Regulatory Programs of B-cell Activation and Germinal Center Reaction Allow B-ALL Escape from CD19 CAR T-cell Therapy

The application of CD19 chimeric antigen receptor (CAR) T-cell therapies has shown remarkable success in treating relapsed CD19⁺ B-cell acute lymphoblastic leukemia (B-ALL; refs. 1, 2). Up to 90% of the heavily pretreated patients who receive this therapy enter into remission (2–6). However, a large fraction of patients relapse (3–6). Loss of CD19 expression via mutations (7, 8), lineage switch (9, 10), and loss of the targeted epitope through alternative splicing (11) are causes of resistance to CD19 CAR T-cell therapy and relapse. Additional studies have shown that intercellular processes, such as trogocytosis whereby membrane fragments transfer between cells, can result in emergence of CD19^low B-ALL cells (12) and low expression of the target antigen can lead to resistance (13) and less effective killing (12, 14). Antigen density thresholds, however, are difficult to define and depend on CAR constructs (14), tumor type (15), and target antigens (13, 16). Other resistance mechanisms that have been described include CD19⁺ resistance through antigen masking (17) and impaired death receptor signaling (18).

CD19 is a coreceptor of the B-cell receptor (BCR) complex and is expressed throughout the B-cell lineage until the plasma-cell stage. CD19 clusters upon antigen binding (19) and forms BCR–antigen complexes, which are internalized and processed thereafter, thus contributing to B-cell activation (19). Although CD19 targeting by and resistance to CAR T cells has been well established (7, 8, 11), the dynamics of CD19 antigen expression and its regulation during the early phases of interactions between CAR T cells and CD19⁺ target cells remain unclear. These early dynamics may have important implications for initial treatment failures.

In this study, we hypothesized that the early interactions of CAR T cells and CD19⁺ target cells trigger discrete regulatory programs that allow for escape of some target cells. By combining live single-cell microscopy with single-cell RNA sequencing (scRNA-seq) and single-cell assay for transposase-accessible chromatin using sequencing (scATAC-seq), we show that the interaction between CD19 CAR T cells and CD19-expressing B-ALL cells caused clustering and internalization of CD19 in all leukemia cells that resulted in reduced surface expression of CD19. We found that a subpopulation of target cells employed physiologic B-cell activation programs reminiscent of germinal center reactions in normal B cells that caused sustained transcriptional downregulation of CD19. Initiation of these programs was inhibited by pretreating leukemic cells with the Bruton's tyrosine kinase (BTK) inhibitor ibrutinib, providing a mechanistic explanation for how ibrutinib increases effectiveness of CD19 CAR T-cell cytotoxicity.

293T and Jurkat cell lines were obtained from ATCC in 2016 and 2015, respectively, NALM-6, REH, and KOPN-8 cell lines were obtained from the Broad Institute's Cancer Cell Line Encyclopedia in 2016. ALL cell lines were cultured in RPMI1640 (Thermo Fisher Scientific, 11875119) supplemented with 10% heat-inactivated fetal bovine serum (FBS) (Gemini, 900–208) and 100 μmol/L penicillin/streptomycin (Invitrogen, 15140122). 293T cells were cultured in DMEM (Life Technologies, 11995073) supplemented with 10% FBS and 100 μmol/L penicillin/streptomycin. Cell line authentication was performed by short tandem repeat profiling and reauthentication was performed within the past year. Cultures were tested for Mycoplasma contamination and found to be negative. Cells were used for experiments before 20 passages.

Ibrutinib (Selleckchem, S2680) was dissolved in DMSO (Thermo Fisher, BP231100) and titrated. IC₁₀ concentration for NALM-6 cells was determined after 72 hours using CellTiter-Glo (Promega, G1781) on a Spectramax M5 microplate reader (Molecular devices).

Two third-generation CD19 CAR constructs containing a single chain variable fragment targeting the FMC63 locus of the CD19 antigen with an intracellular CD3ζ as well as 4–1BB and CD28 costimulatory domains were used. Constructs contained a truncated epidermal growth factor receptor (tEGFR), separated from the CAR by a T2A sequence. Construct 1 was obtained from Creative Biolabs (pCAR-T-h; anti-CD19 scFv–CD28–4–1BB–CD3ζ) and consisted of an anti-CD19 single-chain variable fragment, CD8 hinge, CD28 transmembrane domain, 4–1BB/CD3 zeta signaling module, T2A sequence, and a tEGFR. Construct 2 was designed in-house, custom-ordered at Genewiz and cloned in a pLVX-CMV100 vector backbone (Addgene, 110718). The components of construct 2 are displayed in Supplementary Fig. S1A. Construct 1 was used in the experiments unless stated differently. 293T cells were cotransfected with the CAR lentiviral construct, psPAX2 (Addgene, 12260) and pCMV-VSV-G (Addgene, 8454) packaging vectors using Lipofectamine 3000 (Thermo Fisher Scientific, L3000015) according to the manufacturer's protocol. Lentivirus-containing medium was collected, and fresh medium added after 12 hours, 24 hours, and 36 hours. Virus-containing media was filtered and concentrated by ultracentrifugation for 2 hours at 49,000 g at 4°C.

Human blood from healthy donors was obtained from Research Blood Components, LLC, or the Crimson Core of the Brigham and Women's Hospital. Mononuclear cells (PBMCs) were isolated by Ficoll-Paque PLUS (Global Life Sciences Solutions USA LLC, 45001750). PBMCs were further processed by isolating CD3⁺ T cells with the EasySep Human T Cell Enrichment Kit (STEMCELL Technologies, 19051) with an EasySep Magnet (STEMCELL Technologies, 18000) according to the manufacturer's protocol. For CD4⁺ T-cell purification, selection was performed using EasySep Release Human CD4 Positive Selection Kit (STEMCELL Technologies, 17752) according to the manufacturer's protocol.

