GD2-CAR T cell therapy for H3K27M-mutated diffuse midline gliomas

Diffuse intrinsic pontine glioma (DIPG) and other H3K27M-mutated diffuse midline gliomas (DMGs) are universally lethal paediatric tumours of the central nervous system1. We have previously shown that the disialoganglioside GD2 is highly expressed on H3K27M-mutated glioma cells and have demonstrated promising preclinical efficacy of GD2-directed chimeric antigen receptor (CAR) T cells2, providing the rationale for a first-in-human phase I clinical trial (NCT04196413). Because CAR T cell-induced brainstem inflammation can result in obstructive hydrocephalus, increased intracranial pressure and dangerous tissue shifts, neurocritical care precautions were incorporated. Here we present the clinical experience from the first four patients with H3K27M-mutated DIPG or spinal cord DMG treated with GD2-CAR T cells at dose level 1 (1 × 106 GD2-CAR T cells per kg administered intravenously). Patients who exhibited clinical benefit were eligible for subsequent GD2-CAR T cell infusions administered intracerebroventricularly3. Toxicity was largely related to the location of the tumour and was reversible with intensive supportive care. On-target, off-tumour toxicity was not observed. Three of four patients exhibited clinical and radiographic improvement. Pro-inflammatory cytokine levels were increased in the plasma and cerebrospinal fluid. Transcriptomic analyses of 65,598 single cells from CAR T cell products and cerebrospinal fluid elucidate heterogeneity in response between participants and administration routes. These early results underscore the promise of this therapeutic approach for patients with H3K27M-mutated DIPG or spinal cord DMG.

Diffuse intrinsic pontine glioma (DIPG) and other H3K27M-mutated diffuse midline gliomas (DMGs) are universally lethal paediatric tumours of the central nervous system 1 . We have previously shown that the disialoganglioside GD2 is highly expressed on H3K27M-mutated glioma cells and have demonstrated promising preclinical efficacy of GD2-directed chimeric antigen receptor (CAR) T cells 2 , providing the rationale for a first-in-human phase I clinical trial (NCT04196413). Because CAR T cell-induced brainstem inflammation can result in obstructive hydrocephalus, increased intracranial pressure and dangerous tissue shifts, neurocritical care precautions were incorporated. Here we present the clinical experience from the first four patients with H3K27M-mutated DIPG or spinal cord DMG treated with GD2-CAR T cells at dose level 1 (1 × 10 6 GD2-CAR T cells per kg administered intravenously). Patients who exhibited clinical benefit were eligible for subsequent GD2-CAR T cell infusions administered intracerebroventricularly 3 . Toxicity was largely related to the location of the tumour and was reversible with intensive supportive care. On-target, off-tumour toxicity was not observed. Three of four patients exhibited clinical and radiographic improvement. Pro-inflammatory cytokine levels were increased in the plasma and cerebrospinal fluid. Transcriptomic analyses of 65,598 single cells from CAR T cell products and cerebrospinal fluid elucidate heterogeneity in response between participants and administration routes. These early results underscore the promise of this therapeutic approach for patients with H3K27M-mutated DIPG or spinal cord DMG.

