Neurocognitive and hypokinetic movement disorder with features of parkinsonism after BCMA-targeting CAR-T cell therapy

B-cell maturation antigen (BCMA) is a prominent tumor-associated target for chimeric antigen receptor (CAR)-T cell therapy in multiple myeloma (MM). Here, we describe the case of a patient with MM who was enrolled in the CARTITUDE-1 trial (NCT03548207) and who developed a progressive movement disorder with features of parkinsonism approximately 3 months after ciltacabtagene autoleucel BCMA-targeted CAR-T cell infusion, associated with CAR-T cell persistence in the blood and cerebrospinal fluid, and basal ganglia lymphocytic infiltration. We show BCMA expression on neurons and astrocytes in the patient’s basal ganglia. Public transcriptomic datasets further confirm BCMA RNA expression in the caudate of normal human brains, suggesting that this might be an on-target effect of anti-BCMA therapy. Given reports of three patients with grade 3 or higher parkinsonism on the phase 2 ciltacabtagene autoleucel trial and of grade 3 parkinsonism in the idecabtagene vicleucel package insert, our findings support close neurological monitoring of patients on BCMA-targeted T cell therapies. A progressive movement disorder in a patient with multiple myeloma treated with anti-BCMA CAR-T cells that might have been related to on-target activity in the brain supports prospective neurologic monitoring after BCMA-targeting therapies.

was a candidate for BCMA-targeted CAR-T therapy. The patient underwent a bridging chemotherapy regimen (melphalan 56.25 mg once and two doses of bortezomib 1.3 mg m −2 ). Before CAR-T infusion, he received 300 mg m −2 of fludarabine and 30 mg m −2 of cyclophosphamide daily for 3 days to induce lymphodepletion. The baseline tumor burden was low (3% plasma cells on bone marrow biopsy), but he had multiple extramedullary plasmacytomas.
After cilta-cel infusion (day 1), the patient developed fever on day 9 and hypotension on day 11 (CRS up to maximum grade 3) with the peak of C-reactive protein, ferritin and inflammatory cytokines (TNF-α, IFN-γ, IL-6 and IL-18) 2 weeks after CAR-T infusion. CRS symptoms resolved by day 14 after treatment with tocilizumab and anakinra 17 . Relevant clinical and biochemical parameters, timing of selected therapeutic agents and additional cytokine profiling are shown in Extended Data Figs. 1 and 2. The patient remained afebrile until hospitalization on day 51 with neutropenic fever and pneumonia, treated with empiric antibiotics. Cultures remained negative, but polymerase chain reaction (PCR) detected rhinovirus in respiratory secretions. The patient was discharged at day 57. Disease evaluation (day 79) showed a very good partial response, according to International Myeloma Working Group criteria.
At day 101 after CAR-T cell infusion, the patient was evaluated with complaints of increasing fatigue interfering with daily activities. Initially, we observed slow gait and psychomotor retardation. Subsequent evaluation by two independent neurologists confirmed a clinical syndrome with features of parkinsonism, including bradykinesia, postural instability, hypophonia, hypomimia, micrographia and a mild right-sided (action and resting) tremor, as well as saccadic intrusions on smooth pursuit and impaired short-term memory. There was no cogwheeling, no focal paresis or atrophy, no pathological reflexes or alterations of deep tendon reflexes and no ataxia, and the Romberg test was negative. Sensation was intact. The features were progressive over time with development of increasing hypomimia, rigidity and difficulty initiating movements. There were no recent drug changes or toxin exposure to account for the observed clinical phenotype. The patient was on a benzodiazepine for anxiety disorder, which was discontinued without improvement of the clinical features. Laboratory tests showed fluctuating neutrophil counts due to regular granulocyte-macrophage colony-stimulating factor (GM-CSF) injections (Extended Data Fig. 1c). Magnetic resonace imaging (MRI) of the brain with and without contrast showed only small pre-existing foci of T2/FLAIR signal hyperintensity scattered throughout the periventricular and subcortical white matter (Extended Data Fig. 3a), and lumbar puncture findings were non-explanatory (Supplementary Table 1; additional clinical background is provided in the Methods). A treatment attempt with levodopa because of progressive movement disorder and functional decline was unsuccessful. Fluorodeoxyglucose-positron emission tomography (FDG-PET) of the brain for response evaluation indicated a decreased uptake in the caudate nucleus bilaterally in comparison with imaging of 2 months prior, without any structural abnormalities (Extended Data Fig. 4). Ioflupane (123-I) scan was negative, suggesting a disease mechanism different than Parkinson's disease (Extended Data Fig. 3b).
