Intrathecal Allogeneic B7-H3–Targeted CAR γδ T Cells for Leptomeningeal Metastasis from Solid Tumors: Safety, Efficacy, and Immunologic Dynamics in a Phase I Trial

Abstract

Purpose:

To evaluate intrathecal QH104, an allogeneic B7 homolog 3 (B7-H3)–targeted chimeric antigen receptor (CAR) γδ T-cell therapy, in patients with leptomeningeal metastasis (LM).

Patients and Methods:

In this phase I study (NCT06592092), 3 patients with LM from B7-H3–positive solid tumors were enrolled, including 2 with lung adenocarcinoma and 1 with triple-negative breast cancer. QH104 was administered at a fixed dose of 3 × 10⁷ cells per infusion via lumbar puncture or an Ommaya reservoir. The primary objective was to evaluate clinical response. Secondary objectives included safety, overall survival, quality of life, and pharmacokinetic/pharmacodynamic profiling.

Results:

QH104 was generally well tolerated, with no grade ≥ 4 treatment-related adverse events (TRAE). One patient developed grade 3 immune effector cell–associated neurotoxicity syndrome. All TRAE resolved with supportive care. All patients had stable disease on days 14 and 30 according to the Response Assessment in Neuro-Oncology Leptomeningeal Metastases criteria, with symptom improvement in 2 patients. Cerebrospinal fluid (CSF) cytology converted to and remained negative through day 30 in 1 patient with positive baseline cytology. B7-H3–CAR γδ T cells persisted in CSF for at least 1 week and were accompanied by increased interferon (IFN) γ levels. Single-cell sequencing suggested IFNγ-associated immune remodeling in CSF, including expansion of an inflammatory macrophage state and subsequent lymphocyte recruitment.

Conclusions:

This study provides preliminary proof of concept for intrathecal allogeneic B7-H3–CAR γδ T-cell therapy in LM, with early evidence of safety, clinical activity, local persistence, and IFNγ-associated immune remodeling in CSF.

Translational Relevance.

This phase I study provides preliminary proof of concept for intrathecal administration of allogeneic B7-H3–targeted chimeric antigen receptor (CAR) γδ T cells as an off-the-shelf cellular therapy for leptomeningeal metastasis (LM), a rapidly progressive complication with limited treatment options and a poor prognosis. In addition to preliminary evidence of safety and early clinical activity, the study shows local persistence of CAR γδ T cells and interferon γ–associated immune remodeling in cerebrospinal fluid, including expansion of an inflammatory macrophage state and recruitment of additional lymphocytes. These findings suggest that therapeutic activity in LM may arise not only from direct cytotoxicity but also from remodeling of the immunosuppressive microenvironment. More broadly, this work supports further investigation and development of γδ T-cell–based therapies for LM and potentially other central nervous system metastases from solid tumors.

Introduction

Leptomeningeal metastasis (LM) is an uncommon but devastating complication of advanced solid tumors, affecting approximately 1% to 10% of patients, most commonly in breast cancer, lung cancer, and melanoma (1–3). Once LM develops, neurologic deterioration can be rapid, often with increased intracranial pressure, and prognosis remains poor. No therapy is currently approved specifically for LM. Existing treatments, including systemic chemotherapy, radiotherapy, and neurosurgical interventions, are used primarily to palliate symptoms and delay progression, underscoring the substantial unmet clinical need for more effective therapies (4, 5).

Clinical trial data to date have not shown a consistent survival benefit of immune checkpoint inhibitors (ICI) over other commonly used LM treatments, including chemotherapy and targeted therapy (6, 7). These limitations have increased interest in engineered cellular therapies. Prior studies of intrathecal chimeric antigen receptor (CAR) αβ T-cell therapy in central nervous system (CNS) tumors, including HER2–targeted CAR αβ T cells in recurrent or refractory CNS tumors (8) and epidermal growth factor receptor (EGFR)/IL13Rα2–targeted CAR αβ T cells in glioblastoma (GBM; ref. 9), support the feasibility of regional cellular immunotherapy in the CNS. For rapidly progressive CNS disease, γδ T cells may offer additional advantages because they recognize tumors in an MHC-independent manner, mediate potent cytotoxicity, and may be developed as off-the-shelf products (10–12).

QH104 is an allogeneic B7 homolog 3 (B7-H3)–targeted CAR Vγ9Vδ2 T-cell product with antiproliferative activity against B7-H3–positive tumor cells in vitro (Supplementary Fig. S1). Building on our prior report supporting the safety and activity of intrathecal QH104 in recurrent GBM (13), we initiated a phase I trial in patients with LM. Here, we report the initial clinical and correlative observations from the first 3 treated patients.

Patients and Methods

Study design and participants

This phase I, nonrandomized, single-center, fixed-dose trial was conducted at Cancer Hospital, Chinese Academy of Medical Sciences (NCT06592092; Fig. 1A). The study was approved by the institutional ethics committee (24/023-4303) and conducted in accordance with the Declaration of Helsinki and the International Council for Harmonization guideline for Good Clinical Practice. Written informed consent was obtained from all participants.

Figure 1.

Study design and clinical response to intrathecal QH104. A, Overview of the study design and QH104 manufacturing workflow. B, Swimmer plot showing, for each patient, the interval between LM diagnosis and QH104 infusion, prior and concomitant treatments, and B7-H3 H-scores. C, Longitudinal clinical response assessments for each patient, including CSF cytology, neuroimaging, and neurologic examination, and overall response according to the RANO-LM criteria. D, Representative contrast-enhanced brain and spinal MRI scans obtained before and after QH104 infusion, showing radiographic stability (red arrowheads). mbIL15, membrane-bound IL15; NA, not available; NANO, Neurologic Assessment in Neuro-Oncology; PK/PD, pharmacokinetics and pharmacodynamics; TM, transmembrane. [A, Created in BioRender. Zhou, Y. (2026) https://BioRender.com/jrzrc5a.]

