Transcriptome Analysis Identifies a Myeloid‐Like Signature in Relapsed or Refractory B‐Cell Acute Lymphoblastic Leukemia Patients With Poor Response to CD19 Chimeric Antigen Receptor T Cell Therapy

ABSTRACT

Background

Relapsed or refractory B‐cell acute lymphoblastic leukemia (R/R B‐ALL) patients exhibit heterogeneous responses to CD19 chimeric antigen receptor T‐cell (CAR‐T) therapy. However, current studies primarily focus on the CAR‐T cells or the leukemia blasts, and the patient's intrinsic molecular characteristics before treatment remain poorly understood.

Methods

Fifteen R/R B‐ALL patients enrolled in a phase II, single‐arm clinical trial (NCT04684147) were stratified at a three‐month landmark into 8 good responders and 7 poor responders. Clinical characteristics, genetic alterations and pre‐treatment bone marrow transcriptomics were compared between groups, and an exploratory gene‐expression score was derived.

Results

Differential expression and enrichment analyses identified a myeloid‐like signature and upregulation of myeloid‐related genes in poor responders. Seven myeloid‐related genes were examined in independent patient cohorts, in B‐ALL cell lines and at single‐cell resolution in one patient; three were expressed by the leukemic cells themselves and four predominantly by monocytes and macrophages. A separate, variance‐derived seven‐gene least absolute shrinkage and selection operator (LASSO) score separated response groups in the discovery cohort (apparent area under the curve, AUC 1.00), with bias‐corrected and external estimates of approximately 0.57–0.59 (external cohort AUC 0.592, 95% confidence interval 0.356–0.828).

Conclusions

Pre‐treatment transcriptomic profiling revealed a myeloid‐like signature in R/R B‐ALL patients with poor CD19 CAR‐T response, warranting further investigation as a candidate correlate of treatment outcome.

1. Background

B‐cell acute lymphoblastic leukemia (B‐ALL) is a hematologic malignancy originating from precursor B lymphoid cells [1]. With the incorporation of immunotherapy and evolving treatment strategies, the treatment landscape for B‐ALL has been revolutionized and the outcome of B‐ALL patients has been substantially improved [2, 3]. Notably, CD19‐targeted chimeric antigen receptor T (CAR‐T) cell therapy can induce potent response in B‐ALL patients, with complete remission (CR) rates exceeding over 80% [4, 5, 6, 7]. However, although high initial remission rates can be achieved, a considerable proportion of relapsed or refractory (R/R) B‐ALL patients still experience disease recurrence in several pivotal CD19 CAR‐T clinical trials and 12‐month relapse‐free survival (RFS) is only 35%–55% [6, 7, 8]. In the clinical trial of our in‐house developed CD19 CAR‐T therapy (CNCT19), patients also demonstrated considerable heterogeneity in their response to the CAR‐T therapy, ranging from recurrence within months to sustained remission over 2 years [7]. Once CD19 CAR‐T therapy fails, patients face very limited treatment alternatives and poor prognosis [9]. Therefore, the heterogeneous response of B‐ALL patients, particularly the potential contributors to poor CAR‐T response, warrants further investigation.

Typically, disease relapse or CD19 CAR‐T failure is attributed to CD19 antigen escape or to factors intrinsic to the CAR‐T cells, such as poor persistence or inadequate expansion [8, 10]. However, CD19⁺ B‐ALL relapses are also frequent, and a less common mode of progression is lineage switch to acute myeloid leukemia (AML), in which patients frequently harbor myeloid‐related cytogenetic aberrancies [11, 12]. The non‐response to blinatumomab and high tumor burden have been associated with poor response to CD19 CAR‐T and inferior outcomes [13]. Both leukemia‐intrinsic factors and an immunosuppressive microenvironment can contribute to CAR‐T resistance [14]. Ramos et al. revealed leukemia‐intrinsic determinants of CAR‐T response and emphasized the leukemia blasts' interaction with immune cells in the tumor microenvironment (TME) [15]. The TME of B‐ALL comprises multiple cellular subpopulations including leukemic blasts, T cells, natural killer (NK) cells, myeloid‐derived cells and various supportive cell populations [16]. With advances in sequencing and the emergence of single‐cell omics, myeloid cells within B‐lineage hematological malignancies have been found to play a crucial role in CAR‐T response. Jain et al. found that large B‐cell lymphoma (LBCL) patients' pre‐CAR‐T state of immune dysregulation, particularly tumor interferon signaling and suppressive myeloid cells, was associated with no durable response [17]. Similarly, Stahl et al. identified a myeloid‐monocytic gene signature and CSF1R ⁺ myeloid‐monocytic cells that correlated with poor CD19 CAR‐T response in aggressive B cell lymphoma [18]. However, the correlation between CD19 CAR‐T response and myeloid subpopulations or myeloid signature in B‐ALL remains incompletely characterized.

Here, we collected pre‐CAR‐T bone marrow (BM) samples from 15 B‐ALL patients enrolled in the clinical trial (NCT04684147) of inaticabtagene autoleucel (inati‐cel) and conducted comprehensive clinical, genetic, and transcriptomic analyses. We aimed to characterize the profiles of R/R B‐ALL patients prior to CD19 CAR‐T infusion and identify underlying molecular disparities. We also derived an exploratory gene‐expression score associated with response to CD19 CAR‐T therapy.