As shown in Fig. 1a, CAR-T cells were detectable in the blood in large numbers, starting at day 11 after infusion up to day 156. Notably, 70-90% of all T cells in the peripheral blood were CAR-T cells (Fig. 1b). This observation suggested a role for persistent CAR-T cells in the development of the patient's neurologic complaints. Extensive phenotyping of T cells using a mass cytometry (CyTOF) approach (Fig. 1c) suggested that most CAR-T cells had an effector memory phenotype (that is, CD45RA − CCR7 − ) ( Fig. 1d and Extended Data Fig. 5). Functional assays of peripheral blood CAR-T cells, isolated 128 days after treatment, confirmed their ability to produce inflammatory cytokines (IFN-γ, TNF-α and GM-CSF) upon PMA/ionomycin stimulation in vitro (Extended Data Fig. 6), highlighting their cytotoxic/pro-inflammatory potential. The full list of cytokines tested and comparison with a healthy donor are shown in Extended Data Fig. 6b. The patient's CAR-T cells did not exhibit a Th17 phenotype-a T cell subtype previously associated with immunologic neurodegenerative disorders. Comparative single-cell analysis of CAR-T cells of this patient with three other patients from the same trial (without parkinsonism) by cellular indexing of transcriptomes and epitopes by sequencing (CITE-seq) ( Fig. 1e and Extended Data Fig. 7) showed qualitative transcriptomic differences with significantly higher expression of genes associated with long-term survival (for example, IL7R) and genes encoding inflammatory cytokines (IFNG, TNF and CSF2) and lower expression of anti-inflammatory cytokine genes (for example, IL10).
Microarray data of healthy human brains of the Allen Brain Atlas 18 confirmed localized RNA expression of BCMA in the basal ganglia and, more specifically, in the caudate nucleus in five of six available specimens ( Fig. 2b and Extended Data Fig. 8). We hypothesized that the symptoms could result from CAR-T cell infiltration in the brain targeting BCMA-expressing cells, thereby causing a movement disorder with features of parkinsonism. Analysis of the cerebrospinal fluid (CSF) by fluorescence-activated cell sorting (FACS) confirmed the presence of CAR-T cells in the CSF (0.477 CAR-T cells per µl; Extended Data Fig. 9b). Cytokine profiling of CSF and blood plasma of the patient and a healthy control showed overexpression of multiple cytokines in the patient's CSF associated with T cell chemotaxis (for example, CXCL5, CXCL10 and CXCL11), T cell activation (for example, granzymes, IFN-γ and CD40-L) and blood-brain barrier dysfunction (for example, PDGFB and angiopoetin-1) (Extended Data Fig. 9).
Due to the sustained proliferation of CAR-T cells with spread beyond the blood-brain barrier and progressive decline in the patient's general condition, IV cyclophosphamide (300 mg m −2 ), IT cytarabine (100 mg) and hydrocortisone (50 mg) were given on day 149, after careful consideration, aiming to rapidly reduce circulating CAR-T cells. We observed a decline of the absolute T cell count in the CSF with a stable fraction of CAR-T cells (0.128 CAR-T cells per µl; Extended Data Fig. 9c). A second dose of IV cyclophosphamide (300 mg m −2 ) and IT cytarabine/hydrocortisone were administered on day 156. The patient subsequently developed neutropenic fever with acute respiratory distress syndrome and multi-organ failure and died on day 162.