The primary objective was to evaluate clinical response. Secondary objectives were safety, overall survival (OS), quality of life (QoL), and pharmacokinetic/pharmacodynamic profiling. Exploratory objectives were to characterize postinfusion B7-H3–CAR γδ T-cell dynamics in cerebrospinal fluid (CSF), identify response-associated biomarkers, and explore mechanisms underlying the B7-H3–CAR γδ T-cell–tumor interaction.

Eligible patients had confirmed B7-H3 expression, a Karnofsky performance status (KPS) of ≥ 60, and a life expectancy of > 8 weeks. LM was diagnosed in accordance with the European Association of Neuro-Oncology–European Society for Medical Oncology guidelines, based on neurologic symptoms with supportive neuroimaging and/or positive CSF cytology (14). B7-H3 expression was assessed by immunohistochemistry on primary tumor specimens and quantified by H-score.

QH104 was administered intrathecally at a fixed dose of 3 × 10⁷ cells per infusion via lumbar puncture or an Ommaya reservoir. The dose was selected based on prior phase I data supporting the safety and activity of QH104 in recurrent GBM at doses of 1 × 10⁷, 3 × 10⁷, and 1 × 10⁸ cells (13). No lymphodepleting chemotherapy was given. If the first infusion was tolerated, a second infusion was permitted at the investigator’s discretion for neurologic deterioration or worsening CSF cytology. The full protocol, statistical analysis plan, and Consolidated Standards for Reporting Trials Dose-finding Extension (CONSORT-DEFINE) checklist are provided as Supplementary Data S1–S3, respectively.

QH104 manufacturing

The lentiviral vector encoding the B7-H3–targeted CAR with membrane-bound interleukin (IL) 15 was produced under Good Manufacturing Practice (GMP) conditions using a third-generation, four-plasmid packaging system in suspension-adapted HEK293T/17 cells (ATCC, cat. #CRL-11268, RRID: CVCL_1926). The vector was self-inactivating, with deletion of the U3 region in the 3′ long terminal repeat, resulting in loss of promoter activity after integration. The viral supernatant was harvested and purified by ultrafiltration, nuclease treatment, chromatography, and sterile filtration. Functional titer was determined in Jurkat clone E6-1 cells (ATCC, cat. #TIB-152, RRID: CVCL_0367) by serial dilution transduction followed by flow cytometric assessment of CAR expression at 72 hours. Lentiviral vector release testing included identity, purity, sterility, Mycoplasma, endotoxin, and replication-competent lentivirus.

QH104 was manufactured under GMP conditions from leukapheresis products obtained from healthy donors after infectious disease screening. Peripheral blood (PB) mononuclear cells were isolated via Ficoll-Paque density-gradient centrifugation. αβ T cells were depleted using the TCRα/β T Cell Depletion Kit (Miltenyi Biotec, cat. #130-133-896), and depletion efficiency was confirmed via flow cytometry. γδ T cells were activated with zoledronic acid (Zometa), transduced with the lentiviral vector at a multiplicity of infection of 1 to 3, and expanded for 9 days in culture bags using serum-free medium supplemented with IL2 (100–1,000 IU/mL). Cells were then harvested, washed, formulated in CS10 cryopreservation medium (BioLife Solutions, cat. #210102), cryopreserved, and stored in vapor-phase liquid nitrogen. Final product release testing included identity, purity, potency, and safety (Supplementary Table S1).

Clinical response and safety assessments

Clinical response was evaluated according to the Response Assessment in Neuro-Oncology Leptomeningeal Metastases (RANO-LM) criteria using neurologic examination, neuroimaging, and CSF cytology (15). Assessments were performed at baseline and at prespecified time points from day 14 onward. Neurologic examination was performed using the Neurologic Assessment in Neuro-Oncology scale (16). CSF cytology was performed using a liquid-based ThinPrep test with Papanicolaou staining and independently reviewed by 2 experienced cytopathologists. Cytology was considered positive when malignant cells were definitively identified.

All adverse events (AE) after enrollment were recorded and assessed for incidence, duration, severity, and management. Cytokine release syndrome (CRS) and immune effector cell–associated neurotoxicity syndrome (ICANS) were graded according to the American Society for Transplantation and Cellular Therapy consensus criteria. Graft-versus-host disease (GvHD) was graded according to the Mount Sinai Acute GvHD International Consortium criteria. All other AE were graded according to the Common Terminology Criteria for Adverse Events (v5.0).

Correlative assays

CAR transgene levels were measured using a validated TaqMan-based quantitative PCR (qPCR) assay (forward: GGCTTTCATTTTCTCCTCCTTGTA; reverse: CGGGCCACAACTCCTCATAA; probe: FAM-ATCCTGGTTGCTGTCTC-MGB). Genomic DNA was extracted using a column-based kit (Qiagen, cat. #51185), and qPCR was performed on a QuantStudio 3 Real-Time PCR System (Applied Biosystems). The assay showed a linear dynamic range of 5 to 1 × 10⁶ copies/μL, amplification efficiency of 96.1% to 99.5%, and a correlation coefficient of ≥ 0.9996. Samples were run in duplicate and accepted only when the coefficient of variation was < 15%.

Flow cytometry was performed on a NovoCyte Quanteon instrument (Agilent) using antibodies against CD45 (BioLegend, cat. #304036, RRID: AB_2561940), CD3 (BioLegend, cat. #300428, RRID: AB_893298), Vδ2 (BioLegend, cat. #331428, RRID: AB_2783211), T-cell receptor (TCR) α/β (BioLegend, cat. #306718, RRID: AB_10612569), CD19 (BD Biosciences, cat. #563226, RRID: AB_2744313), and CD56 (BD Biosciences, cat. #564058, RRID: AB_2738569), together with biotinylated B7-H3 protein followed by PE-conjugated streptavidin (BioLegend, cat. #405204).

Cytokine levels in CSF were measured using the LEGENDplex Human Th Cytokine Panel (BioLegend, cat. #741027), and those in PB were measured using a 12-cytokine detection kit (Raisecare Biotech). Protein profiling was performed using the Olink Target 96 Immuno-Oncology Panel (Olink Proteomics, cat. #95311). Olink data were quality-controlled, normalized, and reported as Normalized Protein eXpression (NPX) values on a log₂ scale.

