BioXcell, BE0223), and 250-ug anti-NK1.1 depletion antibody (BioXcell, BE0036) were administered as described previously (42). Antibodies were administered according to treatment diagrams in the figures.

Tissue was processed as described previously (30, 31). Brain tumor dissection began posteriorly with a midline cut in the skull and rongeur removal of skull laterally. Tumor resection extended to gross borders of tumor mass near the site of injection. Tumors were dissociated mechanically with a sterilized razor blade and chemically with papain (Worthington, catalog No. NC9809987) for 30 minutes. Tumors were filtered with a 70-μm cell strainer (BD Biosciences, catalog No. 08-771-2) prior to antibody incubation.

Quantity of RNA was measured using NanoDrop 2000. Each reverse transcription reaction was performed using 250 ng of RNA in a 10-uL volume. cDNA libraries were generated using the SMARTscribe reverse transcriptase kit from total tumor RNA as per the manufacturer's instructions (Clontech, catalog No. 639537). Samples were stored at −80°C for subsequent qPCR analyses. The CFX Connect Real-Time PCR Detection System (Bio-Rad Laboratories, 1855201), TaqMan Universal PCR Master Mix (Applied Biosystems, 4324018), and validated TaqMan probes were used for qPCR analyses. Transcriptional expression of H2k1 (catalog No. 4331182, Mm01612247_mH) and Pdl1 (catalog No. 4331182, Mm00452054_m1) were normalized to Hprt (catalog No. 4331182, Mm03024075_m1) per sample and expressed as fold-change versus untreated tumors. A reaction volume of 10 uL/well was prepared on ice using 2.5 ng of cDNA template per well. Reagent preparation and thermal cycling parameters were followed as per manufacturer's instruction. No template controls and reverse transcriptase negative samples were included to ensure the absence of contamination and genomic DNA.

Flow cytometry was done with the BD FACS Canto-II machine under the management of the University of Florida Cancer Center (Gainesville, FL). Cell sorting was performed using the BD FACS Aria-II. DsRed⁺ mouse-derived cells were detected in FL-2. Analysis was performed with FlowJo version 10 (Tree Star). Results were analyzed using isotype controls after debris and doublets were excluded and target populations were gated on size and granularity. FACS antibodies are listed in Supplementary Table S1.

Statistical tests were performed using GraphPad Prism 8. For in vitro experiments, we utilized the unpaired Student t test and for in vivo experiments we utilized the Mann–Whitney rank-sum test. Correlation studies employed Pearson two-tailed test for correlation. Experiments are powered to include at least five randomized animals per group. For survival experiments, we utilized seven or more animals to attain enough power to distinguish groups after analysis by Mantel–Cox log-rank test. Median survival for KR158B-luc is 42 days and for TOGA1.1 is about 21 days. Significance was determined as P < 0.05.

ACT leads to approximately 30% long-term cures in preclinical murine models including NSC medulloblastoma, K2 brainstem glioma, and KR158B-luc glioma (31). Here, C57BL/6 mice received orthotopic injection of syngeneic KR158B-luc high-grade glioma and were treated with ACT as described previously (29–31). Despite a significant increase in median and overall survival, a proportion of glioma-bearing mice treated with ACT succumb to disease (29–31). Bioluminescent imaging revealed that all ACT-treated mice maintain control of tumor growth up to 21 days post-ACT (P < 0.0001) with a fraction demonstrating long-term survival (Fig. 1A–C).

Tumors from this experiment that escaped ACT treatment after a period of immunologic control were referred to as TOGA1.1 and TOGA1.2. When orthotopically injected into naïve C57BL/6 mice, TOGA1.1 tumor was demonstrably more aggressive than its primary counterpart KR158B-luc (median survival, 21.5 days vs. 41 days; P < 0.0001; Fig. 1D). RNA-Seq revealed global genetic differences between the primary KR158B-luc glioma and the escaped tumor TOGA1.1 (Fig. 1E). Gene expression of TOGA1.1 was also compared with global gene expression of primary murine glioma cell lines KR158B-luc, GL261 glioma, and normal brain tissue (Fig. 1E). We found that between TOGA1.1 and primary KR158B-luc 8,487 genes are differentially expressed by at least twofold, indicating that the selection process of ACT led to genetically distinct gliomas.