Isolated T cells were activated by Dynabeads Human T-Activator CD3/CD28 (Thermo Fisher Scientific, 11132D) and cultured in X-VIVO 15 Media (Lonza, 04–418Q) supplemented with 5% human serum (Sigma-Aldrich, H4522). 50 IU/mL IL2 (Miltenyi Biotec, 130–097–743) was added every other day. One day after isolation, T cells or Jurkat cells were infected by spinoculation at multiplicity of infection (MOI) = 5. After 7 days, infection efficiencies were determined by flow cytometry as described below using a hEGFR-specific antibody (Biotinylated, Cetuximab; R&D Systems, FAB9577B-100) and a secondary allophycocyanin (APC) streptavidin-conjugated antibody (BD Biosciences, 554067). CAR-expressing cells were isolated by magnetic isolation using the EasySep Release Human Biotin Positive Selection Kit (STEMCELL Technologies, 17653) according to the manufacturer's protocol. Persistent expression of the CAR construct (detected using biotinylated anti-hEGFR – Cetuximab, R&D Systems; and secondary APC Streptavidin-conjugated antibody; BD Biosciences) and CD8 percentage (HIT8a, BioLegend, 300912) was assessed by flow cytometry prior to coculture (Supplementary Fig. S1B). Activation beads were removed after 10 days with restimulations according to the manufacturer's protocol. Uninfected T cells from the same donor or uninfected Jurkat cells were maintained in parallel and used as controls.

CAR T cells were stained for CD3 phycoerythrin (PE; HIT3a, BioLegend, 300308), CD8 APC (HIT8a, BioLegend, 300912) and hEGFR (Biotinylated, Cetuximab; R&D Systems, FAB9577B-100) using a secondary APC Streptavidin-conjugated antibody (BD Biosciences, 554067). B-ALL target cells were stained for CD19 PE (HIB19, BioLegend, 302208), CD40 PE (5C3, BioLegend, 334307) or, where applicable, a primary unconjugated CD19 (HIB19, BioLegend, 302202) and a secondary PE-conjugated goat anti-mouse IgG (Poly4053, BioLegend, 405307). Fixation and permeabilization for intracellular staining were performed using the BD Cytofix/Cytoperm Kit (BD Biosciences, 555028) according to the manufacturer's protocol. Cells were stained with 7-AAD (Thermo Fisher Scientific, A1310) to exclude dead cells. Cell nuclei were stained with DAPI (1 μg/mL, Sigma-Aldrich, D9542). B-ALL target cells were labeled with carboxyfluorescein succinimidyl ester (CFSE; Thermo Fisher Scientific, C34554), as indicated.

Samples were analyzed on a BD Accuri C6 Flow Cytometer (BD Biosciences) or BD FACSMelody (BD Biosciences) with compensation being performed by AbC Total Antibody Compensation Bead Kit (Thermo Fisher Scientific, A10497) or single-stained controls. Where applicable, absolute cell counts were measured with Precision Count Beads (BioLegend, 424902).

Flow cytometric analyses were performed with FlowJo V10 (BD Biosciences). CD19 fluorescence signal was normalized to the mean fluorescence intensity of CD19 in control leukemia cells that were not cocultured with any effector cells.

For scRNA-seq and scATAC-seq experiments, NALM-6 cells were stained for CD19 and the 30% highest CD19-expressing cells (CD19^high cells) were sorted and rested overnight. The next day, sorted CD19^high cells were labeled with CFSE (Thermo Fisher Scientific, C34554), counted and cocultured at 1×10⁵ target cells with CAR T cells or uninfected T cells in a 5E:1T ratio for 24 hours in 96-well round bottom plates. After coculture, cells were stained with 7-AAD (Thermo Fisher Scientific, A1310) and anti-CD19 PE (HIB19, BioLegend, 302208) antibody. Subsequently, 7-AAD–negative and CFSE⁺ NALM-6 cells were sorted into 96-well plates (Thermo Fisher, E0030129504) containing TCL buffer (Qiagen, 1031576) and 1% 2-Mercaptoethanol (Life Technologies, 21985023) based on CD19 expression (Supplementary Table S1) using a Sony SH800 sorter. Plates were spun down immediately, flash frozen on dry ice and stored at –80°C until further processing.

NALM-6, REH and KOPN-8 cells were labeled with CFSE (Thermo Fisher Scientific, C34554) and used as target cells (Target, T). 1×10⁵ target cells were cocultured with CAR T cells or CAR Jurkat cells (Effector, E) at an E:T ratio of 1:1, 2:1, and 5:1 in a 96-well round bottom plate for 1, 2, 4, 6, 12, and 24 hours in X-VIVO 15 media supplemented with 5% human serum. After coculture, cells were stained with 7-AAD (Thermo Fisher Scientific, A1310) and an anti-CD19 PE (HIB19, BioLegend, 302208) and analyzed on a BD Accuri C6 Flow Cytometer (BD Biosciences) or BD FACSMelody (BD Biosciences) with compensation being performed by AbC Total Antibody Compensation Bead Kit (Thermo Fisher Scientific, A10497) or single-stained controls. Cytotoxicity was calculated as number of 7-AAD–positive (dead) target cells divided by total number of target cells.

Where applicable, 7-AAD–negative target cells were isolated from effector cells after 24 hours of coculture using a Sony SH800 sorter. Sorted target cells were taken back into culture and viability was assessed by trypan blue staining using a Neubauer cell counting chamber (Fisher Scientific, 026716) after 24 hours, 48 hours, and 72 hours.