2. Methods

2.1. Patients and Samples

This study included 15 adult patients with R/R B‐ALL enrolled in the clinical trial of the in‐house developed CD19 CAR‐T product CNCT19 (inaticabtagene autoleucel; HY001201, NCT04684147) at the Institute of Hematology & Blood Diseases Hospital, Chinese Academy of Medical Sciences & Peking Union Medical College (IHCAMS) [7]. All patients received lymphodepletion with cyclophosphamide 500 mg/m² daily for 2 days and fludarabine 30 mg/m² daily for 4 days, as specified in the trial protocol [7]. Clinical characteristics, including time to relapse and time to allogeneic transplantation, are listed in Table S1. Bone marrow mononuclear cells (BMMCs) were collected before lymphodepletion and used for bulk RNA sequencing (RNA‐seq). Targeted sequencing data were obtained from the next‐generation sequencing center of IHCAMS (Table S2). Response was assessed at a three‐month landmark after infusion, consistent with the primary end point of that trial, in which the overall remission rate (ORR) at the end of month 3 was pre‐specified [7]. No patient had undergone allo‐HSCT before this time point. Patients in continuous remission at that assessment were classified as good responders (n = 8) and those who had relapsed as poor responders (n = 7).

2.2. Transcriptome Sequencing and Data Processing

Total RNA isolation, library preparation and sequencing were performed by two providers, one is Tianjin Union Precision Medical Diagnosis Co. Ltd. (14 samples), and the second one is Beijing Novogene Co. Ltd. (one sample, P07, a good responder) based on the Illumina NovaSeq 6000 platform. The raw data processing pipeline included quality control (FastQC v0.11.8), filtering (Trimmomatic v0.39), alignment to the UCSC hg19 human reference (STAR v2.7.8a), and quantification (Subread v2.0.6, featureCounts). Of 28,265 quantified genes, those detected in more than 75% of samples were retained (18,888 genes), and ribosomal protein, hemoglobin, mitochondrial, microRNA and uncharacterized non‐coding (LINC and LOC) genes were then removed, leaving 16,890 genes for downstream analyses (Table S3). Read counts were converted to fragments per kilobase of transcript per million mapped reads (FPKM), log₂(x + 1) transformed, and quantile normalized across samples with the normalizeBetweenArrays() function of the limma (v3.65.3) R package. The 3000 genes with the highest median absolute deviation were used for principal component analysis (PCA) and unsupervised clustering. PCA was performed using the FactoMineR (v2.12) and factoextra (v1.0.7) R packages, and the hierarchical clustering used the complete agglomeration method.

2.3. Differential Expression Analysis and Enrichment Analysis

Differential expression analysis between good and poor response groups was performed using the DESeq2 (v1.42.1) R package. Genes with |log₂FoldChange| > 1 and unadjusted p value < 0.05 were defined as differentially expressed genes (DEGs; n = 1014) and used for enrichment analysis. Benjamini‐Hochberg adjusted p values are reported for all genes in Table S3, of which 144 remained significant at FDR < 0.05. DEGs were visualized using volcano plot and heatmap. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways enrichment analysis was performed using 1014 DEGs with clusterProfiler (v4.10.1) R package, with Benjamini‐Hochberg correction of enrichment p values (Tables S4 and S5). For comparison, the same analysis was applied to the 144 FDR‐significant genes, and the resulting terms are given in the Results. Gene Set Enrichment Analysis (GSEA) was performed by GSEA software (v4.3.3) on the complete ranked genes. Gene sets were obtained from MSigDB (https://www.gsea‐msigdb.org/gsea/index.jsp) and van Galen et al.'s study [19]. Enrichment score (ES), normalized ES (NES), nominal p‐value, and false discovery rate (FDR) q‐value were calculated to assess the significance of pathways.

2.4. Protein–Protein Interaction (PPI) Network Construction and Hub Genes Identification

The 1014 DEGs were submitted to STRING website (v12.0, https://cn.string‐db.org/) for protein–protein interaction analysis, restricted to Homo sapiens and using a minimum required interaction score of 0.4. The resulting interaction network was further analyzed and visualized in Cytoscape (v3.9.1). CytoHubba plugin was applied to identify the top 20 hubba nodes ranked by maximal clique centrality (MCC) method.

2.5. Single‐Cell RNA Sequencing (scRNA‐Seq) Analysis

Single‐cell RNA sequencing was available for one poor responder (P03) only and was conducted in our previous study [12]; no additional patient was profiled in the present work. BMMCs from this patient underwent 10× Chromium 3′ scRNA‐seq (Beijing Novogene Co. Ltd.), yielding 8666 cells. Clean FASTQ data were processed by Cell Ranger (v6.1.2) software to obtain the gene count matrix, and downstream analysis was performed with Scanpy (v1.11.0) Python package. Cells with fewer than 300 detected genes, genes detected in fewer than 5 cells, cells with more than 7000 detected genes and cells with more than 20% mitochondrial reads were removed. Clustering used the Leiden algorithm at a resolution of 0.4, giving 15 clusters that were assigned to five major cell types (preB‐like blasts, myeloid‐like B blasts, monocytes/macrophages, T/NK cells, Erythrocytes). Automated annotation with CellTypist (Immune All High model) was performed and then refined manually based on marker gene expression: CD19, CD79A and PAX5 for B cell markers; IGLL1, VPREB3 and IGHM for pre‐B cell developmental markers; CD34, CD38 and DNTT for blast cell markers; MKI67, TOP2A and STMN1 for proliferation markers; MPO, CD300E, CD14, LYZ and S100A12 for monocyte markers; CD68, CSF1R and CST3 for macrophage markers; FCER1A and CLEC4C for dendritic cell (DC) markers; CD3D, CD3E and CD2 for T cell markers; GNLY and NKG7 for NK cell markers; GYPA, TFRC and CD36 for erythrocyte markers. The myeloid signature score was calculated with scanpy.tl.score_genes() using the MSigDB gene set GOBP_MYELOID_CELL_ACTIVATION_INVOLVED_IN_IMMUNE_RESPONSE (v2025.1.Hs). Cell–cell communication analysis was conducted using CellChat (v1.6.1) R package according to official recommendations.