After observing genetic differences between primary KR158B-luc and escaped TOGA1.1, we next sought to determine whether the tumors were immunologically discrete. We hypothesized that if the escape variants were immunologically distinct, T cells generated from the primary tumor would no longer provide a remarkable treatment effect. To test this, we administered serial adoptive T-cell transplants specific for the primary KR158B-luc to KR158B-luc-bearing animals. While there was not a significant difference in median survival, giving two serial T-cell transplants did induce a shift from 30% to 40% long-term cures (P = 0.2503, P = 0.5427; Fig. 1G; Supplementary Fig. S1A). Regardless, tumor escape persisted.

We next asked whether T cells generated against the primary KR158B-luc glioma maintain immunologic recognition of the escaped tumor TOGA1.1. To determine whether KR158B-luc–specific T cells recognize cognate antigen on TOGA1.1, KR158B-luc–specific T cells were generated then used as effector cells against either KR158B-luc, TOGA1.1, or B16-F10-OVA melanoma tumor target cells in a functionality assay (Fig. 2A). Supernatant IFNγ secretion was measured as an indication of antitumor T-cell reactivity. KR158B-luc T cells secreted IFNγ upon recognition of KR158B-luc but had markedly diminished recognition of TOGA1.1 or B16-F10-OVA melanoma cells (KR158B-luc: 2911pg/mL, TOGA1.1: 448pg/mL, P = 0.0767; Fig. 2A). This strongly suggests that T cells specific for primary tumor provide very little immune function against escape variant tumor cells.

We next investigated whether we could regenerate a secondary T-cell therapy that was more specific for the TOGA1.1 escape variant and could mediate significant antitumor function. To do this, we generated TOGA1.1-specific T cells by using total tumor RNA isolated from TOGA1.1 cells. TOGA1.1-specific T cells were then tested in a functionality assay to target either KR158B-luc, TOGA1.1, or B16-F10-OVA tumor cells. TOGA1.1-specific T cells secreted significantly more IFNγ upon coculture against TOGA1.1 over primary KR158B-luc tumor (TOGA1.1: 3782pg/mL, KR158B-luc: 1200 pg/mL, P = 0.0128; Fig. 2B). Importantly, this demonstrates that the TOGA1.1 tumor, which was outgrown from a primary KR158B-luc glioma that escaped ACT, is immunologically distinct from its primary counterpart.

Our data thus far demonstrate that escape from ACT results in immunologically distinct tumors from the primary tumor. We next sought to determine whether the escaped tumors after treatment are immunologically distinct from other escape variants originating from the same tumor that received the same ACT treatment.

Because T cells generated against the primary KR158B-luc glioma no longer recognize escaped TOGA1.1 glioma, we then specifically asked whether T cells generated against tumors that have escaped ACT recognize and target each other. Here we used TOGA1.1 and TOGA1.2 which are escaped tumors that both originated from the same KR158B-luc tumor which escaped the same T-cell treatment (Fig. 1D). We generated antigen-specific T cells against TOGA1.1 and used those to target either TOGA1.1 tumor cells or TOGA1.2 tumor cells and supernatant IFNγ was measured as an indicator of T-cell recognition of cognate tumor antigen. While IFNγ was released when TOGA1.1 T cells were cultured with TOGA1.1, minimal IFNγ was detected when TOGA1.1 T cells were cultured with TOGA1.2 tumor cells (TOGA1.1 vs. TOGA1.2, P = 0.0224; Fig. 2C). The converse experiment was conducted when TOGA1.2-specific T cells were generated and used to target either TOGA1.1 or TOGA1.2 glioma cells. The TOGA1.2 T cells failed to recognize TOGA1.1 escaped tumor cells (TOGA1.1 vs. TOGA1.2, P = 0.0313; Fig. 2D). Therefore, there was successful recognition of cognate tumors, yet minimal cross-reactivity between T cells and the converse tumor target (Fig. 2C and D). This indicates that tumor escape variants were at least partly unique.