For reexposure experiments, NALM-6 cells were cocultured with CAR T cells or uninfected T cells at an E:T ratio of 2:1 in a 96-well round bottom plate overnight. Cells were then pooled and NALM-6 cells were isolated from effector cells using the EasySep Human CD3-Positive Selection Kit II (STEMCELL Technologies, 17851) according to the manufacturer's protocol. Successful isolation was assessed by positivity of CFSE-labeling via flow cytometry. NALM-6 cells were subsequently reexposed to fresh effector cells at an E:T ratio of 2:1 in a 96-well round bottom plate for 6 hours. After coculture, cells were stained with 7-AAD and analyzed as described above.

For transwell experiments, cytotoxicity assays were performed in a Corning HTS Transwell-96 system with 0.4 μm pore size (Sigma-Aldrich, CLS3381). For ibrutinib experiments, NALM-6 cells were pretreated with ibrutinib for 72 hours at IC₁₀ dosing. To exclude dead cells, NALM-6 cells were then ficolled, washed, counted, and cocultured with effector cells either in presence of DMSO (Thermo Fisher, BP231100) or ibrutinib at IC₁₀ dosing for 6 hours. After coculture, cells were stained with 7-AAD and analyzed as described above.

Live cell imaging was performed in X-VIVO 15 media without phenol red (Lonza, 04–744Q) supplemented with 5% human serum. NALM-6 cells were labeled with CFSE or an AF647-conjugated anti-CD19 (HIB19, BioLegend, 302220). A stage top incubator was used to maintain constant humidified O₂ and CO₂ flow at 37°C (Okolab). 1×10⁶ cells were seeded on a Petri dish and allowed to settle for at least 30 minutes before starting the timelapse. CAR T cells were carefully added at a 1:1 ratio. Where applicable, SYTOX Blue Dead Cell Stain (Thermo Fisher Scientific, S11348) was added to the media at 1 mmol/L.

For holotomography-based three-dimensional live microscopy, interactions were recorded by measurement of refractive index and CFSE fluorescence signal using a 3D Cell Explorer microscopy system with a 60× magnifying objective at 512 × 512 resolution (Nanolive). Images were further processed with Nanolive's software STEVE v1.6.3496 to display three-dimensional timelapses.

For CD19 antigen movement tracking, a Nikon Eclipse Ti microscope system recorded interaction sites every 10 minutes for 6 hours with a 20× magnifying objective at 2048 × 2048 resolution. Z-stack images were recorded focusing on the middle layer of the cells as well as 2 μm above and below using Nikon's Perfect Focus System. Four-dimensional hyperstacks (x-, y-, z-, and time axis) were assembled using the Fiji distribution of ImageJ 1.52p. For representative still images, effector cells were pseudo-colored in orange. Cells were first analyzed qualitatively with respect to CD19 clustering and internalization. Interactions were characterized as “cluster formation,” when CD19 clustering was observed in at least two sequential frames or if CD19 clustering was followed by CD19 internalization in two sequential frames. Interactions that did not meet these criteria but lasted for at least two sequential frames were characterized as “no cluster.” Dead cells were determined by positivity of SYTOX Blue Stain. Analysis only included NALM-6 cells that were SYTOX Blue–negative at time of analysis unless stated differently. Mean fluorescence intensity of the three Z-stack images was then quantified in the Fiji distribution of ImageJ 1.52p. Photobleaching was accounted for by normalizing to the mean loss of fluorescence after 6 hours in NALM-6 cells that did not interact with any effector cells in the recorded frames during the same experiment. Analysis was performed in a blinded manner.

Full-length scRNA-seq libraries were prepared using the SMART-seq2 protocol (20). As described above, cells were sorted into 96-well plates containing 10 μL TCL buffer (Qiagen, 1031576) and 1% 2-mercaptoethanol. RNA was purified using RNAClean XP beads (Beckman Coulter, A63987) according to the manufacturer's protocol and reverse transcribed using Maxima RNAse H-minus (Thermo Fisher Scientific, FEREP0752) in the presence of oligo-dT30VN (Integrated DNA Technologies, (/5BiosG/AAGCAGTGGTATCAACGCAGAGTACT(30)VN), template-switching oligonucleotides (Qiagen, /5BiosG/AAGCAGTGGTATCAACGCAGAGTACATrGrG+G), and betaine (Sigma-Aldrich, B0300–5VL). cDNA was then amplified using the KAPA Hifi Hotstart ReadyMix (Kapa Biosystems, KK2602) with ISPCR primers (Integrated DNA Technologies, (/5BiosG/AAGCAGTGGTATCAACGCAGAGT). After purification with Agencourt Ampure XP beads (Beckmann Coulter, A63881), product size distribution was assessed on a Bioanalyzer using a High Sensitivity DNA Kit (Agilent Technologies, 5067–4626) and quantified by Qubit (Thermo Fisher Scientific, Q32854), followed by fragmentation and indexing using Nextera XT, and Nextera PCR primers (Illumina, FC-131–1096). Libraries were pooled and paired-end sequencing was performed on a NextSeq 500 (Illumina) aiming for sequencing depth of 1 million reads per single cell.

Single-cell ATAC-seq libraries were prepared following the protocol by Chen and colleagues (21). NALM-6 cells were presorted, cocultured, and processed as described above (see Flow cytometry and single-cell sorting), followed by DAPI-staining (1 μg/mL, Sigma-Aldrich, D9542) and sorting into 96-well plates containing lysis buffer [recipe as described by Chen and colleagues (21): 2X lysis buffer - 100 mmol/L Tris.HCl, pH 8.0 (Boston BioProducts, BM-320); 100 mmol/L NaCl (Sigma-Aldrich, 71386); 40 μg/mL Proteinase K (Ambion, AM2546); 0.4% SDS (Boston BioProducts, BM-230A)] and Nextera XT indexes (Illumina, FC-131–1096). Libraries were pooled and purified using a Qiagen MinElute PCR purification kit (Qiagen, 28004) followed by size selection using Agencourt Ampure XP beads, (Beckman Coulter, A63881). Paired-end sequencing was performed on a NextSeq 500 aiming for sequencing depth of 1 million reads per single cell.