2.6. Prediction Model Construction

A response prediction model was built from the discovery cohort of 15 patients. The 2500 genes with the highest median absolute deviation among genes also quantified in GSE153670 were enrolled as candidates. Because this filtering step used all 15 patients, the apparent performance reported below is optimistic. Least absolute shrinkage and selection operator (LASSO) regression (α = 1) was applied with the glmnet (v4.1.10) R package using a Gaussian family, and the penalty was fixed at λ = 0.1887; exactly seven genes retained non‐zero coefficients. A composite risk score was calculated as the linear combination of the selected genes weighted by their coefficients:

$$\text{Risk score}=\sum i\left({\mathit{\beta }}_i*{\mathit{G}}_i\right)$$

βᵢ represents the LASSO coefficient and Gᵢ represents the log₂‐transformed, quantile‐normalized expression level of gene i.

The linear CD19 CAR‐T response prediction model was given:

$$\begin{array}{cc}\text{Risk score}=\hfill & 0.6457+0.1712\times \mathit{\text{SULT}}1A1+0.0179\hfill \\ \hfill & \times \mathit{\text{MAGEF}}1+0.0011\times \mathit{\text{CSF}}1R-0.0088\hfill \\ \hfill & \times \mathit{\text{LCN}}10-0.0277\times \mathit{\text{PCDHGB}}6-0.1008\hfill \\ \hfill & \times \mathit{\text{FAM}}234B-0.1735\times \mathit{\text{OXNAD}}1\hfill \end{array}$$

Because a Gaussian family was used, the model returns a linear risk score rather than a probability, and scores are comparable in rank but not on an absolute scale across cohorts. Model discrimination was assessed through area under the curve (AUC) analysis using the pROC (v1.19.0.1) R package, with higher scores predicting poor response. Because apparent performance in the discovery cohort is optimistic, two bias‐corrected estimates were obtained by repeating the entire pipeline within each resample: bootstrap optimism correction with 1000 resamples, and 100 repeats of stratified five‐fold cross‐validation. The model was additionally evaluated in the external cohort GSE153670.

2.7. Immune Cell Deconvolution

The relative abundance of 22 immune cell types was estimated with CIBERSORT [20] using the LM22 signature matrix, 1000 permutations and quantile normalization disabled, applied to transcripts per million (TPM) values computed from raw read counts. Cell types estimated as absent in more than half of the samples were excluded, and the remaining 10 were compared between response groups by Wilcoxon rank‐sum test with Benjamini‐Hochberg correction.

2.8. Public Dataset Analyses

Expression data from B‐ALL patients were obtained from GSE270185 (pediatric B‐ALL) and GSE34861 (adult B‐ALL). GSE270185 comprises 249 BM samples, of which the 209 annotated as B cell precursor acute lymphoblastic leukemia (BCP‐ALL) were analyzed; raw counts were converted to counts per million (CPM) and log₂(CPM + 1) transformed with edgeR (v4.7.3). GSE34861 was profiled on platform GPL15088 and the 191 B‐ALL patients with available cytogenetic annotation were analyzed. Single‐sample GSEA (ssGSEA) scores for the granulocyte‐macrophage progenitor (GMP)‐like, monocyte‐like and myeloid activation gene sets were calculated with GSVA (v2.3.1). Patients in these datasets are independent of the in‐house cohort and did not receive CAR‐T therapy. Cell line expression data were retrieved from the Cancer Cell Line Encyclopedia (CCLE); TPM values were log₂(x + 1) transformed and the 15 B‐ALL cell lines were analyzed.

RNA‐seq data from 27 R/R B‐ALL patients profiled before CD19 CAR‐T infusion were obtained from GSE153670 and used as an independent validation cohort. Patients were classified as complete response (CR) and no response (NR) to treatment as in the original report [21]. RNA‐seq data from 24 R/R B‐NHL patients profiled before CD19 CAR‐T therapy were obtained from Stahl et al. [18]. Patients were classified as durable response (DR) and non‐durable response (NDR) as in the original study. CR in GSE153670 was defined as complete remission at 1 year, and DR in Stahl et al. as durable response at 6 months, in contrast to the three‐month landmark used in our in‐house cohort. Differential expression between non‐durable and durable B‐NHL responders was analyzed with DESeq2 and used for pre‐ranked GSEA.

2.9. Statistical Analysis

Statistical analyses were conducted using R software (v4.3.1) and Python (v3.9.1). Two‐group comparisons of continuous variables used the Wilcoxon rank‐sum test, categorical variables were compared by Fisher's exact test, and correlations were assessed by Spearman's rank correlation. Multiple testing was adjusted by the Benjamini‐Hochberg procedure. AUC directions were pre‐specified (higher = poor response) and never reversed, and confidence intervals were obtained by the DeLong method or stratified bootstrap (2000 resamples). Sensitivity analyses used a six‐month response definition (Table S7), sex or age as a covariate (Table S8). Data visualization was primarily performed with the ggplot2 (v4.0.0), pheatmap (v1.0.13) and ComplexHeatmap (v2.18.0) R packages and the Matplotlib (v3.10.0) Python package. Statistical significance was denoted with symbols (*p < 0.05, **p < 0.01, and ***p < 0.001).