It has been previously reported that spectratyping using FACS analysis of T-cell receptor (TCR) Vβ may be used to track the reactivity of specific TCR Vβ families toward target cells (40, 43, 44). For validation of this method, we verified the reactivity of TCR Vβ 5.1, 5.2 T cells for OVA-expressing target cells (45; Supplementary Fig. S1B and S1C). We next evaluated the proportion of TCR Vβ families within the polyclonal pools of KR158B-luc T cells or TOGA1.1 T cells belong after ex vivo expansion and determined they were largely comparable (Supplementary Fig. S1D). In a recent manuscript, we discovered that the primary TCR Vβ family that drives ACT response to KR158B-luc in vivo is TCR Vβ6 (40). We therefore FACS-sorted TCR Vβ6⁺ T cells from the polyclonal KR158B-luc T-cell pool. After sorting, we performed T-cell functional assays against KR158B-luc or TOGA1.1 tumor cells. TCR Vβ6⁺ KR158B-luc–specific T cells were eightfold more reactive toward KR158B-luc when compared with TOGA1.1 tumor cells (KR158B-luc: 771 pg/mL, TOGA1.1: 94 pg/mL, P = 0.0033; Fig. 3A). This indicates that TCR Vβ can be used to broadly demarcate specificity of T cells between KR158B-luc T cells and TOGA1.1 T cells and that TCR Vβ6⁺ KR158B-luc T cells were not as capable of targeting epitopes that were present on TOGA1.1. We therefore tested the reactivity of individual FACS-sorted TCRVβ families in vitro against their cognate tumor. This revealed that the T-cell TCRVβ families that are reactive for TOGA1.1 tumors are largely different than those that are reactive for KR158B-luc and TOGA1.2 tumors (Fig. 3B and C).

We next investigated alternative mechanisms that could be responsible for escape following ACT. Given the dependence of CD8⁺ T cells on tumor MHC-I expression, we investigated the potential for MHC downregulation on ACT-treated tumors. We identified a downregulation of MHC-I by percent and mean fluorescence intensity (MFI) on the TOGA1.1 tumor compared with KR158B-luc (Fig. 4A and B). We then investigated MHC-I expression on tumors of a cohort of animals treated with ACT in a separate experiment (Fig. 4C). In the KR158B-luc–bearing group, four of seven tumors maintain expression of MHC-I after ACT, three of seven tumors displayed a marked decrease in expression of MHC-I (P = 0.0289; Fig. 4D–F). This bifurcation in response was verified by percent expression, MFI, and PCR (Fig. 4D–F).

We additionally tested the impact of ACT on MHC-I expression in TOGA1.1-bearing animals. This revealed that MHC-I was ubiquitously downregulated on TOGA1.1 tumors after TOGA-ACT by percent expression and MFI (MHC-I⁺, 22.59%–10.12%, P = 0.0033; Fig. 4D, G, and H). It should be noted that while only a fraction of ACT-treated KR158B-luc–bearing hosts demonstrated downregulation of MHC-I, all ACT-treated TOGA-bearing animals demonstrated a uniform downregulation of MHC-I by flow cytometry. We then generated TOGA-ACT T-cells from DsRed transgenic animals and tracked T cells in tumors of ACT-treated hosts. At endpoint, we excised tumors and analyzed tumor-infiltrating lymphocyte (TIL) TCR Vβ families. The Vβ families driving at least partial in vitro reactivity against TOGA1.1 tumors (4, 5.1/5.2, 6, 8.1/8.2, and 8.3) also comprised the majority of the TIL fraction of the families tested at endpoint (Supplementary Fig. S1E).

Given recent reports detailing the importance of checkpoint molecules and MHC-I expression in determining the cytotoxic response in peripheral tumors (42), we next investigated PD-L1 expression on tumors. We determined that ACT induces an upregulation of PD-L1 by PCR and flow cytometry (P = 0.0466, P = 0.001; Fig. 5A and B). PD-L1 expression was then compared with MHC-I at the gene level (Fig. 5C) and the protein level (Fig. 5D) within each sample. Within-sample analysis revealed a strong positive correlation between the two molecules (r = 0.9953, P < 0.0001), indicating that tumors that escape by downregulating MHC-I are largely PD-L1^lo. In addition, in concordance with this correlative data, TOGA-ACT treated hosts, which are low for MHC-I, remained low for Pdl1 even after ACT (Fig. 5E). This suggests that TOGA1.1 is a variant that primarily escapes by MHC-I downregulation and not necessarily by PD-L1 upregulation. This may suggest distinct mechanisms of escape whereby brain tumors primarily either upregulate PD-L1 or downregulate MHC-I in response to escape T-cell pressure.