Postmortem analysis of the caudate nucleus revealed the presence of focal gliosis as shown on hematoxylin and eosin (H&E) staining and immunohistochemistry of glial fibrillary acidic protein (GFAP) (Fig. 2c,d). Immunohistochemistry further showed a T cell infiltrate (CD3 + , predominantly CD8 + ) in the periventricular   region of the basal ganglia (Fig. 2e). BCMA staining was performed (Methods), and we found BCMA expression on a subset of neurons and astrocytes in the caudate nucleus as well as on a layer of neurons in the adjacent frontal cortex ( Fig. 2f and Extended Data Fig. 10).
The value of BCMA as a tumor-associated target in MM depends on the selective expression on (malignant) plasma cells. Even though BCMA expression has been extensively characterized on hematopoietic lineages, studies on other tissues are limited 6,7,19 . We found that Orbitofrontal region

Fig. 2 | BCMA is expressed in the caudate nucleus of healthy donors and postmortem in the patient after CAR-T cell therapy. a,
FDG-PET/CT shows decreased uptake in the caudate nucleus after development of neurotoxicity (POST, right, day 134 after CAR-T infusion), compared to previous imaging before development of neurotoxicity symptoms (PRE, left, day 77 after CAR-T infusion). Prior FDG-PET/CT imaging (before CAR-T infusion) was similar to the pre-neurotoxicity scan. The scatter plot on the right illustrates the normalized z-score of different regions of the brain before and after CAR-T infusion.
The caudate is highlighted. The normalized score is calculated using MIMneuro, comparing the image with a library of 43 FDG neurologic controls (41-80 years old). b, Visual representation of the expression of DRD1 and TNFRSF17 (BCMA) in a single patient from the Allen Brain Atlas. Expression of both genes (left, red = high) overlaps with the caudate nucleus region shown in three dimensions (right, purple). Image credit: Allen Institute for Brain Science (2010 microarray data from the Allen Brain Atlas shows BCMA mRNA expression in the basal ganglia and confirmed this in the patient and in a healthy brain (Extended Data Fig. 10) by immunohistochemistry. Assessment of BCMA protein expression on a human tissue array was positive on lymph node, spleen, lung and stomach due to plasma cells present in the bronchus-and mucosa-associated lymphoid tissue, respectively 5 . This tissue array, however, showed some positivity on climbing fibers in the cerebellum and explicitly does not rule out low-density expression in the central nervous system. We acknowledge that these data are, to some extent, conflicting due to lack of standardized protocols for staining tissues, other than bone marrow. A comprehensive evaluation of brain tissue for BCMA protein expression might be warranted to characterize the prevalence and extent of BCMA expression in the central nervous system and confirm the findings of this case report. BCMA expression on neurologic tissues in a subset of patients could affect the applicability of BCMA-targeted adoptive cell transfer in MM. Implications for other BCMA-targeted immunotherapies-for example, antibody-drug conjugates and bispecific antibodies-are unknown. Even though therapeutic antibodies are thought not to cross the blood-brain barrier, their permeability into the CSF should be carefully evaluated. Other tumor-associated targets are being currently studied in MM, including bispecific antibodies and CAR constructs targeting GPRC5D, FcRH5, CD19, CD38, CD56, CD138 and SLAMF7, some of which have a broader expression outside of plasma cells and warrant careful monitoring 20 .
Using chemotherapy to destroy CAR-T cells after infusion is itself associated with toxicity, as this case shows, because the patient died of infectious complications. Other strategies include a modification of CAR structure with engineered suicide genes, the incorporation of inhibitory CAR constructs or usage of a small molecule system as a safety switch to selectively deactivate the CAR-T cells.
Recently, CAR natural killer cells have been proposed as an alternative with off-the-shelf use as a potential advantage 20 .