Single-cell RNA and V(D)J sequencing

Single-cell suspensions were quantified and assessed for viability by acridine orange/propidium iodide staining using a fluorescence cell counter (SeekGene, cat. #M002C, RRID: SCR_028028) before library preparation for single-cell RNA sequencing (scRNA-seq) and single-cell V(D)J sequencing [scV(D)J-seq]. Single cells, reverse transcription reagents, gel beads, and partitioning oil were loaded onto a SeekOne DD Chip S3 (SeekGene, cat. #K00202-0801) to generate emulsion droplets for reverse transcription. Complementary DNA (cDNA) was amplified and used for library construction with the SeekOne DD Single Cell 5′ Library Preparation Kit (SeekGene, cat. #K00501), which incorporates cell barcodes and unique molecular identifiers (UMI). V(D)J transcripts were enriched from 5′ cDNA using the SeekOne DD Single Cell V(D)J Enrichment Kit for human TCR (SeekGene, cat. #K00601). Libraries were purified, quantified, quality-assessed, and sequenced on a NovaSeq X Plus platform (Illumina, cat. #20084804, RRID: SCR_024568) with paired-end 150-bp reads.

Single-cell data processing and analysis

FASTQ files were processed with Cell Ranger (v9.0.1, RRID: SCR_017344). scRNA-seq data were aligned using a custom reference combining the GENCODE GRCh38 human genome assembly (RRID: SCR_014966) with the annotated B7-H3–targeted CAR single-chain variable fragment (scFv) sequence. scV(D)J-seq data were processed to assemble per-cell TCR contigs and recover CDR3 sequences.

Downstream analyses were performed in Seurat (v5.3.0, RRID: SCR_016341; ref. 17). Cells with 400 to 8,000 detected genes, 600 to 120,000 UMI, and < 10% mitochondrial RNA were retained. Doublets were removed per sample using DoubletFinder (v2.0.6, RRID:SCR_018771; ref. 18). After quality control, data were merged, log-normalized, and scaled, and principal component analysis was performed. Batch effects were corrected using Harmony (v1.2.3, RRID: SCR_022206; ref. 19). Uniform Manifold Approximation and Projection embedding used Harmony-corrected dimensions 1 to 30. Graph-based clustering was performed with FindNeighbors and FindClusters, and clustree (v0.5.1, RRID: SCR_016293) was used to guide resolution selection.

Clusters were annotated using marker genes identified with FindAllMarkers and established cell-type markers (Supplementary Table S2; refs. 20–23). Cells were first partitioned into lymphoid and myeloid compartments and then subclustered. B7-H3–CAR γδ T cells were identified by expression of the CAR scFv transcript. In sex-mismatched allogeneic samples, donor versus recipient origin was inferred from expression of the sex-linked transcripts XIST and RPS4Y1. Signature scores were calculated with AddModuleScore using previously reported gene sets (Supplementary Table S3; refs. 24–29).

Differential gene expression analysis was performed using FindAllMarkers or FindMarkers. Genes with |avg_log₂ fold change| > 0.5 and an adjusted P value <0.05 were considered significant and subjected to enrichment analysis. Gene Ontology (GO; RRID: SCR_002811) and Kyoto Encyclopedia of Genes and Genomes (KEGG, RRID: SCR_012773) pathway analyses were performed with clusterProfiler (v4.14.6, RRID: SCR_016884; ref. 30) using enrichGO and enrichKEGG.

Trajectory analysis was performed using Monocle2 (v2.36.0, RRID: SCR_016339; ref. 31). Ordering genes were selected based on high dispersion (mean expression ≥ 0.1 and empirical dispersion ≥ fitted dispersion), and patient identity was included as a covariate during dimensionality reduction. For monocyte and macrophage trajectory analysis, classical monocyte was set as the root state in order_cells based on its established precursor role. Cell–cell communication analysis was performed using CellChat (v1.6.1, RRID: SCR_021946; ref. 32) using the normalized gene expression matrix and CellChatDB.human.

TCR clonotype analysis

V(D)J contig data were analyzed using scRepertoire (v2.2.1, RRID: SCR_025691; ref. 33). For barcodes containing multiple chains, the 2 most highly expressed chains were retained. Clonotype assignments were integrated into the Seurat object. Within each sample, clonotypes were classified as expanded or unexpanded; expanded clonotypes were defined by a frequency of > 0.1% and a count of > 2 (34). Clonotypes were also classified as PB–CSF shared or compartment-restricted based on TCRα or TCRβ clonotype overlap between compartments. Among CSF-restricted clonotypes, those detected at baseline were further classified as CSF-preexisting, whereas the remainder were considered to be of undetermined origin.

Temporal pattern analysis

Temporal patterns in protein levels and scRNA-seq pseudobulk profiles were analyzed using Mfuzz (v2.66.0, RRID: SCR_000523; ref. 35). For Olink data, NPX values were centered to day 0 within each cycle for each patient and scaled by the standard deviation across time points. For scRNA-seq data, pseudobulk profiles were generated using AggregateExpression and normalized with estimateSizeFactors in DESeq2 (v1.46.0, RRID: SCR_000154). The top 2,000 genes ranked by analysis of variance (ANOVA) P values after Benjamini–Hochberg correction were retained. For both analyses, patient-level values were averaged at each time point before clustering.

Statistical analysis

Statistical analyses were performed in R (v4.4.3, RRID: SCR_001905). All tests were two-sided, and P values < 0.05 were considered statistically significant. Unless otherwise specified, analyses were descriptive or unsupervised.