Statistical analyses were performed using GraphPad Prism 9. Normal distribution was tested using the D'Agostino and Pearson test or the Shapiro–Wilk test unless otherwise indicated. For parametric data, the two-tailed unpaired Student t test was used, or an ordinary one-way ANOVA with Tukey multiple comparisons in data with more than two groups. For nonparametric data, the two-tailed Mann Whitney test was used, or Kruskal–Wallis test with Dunn multiple comparisons testing in data with more than two groups.

Raw sequencing reads were trimmed using trimmomatic and aligned to the hg19 genome using STAR (22) with the following parameters ‘– twopassMode Basic –alignIntronMax 100000 –alignMatesGapMax 100000 –alignSJDBoverhangMin 10 –alignSJstitchMismatchNmax 5 -1 5 5. HTSeq (23) and RSEM (24) were used to calculate raw counts and normalized transcript per kilobase million (TPM) values from the aligned bam files.

Four different parameters were used to filter out low quality cells – library size, number of detected genes, percentage of reads mapping to mitochondrial genes, and percentage of reads mapping to house-keeping genes. Cells were distributed according to these parameters and determined as low quality when falling beyond three median absolute deviations.

Clustering of high-quality cells was performed using PAGODA2 (ref. 25; https://github.com/kharchenkolab/pagoda2). Due to batch effects, one cluster was uniquely formed by cells from one well plate, and thus excluded from further analysis. Seven clusters were identified using multilevel graph-based clustering within PAGODA2.

Each individual cell was analyzed for expression of G₁, G₂–M, and S phase markers to predict the cell-cycle phase using Seurat2 (ref. 26; https://github.com/satijalab/seurat). The findMarkers function in the scran package (https://github.com/elswob/SCRAN) was used to identify differentially regulated genes in each of these clusters. The mean log₂FC of differentially regulated genes in each cluster compared with all other clusters was used to perform gene set enrichment analysis (GSEA) on all hallmark gene sets. Monocle2 (refs. 27, 28; http://cole-trapnell-lab.github.io/monocle-release/) was employed to discover distinct transcriptional states within the NALM-6 cells cocultured with CAR T cells. Transcriptional regulons associated with each Monocle2 state were determined using SCENIC (ref. 29; https://github.com/aertslab/SCENIC).

Signatures for hematopoietic cell types were derived from previously published literature comprising bulk RNA-seq studies on purely sorted immune populations (30, 31). The preB signature was taken from the Human Cell Atlas (32). The germinal center signatures were derived from Holmes and colleagues (33). The genes in each of these signatures were scored using the AUCell package (https://github.com/aertslab/AUCell). The mean Z-scores of a cell type of each state were chosen for further visualization.

Gene expression data from bulk-sorted populations of human naïve B cells and germinal center B cells from Holmes and colleagues (33) were analyzed. All three datasets for the naïve B cells and for the germinal center B cells from three independent donors were used. The gene expression count matrix (TPM) was analyzed using R version 4.0.1 and the ggplot2 package (https://github.com/tidyverse/ggplot2).

The reads from scATAC-seq samples were aggregated by each 96-well plate prior to peak calling and analyzed using the nf-core ATAC pipeline (34) with the ‘–narrowPeak’ parameter, employing MACS2 for peak calling and HOMER for annotation of peaks. In total, the union of all peak calls resulted in 33,125 consensus peaks. For gene-enhancer relationships, annotations were derived from GREAT (35). Raw fastq files from individual scATAC-seq samples were processed using the snapATAC pipeline (https://github.com/r3fang/SnapATAC). Briefly, single-cell bam files were processed to obtain read counts in the consensus peaks. Peaks were resized to a fixed width of 500 bp, centered at the summit and overlapping peaks were removed. Filtering was performed using the parameters ‘min_depth = 1,500, min_in_peaks = 0.15'. Clustering of cells was performed using cisTopic (36) (https://github.com/aertslab/cisTopic) and densityClust (37) and visualization of clusters using uniform manifold approximation and projection (UMAP). Differential motif accessibility was investigated using ‘chromVAR’ (38) (https://github.com/GreenleafLab/chromVAR). Motifs were imported from the JASPAR database and differential accessibility was determined using the ‘differentialDeviations’ function. Bias-corrected deviation and mean Z-scores were calculated and used for visualization.

Sequencing data are deposited in GEO under accession number GSE176418. All other data are available within the manuscript and its accompanying Supplementary Data files or from the corresponding author upon reasonable request.

For processing of sequencing data, we used trimmomatic, STAR aligner, HTSeq, RSEM and the nf-core ATAC pipeline, and for downstream analysis R version 4.0.1, python 2.7.2 and the following packages: PAGODA2, MONOCLE2, SingleR, scran, SCENIC, Seurat, chromVAR, cisTopic, all of which are available on github.