Dose level 1
Beginning June 2020, four participants were enrolled on dose level 1 (DL1; 1 × 10 6 GD2-CAR T cells per kg administered intravenously; three participants with DIPG and one participant with spinal cord DMG; 5-25 years of age; one male and three female; Supplementary Table 1); all participants were more than 6 months from completion of standard radiotherapy. GD2-CAR T cells were successfully manufactured and met release criteria for all four patients (Supplementary Table 2). During cell manufacturing, patient 1 with spinal DMG experienced rapid tumour progression and was removed from the study but was treated at DL1 on a single-patient compassionate emergency investigational new drug application (eIND). Results are reported here with a data cut-off of March 2021. Patient 1 with DIPG was a 14-year-old girl with H3K27M + DIPG, with early radiographic signs of tumour progression, right sixth nerve palsy, left facial weakness, dysarthria, bilateral dysmetria and wide-based gait at the time of treatment. On day +6 following treatment with GD2-CAR T cells intravenously (i.v.), she experienced grade 1 cytokine release syndrome (CRS; 40.1 °C fever) together with worsening cranial nerve symptoms consistent with TIAN and was treated with tocilizumab (an IL-6 antagonist) and corticosteroids (Supplementary Table 3). On day +9, she experienced an acute episode of fever, hypertension, decreased responsiveness, hemiplegia and extensor posturing. Her Ommaya was immediately accessed; her ICP was elevated at 22 mmHg, 10 ml CSF was drained, and she returned to her neurological baseline within minutes. Anakinra (an IL-1R antagonist) was started. MRI demonstrated increased pontine oedema. She remained on corticosteroids, with transient improvement in her baseline dysarthria and dysmetria. One month post-infusion, increased disease (approximately 20% enlargement) was evident on MRI, followed by tumour progression resulting in death 3 months after infusion of GD2-CAR T cells, 3 months after radiographic progression was first noted and 13 months after diagnosis (Extended Data Fig. 1a, b). Post-mortem brain examination   . From left to right: haemotoxylin and eosin staining; CD3 immunohistochemistry (brown); GD2-CAR mRNA puncta (pink; haematoxylin counterstain for all cells in blue); and GD2 antigen (blue) immunohistochemistry (H3K27M + nuclei in brown). d, qPCR for GD2-CAR DNA from autopsy samples from DIPG-1 exhibit the presence of GD2-CAR in tumour-involved midbrain and pons, but not her uninvolved cortex or brain tissue from an untreated individual (control). Resected temporal lobe tumour from patient 1 with spinal DMG (spinal DMG-1) following i.v. infusion reveals the presence of GD2-CAR DNA. Data represent mean ± s.e.m., n = 3 technical replicates for each sample. e, MRI scan (axial T2, shown at the level of the lower pons) of participant 2 with DIPG (DIPG-2) before and 4 weeks after i.v. infusion (top row). An MRI scan before and 2 weeks following i.c.v. infusion is also shown (bottom row). A reduction in the tumour size is observed after each infusion (red arrows). f, Photographs of DIPG-2 demonstrate significant improvement in left facial strength 2 weeks after i.c.v. infusion. Photographs were obtained and published with informed consent. Schematics were created with BioRender. com. Article demonstrated substantial infiltration of lymphocytes in the tumour that is uncharacteristic for DIPG 4 ; lymphocytic infiltration was not observed in areas of the normal brain (Fig 1c, Extended Data Fig. 1c). RNAscope identified cells expressing GD2-CAR mRNA transcript in the tumour, but not the unaffected cortex (Fig. 1c, Extended Data Fig. 1d). Similarly, the GD2-CAR transgene was detected by quantitative PCR (qPCR) of DNA from tumour tissue (Fig. 1d). GD2-antigen expression was substantially higher in the tumour than in the normal brain tissue (Fig. 1c, Extended Data Fig. 1e). Tumour microglial and other myeloid cell infiltration was prominent (Extended Data Fig. 1f, g). Patient 2 with DIPG was a 21-year-old man with H3K27M + DIPG exhibiting early signs of clinical and radiographic progression at the time of enrolment, with left sixth nerve palsy, left facial weakness, decreased left facial sensation, trismus with difficulty opening his mouth that limited large bites of food, right lower extremity sensory loss and wide-based gait with inability to tandem walk. On day +7 following i.v. GD2-CAR T cell administration, he developed grade 2 CRS (fever to 39.4 °C and hypotension responsive to fluids) and transient worsening of baseline deficits consistent with TIAN affecting pontine function, including transiently worsened trismus limiting oral intake to eating through a straw, transient right upper extremity sensory loss and mild left hearing loss lasting 2 days. MRI demonstrated increased T2/FLAIR signal in the left trigeminal nucleus (Extended Data Fig. 2a), which controls the muscles of mastication, correlating with the symptom of trismus. He was treated with tocilizumab and anakinra, but no corticosteroids (Supplementary Table 3). By the second week, he experienced marked improvement in trismus (near-normal mouth opening and no limitations to eating), improvement in facial symmetry, resolution of baseline sensory loss and improvement in gait with the ability to tandem walk (Supplementary Table 4). By 1 month following treatment with GD2-CAR T cells, he exhibited a near-normal neurological examination, with only a residual sixth nerve palsy. MRI demonstrated areas of improved T2/FLAIR signal and mildly decreased (approximately 17% smaller) tumour volume (Fig. 1e, Extended Data Fig. 2b).
Clinical and radiographic improvement persisted until 2-3 months, when he experienced recrudescence of previous symptoms, including return of trismus, facial weakness and sensory loss, left hearing loss, increased weakness and sensory loss affecting the right side of his body, and gait instability that required wheelchair use for any distance beyond a few steps. He received a second dose of 30 × 10 6 GD2-CAR T cells administered intracerebroventricularly (i.c.v.) via Ommaya reservoir without lymphodepletion. He developed a fever ( Table 3). Hydrocephalus resolved without need for continued drainage of CSF after 2 days. Corticosteroids were discontinued after 4 days (on day +6). Two weeks post-i.c.v. infusion, his neurological examination had markedly improved, with near-normalization of facial strength (Fig. 1f) and sensation, improved trismus, hearing, right-sided weakness and dramatic improvement in gait that enabled independent walking for long distances (Supplementary Videos 1, 2, Supplementary Table 4). An MRI scan 2 weeks post-i.c.v. GD2-CAR T cell infusion demonstrated 27% reduction in tumour volume compared to his pre-i.c.v. infusion MRI (Fig. 1e, Extended Data Fig. 2b).
After the data cut-off of March 2021, he went on to receive three more i.c.v. infusions (five infusions total). Before the planned sixth infusion, he died due to an intratumoural haemorrhage in a known area of intratumoural vascular anomaly. Such intratumoural haemorrhages are relatively common in DIPG and risk increases with time from diagnosis 5 . He survived 26 months from diagnosis, 12 months after he began to exhibit radiographic and clinical progression, and 10 months after his first GD2-CAR T cell infusion. Patient 3 with DIPG was a 5-year-old girl with H3K27M + DIPG who enrolled before progression. She did not exhibit brainstem enlargement, only areas of infiltrative tumour evident as patchy abnormal T2/ FLAIR signal in the left cerebral peduncle (corticospinal tract motor fibres) of the midbrain (Fig. 2a), anterior pons, medulla and left cerebellar peduncle. At baseline, she exhibited bilateral facial weakness, hypophonic voice, rightward tongue deviation, difficulty controlling oral secretions (drooling), chronic nausea, hypertonic right hemiparesis that limited use of the right arm and hand (her right hand was held in a flexed position; Fig. 2b), dysmetria and a hemiparetic, wide-based gait. On day 7 after i.v. administration of GD2-CAR T cells, she developed grade 1 CRS (fever of 40.4 °C) and transiently increased ataxia, consistent with TIAN. She was treated with tocilizumab and anakinra, but no corticosteroids (Supplementary Table 3). By 2 weeks post-infusion, her chronic nausea (present since diagnosis) had resolved, her right-sided motor function improved and she was using her right hand spontaneously in play, with improved ability to extend her fingers (Fig. 2b). By 1 month post-infusion, she was controlling oral secretions without drooling, taking bigger and stronger bites of food, speaking with a louder voice and showing improved facial strength that enabled a wide smile. These clinical improvements were accompanied by decreased T2/FLAIR MRI signal abnormality in the midbrain (Fig. 2a, c). By 6 weeks after i.v. GD2-CAR T cells, her balance, coordination and use of her right leg had improved, enabling her to ride a scooter using her right leg (Supplementary Table 4). With resolution of chronic nausea, her food intake improved; she increased from the sixth percentile to the thirtieth percentile for weight and grew 2 inches in height in the 2 months following i.v. GD2-CAR T cells.
She was re-treated with 12.9 × 10 6 GD2-CAR T cells i.c.v. (equivalent to 1 × 10 6 per kg) without lymphodepletion 3 months after her i.v. infusion, when improvement in her clinical examination had plateaued and increased cerebellar peduncle disease was evident (Fig. 2d, Extended Data Fig. 2c). She experienced only mild headache. No anti-cytokine agents or steroids were used. Her clinical response was mixed. She demonstrated additional improvement in right-sided motor function (improved right ankle dorsiflexion and wrist extension, and further increased use of right hand) within 1 week of her second GD2-CAR T cell infusion (Supplementary Table 4). By contrast, her left-sided ataxia progressively worsened consistent with increased left cerebellar peduncle disease (Fig. 2d, Extended Data Fig. 2c).
After the data cut-off of March 2021, this participant received one more i.c.v. infusion but her tumour continued to progress. She died from tumour progression 20 months after diagnosis, 4 months after she had begun to exhibit tumour progression, and 7 months after her first GD2-CAR T cell infusion.
Patient 1 with spinal DMG was a 25-year-old woman with an H3K27M + spinal cord DMG centred at thoracic levels T10-T12 who exhibited early signs of clinical and radiographic progression at the time of enrolment. After trial enrolment and during cell manufacturing, she experienced rapid tumour progression resulting in near-complete paraparesis, severe neuropathic pain, sensory loss below the T10 level, bladder dysfunction requiring urinary catheterization, and tumour spread to her temporal lobe. An urgent duraplasty was performed to relieve pressure in the spinal canal due to the rapidly expanding, tumour-infiltrated cord and to mitigate her severe neuropathic pain. Pain was moderately improved after duraplasty, but she still required a lidocaine drip and opiates for pain management. She no longer met criteria for GD2-CAR T cell infusion on-study due to rapid tumour progression with neurological deterioration and corticosteroid use before GD2-CAR T cell infusion and was removed from the protocol. She received GD2-CAR T cells (1 × 10 6 per kg) administered i.v. on an eIND. Beginning day +6, she exhibited grade 3 CRS (fever to 39.6 °C and hypotension requiring vasopressor support), which was treated with tocilizumab and corticosteroids for 3 days (days +8-11; Supplementary Table 3). Despite the trajectory of her rapid tumour progression before GD2-CAR T cell treatment, imaging demonstrated more than 90% reduction in volume of the spinal cord tumour by day +32 (Fig. 2e, f) accompanied by improved lower extremity motor and urinary function, resolution of neuropathic pain and discontinuation of all pain medications by 2 months post-infusion (Supplementary Table 4). By contrast, her brain metastasis did not exhibit improvement (Fig. 2g). Both brain and spinal cord disease progressed by day +75 (Fig. 2e, f). The tumour had spread extensively throughout her brain (Extended Data Fig. 2d), and the bulkiest component in the medial temporal lobe was resected. Flow cytometry of the resected tissue demonstrated robust tumour cell GD2 expression (Fig. 2h) but few infiltrating T cells. Low levels of the GD2-CAR transgene were detected by qPCR in the resected tumour of the temporal lobe (Fig. 1d). With limited therapeutic options, she was re-treated with 50 × 10 6 GD2-CAR T cells administered i.c.v. following increased lymphodepletion, given the concern for immune rejection of the CAR T cells. The chosen dose for CAR T cells was within the range of cell doses administered i.c.v. on other brain tumour trials 6,7 . She was the first patient chronologically to receive GD2-CAR T cells i.c.v., and her course informed the i.c.v. strategy described above. Within 48 h, she developed persistent fevers (up to 39.9 °C) and grade 3 encephalopathy (grade 4 immune effector cell-associated neurotoxicity (ICANS)), associated with no MRI changes in the normal brain and diffuse slowing with triphasic waves on continuous EEG monitoring, consistent with a reversible toxic/metabolic/inflammatory encephalopathy (Extended Data Fig. 3). Given the extensive tumour invasion of brain structures including the thalami, hypothalamus, mammillary bodies, bilateral insula and bilateral frontal lobes, we were unable to distinguish between ICANS and TIAN as the primary cause for her encephalopathy. She was treated with anti-inflammatory therapy (anakinra, siltuximab (an IL-6 antagonist) and corticosteroids (i.v. and i.c.v.)) and oral dasatinib to dampen the activity of CAR T cells 8 Table 3). Within 4 days, the encephalopathy resolved. Steroids were weaned and discontinued by day +24. Regression was again observed   Article in her spinal tumour (more than 80% reduction over 3 weeks; Fig. 2e, f) as well as select areas of her brain disease (Extended Data Fig. 2d). Clinical improvement in spinal cord function was not observed after the second infusion as it was after the first (Supplementary Table 4). Both spinal and brain disease progressed further by 2 months following i.c.v. infusion and she received no further GD2-CAR T cell infusions. After the data cut-off of March 2021, she survived 20 months from diagnosis, 14 months from the beginning of tumour progression and 11 months after her first GD2-CAR T cell infusion.