This case shows the potential of BCMA-targeted CAR-T cells to cross the blood-brain barrier in a subset of patients and cause a progressive neurocognitive and movement disorder, possibly through targeting of BCMA-expressing cells of the basal ganglia. Neurotoxicity in general has been observed in 23 of 128 patients on ide-cel 11 and in 20 of 97 patients on cilta-cel 13 . Non-ICANS neurotoxicity was not addressed specifically in the ide-cel study but was reported in 12 of 97 patients from the phase 2 study of cilta-cel, of which five patients had a cluster of movement and neurocognitive adverse events (three with grade 3 or higher parkinsonism) 13 . The development of this toxicity in the cilta-cel trial was associated with the presence of two or more risk factors (including high tumor burden, previous grade ≥2 CRS, previous ICANS and high CAR-T cell expansion and persistence). The ide-cel package insert also mentions that grade 3 parkinsonism has occurred after treatment, suggesting that this complication is not necessarily specific to one BCMA-targeted CAR-T cell product. We acknowledge that important questions remain unanswered. Our patient developed neutropenic fever at day 51; it is not well studied whether infections after CAR-T infusion might activate CAR-T cells in vulnerable patients and whether more stringent prophylaxis of infection is warranted. Additional studies to confirm the proposed mechanism of neurotoxicity could help delineate the fraction of patients at risk. In conclusion, our findings suggest that anti-BCMA CAR-T cell therapies, although effective in MM, warrant close monitoring for neurotoxicity, especially as such treatments acquire more widespread implementation in patients with MM.

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/ s41591-021-01564-7.

Methods
Trial design. The CARTITUDE-1 trial (https://clinicaltrials.gov/ct2/show/ NCT03548207) is an open-label, single-arm, phase 1b/2 trial that evaluates the safety and efficacy of JNJ-68284528 (cilta-cel), a CAR-T cell therapy directed against BCMA in patients with relapsed or refractory MM. Here we provide the case report of a patient with neurotoxicity enrolled on the CARTITUDE-1 trial. Analysis and reporting follow the CARE guidelines. The CITE-seq experiment includes data on three additional patients with MM who were enrolled on the CARTITUDE-1 trial (a 61-year-old female, a 67-year-old male and a 67-year-old female). Furthermore, all patients with MM included in this work consented to participation in the Multiple Myeloma Biorepository (HSM:18-00456). All patients provided written informed consent for the evaluations. All study protocols were approved by the Program for the Protection of Human Subjects and the Institutional Review Board at the Icahn School of Medicine at Mount Sinai and adhere to the 2008 Declaration of Helsinki.
Sample collection and tissue processing. Peripheral blood was collected in heparin anti-coagulated green tops (10 ml) via venipuncture throughout the course of the patient's treatment in accordance with standard-of-care lab draws. Plasma was isolated from peripheral blood. Peripheral blood mononuclear cells (PBMCs) were Ficoll density separated and cryopreserved. Cryopreserved PBMC samples were used for flow cytometry, mass cytometry and other assays as detailed below. CSF was collected by lumbar puncture in accordance with standard of care. Each sample of 8 ml was centrifuged at 300g at 4 °C for 10 min. Then, 0.5 ml of supernatant was divided into aliquots and frozen at −80 °C. Approximately 200 μl of CSF and plasma from peripheral blood were used for Olink and Ella proteomics analysis as detailed below. Cells from CSF were used immediately for flow cytometry.

Statistics and reproducibility.
Experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment. No statistical method was used to predetermine sample size for the analyses. For clinical and cytokine assays, no data points were removed from the analysis. Cytometry data were gated to relevant populations, as shown in Extended Data Fig. 5a. CITE-seq data were filtered to remove multiplets based on the crossSampleDoublets() and withinSampleDoublets() functions of the CiteFuse package (version 1.2.0) in R (version 3.6.1). No other cells were excluded from the analysis. The non-parametric Mann-Whitney U-test was used to compare gene expression values where appropriate. The Pearson correlation coefficient was used to characterize correlation of cytokine expression between blood and CSF. For all analyses, a two-sided P value of less than 0.05 was considered significant.