Results

Patient characteristics and QH104 characterization

Between June 2024 and February 2025, 3 female patients with LM were enrolled and received intrathecal QH104 (Supplementary Table S4; Supplementary Fig. S2). The patients were 38 to 58 years of age (median, 53 years), all had a KPS of 80, and B7-H3 H-scores of 40 to 70 (median, 60). Patients 01 and 02 had EGFR exon 21 p.L858R–mutant lung adenocarcinoma, and patient 03 had triple-negative breast cancer (TNBC; Table 1). At enrollment, all patients had progressive neurologic deterioration despite stable extracranial disease.

Table 1.

Baseline characteristics and QH104 treatment details.

Patient ID	Diagnosis	Sex	Age, years	Genetic mutation/tumor subtype	B7-H3 H-score	QH104 dose, cells	Administration route
01	LUAD LM	F	53	EGFR exon 21 p.L858R	70	3 × 107	Lumbar puncture
02	LUAD LM	F	58	EGFR exon 21 p.L858R	40	3 × 107	Ommaya reservoir
03	BRCA LM	F	38	TNBC	60	3 × 107	Ommaya reservoir

Abbreviations: BRCA, breast cancer; F, female; ID, identification number; LUAD, lung adenocarcinoma.

Prior systemic and intrathecal treatments are summarized in Fig. 1B. Briefly, patient 01 had received osimertinib, bevacizumab, and intrathecal pemetrexed. Patient 02 had received osimertinib plus intrathecal methotrexate, followed by furmonertinib plus intrathecal pemetrexed. Patient 03 had received sacituzumab govitecan plus bevacizumab, followed by eribulin, intrathecal methotrexate, and intrathecal thiotepa.

Three QH104 batches derived from 3 different male donors were used to treat the 3 female patients. QH104-1 was administered for both infusions in patient 01 and the first infusion in patient 02, QH104-2 for the second infusion in patient 02, and QH104-3 for the single infusion in patient 03. All batches met the predefined release criteria (Supplementary Table S1). Vδ2 T-cell purity ranged from 89.63% to 97.77% (median, 96.59%), cell viability from 80.5% to 86.9% (median, 86.7%), and residual αβ T-cell contamination from 0% to 0.03% (median, 0.01%). Transduction efficiency ranged from 20.44% to 69.13% (median, 46.56%), and the vector copy number ranged from 0.64 to 1.47 copies per transduced cell (median, 1.47).

Toxicity

All 3 patients experienced at least 1 treatment-related adverse event (TRAE; Table 2). The most common TRAE were nausea, headache, and lethargy. No grade ≥ 4 TRAE were reported. Patient 02, who had active epilepsy within the preceding 3 months, experienced a transient absence seizure on day 1. The event was managed with oral levetiracetam, and no further seizures occurred. Because of its temporal association with the infusion, it was managed as neurotoxicity and graded as grade 3 ICANS. However, attribution to QH104 infusion was confounded by the patient’s baseline seizure risk. CRS occurred in all patients, and the highest grade was 2. All other TRAE were grade 1 or 2 and included fatigue, fever, and vomiting. All events returned to baseline within 2 weeks with supportive care, including acetaminophen, mannitol, aprepitant, and intravenous fluids. No GvHD was observed.

Table 2.

Summary of treatment-related adverse events.

TRAE	Any grade	Grade 1	Grade 2	Grade ≥ 3	Highest grade
Any TRAE	40	23	16	1	3
TRAE of special interest	​	​	​	​	​
CRS	3	1	2	0	2
ICANS	3	0	2	1	3
Blood and lymphatic system disorders	​	​	​	​	​
Leukopenia	1	1	0	0	1
Anemia	1	1	0	0	1
Gastrointestinal disorders	​	​	​	​	​
Nausea	3	0	3	0	2
Vomiting	2	0	2	0	2
Nervous system disorders	​	​	​	​	​
Headache	3	0	3	0	2
Lethargy	3	3	0	0	1
Dizziness	2	2	0	0	1
Paresthesia	1	1	0	0	1
Seizure	1	0	1	0	2
Memory impairment	1	1	0	0	1
Metabolism and nutrition disorders	​	​	​	​	​
Hypokalemia	3	3	0	0	1
Hyponatremia	3	2	1	0	2
Hyperglycemia	2	2	0	0	1
Hypoalbuminemia	2	2	0	0	1
Hypophosphatemia	1	1	0	0	1
Others	​	​	​	​	​
Fatigue	2	0	2	0	2
Creatinine increased	1	1	0	0	1
Alanine aminotransferase increased	1	1	0	0	1
Aspartate aminotransferase increased	1	1	0	0	1

AE attribution to QH104 was assessed by the investigators. AE were graded primarily according to the Common Terminology Criteria for Adverse Events (v5.0). CRS and ICANS were graded according to the American Society for Transplantation and Cellular Therapy consensus criteria.

Clinical outcomes

All 3 patients had stable disease (SD) on days 14 and 30. After the first infusion, neurologic symptoms improved in patients 01 and 02, whereas patient 03 showed transient improvement, followed by worsening on day 30. CSF cytology converted to and remained negative through day 30 in patient 02 with positive baseline cytology. The median OS was not reached, and all patients were alive at a median follow-up of 12 months. Changes in QoL are summarized in Supplementary Table S5.

At baseline, patient 01 had positive CSF cytology and leptomeningeal enhancement on magnetic resonance imaging (MRI), with impaired verbal expression, reduced facial strength affecting swallowing, mild behavioral change, grade 4 strength in the upper limbs, and an ataxic gait with lateral instability. She continued oral osimertinib 160 mg once daily for the primary lung cancer and discontinued bevacizumab and intrathecal pemetrexed 21 days before QH104 infusion. On day 14, upper-limb strength normalized, speech clarity improved, and gait improved from assisted ambulation to independent walking. However, reduced masticatory function persisted, and a new focal sensory deficit in the right inguinal region was noted. CSF cytology remained positive, although fewer tumor cells were seen after infusion (Fig. 1C). MRI showed no meaningful change (Fig. 1D; Supplementary Fig. S3A). She received a second infusion on day 80 for subsequent neurologic deterioration, without clear benefit (Fig. 1C; Supplementary Fig. S3A).