Correlative findings
Serum and CSF samples were obtained routinely and at times of clinical intervention to assess markers of inflammation and correlates of response. Patient 1 with spinal DMG had CSF samples collected following only i.c.v. administration but not following i.v. administration, and limited CSF samples were obtained from patient 3 with DIPG due to her young age. Correlating with tumour response, serum levels of lactate dehydrogenase increased in all patients after GD2-CAR T cell treatment and tracked with evidence of inflammation (Extended Data Fig. 4a). Using digital droplet PCR for the tumour-specific H3K27M mutation, cell-free tumour DNA was detected in CSF in two of four patients (Extended Data Fig. 4b-f). CSF cell-free tumour DNA was elevated during peak inflammation post-i.c.v. treatment, a pattern expected with tumour cell killing 9 . Additional sampling may further elucidate the kinetics of tumour cell killing in future studies.
GD2-CAR T cell expansion and persistence was monitored using GD2-CAR transgene qPCR and flow cytometry of cell-surface CAR expression. GD2-CAR T cell expansion in blood assessed by qPCR was similar in magnitude to expansion of highly active CAR T cells for haematological malignancies [10][11][12] (Fig. 3a). Flow cytometry-based GD2-CAR T cell detection in blood was limited, potentially due to decreased surface expression of GD2-CAR after activation 13 . GD2-CAR T cells were detected in CSF by flow cytometry, particularly following i.c.v. administration (Fig. 3b, Extended Data Fig. 6b).
Single-cell RNA sequencing was performed on cells isolated from CSF (Extended Data Fig. 6a) as well as flow-sorted GD2-CAR-positive and GD2-CAR-negative T cells from the manufactured products (Fig. 4a, b). We analysed 65,598 single cells: 25,598 cells from CSF and 40,000 cells from manufactured products. The cellular composition of CSF included lymphocytes, myeloid cells, rare astrocytes and B cells (Fig. 4b). We observed a population of lineage doublets, raising the possibility of phagocytosis of T cells by myeloid cells (Fig. 4b). We observed more GD2-CAR T cells in CSF at times of peak inflammation, and more CSF GD2-CAR T cells after i.c.v. versus i.v. administration for patient 2 with DIPG, the only patient in whom this comparison was possible (Extended Data Fig. 6b). By single-cell RNA sequencing, we did not identify GD2-CAR T cells in the CSF of patients 1 and 3 with DIPG. Comparing manufactured cell products for patient 1 with DIPG, who did not have a response, to patient 2 with DIPG, who exhibited radiographic and clinical benefit, revealed numerous transcriptional differences. These included baseline interferon and PD1 signalling activation pathways in patient 1 with DIPG, potentially indicating  c, UMAP highlighting FOXP3 + regulatory T (T reg ) cells (left). T reg cell population (defined by CD4, FOXP3 and CD25 expression) identified in CSF samples at peak inflammation timepoints following i.v. or i.c.v. administration (right). n = 523 T reg cells were analysed from a total of 17,699 CSF T cells from 3 participants after i.v. and 2 participants after i.c.v. administration. With a Bayesian model-based single-cell compositional data analysis (scCODA) framework, the log fold change was 0.6, the inclusion probability was 0.75 and the false discovery rate was less than 0.05. d, CSF sample single-cell RNA sequencing was filtered to isolate myeloid cells. Clustering was conducted after data integration by Harmony. n = 6,497 myeloid cells. e, Pie charts represent myeloid cluster proportions at different timepoints. A single UMAP was generated with myeloid cells from the CSF, and then individually visually represented based on timepoint of the sample, coloured by patient. Note the alterations in the presence of myeloid clusters over time. f, Expression signature of 'immune activating', 'immune suppressive', 'disease-associated myeloid (DAM) stage 1', 'DAM stage 2', 'myeloid-derived suppressor cell (MDSC)' and 'axon tract-associated microglial (ATM)' cell states and associated representation within clusters based on single-cell expression scores (Z score).