To investigate possible early escape of CD19-expressing B-ALL leukemic cells from CD19 CAR T cells, we first analyzed interactions between CD19 CAR T cells and NALM-6, a CD19⁺ B-ALL cell line. For these studies, we used commercially available or in-house designed third-generation CD19 CAR constructs containing 4–1BB and CD28 costimulatory domains, which we expressed in T cells isolated from blood of normal donors. Using holotomography-based three-dimensional live microscopy, CAR T-cell engagement with the NALM-6 cells was identified by cell–cell contacts suggesting immunologic synapse formation, as has previously been reported (39). We identified three major types of interactions between CAR T cells and B-ALL cells (Fig. 1A; Supplementary Movies S1–S3): (i) Effective cytotoxic killing was shown by the leukemia cell blebbing after approximately 20 minutes and subsequent nuclear condensation within 1 hour (Supplementary Movie S1); (ii) “Scanning interactions” were characterized by CAR T-cell–leukemia cell interactions in which contacts were sustained over at least two hours without subsequent cytotoxic activity (Supplementary Movie S2); (iii) “Escape interactions” occurred when leukemia cells moved away within an hour of initial contact with a CAR T cell (Supplementary Movie S3).

We hypothesized that these differences in interactions may be related to heterogeneous expression of the target antigen CD19. To address this question, we used flow cytometry to measure CD19 expression by NALM-6 cells after coculture with CD19 CAR T cells and compared CD19 expression levels with cytotoxicity. Increased time of coculture and increased E:T ratio led to more killing of leukemia cells (Fig. 1B and C). However, among the remaining live leukemia cells in these conditions, we detected progressively reduced levels of CD19 expression with higher E:T ratios and longer periods of coculture, resulting in more than 95% downregulation of CD19 expression compared with CD19 expression of control NALM-6 cells that had not been cocultured with any effector cells after 24 hours (Fig. 1D and E). To confirm that the remaining leukemia cells were viable, we sorted live NALM-6 cells and monitored cell viability for 24 hours, 48 hours, and 72 hours after removal from coculture. Viability of CAR T-cell exposed leukemia cells remained relatively stable over the course of 72 hours (50% after 48 hours, 58% after 72 hours), and the expression of CD19 recovered, underscoring that a significant fraction of the leukemia cells that survive CAR T-cell cytotoxicity are viable (Supplementary Fig. S2A and S2B). We next assessed whether these CD19^low expressing leukemic cells were refractory to CAR T-cell killing. Indeed, NALM-6 cells that had been cocultured with CD19 CAR T cells overnight and were reexposed to fresh CD19 CAR T-cells escaped CAR T cell–mediated killing (viability 98.85% vs. 63.91%; Supplementary Fig. S2C).

To ensure that CD19 downregulation was induced by CAR T-cell – leukemia cell interactions and not due to selection of preexisting CD19^low leukemia cells, we performed coculture experiments using either CD4⁺ CAR T cells or Jurkat cells transfected with the CAR construct, both of which are less cytotoxic than CD8⁺ CAR T cells. Using a mixture of CD8⁺ and CD4⁺ CAR T cells at a 5:1 E:T ratio, only 10% of NALM-6 cells survived after 24 hours of coculture (Fig. 1C). Using the same coculture conditions, 64% and 71% of leukemia cells survived in the CD4⁺ CAR T cell and CAR Jurkat cell cocultures, respectively (Fig. 1F and G). Although the majority of leukemia cells survived in these cocultures, the degree of CD19 downregulation was similar to that observed among leukemia cells cocultured with a mixture of CD8⁺ and CD4⁺ CAR T cells (Fig. 1F and G compared with Fig. 1E) and leukemia cells exhibited an 85% and 84% downregulation of CD19 expression when exposed to CD4⁺ CAR T cells or CAR Jurkat cells, respectively, (Fig. 1F and G right) compared with CD19 expression of NALM-6 cells without coculture of any effector cells (control). Experiments using two other B-ALL cell lines, REH and KOPN-8, provided similar results (Supplementary Fig. S3). These data suggest that the occurrence of CD19^low expressing cells is mediated by active processes induced by CAR T cells causing CD19 downregulation rather than by selection of preexisting CD19^low cells.

To analyze the kinetics of decreased CD19 expression, we recorded time-lapse images of CAR T-cell–leukemia cell interactions using live microscopy tracking the CD19 protein in the leukemia cells. When interacting with the CAR T cells, CD19 first clustered at the interface of the interaction with CAR T cells (Fig. 2A; Supplementary Movie S4; Supplementary Image File S1). These clusters were primarily seen in leukemia cells interacting with CAR T cells; in cocultures with uninfected T cells, there was no change in CD19 antigen distribution (Fig. 2B; Supplementary Movie S5).

During and after interaction with CAR T cells, CD19-specific antibody fluorescence was detectable inside the leukemia cell, indicating that CD19 is internalized after cluster formation (Fig. 2A). To further evaluate this hypothesis, we stained NALM-6 cells with an anti-CD19 prior to coculture with subsequent staining using a fluorescently tagged secondary antibody either before or after the 24 hours coculture (Fig. 2C). Staining with the secondary antibody after coculture led to significantly decreased surface fluorescence intensity compared with staining before coculture (58.16% vs. 6.61%, P value = 0.0001), consistent with CD19 internalization (Fig. 2C), which we confirmed by intracellular flow cytometry for CD19 (Supplementary Fig. S4). Cytotoxicity of the CAR T cells was not affected by the presence of the anti-CD19 during coculture, thus excluding interference of the CD19-specific antibody with target epitope recognition by the CAR T cells (Supplementary Fig. S5).

When accounting for antibody bleaching, we observed a significant reduction of CD19 signal within the cells that formed clusters when interacting with CAR T cells (Fig. 2D), suggesting that these leukemia cells may degrade CD19 after internalization. In contrast, leukemia cells that interacted with CAR T cells without cluster formation or leukemia cells that interacted with uninfected T cells did not show changes in fluorescence intensity on either the cell surface or intracellularly.

Leukemia cells that were killed by CAR T cells demonstrated reduced CD19 expression after six hours of coculture (Supplementary Fig. S6), implying that internalization of CD19 is triggered by CD19–CAR interactions independent of cytotoxic or apoptotic outcomes.