Additional clinical information on the patient. There was no documented family history of movement disorders for the patient of the case report. In terms of neuro-psychiatric history, the patient had a remote history of migraines, documented in 2009, for which he received sumatriptan 100 mg as needed. In 2014, the patient was diagnosed with a mood disorder and started on sertraline 150 mg. He had been taking lorazepam and alprazolam; these were discontinued at that time, and clonazepam 1 mg daily was started instead. No anti-dopaminergic medications were taken by the patient around this episode or later. Due to recurring anxiety with panic attacks, the sertraline dose was increased to 200 mg, and clonazepam was gradually increased to a maximum daily dose of 4 mg as needed. Sertraline was discontinued in 2016. The clonazepam dose was maintained for recurring anxiety with panic attacks. In addition, the patient was seen at an outside hospital in 2018 after a traffic accident. All documented neurological examinations at that time were normal. He received an MRI of the brain, which noted non-specific punctate foci T2/FLAIR hyperintensity in the periventricular and subcortical white matter, likely secondary to chronic microvascular ischemic disease, but no other intracranial abnormalities. Six months before the CAR-T trial, the patient was seen by a neurologist for the evaluation of weakness in the right hand. The neurological examination and tests of motor function in the limbs were normal, with the exception of portions of the right arm. Symptoms were thought to be suggestive of a radial nerve irritation at the spiral groove related to a work-related overuse problem. Electromyography confirmed mild acute denervation showing a radial nerve injury with mild acute axonal involvement, and the patient's symptoms resolved with rest. During the screening visit for the CAR-T trial, the patient reported grade 1 fatigue, a remote syncope (around 2011) as well as grade 1 peripheral sensory neuropathy (which affects the soles of the feet, toes, calves and fingers), described as cramping without numbness. The neuropathy complaints did not interfere with walking, balance or fine motor movements and developed after bortezomib treatment. During the hospitalization at the time of CAR-T cell infusion, which includes the CRS period, the patient received a neurological evaluation every day. The patient was specifically monitored for neurotoxicity according to the specifications of the clinical trial protocol. Neurological examination was documented at every subsequent study visit (every 28 d) after CAR-T infusion. Additionally, handwriting logs for dexterity were performed as specified. In conclusion, the patient was evaluated by a neurologist 6 months before CAR-T therapy (by the same physician who evaluated the patient when he presented with the described neurotoxicity after CAR-T). He received other neurological examinations immediately before and after CAR-T cell infusion according to the clinical trial protocol. No pre-existing signs of parkinsonism were present during evaluation before CAR-T infusion.
Flow cytometry. Cryopreserved Ficoll density separated PBMCs were thawed by standard technique. Cells in the CSF were used within 3 h of collection after isolation. CD3 + /CD4 + /CD8 + T cell, CD19 + B cell and anti-BCMA-directed T cells were measured by multiple-color flow cytometry with human monoclonal ACROBiosystems anti-BCMA (FITC) (cat. no. BCA-HF254-25µg) and BioLegend human monoclonal anti-CD3 (cat. no. 300472 and cat. no. 344842), human monoclonal anti-CD4 (cat. no. 317434), human monoclonal anti-CD8 (cat. no. 344742) and human monoclonal anti-CD19 (cat. no. 561121). All cell surface antibodies were used at a 1:20 dilution following the manufacturer's recommendations. The FITC-labeled human BCMA was used at a 1:100 dilution. The samples were acquired on a FACS LSRFortessa flow cytometry system (BD Biosciences). Data were visualized and analyzed using Cytobank 21 .
Olink multiplex proteomics assay. Relative protein expression was measured in the CSF and peripheral blood plasma using Olink proximity extension technology, a high-throughput multiplex proteomic immunoassay, following the manufacturer's protocols. The commercially available Immuno-Oncology (article no. 95310), which includes 92 immune-and oncology-related, proteins was used. A table with all cytokines measured is included below as Supplementary Table 2. Olink uses marker-specific binding and hybridization of a set of paired oligonucleotide antibody probes that is subsequently amplified using a quantitative PCR. Protein expression values are reported as normalized protein expression values on a log 2 scale. Analysis was conducted in R (version 3.6.1), and figures were produced using the package pheatmap 22 .