Article
increased tonic signalling associated with reduced GD2-CAR T cell activity 14 (Extended Data Fig. 6c, d). Comparisons of other products are shown in Extended Data Fig. 6e, f. Assessing all lymphocytes from the CSF of all participants revealed that the T cell CD4:CD8 ratios in CSF differed over time between patients (Extended Data Fig. 7a). Increased numbers of regulatory T (T reg ) cells were found in patient CSF after i.v. GD2-CAR T cell administration compared to i.c.v. administration (Fig. 4c, Extended Data Fig. 7b).
Examination of all CSF myeloid cells from all patients after each administration route revealed seven distinct clusters of cells, including monocytes, microglia and macrophages expressing various functional signatures, and proliferating myeloid cells (Fig. 4d, Extended  Data Fig. 8a, Supplementary Table 5). We identified a myeloid population characterized by interferon response, present primarily at peak inflammation timepoints after i.c.v. administration (Fig. 4d, e). This cluster exhibited a distinctly immune-activating signature (Fig. 4f, Extended Data Fig. 8b), concordant with the pro-inflammatory CSF cytokine profile observed after i.c.v. administration (Fig. 3c, d). By contrast, i.v. administration and late timepoints had CSF myeloid cell subpopulations that expressed prominent phagocytosis and lipid metabolism gene programmes. These cells also exhibited a strong immune-suppressive profile, which aligned with transcriptional signatures of disease-associated microglia 15 , myeloid-derived suppressor cell and axon tract-associated microglia 16 cell states (Fig. 4e, f, Extended Data Fig. 8). Comparison of the CSF myeloid cell fraction in patient 1 with DIPG to other patients at the time of peak inflammation highlighted differentially increased interleukin, chemokine and neutrophil degranulation processes (Extended Data Fig. 9a-c). Similarly, comparison of the CSF myeloid cell fraction in patient 2 with DIPG at peak inflammation following i.v. versus i.c.v. administration revealed increased myeloid activation processes following i.v. administration (Extended Data Fig. 9d).
Together, these correlative studies provide further evidence of GD2-CAR T cell activity against H3K27M + DIPG and spinal cord DMG and begin to elucidate the heterogeneity observed in antitumour responses.

Discussion
H3K27M + DIPG and spinal cord DMGs are extremely aggressive, universally fatal tumours with few therapeutic options. The average life expectancy is 10 months from diagnosis, and the 5-year survival is less than 1% 17 . Palliative radiotherapy is the only established treatment, and neither cytotoxic nor targeted pharmacological approaches have demonstrated improved prognosis to date 18 . Implementation of immunotherapy as treatment for tumours in these precarious neuroanatomical locations is both promising and dangerous. This phase I study was carefully designed to manage potential consequences of inflammation-induced swelling of an already expanded brainstem. The toxicities associated with GD2-CAR T cell infusions were manageable and reversible with intensive supportive care in the inpatient setting. CRS and ICANS were similar to that described with other CAR T cell therapies 19,20 , but patients also developed signs and symptoms consistent with CAR T cell-mediated inflammation in sites of CNS disease, which we have termed TIAN. TIAN most often manifested as transient worsening of existing deficits but also resulted in episodes of increased ICP secondary to brainstem oedema and consequent obstructive hydrocephalus, which would have been life-threatening unless immediately and appropriately managed. Our toxicity management algorithm incorporated several pre-emptive measures that enabled safe delivery of this potent therapy. As increasingly efficacious CAR T cell and other immunotherapies are deployed for CNS tumours, TIAN is likely to emerge as an important axis of toxicity in neuro-immuno-oncology.
Two categories of TIAN were evident in these patients: the first category of TIAN relates to ICP and intracranial space constraints secondary to inflammation-induced tissue oedema and/or obstruction of CSF flow. This can require urgent and/or emergent management utilizing principles of neurocritical care. The second category of TIAN relates to primary dysfunction of brain or spinal cord structures affected by inflammation, is typically transient, frequently manifests as worsening of pre-existing symptoms and can often be managed conservatively unless the neurological dysfunction involves critical functions such as respiratory drive.
Although patients developed symptoms of on-tumour neurotoxicity, they did not manifest any signs or symptoms of on-target, off-tumour toxicity. GD2 is expressed on normal neural tissues, including the brain and peripheral nerves, and treatment with the US Food and Drug Administration-approved anti-GD2 antibody dinutuximab for neuroblastoma is associated with transient painful neuropathy in most children 21,22 . None of these four patients developed painful neuropathies or any other clinical or radiographic indication of on-target, off-tumour toxicity. This study adds to the growing evidence that GD2 can be safely targeted with CAR T cells [23][24][25][26] , and is consistent with evidence that CAR T cells require high antigen density for full effector function 10,13,27,28 . Indeed, autopsy tissue from patient 1 with DIPG confirmed robustly differential GD2 expression levels and T cell infiltration between the tumour and the uninvolved brain.
Although the clinical experience reported here is early and the number of participants treated to date is small, three of four patients derived radiographic and clinical benefit after i.v. administration of GD2-CAR T cells. The improvement observed in neurological function highlights the extent to which this diffusely infiltrating tumour chiefly integrates with and disrupts-rather than destroys-neural circuits 29 , underscoring the potential that a tumour cell-specific therapy offers for functional recovery. i.c.v. administration of a second dose of GD2-CAR T cells provided additional radiographic and/or clinical benefit in three of three patients treated. In comparison to i.v. infusions, i.c.v. administrations were associated with less systemic toxicity such as CRS, and correlated with enhanced levels of pro-inflammatory cytokines and reduced immunosuppressive cell populations in CSF. The promising early experience with GD2-CAR T cells for DIPG and spinal cord DMG described here sets the stage for further optimization of this approach for this historically lethal CNS cancer. This clinical trial will continue to treat patients with H3K27M + DIPG and spinal DMG using GD2-CAR T cell therapy to determine optimal dose, route and schedule, and to determine efficacy.

Online content
Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-022-04489-4.

Clinical trial design
This phase I dose-escalation trial of GD2-CAR T cells in children and young adults with pontine and spinal H3K27M + DMG uses a 3 + 3 design with the primary objectives of assessing feasibility of manufacturing, safety and tolerability, and identifying the maximally tolerated dose or recommended phase II dose. Assessment of clinical activity is the secondary objective and identifying correlative biomarkers of response is an exploratory objective. Patients were eligible for enrolment if they had a pathologically confirmed diagnosis of H3K27M-mutated DIPG or spinal cord DMG, had completed standard radiotherapy and were not receiving corticosteroid therapy (additional eligibility and exclusion criteria below).