To exclude the possibility that soluble mediators such as cytokines were causing CD19 internalization, we performed coculture assays in transwells, where leukemia cells that did not interact with CAR T cells shared the same media with leukemia cells cocultured with CAR T cells (Fig. 2E). In this scenario, CD19 expression by the leukemia cells not able to contact CAR T cells did not change.

Taken together, these data demonstrate that direct interaction between the CAR and CD19 is necessary for CD19 clustering, internalization, and downregulation.

CD19 cluster formation is a well-described phenomenon of B-cell activation in response to BCR stimulation by ligand (19). We hypothesized that CD19 clustering and internalization may lead to transcriptional programs of B-cell activation, similar to what has been described upon BCR stimulation–induced CD19 clustering in normal B cells. To address this question, we performed scRNA-seq of surviving leukemic cells after coculture with CAR T cells for 24 hours and compared these data with scRNA-seq of leukemic cells cocultured with uninfected T cells. For these experiments, leukemic cells were presorted for high CD19 expression to exclude preexisting cells expressing low CD19 prior to coculture.

A total of 1,039 cells were sorted and sequenced based on CD19 expression after coculture, with 831 cells remaining after quality filtering steps (Supplementary Fig. S7A and S7B). Pagoda2 clustering and visualization by t-Stochastic Neighbor Embedding (t-SNE) revealed a total of 7 clusters. Clusters 2 to 7 contained a mixture of leukemic cells cocultured with CAR T cells and uninfected T cells (Fig. 3A; Supplementary Table S2). In contrast, cluster 1 solely consisted of leukemic cells that had been cocultured with CAR T cells (Fig. 3A).

To understand how the clusters differed from each other, we first assessed differences in cell cycle, a known driver of cluster formation in single-cell analysis (40) using Seurat2 (26). We found that clusters 2 and 3 were enriched for cells in S-phase and clusters 4 and 5 for cells in G₂–M phase (Supplementary Fig. S8). Cluster 7 was enriched for cells in G1, whereas clusters 1 and 6 showed a mixture of cell-cycle phases. Accordingly, marker gene analysis of the respective clusters revealed expression of genes involved in cell-cycle regulation, such as CDK1, CDC20, CCNB1/B2, CCNF, CDCA2/3/5/8, MKI67, HMGB2, TOP2A in cluster 4, and MYC and RRP15 in cluster 5 (Fig. 3B; Supplementary Table S3). Consistent with cells in S-phase, clusters 2 and 3 showed high expression of DNA synthesis and repair genes such as RRM2, MCM5/10, PCNA, and POLE2, BRCA1, MSH2. Cells in cluster 6 expressed various genes of the NADH:ubiquinone oxidoreductase and ATP synthase subunits, both of which are involved in the mitochondrial respiratory chain. In addition, clusters 2 to 6 expressed genes that are part of several known signaling pathways in B cells – SYK and BTK expression were highest in cluster 2; VPREB1/3, IGLL1, BLK, and JCHAIN highest in cluster 3; CD72 highest in cluster 4; BCR highest in cluster 5; CD79B highest in cluster 6; and BLNK, CD19 and CD22 highest in cluster 7 (Supplementary Table S3).

Cluster 1, which was the only cluster consisting exclusively of leukemic cells exposed to CD19 CAR T cells, demonstrated the lowest CD19 expression of all clusters, both at the protein (Fig. 3C, left; Supplementary Table S4) and transcript level (Fig. 3C, right). We therefore hypothesized that this cluster might represent a unique population of leukemic cells refractory to CAR T cells. This cluster demonstrated strong enrichment of B-cell activation signatures with expression of CD69, CD40, and CCR7. To test whether these transcriptional changes reflected protein expression, we assessed CD40 protein expression by flow cytometry. We observed increased CD40 expression in NALM-6 cells that survived CAR T-cell coculture (Supplementary Fig. S9). Cluster 1 also expressed genes involved in the NFkB- and CD40-signaling pathway (NFKB2, NFKBIA/D/Z, REL, RELB, TRAF1, TNF, TNFAIP3/8, BCL2A1; Fig. 3B and D). GSEA revealed enrichment for NFkB signatures, several other immune response signatures related to cytokines (IFNγ response, IL2–STAT5 signaling, and IL6–JAK–STAT3 signaling) and inflammatory response signatures, indicating the activation of immunomodulatory pathways in cluster 1 cells (Fig. 3D; Supplementary Table S5). These data suggest that CAR-mediated CD19 crosslinking causes transcriptional programs of B-cell activation in a subset of B-ALL cells.

To gain further insights into the transcriptional dynamics of CAR T cell–exposed B-ALL cells, we employed pseudotime analysis by Monocle2. This analysis revealed seven distinct states distributed along a single common trajectory (Fig. 3E, top). State 5 cells at one end of the trajectory showed high expression of two genes, VPREB1 and VPREB3, that are normally expressed in preB cells (Fig. 3F; ref. 41). Genes informing the pseudotemporal distribution were also involved in cell-cycle regulation (CCNA2, CCNF, TOP2A, CDK1). State 1 cells, which consisted of Pagoda2 cluster 1 cells, were located at the opposite end of the trajectory (Fig. 3E, bottom). Cells towards this end of the pseudotime distribution exhibited increased expression of genes involved in NFkB signaling (NFKBIZ, NFKBIA, TNFAIP3, NFKBID, NFKB2, RELB, NFKB1, TNFAIP9, FCER2) and CD40 signaling (CD40, CD80, TRAF1, ICAM1, CFLAR, BCL2A1, EBI3), consistent with the identified marker genes of cluster 1 (Fig. 3F; Supplementary Table S6). These analyses suggest a continuous differentiation process of leukemic cells with state 1 cells representing the most differentiated state, exclusively containing CAR T cell–exposed CD19⁺ B-ALL cells.