Patient 02 developed suspected LM after an epileptic seizure and lower-limb weakness, 3 years after surgery for right lung adenocarcinoma. Despite EGFR-targeted therapy and intrathecal chemotherapy, symptoms persisted and progressed. LM was confirmed 1 year later. At baseline, she had dizziness, gait instability, reduced vision and visual fields, and right facial hypesthesia. She continued oral osimertinib 80 mg once daily for the primary lung cancer and discontinued intrathecal pemetrexed 21 days before QH104 infusion. On day 30, right facial sensation recovered and dizziness and visual impairment resolved, whereas other symptoms were unchanged. CSF cytology converted to and remained negative through day 30, although MRI showed no meaningful change (Fig. 1C and D; Supplementary Fig. S3B). She received a second infusion on day 90 for recurrent CSF cytology positivity, without clear benefit (Fig. 1C; Supplementary Fig. S3B).

Patient 03 developed suspected LM after lower-limb numbness, 4 years after right modified radical mastectomy for TNBC with adjuvant chemotherapy and 1 year after right lung metastasectomy for breast cancer metastasis. Positron emission tomography/computed tomography (PET/CT) showed spinal leptomeningeal involvement, whereas CSF cytology remained consistently negative before enrollment. At baseline, symptoms persisted despite prior therapies and included lower-limb numbness, constipation, mild urinary retention, and numbness of the left-hand fingers. She discontinued prior treatment 21 days before QH104 infusion. Lower-limb numbness partially improved on day 14 but returned to baseline on day 30, with worsening lower-limb weakness. CSF cytology became transiently positive within 1 week after infusion but returned to negative on days 14 and 30 (Fig. 1C). Given the consistently negative baseline cytology, this finding should be interpreted cautiously and may reflect sampling variability, treatment-associated tumor cell shedding, or a transient fluctuation in tumor burden. MRI showed no meaningful change (Fig. 1D; Supplementary Fig. S3C).

CAR γδ T-cell persistence and immunologic dynamics in CSF after QH104 infusion

To assess CAR γδ T-cell persistence and associated immune changes after QH104 infusion, serial CSF and PB samples were collected on days 0, 1, 3, 7, and 14 of each cycle (Fig. 2A). qPCR and flow cytometry consistently showed that CAR transgene and CAR γδ T cells, respectively, remained detectable in CSF for at least 1 week after infusion in all patients (Fig. 2B; Supplementary Table S6). In the infused product (IP), CAR γδ T cells comprised 20.44% to 69.13% (median, 46.56%) of total Vδ2 T cells. During the first week after infusion, the corresponding proportion in CSF was 1.15% to 91.51% (median, 36.57%; Supplementary Table S6). CAR γδ T cells were not detected in PB (Supplementary Table S7; Supplementary Fig. S4A).

Figure 2.

Persistence of B7-H3–CAR γδ T cells and longitudinal changes in CSF immune cells, cytokines, and proteins after QH104 infusion. A, Sampling schedule for CSF and PB collection relative to each QH104 infusion. Day 0 denotes the sample collected before infusion. B–D, Longitudinal changes in the CSF CAR transgene copy number (B; per μg DNA); total T-cell, Vδ2 T-cell, NK-cell, and B-cell counts (C); and cytokine levels (D). E, Longitudinal changes in CSF protein levels with Mfuzz clustering. Heatmaps (left) show NPX values for proteins in clusters C1 and C2, centered on day 0 within each cycle for each patient and scaled by the standard deviation across time points. Line plots (top right) show representative clusters. Bubble plot (bottom right) shows enrichment analysis of proteins in clusters C1 and C2.

After each infusion, CSF lymphocyte counts peaked on day 1 and then declined with a modest secondary increase before returning to baseline on day 30. In parallel, CSF immune-cell composition shifted from γδ T-cell predominance during the first week to non–γδ T-cell predominance during the second week (Fig. 2C; Supplementary Table S6). Seven of the 12 cytokines measured in CSF increased on day 1, including IL5, IL6, IL9, IL13, IL22, tumor necrosis factor (TNF) α, and interferon (IFN) γ, and generally returned toward baseline on day 7 (Fig. 2D). By contrast, cytokine levels in PB remained largely unchanged (Supplementary Table S7; Supplementary Fig. S4B).

To characterize temporal patterns in CSF protein levels, we applied Mfuzz clustering to 92 proteins profiled in patients 01 and 02. Four clusters were identified, with clusters 1 and 2 showing an early peak on day 1 (Fig. 2E; Supplementary Fig. S4C). Cluster 1 included TNFα and TNF-related apoptosis-inducing ligand (TRAIL) and decreased rapidly after day 1. In contrast, cluster 2 included IFNγ, C–X–C motif chemokine ligand 10 (CXCL10), and programmed death ligand 1 (PD-L1) and declined gradually. GO and KEGG enrichment analyses of clusters 1 and 2 highlighted leukocyte migration and activation, humoral immune response, and innate inflammatory pathways, including Toll-like receptor and NF-κB signaling. JAK–STAT signaling was enriched specifically in cluster 2.

Single-cell immunologic dynamics in CSF after QH104 infusion

Given CAR γδ T-cell persistence and associated immune changes in CSF, we performed scRNA-seq on the IP, PB, and CSF samples and scV(D)J-seq on PB and CSF samples from all 3 patients. PB samples were collected on day 0, and CSF samples were collected on days 0, 3, 7, and 14. The baseline CSF sample from patient 03 was excluded because of insufficient cells. In total, 17 samples were profiled (Fig. 3A).

Figure 3.