Toxicity management
In preclinical studies, a fraction of the mice exhibited brainstem oedema as a result of on-tumour inflammation that led to obstructive hydrocephalus due to the neuroanatomical location of these tumours 2 . We anticipated the development of neurological symptoms related to CAR T cell-mediated inflammation in sites of CNS disease, which we have termed TIAN. To mitigate the risks associated with TIAN, we excluded patients with bulky thalamic or cerebellar tumours, required placement of an Ommaya reservoir in patients with DIPG, conducted both routine and symptom-prompted ICP measurements and instituted a toxicity management plan that included TIAN management (CSF removal via Ommaya, hypertonic saline and/or corticosteroids) and CRS management using anti-cytokine agents (tocilizumab, an IL-6R antagonist, and anakinra, an IL-1R antagonist) and corticosteroids. In addition, the CAR construct incorporated an iCasp9-inducible suicide switch, which could be activated with rimiducid in the event of life-threatening toxicity that is not manageable with the approaches and agents discussed above 30 .

Oversight and informed consent
The clinical study and all amendments were approved by the Stanford University Institutional Review Board. An external data and safety monitoring board (DSMB) reviewed the protocol, amendments and all important patient events and outcomes. Before enrolment in the clinical trial, patients and/or their parents provided written, informed consent, and minor assent was obtained when appropriate. In addition, patients and/or their parents provided written, informed consent for use of photographs and videos that may be used in scientific presentations and publications in print and electronic formats.

CAR T cell manufacturing
Autologous peripheral blood mononuclear cells (PBMCs) were collected via apheresis and GD2-CAR T cells were manufactured using the closed-system CliniMACs Prodigy (Miltenyi). In brief, CD4 and CD8 T cells were selected and transduced with one bicistronic retroviral vector encoding an iCasp9 domain (Bellicum Pharmaceuticals, Inc.), and a GD2-4-1BB-CD3Ζ CAR (Fig. 1a) containing the 14g2a scFv under control of an MSCV promoter. GD2-CAR T cells were expanded for 7 days in the presence of recombinant human IL-7 and IL-15, plus the addition of the tyrosine kinase inhibitor dasatinib on days 3 and 5 to improve T cell fitness 31 .

CAR T cell administration
Lymphodepleting chemotherapy (cyclophosphamide 500 mg m 2 daily and fludarabine 25 mg m 2 daily on days −4, −3 and −2) was administered followed by i.v. CAR T cell infusion on day 0 (Fig. 1b). Patients were monitored closely, including routine ICP measurements, in the inpatient setting through toxicity resolution. For patients receiving second infusions, CAR T cells were infused i.c.v. via the Ommaya. The first patient to be treated with i.c.v. administered CAR T cells (patient 1 with spinal DMG) received a flat dose of 50 million CAR T cells after increased lymphodepleting chemotherapy (cyclophosphamide 600 mg m 2 daily and fludarabine 30 mg m 2 daily on days −5, −4, −3 and −2) under a single-patient eIND (see below); subsequent patients received 30 million CAR T cells or the equivalent of their weight-based dose at DL1, whichever was lower, i.c.v. without lymphodepletion. AST) ≤ 3.0 upper limit of normal (ULN;grade 1) (viii ) Total bilirubin ≤ 1.5 mg dl−1, except in participants with Gilbert's syndrome (ix) Cardiac ejection fraction ≥ 45%, no evidence of physiologically significant pericardial effusion as determined by an echocardiogram, and no clinically significant electrocardiogram findings (x) Baseline oxygen saturation > 92% on room air (6) Female individuals of childbearing potential must have a negative serum or urine pregnancy test (female individuals who have undergone surgical sterilization are not considered to be of childbearing potential). (7) Contraception: participants of childbearing or child-fathering potential must be willing to practice birth control from the time of enrolment on this study and for 4 months after receiving the preparative lymphodepletion regimen or for as long as GD2-CAR T cells are detectable in peripheral blood or CSF. (8) Ability to give informed consent. Paediatric participants will be included in age appropriate discussion and written assent will be obtained for those 7 years of age or older, when appropriate.

Exclusion criteria
(1) Bulky tumour involvement of the cerebellar vermis or hemispheres (pontocerebellar peduncle involvement is allowed), thalamic lesions that in the investigator's assessment place the participant at unacceptable risk for herniation. or hepatitis C virus (anti-HCV positive). A history of hepatitis B or hepatitis C is permitted if the viral load is undetectable per qPCR and/or nucleic acid testing. (7) Clinically significant systemic illness or medical condition (for example, significant cardiac, pulmonary, hepatic or other organ dysfunction), that in the judgement of the principal investigator is likely to interfere with assessment of safety or efficacy of the investigational regimen and its requirements. (8) In the investigator's judgement, the participant is unlikely to complete all protocol-required study visits or procedures, including follow-up visits, or comply with the study requirements for participation. (9) Known sensitivity or allergy to any agents and/or reagents used in this study.

Response assessment
Radiographic responses were evaluated by gadolinium-enhanced MRI scans of the brain and/or spinal cord. Because DMGs are diffusely infiltrative of CNS structures and therefore are difficult to measure, volumetric tumour segmentation of T2/FLAIR signal abnormality was performed by a neuroradiologist (K.W.Y.) to measure radiographic change. Clinical response was assessed by neurological examination. To quantify clinical change, a clinical improvement score (CIS) was calculated. The CIS represents an unweighted quantification of the neurological examination and is conducted by a neurologist who examines the participant before infusion and then at designated timepoints after therapy. For each neurological symptom or sign, one point is assigned. Improvement in a given symptom or sign from pre-infusion baseline add one point to the score, no change adds zero and one point is subtracted for symptom or sign worsening. The reported score is a sum of the positive and negative points. For example, a participant with improved right upper extremity strength and improved left sixth nerve palsy would receive a score of +2. A participant with improved right upper and lower extremity strength, improved left sixth nerve palsy but worsened left facial sensation would also receive a score of +2 (+3 − 1 = +2). The assessment cannot be conducted within 7 days of corticosteroid administration. Sample forms are available in Supplementary Tables 6, 7.
Cell-free DNA rare mutation detection Patient CSF samples were collected at pre-determined and trigger timepoints throughout the treatment course. To gently separate the cellular compartment, samples were spun within 1 h of collection at 1,800g for 10 min at 4 °C followed by 20,000g for 10 min. Supernatant was frozen at −80 °C. DNA extraction was performed on 0.5-3 ml CSF supernatant aliquots using the QIAmp Circulating Nucleic Acids Kit (Qiagen) into 25 µl extraction buffer. All samples were assessed for cell-free DNA concentration by nanodrop and fragment size distribution with the BioAnalyzer (Agilent).
Forward and reverse primers and wild-type or mutant probes were designed to optimally detect the H3F3A K27M mutation through digital droplet polymerase chain reaction (ddPCR), and a dilution series of mutant H3K27M g-block against a wild-type H3 background was performed to validate linearity and confirm limit of detection. Seven PCR cycles for pre-amplification were performed on 10 µl of extracted cell-free DNA with Q5 Hot Start High-Fidelity Master Mix (New England Biolabs) for each timepoint. Preamplified reactions were immediately placed on ice after the final extension and diluted with 100 µl TE buffer, pH 8.0, to inactivate the Q5 polymerase. Samples were used immediately for ddPCR (40 cycles; Bio-Rad) and run as 4-6 technical replicates. The raw spectral data were analysed using the Bio-Rad software to plot the signal for mutant and wild-type alleles. All runs contained positive and negative controls. Results were analysed in R to plot variant allele frequency (VAF) as well as mutation count per volume over time.