3. Results

3.1. The Landscape of R/R B‐ALL Patients Receiving CD19 CAR‐T Therapy at IHCAMS

To identify the potential factors influencing the CD19 CAR‐T response, a cohort of 15 adult R/R B‐ALL patients receiving our in‐house developed CD19 CAR‐T therapy (CNCT19) was enrolled (Figure 1A). Key clinical events after CAR‐T infusion are shown in Figure 1B. In the cohort, 14 of 15 patients (93.3%) achieved complete remission (CR) at day 28, except patient P03, who developed myeloid lineage switch under CD19 CAR‐T treatment [12]. Bone marrow minimal residual disease (MRD) assessed by B‐lineage flow cytometry was negative in all 15 patients at day 28 (Table S1). Eleven patients (73.3%) experienced a disease relapse and seven (46.7%) underwent allogeneic hematopoietic stem cell transplantation (allo‐HSCT) in CR state. Patients with a favorable outcome remained in remission for up to 1957 days (P01), whereas the earliest relapse occurred at 39 days (P03). Response was classified at a three‐month landmark after infusion: patients in continuous remission at that assessment were defined as good responders (n = 8) and those who had relapsed as poor responders (n = 7). Three months was chosen because no patient had undergone allo‐HSCT before this time point (Table S1, earliest transplantation, P01 on day 104), so that response at this landmark was mainly attributable to CAR‐T therapy. A 6‐month landmark reclassifies only one patient (Tables S1 and S7). Several mutations in genes recurrently altered in myeloid malignancies were confined to poor responders, each in a single patient: BCORL1 and an exon‐14 FLT3 internal tandem duplication (ITD) in P03, WT1 in P11, EZH2 in P12, and concurrent NRAS and KRAS mutations in P14. FLT3 alterations were not restricted to this group: the three remaining FLT3 variants in the cohort were point mutations (p.N676K, p.A680V and p.N841Y), all in good responders. The overall frequency of genetic abnormalities did not differ significantly between groups (Figure 1C, Tables S2 and S6). CD19 CAR‐T dynamics generally exhibited marked expansion during the first month after infusion, followed by reduction and persistence both in the bone marrow and peripheral blood (Figure 1D,E).

FIGURE 1.

The landscape of R/R B‐ALL patients receiving CD19 CAR‐T therapy at IHCAMS. (A) Bioinformatic analysis workflow for identifying key factors affecting CAR‐T response and constructing the prediction model. All R/R B‐ALL patients were enrolled in the clinical trial HY001201/NCT04684147. Created in BioRender. Wang, M. (2026) https://BioRender.com/x27aksd. (B) Swimmer plot showing crucial events experienced by 15 in‐house R/R B‐ALL patients after CD19 CAR‐T infusion. Dashed lines mark the three‐month (M3) and six‐month (M6) landmarks; response was classified at M3, with patients in continuous remission defined as good responders and those who had relapsed as poor responders. CR: Complete remission in bone marrow morphology. Allo‐HSCT: Allogeneic hematopoietic stem cell transplantation. Relapse: Relapse detected in bone marrow morphology or minimal residual disease (MRD) measurement. Last follow‐up: The patient's final follow‐up. (C) Heatmap summarizing fusion genes and gene mutations in the 15 R/R B‐ALL patients, ordered by response group. Variant classes are distinguished by color. The FLT3 row combines an exon‐14 internal tandem duplication in P03 (shown as a non‐frameshift insertion) with point mutations in three good responders. Variant‐level details are given in Table S2. (D, E) Line graphs illustrating the dynamics of CD19 CAR‐T cells over time in peripheral blood (PB) (D) and bone marrow (BM) (E) after infusion. Upper: CD3⁺ CAR‐T cells count per μL blood. Lower: CD3⁺ CAR‐T cells percentage among white blood cells (WBC).

3.2. Differential Expression and Enrichment Analysis Identified Factors Potentially Associated With CAR‐T Response

Before differential analysis, we assessed data quality and technical variation. Library size and the number of detected genes did not differ between response groups, and the single sample sequenced by a second provider (P07, a good responder) was not an outlier by inter‐sample correlation or by PCA (Figure S1A–E). To explore the intrinsic differences between response groups, we compared the transcriptomic profiles of the 8 good and 7 poor responders. PCA and unsupervised clustering showed substantial inter‐patient variability, and good and poor responders did not form clearly separated clusters (Figure S1F,G). Then, differential expression analysis was performed between good and poor response groups. A total of 1014 DEGs were identified, of which 534 genes were up‐regulated and 480 genes were down‐regulated in the poor response group; 144 of these remained significant after Benjamini‐Hochberg correction (Figure 2A,B, Table S3). Two myeloid‐related genes were among the top 10 up‐regulated genes (LILRB4, log₂FoldChange = 3.82, padj = 3.81e‐06; SIGLEC1, log₂FoldChange = 3.12, padj = 6.85e‐04).

FIGURE 2.

Differential expression and functional enrichment analysis. (A) Volcano plot illustrating the differentially expressed genes (DEGs) between good and poor responders. Dark points, genes passing Benjamini‐Hochberg (BH) FDR < 0.05 with |log₂FoldChange| > 1; light points, genes reaching nominal p < 0.05 only. The 10 most significant up‐regulated genes in the poor response group are labeled. (B) Heatmap showing DEGs (BH FDR < 0.05) expression between good and poor responders. (C) Representative GO terms and KEGG pathways enriched by DEGs, grouped by GO biological process (GO‐BP), cellular component (GO‐CC), molecular function (GO‐MF) and KEGG. The blue color scale indicates the count of genes enriched in the terms and pathways. Bars show −log10 of the unadjusted p value; all terms shown had a BH‐adjusted p < 0.05. Complete enrichment results are given in Tables S4 and S5. (D, E) Gene set enrichment analysis (GSEA) analysis of the monocyte‐like and granulocyte‐macrophage progenitor (GMP)‐like signatures in the in‐house R/R B‐ALL patient cohort receiving CD19 CAR‐T therapy (D, n = 15) and in the R/R aggressive B‐NHL patient cohort of Stahl et al. [18] (E, n = 24). Nominal p‐value, FDR q‐value, enrichment score (ES), and normalized enrichment score (NES) were labeled; FDR q < 0.25 was considered significant.