To determine whether ACT induced a cellular immune response, we investigated cytotoxic immune cell infiltration in the tumor-draining cervical lymph nodes (TDLNs) and tumor microenvironment during ACT. For these analyses, we specifically investigated the presence and activation of CD8⁺ T cells and CD3⁻CD19⁻F4/80⁻Ter-119⁻NK1.1⁺NKp46⁺ NK cells. We studied expression of CD107a as a marker of degranulation of activated cytotoxic cells as well as PD-1, a marker of T-cell activation that functions as a regulatory receptor (46–48). These analyses revealed that ACT induced a 22-fold increase in CD8⁺ T cells and a threefold increase in NK cells in TDLN by percent and absolute counts (P = 0.001, P = 0.001; Fig. 5F and G). Importantly, NK cells are not derived from the adoptive T-cell transfer (Supplementary Fig. S2). This demonstrates an important link between T-cell therapy and NK-cell engraftment in TDLN. A concomitant increase of both absolute number and percent of CD107a⁺ and PD-1⁺ T cells and NK cells were also seen with ACT (Fig. 5H and I). ACT induced a 2.5-fold greater expression of CD107a on T cells and a 4.5-fold greater expression on NK cells (P = 0.001, P = 0.001; Fig. 5H and I). In addition, ACT induces a sixfold greater expression of PD-1 on T cells and a fivefold greater expression on NK cells (P = 0.001, P = 0.001; Fig. 5H and I). When we investigated MFI expression on tumor-infiltrating T cells and NK cells, we determined that ACT induced greater expression of PD-1 on T cells and NK cells (P = 0.001, P = 0.042; Supplementary Figs. S2–S4) while inducing greater expression of CD107a on NK cells (P = 0.001; Supplementary Figs. S2–S4). This data demonstrates that ACT mediates T-cell and NK-cell engraftment and activation.

We next performed a series of survival experiments to test the capacity of ACT to overcome the three described immune escape mechanisms: tumor antigen changes, MHC-I downregulation, and checkpoint molecule upregulation on immune cells and tumor cells. On the basis of the in vitro data, we anticipated that tumor-specific ACT would overcome the shift in immunogenic tumor antigens and allow for tumor-specific targeting. We also anticipated MHC-I downregulation could be overcome because tumor-specific T cells still targeted escape variants in vitro and simultaneously enhanced infiltration and activation of NK cells and T cells in vivo. Finally, we anticipated combinatorial addition of PD-1 checkpoint blockade could prevent immune checkpoint-mediated escape. Administration of TOGA1.1-specific ACT prolonged median survival of TOGA1.1-bearing hosts by 60% (median survival, 24–33 days, P = 0.0003; Fig. 6A and B). Because we identified upregulated PD-1 on NK cells and T cells previously during ACT, we tested the ability of PD-1 to enhance the antitumor benefit of TOGA-ACT (Fig. 6C). When we administered TOGA-ACT+PD-1 to TOGA-bearing animals, this yielded 110% prolongation of median survival (UnTx vs. ACT+PD-1, 20–42 days, P = 0.0002, ACT vs. ACT+PD-1, 32–42 days, P = 0.1522; Fig. 6D).

We previously demonstrated an ACT-induced engraftment and activation of NK cells and T cells in TDLN and tumors but had not yet investigated their impact on antitumor efficacy. To test the functional impact of NK cells and T cells during therapy, we utilized depleting antibodies before and during TOGA-ACT+PD-1. When depleting antibodies were administered to deplete NK1.1⁺ NK cells or CD8⁺ T cells from animals during ACT+PD-1 therapy, the survival benefit was significantly diminished (anti-NK1.1, 42–31 days, P = 0.0122; anti-CD8, 42–27 days, P = 0.0038; Fig. 6E). Therefore, both cell types are required for optimal efficacy.