qPCR measurement of in vivo GD2-CAR expansion
Patient blood and tumour tissue samples were processed and mononuclear cells were viably cryopreserved. DNA was extracted from whole blood (2 × 10 6 -5 × 10 6 PBMCs) using QIAmp DNA Blood Mini Kit (51306, Qiagen) at baseline and multiple timepoints following CAR administration and from tumour tissue samples using DNeasy 96 Blood & Tissue Kit (69582, Qiagen). CAR presence was measured by qPCR using the primer and probe sequences provided as below and in Supplementary  Table 8. For the standard curve, a custom Minigene plasmid (IDT) was designed containing a partial GD2-4-1BB-CD3Ζ sequence and a partial albumin sequence, which served as a control for normalization (Supplementary Table 8). The standard curve contained a tenfold serial dilution of plasmid between 5 × 10 8 and 5 × 10 0 copies. Both plasmid and patient DNA from each timepoint were run in triplicate, with each reaction containing 5 µl of DNA (50 ng total), 200 nM forward and reverse albumin primers (or 300 nM forward and reverse GD2-4-1BB-CD3Ζ primers), 150 nM probe suspended in 10 µl of TaqMan Fast Universal PCR Master Mix (2X), no AmpErase UNG (Thermo Fisher Scientific) and 24.5 µl (albumin) or 22.5 µl (GD2-4-1BB-CD3Ζ) of TE buffer (AM9935, Invitrogen). The Thermo Fisher Scientific QuantStudio 6 Pro Real-time qPCR Instrument was used for qPCR with 20 µl per reaction. The quality metrics for all qPCR standard curve results were R 2 > 0.95 and efficiency was 70-110%.

Real-time flow cytometry assay
A high-dimensional (Hi-D) immuno-phenoptyping flow cytometry panel was designed for immune profiling of CAR T cells in real time. PBMCs were isolated from fresh whole blood by gradient centrifugation on ficoll (Ficoll-Paque Plus, GE Healthcare, Sigma-Aldrich). Two to five million PBMCs were stained with fixable live/dead aqua (Invitrogen) Article amine-reactive viability stain. Cells were then preincubated with Fc block (trustain, BioLegend) for 5 min, then stained at room temperature with the following fluorochrome-conjugated mAb in a 15-colour, 17-parameter staining combination (Supplementary Table 9).
CAR-tranduced T cells were used as positive control included in daily staining experiments. Stained and fixed cells were acquired on a LSR (BD BioSciences) five-laser (blue 488 nm, violet 405 nm, UV laser 355 nm, red 640 nm and green 532 nm lasers) analyser. At least 10 6 cells were acquired unless restricted by the number of cells isolated from 8 ml of whole blood or when acquiring isolated cells from CSF. The assay limit of detection for cells calculated as 1 in 10 4 of total acquired PBMCs. Representative gating is shown in Extended Data Fig. 10a.

Luminex cytokines
Patient blood and CSF samples were collected at pre-determined and trigger timepoints throughout treatment. Samples were spun at 250g for 6 min. Supernatant was frozen at −80 °C until batched for assessment. Cytokine assessment was performed by the Immunoassay Team-Human Immune Monitoring Center at Stanford University. Panels include Luminex-EMD Millipore HIMC H80 (panel 1 is Milliplex HCYTA-60K-PX48; panel 2 is Milliplex HCP2MAG-62K-PX23; panel 3 includes the Milliplex HSP1MAG-63K-06 and HADCYMAG-61K-03 (resistin, leptin and HGF) to generate a 9 plex) and TGFβ (TGFBMAG-64K-03). Kits were purchased from EMD Millipore Corporation and used according to the manufacturer's recommendations with modifications described. The assay setup followed recommended protocol. In brief, samples were diluted threefold (panels 1 and 2) or tenfold (panel 3). Of the diluted sample, 25 µl was mixed with antibody-linked magnetic beads in a 96-well plate and incubated overnight at 4 °C with shaking. Cold and room temperature incubation steps were performed on an orbital shaker at 500-600 rpm. Plates were washed twice with wash buffer in a Biotek ELx405 washer (BioTek Instruments). Following 1-h incubation at room temperature with biotinylated detection antibody, streptavidin-phycoerythrin (PE) was added for 30 min with shaking. Plates were washed as described above and PBS was added to wells for reading in the Luminex FlexMap3D Instrument with a lower bound of 50 beads per sample per cytokine. Custom Assay Chex control beads were purchased and added to all wells (Radix Biosolutions). All wells collected met quality control metrics with a bead count >50. Data are represented in pg ml −1 based on standard curves or heat maps of fold change from baseline timepoint. All samples were run in technical duplicate.