Enrichment analysis was performed to explore potential biological factors associated with CAR‐T therapy response. GO terms and KEGG pathways were enriched using the 1014 DEGs (Figure 2C, Tables S4 and S5). In addition to terms related to cell adhesion and membrane organization, several immune‐related processes were enriched, including T cell function regulation, chemokine production, and mononuclear cell differentiation. When gene selection was instead based on the FDR‐corrected list, only two GO biological process terms (homophilic and cell–cell adhesion via plasma‐membrane adhesion molecules) and one KEGG pathway (cadherin signaling) remained significant. The myeloid‐related DEGs and GO terms prompted us to examine specific myeloid‐related gene sets by GSEA, which is pre‐ranked on the total gene list and therefore does not depend on any significance threshold. The monocyte‐like signature (NES = 2.1613, FDR q‐value = 7e‐04) and GMP‐like signature (NES = 2.0056, FDR q‐value = 0.0016) were significantly enriched in the poor response group (Figure 2D). In an independent cohort of R/R aggressive B non‐Hodgkin lymphoma (B‐NHL) patients receiving CD19 CAR‐T therapy [18], the monocyte‐like signature was likewise enriched in patients with non‐durable response (NES = 1.6374, FDR q‐value = 0.0202), whereas the GMP‐like signature was not (NES = 1.2620, FDR q‐value = 0.2600) (Figure 2E). Gene sets relating to regulation of immune response, myeloid cell activation and leukocyte proliferation were also found to be up‐regulated in poor responders (Figure S2A–F). Immune cell fractions estimated by CIBERSORT did not differ significantly between response groups for any cell type, including monocytes and macrophages (Figure S2G).

3.3. Expression Profiles of Myeloid‐Related Genes and Myeloid Signatures

We further explored the myeloid‐related genes associated with CAR‐T therapy response. Seven myeloid‐related genes (LILRB4, SIGLEC1, LGALS2, CSF1R, MPEG1, CD300LB, CLEC10A) were among the top 50 up‐regulated DEGs and were more highly expressed in poor responders (Wilcoxon rank‐sum test: LILRB4, p = 0.00031; SIGLEC1, p = 0.00031; LGALS2, p = 0.00591; CSF1R, p = 0.01399; MPEG1, p = 0.02051; CD300LB, p = 0.04009; CLEC10A, p = 0.07211) (Figure 3A). All seven were also significant by DESeq2 after Benjamini‐Hochberg correction (Table S3). Adjusting the model for sex or for age left all seven genes significant with unchanged direction and near‐identical effect sizes. Adjustment for the pre‐lymphodepletion blast fraction, which itself differs between response groups (Wilcoxon p = 0.056), preserved the direction of all seven genes but attenuated the effect sizes, with only LILRB4 remaining significant (Table S8). We next examined these genes in B‐ALL cell lines, which lack a bone marrow microenvironment and therefore report only what the leukemic cells themselves express. Across 15 B‐ALL cell lines from the CCLE, LILRB4, CSF1R and MPEG1 were expressed [mean log₂(TPM + 1) 2.70, 2.33 and 1.58; detectable in 73%, 80% and 47% of lines, respectively], whereas SIGLEC1, CLEC10A, CD300LB and LGALS2 were essentially undetectable [mean log₂(TPM + 1) ≤ 0.10] (Figure 3B). We then examined the myeloid genes and signature scores in two independent cohorts of B‐ALL patients harboring different cytogenetic and molecular alterations (Figure 3C,D). Pediatric B‐ALL patients with ETV6::RUNX1 fusion exhibited lower myeloid‐related genes expression and signature scores, whereas adult B‐ALL patients with BCR::ABL1 fusion demonstrated higher myeloid signature scores. A PPI network was constructed with 1014 DEGs, and the top 20 hub genes were identified (Figure 3E). In addition to genes related to cellular immunity and chemotaxis (CXCL8, CCR2), myeloid‐associated genes (CD163, CD68, CSF1R, SIGLEC1) were also among the hub genes.

FIGURE 3.

Expression profiles of myeloid‐related genes. (A) Box plot comparing the seven myeloid‐related genes between good (n = 8) and poor (n = 7) responders in the in‐house cohort; p values are from Wilcoxon rank‐sum tests. (B) Heatmap displaying the absolute expression (log₂(TPM + 1)) of seven myeloid‐related genes across 15 B‐ALL cell lines from the Cancer Cell Line Encyclopedia (CCLE). (C, D) Heatmaps displaying myeloid‐related genes' expression profiles and myeloid signatures scores across pediatric (C, GSE270185, n = 209) and adult (D, GSE34861, n = 191) B‐ALL patients with different genetic alterations. Signatures are ssGSEA scores for the GMP‐like, monocyte‐like, and myeloid activation gene sets. In (D, GSE34861), only five of the seven genes are shown, as MPEG1 and LILRB4 were not quantified on this platform. (E) PPI network illustrating the interactions among the top 20 hub genes among the 1014 DEGs, ranked by maximal clique centrality.

Together, these data identified a distinctive myeloid signature in the poor response group, manifested by enrichment of myeloid‐related pathways, high expression of myeloid‐related genes, and a myeloid‐associated regulatory network, and showed that B‐ALL patients with different genetic alterations exhibited different myeloid features.