Ella cytokine detection. The ProteinSimple Ella cytokine detection system uses microfluidics ELISA assays in a multi-analyte chip that were run within cartridges in triplicate following the manufacturer's instructions. Human analytes of IL-6, IL-8, TNF-α, IL18, IFN-γ and IL-10 were performed by the Mount Sinai Human Immune Monitoring Center using 25-30 µl of plasma or CSF from the patient. Analysis was conducted in R, and figures were generated using the package ggplot2 (ref. 23 ). Supplementary Tables 3 and 4. Antibodies used were either purchased pre-conjugated with metals from Fluidigm or purchased unconjugated and metal conjugated in-house at the Human Immune Monitoring Center, Icahn School of Medicine. All in-house conjugated antibodies were titered and validated on healthy donor PBMCs. All antibodies for CyTOF listed in Supplementary Tables 3 and 4 were used at a dilution of 1:100. For longitudinal monitoring of phenotypic changes, cells from selected time points were thawed and labeled with Rh103 intercalator (Fluidigm) as a viability dye and cell proliferation marker IdU (Cell-ID 127 5-Iodo-2′-deoxyuridine, Fluidigm). Cells were initially stained with a cocktail of surface antibodies that included BCMA-FITC (ACROBiosystems) (Panel 1). Surface-stained cells were further stained with polyclonal anti-FITC-159Tb (source) and fixed with 1.6% formaldehyde. Each time point was then barcoded with CyTOF Cell-ID 20-Plex Palladium Barcoding Kit (Fluidigm). Barcoded cells were fixed and permeabilized with Fix-Perm buffer (BD Biosciences) and stained with the remaining intracellular antibodies from CyTOF Panel 1. Intracellular cytokine expression was monitored using CyTOF Panel 2. Cells from selected time points were activated with PMA/ionomycin (BioLegend) in the presence of brefeldin-A (BioLegend) for 6 h. After activation, cells were stained with Rh103 intercalator, stained with BCMA-FITC and fixed with 1.6% formaldehyde. Fixed cells were palladium barcoded with CyTOF Cell-ID 20-Plex Palladium Barcoding Kit and pooled and stained with surface markers from CyTOF Panel 2, including polyclonal anti-FITC-169Tm. Cells stained with surface antibodies were fixed and permeabilized with Fix-Perm buffer and stained with cytokine antibodies. Samples stained with either CyTOF antibody Panel 1 or CyTOF antibody Panel 2 were finally fixed in freshly diluted 2.4% formaldehyde containing 125 nM intercalator-Ir (Fluidgm) and 300 nM OsO 4 (Acros Organics) and stored at 4 °C in cell staining buffer containing (Fluidigm) 125 nM intercalator-Ir until acquisition. Samples for CyTOF acquisition were washed with CAS buffer (Fluidigm) and re-suspended in CAS buffer containing EQ normalization beads (Fluidigm) and acquired on CyTOF2 (Fluidigm). After acquisition, the data were normalized using the bead-based normalization algorithm in the CyTOF software (Fluidigm). Normalized data were de-barcoded using methods and software developed by Gary Nolan's group at Stanford University School of Medicine 24 . Normalized and de-barcoded data were uploaded to Cytobank 21 for final analysis, as detailed below.

Mass cytometry. Cells were stained with either CyTOF antibody Panel 1 or CyTOF antibody Panel 2 listed in
Mass cytometry data analysis. Data in FCS file format were downloaded from Cytobank 21 . For analysis of mass cytometry data, we used a workflow based on the example by Nowicka et al. 25 using the diffcyt 26 and CATALYST 27 packages in R (version 3.6.1). In brief, data were imported and transformed for analysis using the read.flowSet() function from the flowCore package 28 and the prepData(…, cofactor = 5) function from the CATALYST package, respectively. Clustering was based on the FlowSOM algorithm 29 using all protein markers from the panel on a 10 × 10 grid size with a maximum of K = 20 clusters. These clusters were visualized using uniform manifold approximation and projection dimension reduction and subsequently annotated based on canonical protein markers and the FITC-BCMA tag to identify CAR-T cells.