In summary, brain tumor escape from ACT occurs through at least three mechanisms including a shift in immunogenic tumor antigens, MHC-I downregulation, and upregulation of checkpoint molecules. However, ACT promotes the infiltration of both NK cells and T cells, two populations that can cytotoxically target tumors regardless of MHC-I status. When we regenerated T cells specific for tumor escape variants, they were more specific for their cognate escape variant tumor cells when compared with primary glioma cells. When this escape variant-specific T-cell approach was applied in vivo with an ACT regimen, it significantly prolonged median survival. PD-1 blockade during ACT enhanced this benefit and depletion of NK cells and CD8⁺ T cells highlighted the requirement for both cell types for optimal efficacy.

This report highlights multiple mechanisms of escape during immunotherapy for malignant gliomas including a shift in immunogenic tumor antigens, downregulation of MHC-I, and upregulation of checkpoint molecules. This study also demonstrates a translatable method of analyzing tumor that has escaped immunotherapy in situations where rebiopsy is feasible. This mechanistic investigation which was done in recalcitrant primary and recurrent murine gliomas demonstrates that tumor immunity is complex and that single pathways are not solely responsible for escape, which is highly relevant to cancer immunology. For instance, when tumors are targeted with antigen-specific CAR-T cells or other single-antigen targeting modalities, antigen loss provides a tumor escape route (4, 7, 8, 20, 49). Alternatively, checkpoint blockade strategies may promote antitumor immunity, but can encourage escape from immune surveillance (17, 50–54). In another pathway altogether, the most advanced adaptive immunotherapy strategy can be immobilized or stripped of activation by the network of protumor myeloid cells that are endemic to all antitumor immune responses (55, 56). These are all part of a larger coordinated immune system that self-regulates to the extent that opportunistic tumor cells can benefit in the fray.

The role of CD8⁺ T cells in antitumor immunity is well established and was been recapitulated in this report. They were required for optimal antitumor immunity even against TOGA tumors, which express relatively lower MHC-I. However, the role of NK cells in brain tumor immunity is not as well appreciated. While much of the immunotherapy field has focused on checkpoint molecules on T cells, recent evidence has also implicated checkpoint molecules on NK cells. In lymphoma models, PD-1 checkpoint blockade can promote activation of NK cell–mediated immunity, opening the possibility for simultaneous activation of T and NK cells (42). However, PD-1 blockade alone does not mediate remarkable efficacy in primary glioma tumor models (30). What we demonstrated here is that ACT+PD-1 blockade can activate T cells and NK cells and promote antitumor efficacy. We anticipate that NK cells may follow parallel pathways of cytokine and chemokine-mediated migration and activation that promote T-cell immunity. We previously demonstrated that HSPCs in the tumor microenvironment release MIP-1α, which promotes T-cell migration to tumors in coordination with ACT that induces upregulation of Ccl5, Ifng, and other T-cell cytokines and chemokines (10, 29, 31). There are multiple reports that NK cells can migrate and become activated through the same molecules and future investigation will explore this further (57, 58). Regardless, through this combined engagement of T cells and NK cells during ACT+PD-1, CD8⁺ T cells can kill any tumor cell that has MHC-I while NK cells can kill MHC^lo tumors that may appear in response to T-cell pressure. Perhaps, this encourages tumors into “escaping” into the cytotoxic snare of either NK cells or T cells in MHC-pliable tumors.

After ACT, some tumors displayed preferential escape through PD-L1 upregulation or MHC-I downregulation. In the within-sample analysis, there were largely no tumor samples that expressed low MHC-I and high PD-L1 or high MHC-I and low PD-L1. Previous experiments that demonstrated the impact of NK cells during PD-1/PD-L1 blockade used a cell line that was MHC^lo and transduced with Pdl1 (42). Then with increased tumor Pdl1, NK cells were engaged by PD-1 or PD-L1 blockade to generate antitumor efficacy against MHC-I^lo tumors. In our study, PD-L1 was only highly expressed when MHC-I was highly expressed. Because CD8⁺ T cells rely on MHC-I and NK cells depend on its absence, our study suggests that the PD-1/PD-L1 axis is more relevant in the function of CD8⁺ T-cell immunity during ACT. This may partially explain why PD-1 blockade did not induce significant activation of NK cells in TDLN, whereas PD-1 blockade did induce CD8⁺ T-cell activation. However, NK cells and CD8⁺ T cells were both required for optimal efficacy during combined ACT+PD-1. In total, these data indicate a potential bifurcation in the escape mechanisms of brain tumors that is reminiscent of a mechanistic model that has been described in other tumors as escaping by “natural selection” or “acquired resistance” (22, 59). Future studies in brain tumors should perform single-cell analysis to longitudinally track single cells and their predetermined or acquired resistances.