3.4. Single‐Cell Transcription Analysis of One Poor CD19 CAR‐T Therapy Responder

To investigate the myeloid signature at single‐cell level, we analyzed the BMMCs of patient P03 prior to CD19 CAR‐T treatment; no other patient in this cohort was profiled by scRNA‐seq. A total of 8666 cells were annotated into five cell subtypes: PreB‐like blast cells, myeloid‐like B blast cells, monocytes and macrophages (Mono/Macro), T and NK cells (T/NK) and erythrocytes (Eryth) (Figure 4A,B). Myeloid signature was scored for all cells (Figure 4C,D). It is noteworthy that myeloid signature was not confined to the myeloid cluster: myeloid‐like B blasts also showed intermediate scores, whereas preB‐like blasts and T/NK cells showed comparable, lower scores. Examining the seven myeloid genes across these populations showed that they were not uniformly distributed (Figure 4E). LILRB4 was more highly expressed in myeloid‐like B blasts than in monocytes and macrophages (mean 0.368 versus 0.175; detected in 43.8% versus 18.7% of cells), and CSF1R and MPEG1 were expressed at comparable levels in the two populations. In contrast, SIGLEC1, CLEC10A, CD300LB and LGALS2 were expressed predominantly by monocytes and macrophages. LILRB4, CSF1R and MPEG1 were the same genes that were detectable in B‐ALL cell lines (Figure 3B), which lack a marrow microenvironment, indicating that part of the myeloid signature was intrinsic to the leukemic cells and part derived from non‐leukemic myeloid cells. Additionally, CellChat analysis revealed active communication between cell subpopulations, particularly among B blast cells, myeloid cells, and T/NK cells (Figure S3A,B). The macrophage migration inhibitory factor (MIF) signaling pathway (CD74‐CXCR4, CD74‐CD44) was active, with the two blast populations as the dominant senders and monocytes, macrophages and T/NK cells as the principal receivers (Figure S3C–E). As these observations derive from a single patient, they are illustrative rather than evidence of a generalizable pattern.

FIGURE 4.

Expression of myeloid‐related genes and signatures at the single‐cell level in one poor responder. Single‐cell RNA sequencing was available for one patient (P03) only. (A) uniform manifold approximation and projection (UMAP) dimension reduction distributed 8666 BMMCs of P03 into 5 clusters: PreB‐like blast cells, myeloid‐like B blast cells, monocytes or macrophage (Mono/Macro), T natural killer cells (T/NK), and erythrocytes (Eryth). (B) Dot plot showing the expression of marker genes in each cell subset. Populations were annotated from these markers independently of the myeloid signature score. Dot size indicates the fraction of cells expressing the gene and color indicates mean expression within the population. (C) UMAP plot colored by myeloid signature score. (D) Myeloid signature score by cell population. (E) Dot plot showing expression of the seven myeloid‐related genes across the five populations. Values are absolute mean expression without scaling across populations.

3.5. Prediction of CD19 CAR‐T Therapy Response in R/R B‐ALL Patients

To predict CD19 CAR‐T therapy response, a prediction model was constructed following the workflow shown in Figure 5A. Using the discovery cohort of 15 patients, LASSO regression (α = 1) was applied to the 2500 most variable genes at a penalty of λ = 0.1887; seven genes retained non‐zero coefficients (SULT1A1, MAGEF1, CSF1R, LCN10, PCDHGB6, FAM234B and OXNAD1; Figure 5B,C). The resulting risk score was given in the Methods section and separated good from poor responders in the discovery cohort (Wilcoxon rank‐sum test, p = 0.0003; Figure 5D). The predictive performance was assessed by receiver operating characteristic (ROC) analysis (Figure 5E,F). Apparent performance in the discovery cohort was perfect (AUC = 1.00, Figure 5E), but this estimate was optimistic because gene filtering and model fitting used all 15 patients. Two bias‐corrected estimates were therefore obtained by repeating the entire pipeline within each resample: bootstrap optimism correction gave an AUC of 0.911, and 100 repeats of stratified five‐fold cross‐validation gave 0.567 (2.5th‐97.5th percentile 0.375–0.732). In the external cohort GSE153670 (n = 27), the model achieved an AUC of 0.592 (95% CI 0.356–0.828), a confidence interval that included 0.5 (Figure 5F). The cross‐validated and external estimates were therefore concordant at approximately 0.57–0.59, and we regard the signature as exploratory rather than as a validated predictor (Figure 5G).

FIGURE 5.

A seven‐gene LASSO risk score for response to CD19 CAR‐T therapy in R/R B‐ALL. (A) Workflow for construction and evaluation of the risk score. (B) LASSO (α = 1) regularization paths for the 2500 most variable genes; each line is one gene, and the seven retained at λ = 0.1887 (dashed line) are colored. (C) Coefficients of the seven genes in the final model. Red, positive; blue, negative. (D) Risk score by response group in the discovery cohort (Wilcoxon rank‐sum test). (E, F) ROC curves for the risk score, with AUC and 95% confidence intervals: (E) discovery cohort HY001201 (n = 15; apparent performance); (F) external validation cohort GSE153670 (n = 27). (G) Comparison of performance estimates. Bias‐corrected values were obtained by repeating the entire pipeline (variance filtering, penalty selection and model fitting) within each resample: Bootstrap optimism correction (1000 resamples) and 100 repeats of stratified fivefold cross‐validation. Error bars: 95% CI for GSE153670; 2.5th‐97.5th percentile across 100 cross‐validation repeats. Dashed line, AUC = 0.5.

We next compared the risk score with clinical variables recorded for the same patients (Table S6). None of the clinical or MRD variables discriminated response (all p > 0.05), including bone marrow blast fraction before lymphodepletion (AUC 0.196, p = 0.056) and white cell count at screening (AUC 0.750, p = 0.118). Day‐28 MRD was negative in all patients and none had received prior CD19‐ or CD22‐directed therapy, so neither could be evaluated, and the duration of B‐cell aplasia was identical to the time to relapse in six of seven poor responders. The risk score was inversely correlated with pre‐treatment blast burden (Spearman ρ = −0.53 and −0.55, both p < 0.05), indicating that the score and disease burden were not independent in this cohort.