B-cell activation normally leads to B-cell differentiation and initiation of the germinal center reaction, processes that are tightly regulated by master transcription factors (42, 43). Given their distinct transcriptional programs, we postulated that the subset of leukemic cells that clustered in state 1 were employing physiologic programs of B-cell differentiation and activation. To assess a possible involvement of B-cell activation programs, we analyzed transcriptional regulon activity using SCENIC. We found that pseudotime states 2, 4, and 7 cells were enriched for transcriptional regulons involved in cell-cycle regulation such as E2F1/2/4/8, TP53, MYC, BRCA1, STAT3. States 2–5 demonstrated enrichment of the early B-cell regulators PAX5, POU2F1, BACH1, BACH2, and KLF3, consistent with their developmental state. State 1 cells showed high activity of IRF4, BCL6, and ZBTB7A regulons, which are known master transcription factors of the germinal center reaction (Fig. 4A, left; Supplementary Table S7).

As further proof of transcriptional reprogramming we assessed transcription factor–motif enrichment by scATAC-seq in NALM-6 cells after 24 hours of coculture with CD19 CAR T cells. From 288 cells sorted and sequenced, 245 passed quality control (Supplementary Fig. S10A and S10B). Clustering of ATAC-peaks using cisTopic and visualization using UMAP identified 2 major clusters (Fig. 4A, right) with 99 differentially regulated motifs (adjusted P value < 0.05; Supplementary Table S8) that were chosen for further analysis. Differential motif–enrichment analysis showed that germinal center motifs such as ETV5, CTCF, ETS1, BHLHE40, YY1, FOXP1, TFEB, JUND, ZBTB7A, MEF2A/C/D, EGR1/2/3, MYBL1 were more accessible in peaks in cluster 1 (Fig. 4A, right). In addition, we found that motifs of other regulators of B-cell function and differentiation were enriched in peaks in cluster 1, including ERG, PAX5, FLI1, NFYA, NFYB and transcription factors of the ETS and Sp/KLF family. In contrast, cluster 2 peaks were enriched for motifs of SNAI2, a transcription factor active in early lymphoid cell development (44), and MSC, an inhibitor of the critical B-cell development regulator E47 (45, 46). We also detected motifs of the nuclear proapoptotic factor THAP1 (47) and the tumor/growth suppressor MZF1 (48), which were enriched in cluster 2 peaks. In addition, motifs of the known cell-cycle regulators E2F2 and E2F4 were enriched in cluster 1 peaks, corresponding to the G₂–M-phase signatures identified in our RNA-seq data in clusters 4 and 5 (Fig. 3B and D and Fig. 4A, left).

To further define which B-cell differentiation programs leukemic cells employed in response to interaction with CD19 CAR T cells, we assessed expression of validated B cell–differentiation markers, including hematopoietic stem cell (HSC), multipotent progenitor (MPP), common lymphoid progenitor (CLP), pro-B, pre-B, naïve B, germinal center dark zone (GC DZ), germinal center intermediate (GC INT), germinal center light zone (GC LZ), prememory, memory, class-switched memory, plasmablast, and plasma cell signatures. We compiled these signatures from several datasets [BLUEPRINT (30), Human Cell Atlas (32)] and recently published signatures of the germinal center reaction from Holmes and colleagues (2020; ref. 33; Fig. 4B, left; Supplementary Table S9). Most states showed enrichment for several immature signatures of the B-cell lineage, consistent with their B-ALL developmental state. For example, NALM-6 cells cocultured with uninfected T cells were enriched for the proB signature, which is consistent with NALM-6 cells arrested at this developmental stage. In contrast, immature signatures were downregulated in state 1 cells and these cells demonstrated enrichment of mature B–cell signatures including the light zone a signature (Fig. 4B, left). Furthermore, state 1 cells expressed the light zone marker CD83 and other germinal center markers like CD40, BCL2A1, SLAMF1, NFKBIA (Fig. 4B, right). As the physiologic B-cell germinal center reaction is accompanied by decreased CD19 expression [Fig. 4C, analyzing RNA-seq data from Holmes and colleagues (33)], these data suggest that state 1 leukemia cells employ germinal center reaction signatures to transcriptionally downregulate CD19 and maintain lower CD19 expression for escape from CD19 CAR T cells.

As the tyrosine kinase BTK is required for normal B-cell activation (49, 50), we next tested whether inhibiting BTK prevented B-cell activation induced by CD19 CAR T cells and inhibited escape from CAR T cell–mediated killing. To this end, we pretreated leukemic cells with the BTK inhibitor ibrutinib for 72 hours at IC₁₀ dosing (Supplementary Fig. S11A) prior to coculture with CAR T cells. Ibrutinib-treated leukemic cells exposed to CAR T cells were killed at higher rates compared with non-ibrutinib–treated leukemic cells (43.17% vs. 25.45% at E:T ratio of 1:1, P value = 6E-08; 66.62% vs. 45.11% at E:T ratio of 5:1, P value = 0.0131; Fig. 4D). Increased cytotoxicity was also observed after ibrutinib wash-out before coculture and we did not observe any reduced viability of either CAR T cells or leukemic cells with the ibrutinib dosing used in these assays (Fig. 4D; Supplementary Fig. S11B and S11C). Thus, these data demonstrate that leukemic cells may employ B-cell activation and germinal center reaction programs to escape from CAR T cell–mediated killing, a process that can be suppressed by inhibiting B cell–activation programs.