CITE-seq.
For each sample, cell suspensions were split and barcoded using 'hashing antibodies' staining β-2-microglobulin and CD298 and conjugated to 'hash-tag' oligonucleotides (HTOs). Before hashing, each of the five samples was split into two aliquots and either 'stimulated' or 'unstimulated' . Stimulated aliquots were incubated for 3 h at 37 °C with PMA/ionomycin. Unstimulated aliquots were incubated for 3 h at 37 °C with cRPMI. After these incubations, the ten aliquots were hashed and pooled. Hashed samples were pooled and stained with CITE-seq antibodies purchased from the BioLegend TotalSeq catalog; the FITC antibody was a custom conjugate from BioLegend. All commercial antibodies were diluted at 1:100 according to the manufacturer's instructions. The custom conjugate is titered to find the optimal volume to stain PBMCs. The CITE-seq panel is detailed in Supplementary Table 5. Stained cells were then encapsulated for single-cell reverse transcription using the 10x Chromium platform (5′, version 1.0), and libraries were prepared according to the manufacturer's instructions with minor modifications summarized hereafter. In brief, cDNA amplification was performed in the presence of 2pM of an antibody oligo-specific primer to increase yield of antibody-derived tags (ADTs) and 3pM of specific primer to increase the yield of HTOs. The amplified cDNA was then separated by SPRI size selection into cDNA fractions containing mRNA-derived cDNA (>300 bp) and ADT-derived cDNA (<180 bp), which were further purified by additional rounds of SPRI selection. Independent sequencing libraries were generated from the mRNA and ADT cDNA fractions, which were quantified, pooled and sequenced together on an Illumina NextSeq/ NovaSeq to a targeted depth of 25-750 million reads per gene expression library and 1,000-30,000 targeted reads per cell.

CITE-seq data analysis.
Illumina sequencer base call files were de-multiplexed into FASTQ files using the cellranger (version 3.0.1) mkfastq and count pipeline. CITE-seq data were analyzed using R (version 3.6.1) and the CiteFuse package 30 , using the proposed analysis pipeline with minor modifications. In brief, matrices with counts representing RNA, ADT and HTO data, respectively, were read into R separately and combined into a SingleCellExperiment object 31 using the preprocessing() function. Metadata (including patient ID and experimental condition (stimulated versus unstimulated)) were added based on known experimental design and corresponding HTOs. HTO expression was normalized using the log-transform method and the normaliseExprs() function. Cross-sample doublets and within-sample doublets were identified and removed using the crossSampleDoublets() and withinSampleDoublets(…, minPts = 10) functions, respectively. Similarity network fusion (SNF) was used to integrate RNA and ADT matrices after calculating log-transformed normalized expression values with the CiteFuse() function. Both spectral clustering with K = 25 and Louvain clustering were attempted, and t-distributed stochastic neighbor embedding (t-SNE) was used to visualize dimension reduction. Manual inspection and canonical gene and protein expression were used to identify clusters corresponding to CD4 + and CD8 + CAR-T cells. These cells were isolated into distinct SingleCellExperiment objects for downstream analysis. SNF, clustering and dimension reduction of CAR-T cells was done in a similar fashion as detailed above. The DEgenesCross() function with standard parameters was used to determine differentially expressed genes between the patient with neurotoxicity and all other patients. Differential expression was determined with a two-sided Mann-Whitney U-test, and P values were corrected using the Benjamini-Hochberg method.