Single-cell RNA sequencing
Single-cell RNA sequencing (scRNA-seq) was performed using 5′ v2 Single Cell Immune Profiling technology (10X Genomics) according to the manufacturer's protocol. In brief, cells from CSF samples collected at indicated timepoints before and after CAR T cell administration, as well as sorted CAR + and CAR − cells from the CAR T cell infusion products, were counted, resuspended to 700-1,200 cells per µl, and captured using Single Cell Chip A on the 10x Chromium Controller (10X Genomics) to generate gel bead-in emulsions (GEMs). Reverse transcription inside GEMs was performed using a C1000 Touch Thermal Cycler (Bio-Rad). Barcoded complementary DNA (cDNA) was recovered through post-GEM-RT cleanup and PCR amplification. Recovered cDNA was amplified and used to construct 5′ whole-transcriptome libraries. Quality of cDNA and each library was assessed using Agilent 2100 Bioanalyzer. The libraries were indexed using a Chromium i7 Sample Index Kit, pooled and sequenced on NovaSeq 6000 System (Illumina) by Novogene.
Raw sequencing data were processed using the Cell Ranger software version 6.0.0 or higher (10X Genomics). Sequencer's base call files (BCLs) were demultiplexed into FASTQ files using the cellranger mkfastq pipeline. To optimize CAR sequence detection in the scRNA-seq data, we probed the FASTQ files using SeqKit 32 in 50-base-pair segments from the GD2-CAR plasmid sequence. Once we identified the segments with the most hits, we created a custom reference by appending these segments to the GRCh38 5.0.0 human genome reference with the cellranger mkref pipeline. FASTQ files were aligned to this custom reference using the cellranger count pipeline.
Unique molecular identifier (UMI) count matrices from Cell Ranger were analysed using Seurat 33 . Dead cells and cell debris with more than 15% of UMI counts mapping to mitochondrial genes or less than 300 genes detected were excluded from the analysis. Cell doublets containing more than 10,000 genes or more than 50,000 UMI counts were also excluded. Next, scRNA-seq data were subsampled to 5,000 cells per sample, and all data were normalized together to the sequencing depth using the SCTransform pipeline 34 . Data between patients were integrated using reciprocal principal component analysis (RPCA). PCA was performed using all variable genes with the exception of TCR and BCR genes to prevent clonotypes from driving the final layout. In addition, a set of curated genes relevant to T cell function or cell-type identification were included into the list of variable genes. UMAP embedding was performed using the first 50 principal components. Cell types were assigned to all cells based on canonical lineage marker expression. Cell-type doublets ('lineage doublets') were identified as cells expressing markers of more than one lineage: T cells (CD2, CD3E, CD3D, CD3G, CD247, CD7 and GD2-CAR), B cells (CD19, CD22 and PAX5), microglia/myeloid cells (CD14, CD68, CD163, CSF1R and AIF1) or astrocytes (GFAP). Differential expression analysis was performed on SCT counts modelled as a negative binomial distribution using the FindMarkers pipeline.
Further analysis of patient 1 with DIPG and other participant CSF samples was performed using the R-based 'Seurat' package 35 to assess transcriptional differences preceding and during CAR T cell administration. First, the large Seurat object was subset into various smaller objects to enable multiple comparisons across different DIPG CSF samples (see Fig. 4). Seurat objects subsequently underwent scaling and dimensionality reduction based on the UMAP algorithm 36 . Myeloid fractions of the dataset were identified using the 'AddModuleScore' function in Seurat to isolate cells with relatively high expression of myeloid genes (AIF1, CSF1R, CX3CR1, CD14, CD68 and CD163), compared to their expression of B cell-related (CD19, CD22 and PAX5) and T cell-related (CD2, CD3E, CD3D, CD3G, CD247 and CD7) genes. Scoring of cells based on disease-associated myeloid (DAM) or myeloid-derived suppressor cell (MDSC) gene expression was implemented using the 'AddModuleScore' and 'FeaturePlot' functions in Seurat. DAM genes included SPP1, GPNMB, IGF1, CLEC7A, LPL, CD9, CD63, LGALS3, FABP5, ITGAX, APOE and TYROBP, based on previous noteworthy characterizations of myeloid cells in disease 15,16,37,38 . MDSC genes included CD33, CD14, CD15 and IL4RA 39 . Volcano plot analysis was conducted using the R-based 'EnhancedVolcano' package 40 . Gene ontology (GO) analysis was performed using the R-based 'ReactomePA' package 41 .
Single-cell expression scores were computed in a similar way as previously described 42 . Given a set of genes (Gj) for a gene set (for example, a DAM signature), a score, SCj(i), which quantifies the scaled expression (Z-score) of Gj for each cell i, was computed as the average scaled expression (Er) of the genes in Gj, compared to the average scaled expression of a control gene set Gcont: For each gene of the gene set, its control gene set contains 100 genes with the most similar aggregated expression level to that gene. Therefore, the control gene set represents a 100-folder larger but comparable distribution of expression levels to that of the considered gene set.
Graph-based clustering with data integration was adapted to identify cellular clusters. We selected highly variable genes (HVGs) using the FindVariableFeatures function in Seurat and used the scaled expression values of these HVGs for PCA. To disentangle sample-specific biological variations from cell subpopulation-specific variations and integrate multiple samples, we applied a linear adjustment method called Harmony to the first 30 principle components with default parameters to generate a corrected embedding 43 . We chose the first 20 Harmony corrected dimensions for constructing UMAP embeddings by RunUMAP in Seurat and clustered cells by the Louvain algorithm based FindClusters in Seurat. Cells that were from different samples, but expressed similar gene programmes, were well mixed. We next identified differentially expressed genes (DEGs) by FindAllMarkers in Seurat and tested genes that were detected in a minimum of 30% of the cells within each cluster and that showed at least a 0.5-fold mean log difference. We utilized Wilcoxon rank-sum test with Bonferroni correction for multiple testing and only kept genes with adjusted P < 0.05. Cell clusters were annotated with manual inspection of their top DEGs. In addition, we tested for enrichment of described gene sets (GO biological processes) and compared expression programmes of each cell cluster with those of a published glioblastoma-associated myeloid cell dataset 44 .