Single-cell profiling of CSF immune-cell composition and CAR γδ T-cell dynamics after QH104 infusion. A, Sampling and sequencing scheme for the IP, PB, and CSF samples. Day 0 denotes the sample collected before infusion. B, Uniform Manifold Approximation and Projection (UMAP) plot showing the distribution of annotated immune-cell subsets in the IP, PB, and CSF samples. C, Longitudinal changes in CSF immune-cell composition. Marker genes used for cell subset annotation are listed in Supplementary Table S2. D, UMAP plots showing the distribution of all γδ T cells (left), CAR γδ T cells (top right), and non-CAR γδ T cells (bottom right). E, Volcano plot (left) showing differentially expressed genes between CAR and non-CAR γδ T cells, with representative genes labeled, and bar chart (right) showing the corresponding enrichment analysis. F, Dot plot showing marker gene expression and signature scores across γδ T-cell subsets. G, Longitudinal changes in γδ T-cell composition. Numbers above the bars indicate the number of γδ T cells captured at each time point. H, Mfuzz clustering analysis of CAR γδ T cells. Heatmap (left) showing gene expression in each cluster, with representative genes labeled, and bar chart (right) showing the corresponding enrichment analysis. cDC1, conventional type 1 dendritic cell; cDC2, conventional type 2 dendritic cell; cMono, classical monocyte; migDC, migratory dendritic cell; ncMono, non-classical monocyte; pDC, plasmacytoid dendritic cell; Tn, naïve T cell.

After quality control and integration, 124,596 cells were retained for downstream analyses. Unsupervised clustering identified 34 immune subsets, including 24 lymphoid and 10 myeloid subsets (Fig. 3B; Supplementary Fig. S5A). Lymphoid subsets included 9 γδ T-cell subsets spanning CAR and non-CAR compartments; 5 CD4⁺ T-cell subsets corresponding to naïve, central memory, effector (GZMK and GNLY), and regulatory states; 5 CD8⁺ T-cell subsets corresponding to naïve, central memory, effector (GZMK and GNLY), and exhausted states; 2 NK-cell subsets (CD56^dim and CD56^bright); 2 B-cell subsets (naïve and memory); and plasmablasts/plasma cells. Myeloid subsets included classical (CD14⁺) and nonclassical (CD16⁺) monocytes; 4 dendritic-cell (DC) subsets including conventional type 1 dendritic cells (cDC1), conventional type 2 dendritic cells (cDC2), migratory dendritic cells (migDC), and plasmacytoid dendritic cells (pDC); and 4 macrophage subsets including Mφ-FOLR2, Mφ-SPP1, Mφ-CXCL10, and microglia. Paired TCR sequences were identified in 48,768 of 56,158 αβ T cells (86.8%; Supplementary Fig. S5B). Consistent with flow cytometry, scRNA-seq captured a transient increase in CSF immune cells after infusion, with accompanying changes in immune subset composition. These changes were pronounced in CAR γδ T cells and in selected macrophage and CD8⁺ T-cell subsets, including Mφ-CXCL10 and CD8⁺ exhausted (Tex) cells (Fig. 3C; Supplementary Fig. S5C).

To further characterize γδ T-cell dynamics, we analyzed 46,700 γδ T cells from the IP and CSF samples and identified 5 subsets, namely, cycling (MKI67), biosynthetic (MYC), cytotoxic (GNLY), regulatory-like (FOXP3), and memory-like (IL7R) subsets (Fig. 3D; Supplementary Fig. S6A and S6B). Based on transgene expression, γδ T cells were classified as CAR γδ T cells (n = 10,879) or non-CAR γδ T cells (n = 35,821), with the memory-like subset detected exclusively in the non-CAR compartment. CAR γδ T cells were enriched for pathways related to proliferation, migration, IFNγ production, and cytotoxic activity, whereas non-CAR γδ T cells were enriched for antigen processing and presentation (Fig. 3E). RPS4Y1 was broadly expressed across most γδ T-cell subsets, consistent with donor origin, whereas XIST expression was largely restricted to the memory-like non-CAR subset (Fig. 3F). In CSF, CAR γδ T cells were detected on days 3 and 7 but were rare or absent on day 14. The IP was dominated by the cycling and biosynthetic subsets, whereas postinfusion CAR γδ T cells were largely represented by the cytotoxic subset (Fig. 3G; Supplementary Fig. S6C and S6D). Consistently, Mfuzz clustering of pseudobulk scRNA-seq profiles of CAR γδ T cells showed a transition from a hypoxia- and biosynthesis-associated state in the IP to an IFNγ-associated state on day 3 and a cytotoxic state on day 7 (Fig. 3H).

Myeloid cells represented a prominent innate immune compartment in CSF during LM (20, 36, 37). DCs were captured at low frequency and were not analyzed further. Monocyte and macrophage subsets are shown in Fig. 4A. At baseline, macrophages were dominated by M2-like, CNS border-associated macrophage–like (BAM-like) subsets, including Mφ-FOLR2, Mφ-SPP1, and microglia. After infusion, Mφ-CXCL10 accumulated and became the dominant macrophage subset. Compared with other macrophage subsets, Mφ-CXCL10 showed an M1-like signature, with higher expression of CXCL10, CCL2, CCR2, and ISG15, and enrichment of pathways associated with IFNγ response, antigen processing and presentation, and leukocyte migration (Fig. 4B and C).

Figure 4.

B7-H3–CAR γδ T cells remodel the CSF immune microenvironment, with IFNγ-associated inflammatory macrophages and subsequent lymphocyte recruitment. A, Uniform Manifold Approximation and Projection (UMAP) plot showing the distribution of monocyte and macrophage subsets. B, Dot plot showing marker gene expression and signature scores across monocyte and macrophage subsets. Tcm, central memory T. C, Volcano plot (left) showing differentially expressed genes between Mφ-CXCL10 and other macrophage subsets, with representative genes labeled, and bar chart (right) showing the corresponding enrichment analysis. D, UMAP plot showing the distribution of αβ T-cell subsets. E, Longitudinal changes in clonal expansion and clonotype overlap among αβ T cells, CD4⁺ conventional T cells (Tconv), CD4⁺ Tregs, and CD8⁺ T cells. Overlaid line plots indicate the proportion of PB–CSF shared clonotype cells among PB αβ T cells. F, UMAP plots showing αβ T cells colored by clonal expansion and clonotype overlap, with representative subsets outlined (CD8⁺ Teff-GZMK, CD8⁺ Teff-GNLY, CD8⁺ Tex, and CD4⁺ Teff-GNLY). G, Subset composition of PB–CSF shared clonotypes in CSF (top), CSF-preexisting clonotypes in CSF (top middle), clonotypes of undetermined origin in CSF (bottom middle), and expanded CSF-preexisting clonotypes on day 14 (bottom). H, Chord diagram showing inferred IFNG–(IFNGR1 + IFNGR2) interactions from γδ T cells and αβ T cells to monocytes and macrophages. I, Monocle2 trajectory analysis of monocytes and macrophages (left) and IFNγ response scores along pseudotime for each trajectory branch (top right). J, Chord diagram showing inferred CCL2–CCR2 interactions within monocyte and macrophage subsets (top) and CXCL10–CXCR3 interactions from monocyte and macrophage subsets to αβ T-cell and NK-cell subsets (bottom). cMono, classical monocyte; ncMono, nonclassical monocyte; Tn, naïve T cell.