4. Discussion

The current study characterized the pre‐treatment profiles of R/R B‐ALL patients with divergent responses to CD19 CAR‐T therapy. Patients with poor response exhibited a distinct myeloid‐like signature, and the seven myeloid‐related genes underlying it were examined in independent patient cohorts, B‐ALL cell lines and at the single‐cell level in one patient. Separately, we derived an exploratory seven‐gene risk score by penalized regression on the most variable genes. These genes were selected independently of the myeloid analysis and share only CSF1R with it. Together, these findings suggest that a myeloid‐like transcriptional state in B‐ALL is associated with poor CD19 CAR‐T response and may warrant further investigation as a candidate biomarker.

The long‐term outcome of CAR‐T immunotherapy is determined by multiple factors including disease status, age, pre‐treatment regimen, and the intrinsic design of the CAR‐T cells [22, 23, 24]. Numerous studies have explored CAR‐T cells‐intrinsic determinants of durability, such as costimulatory domains, CAR affinity, apoptosis, persistence, exhaustion and senescence [25, 26, 27, 28]. Leukemic‐intrinsic resistance and immunosuppressive TME are also recognized contributors to poor CAR‐T therapy outcome [15, 21, 29, 30]. By contrast, we examined patient characteristics prior to CAR‐T therapy and found that a myeloid‐like signature in B‐ALL was associated with poor CAR‐T response. The determinants of poor response to CAR‐T immunotherapy reported to date are summarized in Figure 6.

FIGURE 6.

Summary of determinants of poor response to CD19 CAR‐T immunotherapy. Factors contributing to poor CAR‐T response include: Patient clinical characteristics, CAR‐T intrinsic factors, leukemia‐intrinsic factors, leukemia microenvironment and myeloid‐like signature, etc. Treg: Regulatory T cell. NKT2: Type II natural killer T cell. NK2: Type II natural killer cell. Th2: Type II helper T cell. MDSC: Myeloid‐derived suppressor cell. M2: M2‐polarized macrophage. N2: N2‐polarized neutrophil. ILC2: Group 2 innate lymphoid cell. Created in BioRender. Wang, M. (2026) https://BioRender.com/8d7wia3.

The myeloid‐like signature we identified in poor responders was defined at the bulk transcriptional level. Such a distinctive signature could in principle arise from the blast cells themselves or from myeloid subpopulations within the TME. Our data indicated that both contributed. Across 15 B‐ALL cell lines, which lack a marrow microenvironment, LILRB4, CSF1R and MPEG1 were expressed whereas SIGLEC1, CLEC10A, CD300LB and LGALS2 were essentially undetectable. At single‐cell resolution in one poor responder, the same three genes were expressed by myeloid‐like B blasts, while the remaining four were expressed predominantly by monocytes and macrophages. Therefore, part of the signature appears intrinsic to the leukemic cells, and part derives from non‐leukemic myeloid cells.

Lineage switch is an established mechanism of CD19 CAR‐T failure in B‐ALL [31, 32, 33, 34], and one patient in this cohort (P03) underwent switch to myeloid disease [12]. Because the signature was measured before lymphodepletion, it is compatible with a pre‐existing subpopulation capable of myeloid differentiation rather than with a change induced by CD19‐directed pressure. As switch was documented in only one patient, the signature is better regarded as myeloid‐associated transcriptional priming, of which overt lineage switch is one uncommon outcome.

In solid tumors, myeloid cells within the TME can be reprogrammed by tumor cells into more potent immunosuppressive cells, such as tumor‐associated macrophages (TAM) or myeloid‐derived suppressor cells (MDSC) [35], which create an unfavorable microenvironment for cancer immunotherapy [36, 37]. A suppressive bone marrow microenvironment is a well‐documented contributor to poor response to immunotherapy in AML [38, 39], but myeloid cells have rarely been implicated in B‐ALL. Recently, elevated CD10⁻ neutrophils in B‐ALL were reported to be associated with non‐response and poor prognosis after CD19 CAR‐T therapy [40]. In our study, the enrichment of genes associated with monocyte and macrophage activation, including CSF1R, CD68, CD163, and SIGLEC1, together with the enrichment of GMP‐like and monocyte‐like signatures, was consistent with an immature and potentially immunosuppressive myeloid state [41]. Immune deconvolution did not detect a difference in estimated myeloid cell proportions between response groups. However, with 15 patients the power to detect such a difference is limited, and whether the signature reflects a shift in myeloid programs or in myeloid cell abundance warrants investigation in larger cohorts. In the single‐cell profile of P03, the blast populations were the dominant senders in the MIF signaling network, with monocytes, macrophages and T/NK cells as the principal receivers. MIF‐CD74 signaling has been implicated in the polarization of myeloid cells by tumor cells [42], and LILRB4 expressed on the blasts is an inhibitory receptor that suppresses T cell responses [43]. Blasts carrying a myeloid‐associated program may therefore not only evade CD19‐directed killing but also help sustain the myeloid‐related suppressive environment in which that killing is less effective.