Immunohistochemistry. Slides with 5-µm sections from paraffin-embedded tissues from autopsies were stained with CD3 (LN10) and GFAP (GA5) pre-diluted BOND reagents from Leica Biosystems, heat-induced epitope retrieval for 20 min with ER2 (Bond Epitope Retrieval Solution 2), MSMC DAB detection and counterstained per established staining protocol on the automated Leica Biosystems BOND-III platform. Immunohistochemistry for BCMA was performed using Ventana DISCOVERY ULTRA from Roche. This system allows for automated baking, de-paraffinization and cell conditioning. Semi-automatic staining was performed using BCMA antibody (cat. no. B0807 from USBiological) at 1:10 dilution during 60 min. As secondary antibody, Discovery OmniMap anti-rabbit-HRP from Roche (760-4310) was used, and the signal was obtained using DISCOVERY ChromoMap DAB RUO from Roche (760-2513) (brown signal). Tissues were counterstained with hematoxylin (in blue).
Analysis of public datasets. The mRNA expression data from the Allen Brain Atlas were last accessed on 25 April 2021. The heat map can be found at http:// human.brain-map.org/ (human brain data) when doing a Gene Search for TNFRSF17 and DRD1 and selecting 'View Selection Thumbnails' . Raw expression data of the heat map used as part of the figures were downloaded by the authors from the Allen Brain Atlas data portal and are included as Source Data of Extended Data Fig. 8.
Reporting Summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
All requests for raw and analyzed data and materials will be promptly reviewed by the Icahn School of Medicine at Mount Sinai and Mount Sinai Hospital to determine if the request is subject to any confidentiality and data protection obligations. Requests for data should be addressed to the corresponding author via e-mail, and a reply will be sent within ten business days. Any data and materials that can be shared will be released via a material transfer agreement. Raw and analyzed CITE-seq data are available through the National Center for Biotechnology Information's Gene Expression Omnibus (accession no. GSE182527). Mass cytometry and intracellular cytokine data are available through the FlowRepository website (ID FR-FCM-Z4KB). The images derived from the Allen Human Brain Atlas can be accessed at https://human.brain-map. org/. Specific URLs to recreate the following figures are provided: Fig. 2b (https:// human.brain-map.org/static/brainexplorer), Extended Data Fig. 8a (https://human. brain-map.org/microarray/search/show?search_type=user_selections&user_ selection_mode=1) and Extended Data Fig. 8b (https://human.brain-map. org/microarray/gene/show/605), and source data are available. For all clinical measurements and cytokine levels (Extended Data Figs. 1, 2, 6 and 9d-f), source data are available. Source data are provided with this paper. Fig. 1 | Clinical course and biochemical parameters after CAR-T cell treatment. The time periods associated with cytokine release syndrome (CRS), neutropenic fever and neurotoxicity are annotated in the individual subplots. All cytokine levels were determined in the peripheral blood. CAR-T cell phenotype, as determined by expression of CCR7 and CD45RA, illustrating a high fraction of effector-memory T cells at all time points. Each bar corresponds to N = 1 sample collected from the patient. The UMAP plots visually illustrate the clustering of T cells and confirm low CCR7 and CD45RA expression on CAR-T cells. Fig. 7 | Expression of canonical markers on CiTE-seq data identifies and clusters major immune cell types. (a) t-SNE plot representation of CITE-seq analysis of peripheral blood mononuclear cells before and after PMA/ionomycin stimulation. Clustering was determined by similarity network fusion (SNF) and Louvain clustering algorithm. Individual cells are colored by subject (healthy donor (HD), neurotoxicity patient (NEUROTOX) and 3 other patients on the same clinical trial without neurotoxicity (MM1, MM2, MM3). Highlighted are the major immune cell types (B cells, NK cells, CD8 + T cells, CD4 + T cells, CAR-T cells and monocytes). There is a small cluster of events that corresponds to multiplets or debris (centrally, not highlighted). (b) Expression level of canonical genes: CD8A, CD4, CD14, FCGR3A (CD16), CD19 and NCAM1 (CD56). In each case showing both mRNA (top) and ADT (antibody-derived tag, representation of protein level) (high = red, low = blue). Expression levels are normalized as described in the Methods.