We then assessed αβ T-cell clonotypes by scV(D)J-seq, with αβ T-cell subsets shown in Fig. 4D. After infusion, PB–CSF shared clonotypes peaked on day 3 and then declined. In contrast, clonotypes of undetermined origin accounted for more than half of αβ T cells after infusion, indicating that the postinfusion increase in CSF αβ T cells was not explained solely by overlap with PB. PB–CSF shared clonotypes were observed mainly in CD8⁺ T cells, limited in CD4⁺ conventional T cells, and largely absent from CD4⁺ regulatory T cells (Treg), which were predominantly of undetermined origin (Fig. 4E; Supplementary Fig. S7A). In PB, PB–CSF shared clonotypes mapped mainly to effector states, including CD8⁺ effector T (Teff)–GZMK, CD8⁺ Teff-GNLY, and CD4⁺ Teff-GNLY. In CSF, they were predominantly CD4⁺ Teff-GNLY at baseline, whereas CD8⁺ Teff-GZMK and CD8⁺ Teff-GNLY also became prominent after infusion. By contrast, CSF-restricted clonotypes mapped mainly to CD4⁺ central memory T, CD4⁺ Treg, and CD8⁺ Tex (Fig. 4F and G; Supplementary Fig. S7B). On day 14, expanded CSF-preexisting clonotypes became prominent and mapped mainly to CD8⁺ Teff-GZMK and CD8⁺ Tex (Fig. 4G; Supplementary Fig. S7B).

To place these changes in a signaling context, we used CellChat to infer candidate ligand–receptor interactions (Supplementary Fig. S7C–S7E). Predicted outgoing signals from γδ T cells included IFNγ, CCL, and MIF, with macrophages and monocytes as the major recipients. These myeloid cells were in turn predicted to communicate with αβ T cells and NK cells through CXCL, CCL, and MIF. Infused γδ T cells, particularly CAR γδ T cells, were predicted to be the main sources of IFNγ signaling, consistent with their transcriptional profiles (Fig. 4H). Trajectory analysis rooted in classical monocyte showed increasing IFNγ response scores along the Mφ-CXCL10 branch relative to the alternative branch (Fig. 4I). In addition, Mφ-CXCL10 was predicted to be the major source of CCL2–CCR2 signaling to monocytes and of CXCL10–CXCR3 signaling to αβ T cells and NK cells (Fig. 4J). Together, these results suggest an IFNγ-associated signaling axis linking CAR γδ T cells, Mφ-CXCL10 accumulation, and downstream chemokine signaling that may promote lymphocyte recruitment in CSF.

Discussion

Here, we report the initial clinical and correlative observations from a phase I study of intrathecal QH104, an allogeneic B7-H3–targeted CAR γδ T-cell therapy, in patients with LM from solid tumors. Intrathecal QH104 administration was feasible and generally well tolerated, with no grade ≥ 4 TRAE. These safety observations are consistent with prior experiences with intraventricular immunotherapy in LM, including recombinant IL2 and lymphokine-activated killer cell infusions (38–40), and support the tolerability of intrathecal administration in this setting.

A key question for intrathecal cellular immunotherapy in LM is whether infused cells can persist and remain active in CSF despite rapid CSF turnover and an immunosuppressive microenvironment (20). In our cohort, several observations were consistent with bioactivity in CSF. All patients met the RANO-LM criteria for SD on days 14 and 30, with symptom improvement in 2 patients and clearance of CSF cytology in 1 patient. In parallel, CAR γδ T-cell persistence in CSF was demonstrated for at least 1 week by concordant findings from flow cytometry and a validated qPCR assay, with minimal evidence of systemic dissemination. Single-cell profiling indicated a transition from proliferative and biosynthetic states in the IP toward a cytotoxic state in CSF after infusion, consistent with in situ functional adaptation rather than passive carryover of an ex vivo phenotype. Sustained IFNG expression in CAR γδ T cells, together with increased IFNγ levels in CSF, supports IFNγ as a persistent feature of the intrathecal CAR γδ T-cell response.

Our data further suggest that this local response was accompanied by remodeling of the CSF macrophage compartment. At baseline, CSF macrophages were predominantly BAM-like and showed immunoregulatory features, including high IL10 expression, consistent with prior evidence that BAM can suppress antitumor immunity through IL10 in LM (25). After infusion, this compartment became dominated by Mφ-CXCL10, an inflammatory macrophage state with a prominent IFNγ response signature. Inferred ligand–receptor interactions identified CAR γδ T cells as a potential upstream source of IFNγ, consistent with their contribution to Mφ-CXCL10 accumulation through IFNγ signaling. Inferred CCL2–CCR2 signaling from Mφ-CXCL10 to monocytes, together with higher CCR2 expression in Mφ-CXCL10 than in baseline macrophages, further suggests a feed-forward circuit in which Mφ-CXCL10 may promote monocyte influx and help sustain this inflammatory macrophage state in an IFNγ-conditioned microenvironment. In parallel, inferred CXCL10–CXCR3 signaling provides a plausible explanation for the postinfusion increase in CSF αβ T cells observed by flow cytometry. Together, these findings suggest that, beyond direct cytotoxicity, CAR γδ T cells may help establish an IFNγ-conditioned inflammatory microenvironment in CSF. This environment may promote monocyte influx and subsequent lymphocyte recruitment, thereby supporting local antitumor immunity.