In the cohort of tumors that escape by PD-L1 upregulation and we anticipate that TOGA-ACT+PD-1 may provide considerable benefit and perhaps long-term tumor control. Given recent articles on neoadjuvant PD-1 blockade in glioblastoma (60, 61), we anticipate robust development in neoadjuvant or adjuvant ACT+PD-1 combinations in primary and resistant tumors. With ACT inducing a subset of tumors to upregulate PD-L1, perhaps some escaped tumors are more readily treatable with PD-1 blockade after stratification in the PD-L1^hi escape variant subtype. On the contrary, the addition of PD-1 to TOGA-ACT in the TOGA1.1 model was beneficial but limited. We anticipate this may be due to TOGA1.1 demonstrating a preference for MHC-I downregulation (a potential form of “natural selection”), not PD-L1 upregulation. Even though ACT induces considerable T-cell engraftment (22-fold increase) and significant but limited NK-cell engraftment (threefold increase), we generated a significant survival benefit. We anticipate that for tumors that escape by MHC-I downregulation, additional combinatorial NK-cell activation strategies in addition to ACT+PD-1 may be beneficial. MHC-I was downregulated on TOGA1.1, but it is important to recognize that MHC-I does not appear completely lost on TOGA1.1. It was detectable by PCR and flow cytometry and our functional assays indicate T-cell targeting despite lower MHC-I. However, future studies should investigate the single cell-level kinetics of MHC-I changes and antigen expression levels in escape variants and determine the relative requirement of MHC-I and antigen presence for adequate T- and NK-cell function.

The strengths of this study include the use of multiple therapeutic brain tumor models including the generated escape variant models. Here we laid the groundwork for three primary mechanisms of escape in malignant gliomas. Even after escape, we generated a significant benefit through novel generation of escape variant-specific ACT. Future directions of these studies include the stratification and early detection of escape variant subtypes. With that classification, future cancer regimens can be diversified and tailored with attention to how and when specific tumors escape.

D.A. Mitchell reports grants from NIH/NCI during the conduct of the study; other from iOncologi, Inc. (equity), Annias Immunotherapeutics, Inc. (licensed IP), Bristol Myers Squibb (advisor), and Tocagen, Inc. (advisor) outside the submitted work; has a patent for US20180153982A1 issued and licensed to Immunomic Therapeutics, Inc., a patent for US10660954B2 issued and licensed to iOncologi, Inc., and a patent for WO2018119243A1 pending and licensed to iOncologi, Inc.; and is cofounder and holds equity stake in iOncologi, Inc., a biotechnology company specializing in immuno-oncology. C.T. Flores reports other from iOncologi (cofounder of biotech start-up) outside the submitted work; and a patent for US10660954B2 issued and licensed to iOncologi, Inc. No potential conflicts of interest were disclosed by the other authors.

T.J. Wildes: Conceptualization, data curation, writing-original draft. K.A. Dyson: Data curation. C. Francis: Data curation. B. Wummer: Data curation. C. Yang: Formal analysis. O. Yegorov: Formal analysis. D. Shin: Data curation, formal analysis. A. Grippin: Data curation. B. DiVita Dean: Data curation. R. Abraham: Data curation. C. Pham: Data curation. G. Moore: Data curation. C. Kuizon: Data curation. D.A. Mitchell: Funding acquisition, project administration. C.T. Flores: Conceptualization, resources, data curation, formal analysis, supervision, writing-original draft.

This research was supported by the University of Florida Health Cancer Center Predoctoral Award (to T.J. Wildes); American Brain Tumor Association Research Collaboration grant (to C.T. Flores); Alex's Lemonade Stand Young Investigator grant (to C.T. Flores); Florida Center for Brain Tumor Research grant (to C.T. Flores); Wells Foundation; and University of Florida Clinical and Translational Sciences Award (UL1TR001427).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.