For H3K27M and GD2 immunohistochemistry, primary tumour samples from patients were transferred to cryomolds and embedded in optimal-cutting temperature (OCT) compound (TissueTek). Cryosections (10 µm) were generated on a cryostat (Leica). Tissue was fixed with 4% PFA at 4 °C for 20 min then washed and endogenous peroxidase activity was neutralized (Bloxall, Vector Laboratories; 10 min at room temperature) before permeabilization (0.3% Triton X-100, TBS, 15 min at room temperature) and blocking (5% horse serum, Vector Laboratories; 20 min at room temperature). Sequential double-staining immunohistochemistry was conducted for H3K27M (ab190631, Abcam; 1:1,000, 1 h at room temperature) and GD2 (14g2a, BD; 1:500, 1 h at room temperature). H3K27M was developed with a peroxidase secondary (ImmPRESS VR anti-rabbit IgG, Vector Laboratories; 30 min at room temperature) and DAB substrate (BD). After quenching the DAB substrate development in TBS and staining with the 14g2a primary antibody, the GD2 signal was developed using a polymer-based alkaline phosphatase secondary antibody (ImmPRESS AP anti-mouse IgG, Vector Laboratories; 30 min at room temperature) and blue alkaline phosphatase substrate (Vector Blue AP substrate kit, Vector Laboratories; 5 min at room temperature). Alkaline phosphatase development was quenched in TBS, and samples were mounted and imaged by light microscopy (Zeiss Axio Imager M2).
For immunofluorescence, primary tumour samples from patients were fixed in 4% PFA overnight and then transferred to 30% sucrose until samples sunk (3-4 days). Serial 40-µm cryosections were generated on an automatic freezing microtome (HM450, Thermo Fisher) then incubated with 3% normal donkey serum in 0.3% Triton X-100 in TBS blocking solution at 1 h at room temperature. Sections were stained overnight at 4 °C in primary antibody. Antibodies used were mouse anti-H3K27M (ab190631, Abcam; 1:1000) and rabbit anti-IBA1 (019-19741, Wako; 1:500) diluted in 1% normal donkey serum in 0.3% Triton X-100 in TBS blocking solution. Sections were incubated in secondary antibody conjugated with either 594 or 488 for 2 h at room temperature (715-585-150 and 711-545-152, Jackson Immunoresearch; 1:500). Samples were mounted with ProLong Gold mounting medium (P36930, Life Technologies) and were imaged using confocal microscopy (LSM710, Zeiss). For RNAscope in situ hybridization to visualize the GD2-CAR construct, primary tumor samples from patients were transferred to cryomolds and embedded in optimal-cutting temperature (OCT) compound (TissueTek). 10-µm cryosections were generated on a cryostat (Leica). Slides were fixed with cold 4% PFA for 15 minutes at 4 o C and then dehydrated in increasing ethanol rinses. To prepare for staining, tissue was pretreated with RNAscope hydrogen peroxide for 10 min at room temperature, then treated with RNAscope protease IV and incubated for 30 min at room temperature (ACD 322381). Tissue was then stained following the kit protocol for RNAscope 2.5 HD Duplex Assay (ACD, 322500). To stain control cells for GD2-CAR RNAscope, GD2-CAR T cells were thawed from frozen and prepared for staining according to the RNAscope Multiplex v2 Assay (ACD document MK-50 010). In short, cells were seeded in growth medium on chamber slides for 24 h and then fixed with 4% PFA for 30 min at room temperature. Cells were then dehydrated with ethanal and stored in 100% ethanol at -20 o C until staining. On the day of staining, cells were rehydrated with ethanol rinses and then pretreated with RNAscope hydrogen peroxide for 10 min at room temperature followed by RNAscope protease III diluted 1:15 with PBS for 10 min at room temperature (ACD 322381) and then stained alongside the tissue samples. A custom probe was designed to target the GD2-CARconstruct using the following sequence :  AUGCUGCUGCUCGUGACAUCUCUGCUGCUGUGCGAGCUGCCCCACC  CCGCCUUUCUGCUGAUCCCCGAUAUCCUGCUGACCCAGACCCCUCUG  AGCCUGCCUGUGUCUCUGGGCGAUCAGGCCAGCAUCAGCUGCAGAU  CCAGCCAGAGCCUGGUGCACCGGAACGGCAACACCUACCUGCACU  GGUAUCUGCAGAAGCCCGGCCAGAGCCCCAAGCUGCUGAUUCACAAG  GUGUCCAACCGGUUCAGCGGCGUGCCCGACAGAUUUUCUGGCAGC  GGCUCCGGCACCGACUUCACCCUGAAGAUCAGCCGGGUGGAAGCCGA  GGACCUGGGCGUGUACUUCUGCAGCCAGUCCACCCACGUGCCCCCC  CUGACAUUUGGCGCCGGAACAAAGCUGGAACUGAAGGGCAGCACA  AGCGGCAGCGGCAAGCCUGGAUCUGGCGAGGGAAGCACCAAGGGCGA  AGUGAAGCUGCAGCAGAGCGGCCCCUCUCUGGUGGAACCUGGCGCC  UCUGUGAUGAUCUCCUGCAAGGCCAGCGGCAGCUCCUUCACCGGCU  ACAACAUGAACUGGGUGCGCCAGAACAUCGGCAAGAGCCUGGAAUGG  AUCGGCGCCAUCGACCCCUACUACGGCGGCACCAGCUACAACCAGA  AGUUCAAGGGCAGAGCCACCCUGACCGUGGACAAGAGCAGCUCCACC  GCCUACAUGCACCUGAAGUCCCUGACCAGCGAGGACAGCGCCGUGUA  CUACUGCGUGUCCGGCAUGGAAUACUGGGGCCAGGGCACAAGCGUG  ACCGUGUCCUCUGCGGCCGCAACCACGACGCCAGCGCCGCGACCACC  AACACCGGCGCCCACCAUCGCGUCGCAGCCCCUGUCCCUGCGCCCA  GAGGCGUGCCGGCCAGCGGCGGGGGGCGCAGUGCACACGAGGGG  GCUGGACUUCGCCUGUGAUAUCUACAUCUGGGCGCCCUUGGCCGGG  ACUUGUGGGGUCCUUCUCCUGUCACUGGUUAUCACCCUUUACUGC  AAACGGGGCAGAAAGAAACUCCUGUAUAUAUCAAACAACCAUUUAUG  AGACCAGUACAAACUACUCAAGAGGAAAUGGCUGUAGCUGCCGAUUU  CCAGAAGAAGAAGAAGGAGGAUGUAACUGAGAGUGAAGUUCAGCAGG  AGCGCAGACGCCCCCGCGUACCAGCAGGGCCAGAACCAGCUCUAUA  ACGAGCUCAAUCUAGGACGAAGAGAGGAGUACGAUGUUUUGGACAAG  AGACGUGGCCGGGACCCUAGAUGGGGGGAAAGCCGAGAAGGAAGA  ACCCUCAGAAGGCCUGUACAAUGAACUGCAGAAAGAUAAGAUGGCG  GAGGCCUACAGUGAGAUUGGGAUGAAAGGCGAGCGCCGGAGGGGC  AAGGGGCACGAUGGCCUUUACCAGGGUCUCAGUACAGCCACCAAGG ACACCUACGACGCCCUUCACAUGCAGGCCCUGCCCCCUCGC. After probe hybridization, amplification and detection according to the kit manual, slides were counterstained with 50% hematoxylin for 30 sec, followed by 0.02% ammonia water for blueing. Slides were dried at 60 °C for 15 min, then mounted and imaged by light microscopy (Zeiss Axio Imager M2).

Flow cytometry of resected tumour tissue
Tumour resection material was digested into single-cell suspensions as previously described 45 . Single-cell suspensions were stained with the following antibodies (Supplementary Table 10 Last updated by author(s): Dec 2, 2021 Reporting Summary Nature Portfolio wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency in reporting. For further information on Nature Portfolio policies, see our Editorial Policies and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.

n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly The statistical test(s) used AND whether they are one-or two-sided Only common tests should be described solely by name; describe more complex techniques in the Methods section.
A description of all covariates tested A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals) For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Portfolio guidelines for submitting code & software for further information.

Data
Policy information about availability of data All manuscripts must include a data availability statement. This statement should provide the following information, where applicable: -Accession codes, unique identifiers, or web links for publicly available datasets -A description of any restrictions on data availability -For clinical datasets or third party data, please ensure that the statement adheres to our policy All raw data are provided in the source data file. The single cell RNA-sequencing data has been uploaded to GEO (GSE186802).