TCR clonotype analysis indicates that the postinfusion increase in CSF αβ T cells was not explained solely by peripheral trafficking but likely also involved CNS-resident T cells (41, 42). Clonotypes shared between PB and CSF mapped mainly to effector states, whereas clonotypes detected only in CSF were predominantly central memory, regulatory, and exhausted states. These observations suggest that the contribution of lymphocyte recruitment to clinical response may depend not only on its magnitude but also on the origin and state of the recruited cells. Given the small cohort, however, this inference remains descriptive and should be interpreted cautiously.

Attenuation of clinical benefit after a second infusion raises the possibility of adaptive resistance in CSF. Although an IFNγ-dominant response may initially enhance antigen presentation and immune-cell recruitment, it may also induce counter-regulatory programs in tumor and myeloid cells (43–45). Consistent with this possibility, Mφ-CXCL10 showed increased CD274 and IDO1 expression, and exhausted CD8⁺ T cells accumulated over time, indicating a shift toward a more constrained effector microenvironment. These observations suggest that future studies should refine dose and schedule to preserve early pro-inflammatory activity while limiting compensatory immunosuppression. They also provide a rationale for cautious exploration of selected intrathecal combination strategies, including ICIs, given the unique safety considerations of CNS-directed treatment.

This study has several limitations. First, B7-H3 status was assessed in archived primary tumor tissue rather than in leptomeningeal lesions because LM tissue is rarely available and CSF cytology specimens usually yield insufficient cells for target assessment. As a result, target status may not have been fully captured, owing to discordance between primary and metastatic sites and dynamic antigen modulation during treatment (46–48). Development of CSF-based assays for longitudinal assessment of B7-H3 expression in LM may improve biomarker-guided patient selection in future studies. Second, response assessment in LM remains challenging. Current criteria, including neurologic examination, neuroimaging, and CSF cytology, may not sensitively reflect changes in tumor burden. CSF circulating tumor DNA (ctDNA) monitoring, which has shown improved sensitivity for LM diagnosis and longitudinal tracking (49, 50), may complement existing response assessments by providing a more sensitive and quantitative measure of disease dynamics.

In summary, this phase I study provides preliminary proof of concept for intrathecal allogeneic B7-H3–targeted CAR γδ T-cell therapy in LM, with early evidence of safety, clinical activity, local persistence, and IFNγ-associated immune remodeling in CSF, including expansion of an inflammatory macrophage state and subsequent lymphocyte recruitment (Supplementary Fig. S8). Although the cohort was small and mechanistic inferences remain preliminary, these findings support further evaluation in larger studies with longer follow-up. Future work should refine response assessment, identify response-associated biomarkers, and optimize strategies to prolong cell persistence and sustain antitumor activity in LM.

Supplementary Material

Supplementary Table S1

Release testing criteria and results for the three QH104 product batches

Supplementary Table S2

Marker genes used for cell subset annotation

Supplementary Table S3

Gene sets used for signature score calculation

Supplementary Table S4

Representativeness of study participants

Supplementary Table S5

Longitudinal MD Anderson Symptom Inventory–Brain Tumor symptom severity and interference scores after QH104 infusion

Supplementary Table S6

Longitudinal cerebrospinal fluid CAR transgene vector copy number, immune cell counts, and cytokine levels after QH104 infusion

Supplementary Table S7

Longitudinal peripheral blood CAR transgene vector copy number, immune cell counts, and cytokine levels after QH104 infusion

Supplementary Figure S1

Flow cytometry gating and in vitro functional characterization of QH104

Supplementary Figure S2

CONSORT flow diagram

Supplementary Figure S3

Additional serial MRI scans

Supplementary Figure S4

Peripheral blood CAR transgene and cytokine dynamics and additional Mfuzz clusters of cerebrospinal fluid protein levels after QH104 infusion

Supplementary Figure S5

Single-cell annotation, clonality, and longitudinal changes in cerebrospinal fluid immune-cell composition after QH104 infusion

Supplementary Figure S6

Additional characterization of γδ T-cell states

Supplementary Figure S7

Additional clonotype analyses and inferred cell-cell communication

Supplementary Figure S8

Schematic summary of the proposed safety, efficacy, PK/PD, and immunological dynamics after QH104 infusion

Supplementary Data S1

Protocol

Supplementary Data S2

Statistical analysis plan

Supplementary Data S3

CONSORT-DEFINE checklist

Data Availability

The single-cell datasets generated in this study are publicly available in the Genome Sequence Archive for Human (GSA-Human) at HRA016980. All other data are available within the article and its Supplementary Data or from the corresponding author (N. Li) upon reasonable request.

Authors’ Disclosures

Y. Zhou reports a grant from the Beijing College Student Innovation and Entrepreneurship Training Program during the conduct of the study. S. Wang reports a grant from the National Natural Science Foundation of China during the conduct of the study. N. Li reports grants from the Chinese Academy of Medical Sciences Innovation Fund for Medical Sciences, the National Natural Science Foundation of China, the National Key Research and Development Program of China, and the National High Level Hospital Clinical Research Funding during the conduct of the study. No disclosures were reported by the other authors.

Authors’ Contributions

P. Ma: Data curation, formal analysis, investigation, methodology, writing–original draft, project administration, writing–review and editing. Y. Zhou: Data curation, formal analysis, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. W. Ma: Resources, investigation, methodology, project administration. Y. Wang: Formal analysis, project administration, writing–review and editing. H. Cai: Project administration. Y. Sui: Investigation, methodology, writing–original draft, project administration. L. Zhao: Investigation, writing–original draft, project administration. Y. Zhang: Conceptualization, resources, supervision. S. Wang: Conceptualization, supervision, funding acquisition. N. Li: Conceptualization, supervision, funding acquisition.