It is well established that target antigen recognition by CAR T cells is dependent on antigen density and that efficacy of CAR T cell–mediated cytotoxicity decreases with lower antigen expression on target cells (12, 14, 16, 51). Therefore, low expression or nearly complete loss of target antigen is a challenge for CAR T-cell treatments and other immunotherapies (13, 51, 52). Trogocytosis, a process by which membrane fragments are transferred from target to effector cells, has been identified as a possible mechanism for lowering target epitope expression (12). Enhancing CAR signaling has been proposed to tune CAR T-cells’ antigen recognition and counteract low target epitope expression (14, 53). However, little is known about the dynamic regulation of CD19 expression by target cells during their early interactions with CAR T cells and how these interactions may affect cytotoxic efficacy.

Here, we show that CD19 CAR T-cell therapy induces rapid clustering of CD19 on the leukemia cell–T cell interface and internalization of CD19 by B-ALL leukemic cells, resulting in decreased CD19 surface antigen expression among CAR T cell–exposed B-ALL cells. Among the initially CD19⁺ leukemic cells that survived CAR T-cell killing, we identified a distinct subpopulation in which low CD19 expression was sustained through transcriptional rewiring and engagement of unique transcriptional regulators associated with B-cell differentiation. Specifically, we show that these cells engage regulatory programs of normal B-cell activation such as the germinal center reaction, and the NFkB- and CD40-signaling pathways.

Our studies raise important issues about CAR-mediated receptor engagement on cancer cells and their consequences. This is particularly important as CAR T cells may lose efficacy over time, such as through T-cell exhaustion (54). As their killing efficacy decreases, those CAR T cells may still maintain their state change–inducing effects on the cancer cells, which may propagate a greater number of surviving cancer cells over time. Further work is needed to characterize the functional effects of CAR-mediated receptor engagement in relevant preclinical disease models in vivo and how they translate to treating relapsed disease in patients.

Our findings also raise questions about the effects of T cell–targeting immunotherapies other than CAR T cells on their cancer cell targets. In this context, treatment with the CD3/CD19 bispecific antibody blinatumomab has been associated with CD19^low/negative relapses (55, 56). Further studies are needed to understand the relative impact of bispecific immunotherapy agents and CAR T cells on CD19^low/negative relapses and the optimal sequence of administering different immunotherapies.

Cellular plasticity is a basis for treatment failure in several cancers (57, 58). Cancer cells may engage transcriptional programs that allow for dedifferentiation or cross-differentiation into other lineages (9, 10). Our data demonstrate that rather than de-differentiating into more stemlike states, engagement of lineage-appropriate differentiation and activation programs may allow for less effective recognition by CD19 CAR T cells. This is reminiscent of what has been shown for key signaling pathways that operate in normal B cells and remain essential for leukemia cell survival in BCR signaling–dependent precursor B-ALL, thus providing a rationale for targeting these pathways (59–61).

Our studies highlight the importance of combining CAR T-cell therapy with other targeted therapies to overcome the plasticity of leukemia cells. More work is needed to identify the exact epigenetic mediators that orchestrate the engagement of programs of normal B-cell activation and differentiation in B-ALL cells in response to CD19 CAR T cell interactions, as well as their dependencies on B cell–signaling pathways. While other groups have shown favorable combination of CAR T cells and ibrutinib in other B-cell malignancies, such as chronic lymphocytic leukemia (62) and mantle cell lymphoma (63), previous studies of combination treatment in B-ALL have been inconclusive (64, 65). Our data show that inhibiting BTK, an essential mediator of B-cell activation (49, 50), increases CD19 CAR T cell–mediated B-ALL cytotoxicity, presumably through inhibiting the engagement of normal B-cell differentiation and activation programs.

In summary, our studies highlight how CAR T cells might trigger engagement of physiologic regulatory programs in their target cells, and how CAR T cells’ failure to kill their targets at first interaction may promote target-cell survival.

N.G. Im reports grants from Kind-Philipp-Stiftung fuer paediatrisch-onkologische Forschung during the conduct of the study. J.M. Waldschmidt reports grants from German Research Foundation, International Myeloma Society; personal fees from Sanofi; and personal fees from Janssen outside the submitted work. C. Grassberger reports Nanolive SA - technical advisory board. J.G. Lohr reports grants from NCI during the conduct of the study; grants from BMS/Celgene; and personal fees from T2 Biosystems outside the submitted work. B. Knoechel reports grants from NCI during the conduct of the study. No disclosures were reported by the other authors.

N.G. Im: Conceptualization, data curation, software, formal analysis, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. A. Guillaumet-Adkins: Resources, data curation, formal analysis, validation, investigation, methodology, writing–review and editing. M. Wal: Resources, methodology. A.J. Rogers: Resources, data curation, software, formal analysis, investigation, methodology. J. Frede: Resources, data curation, software, formal analysis, validation, investigation, methodology. C.C. Havig: Formal analysis, validation, investigation, methodology. J. Yang: Formal analysis, validation, investigation, methodology. P. Anand: Resources, data curation, software, formal analysis, validation. S.K. Stegmann: Data curation, formal analysis, validation, methodology, writing–review and editing. J.M. Waldschmidt: Validation, methodology. N. Sotudeh: Data curation, software, formal analysis. L. Niu: Resources, validation. J. Voisine: Resources, data curation, investigation, methodology. M.R. Schweiger: Resources, supervision. C. Grassberger: Resources, visualization, methodology. J.G. Lohr: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. B. Knoechel: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing.

This research was supported by the NCI (K08CA191091 to B. Knoechel and K08CA191026 to J.G. Lohr), and the Ambrose Monell Foundation (to B. Knoechel). N.G. Im was supported by the Kind-Philipp-Stiftung für pädiatrisch-onkologische Forschung.

We thank the Knoechel and Lohr lab members for discussions and helpful comments on the manuscript. We thank the Jimmy Fund DFCI Flow Cytometry core for their support with flow sorting. We thank Min Gyu Im for his help with editing the movies.

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