Given the heterogeneity of patient responses to CAR‐T therapy, an accurate prediction model would be valuable to guide clinical management and subsequent treatment strategies. Maurer et al. found that the baseline immune state of R/R LBCL patients, particularly a lower circulating lymphocyte‐to‐monocyte ratio, was associated with a reduced probability of response to CD19 CAR‐T [44]. Many researchers have utilized various clinical characteristics to construct CAR‐T response prediction models. For instance, Raj et al. integrated multiple laboratory measurements and cytokine release data to construct an inflammation mixture model that predicted CAR‐T treatment failure in non‐Hodgkin lymphoma (NHL) [45]. Gagelmann et al. screened key clinical variables and constructed a simple scoring model that predicted early relapse, response, and survival of R/R multiple myeloma (MM) after B‐cell maturation antigen (BCMA) CAR‐T [46]. Our seven‐gene score should be regarded as exploratory rather than as a comparable prediction tool. With 15 patients and 7 events, apparent discrimination was complete but bias‐corrected and external estimates were considerably lower (AUC approximately 0.57–0.59). Discrimination in the external cohort was modest, and one likely contributor is the difference in endpoint: the model was derived against relapse within 3 months, whereas response in GSE153670 was defined by complete remission at 1 year [21]. The score should therefore not be read as a quantitative version of the myeloid signature. It was derived by unsupervised variance filtering followed by penalized regression, and CSF1R, its only myeloid‐related member, carries a coefficient close to zero (0.0011). Prediction models integrating transcriptomic features with clinical characteristics in larger cohorts would be required before any clinical application.

Multiple approaches are known to enhance CAR‐T cell persistence and function, including CAR design optimization, additional gene editing, or co‐culture strategies [47]. Targeting the TME to improve CAR‐T efficacy is increasingly being investigated. Gong et al. used LILRB4 to simultaneously target tumor cells and MDSCs within the TME of MM patients [48], and Mavuluri et al. demonstrated that macrophages in the TME were reprogrammed toward an M2 phenotype during B‐ALL progression in a murine model, with macrophage depletion restoring CAR‐T responsiveness [49]. Whether targeting myeloid populations in B‐ALL patients can enhance CD19 CAR‐T response remains largely unexplored.

Several limitations should be acknowledged. First, with 15 patients and 7 events from a single center, statistical power is limited and the seven‐gene score should be regarded as exploratory. Second, poor responders had a lower marrow blast fraction before lymphodepletion, so part of the bulk myeloid signal may reflect sample composition. Adjusting for blast fraction preserved the direction of all seven genes but attenuated their effect sizes, leaving only LILRB4 significant (Table S8). LILRB4 is also expressed by B‐ALL cell lines and by myeloid‐like blasts, so its signal is not compositional. However, a compositional contribution to the other genes cannot be excluded. Third, single‐cell profiling was available for one patient only, so the cellular attribution of the myeloid signature is illustrative rather than established. Fourth, response was defined at a three‐month landmark, chosen because no patient had undergone transplantation by that time. A later landmark would capture additional relapses but would allow transplantation to confound the assessment. Finally, this study focused on pre‐treatment characteristics. Following the introduction of inati‐cel into front‐line treatment of B‐ALL at IHCAMS, samples at key time points have been systematically collected. Future work with a larger cohort will integrate longitudinal multi‐omics data, functional validation, and clinical characteristics.

5. Conclusions

In summary, we characterized the pre‐treatment profiles of R/R B‐ALL patients with divergent responses to CD19 CAR‐T therapy and identified an upregulated myeloid‐like signature in poor responders. This signature arises in part from the leukemic cells themselves and in part from non‐leukemic myeloid cells, and warrants further investigation as a candidate correlate of CD19 CAR‐T response.

Author Contributions

Shisen Wang: investigation, validation, writing – original draft, writing – review and editing, data curation, software, methodology, formal analysis, visualization. Wenbing Liu: methodology, software, data curation, investigation, validation, writing – review and editing, visualization, formal analysis. Ying Wang: supervision, methodology. Yingxi Xu: methodology, supervision. Haiyan Xing: methodology, supervision, resources. Kejing Tang: methodology, supervision, resources. Zheng Tian: methodology, supervision, resources. Qing Rao: methodology, supervision. Shaowei Qiu: methodology, supervision. Hui Wei: methodology, supervision. Runxia Gu: conceptualization, funding acquisition, writing – review and editing, project administration, supervision. Min Wang: conceptualization, funding acquisition, writing – review and editing, project administration, supervision. Jianxiang Wang: conceptualization, funding acquisition, writing – review and editing, project administration, supervision.

Funding

This work was supported by National Natural Science Foundation of China (82341213, 82570278), CAMS Innovation Fund for Medical Sciences (2025‐I2M‐XHZY‐033), National Clinical Research Foundation Project (2023NCRCA0102) and Tianjin Municipal Science and Technology Commission Grant (23JCYBJC01060).

Ethics Statement

The study was approved by the Ethics Review Board of IHCAMS (approval no. XY2020045‐EC‐2).

Consent

All patients signed the informed consents in accordance with the Declaration of Helsinki.

Conflicts of Interest

Although Juventas Cell Therapy Ltd. sponsored the clinical trial, the company did not participate in conducting the study or performing any analyses for the current manuscript. The authors declare no conflicts of interest.

Supporting information

Figure S1: Quality control and technical assessment of the 15‐patient RNA‐seq cohort. (A) Library size (million reads) per patient, colored by 3‐month response group. Dashed line, cohort median. The label indicates the robust z‐score for the single depth outlier, P14, which had an unusually deep library but a typical median inter‐sample correlation (0.888). (B) Number of genes detected (count > 0) per patient. Library size (p = 1.000) and genes detected (p = 0.867) did not differ between groups. (C) Distribution of log₂(FPKM+1) across samples before (top) and after (bottom) quantile normalization. (D) Sample‐to‐sample Spearman correlation of normalized expression, annotated by response group and sequencing provider. (E‐F) Principal component analysis (PCA) of normalized expression, colored by sequencing provider (E) and response group (F). All samples except P07 (a good responder) were sequenced by the same sequencing provider. Provider was not associated with response group (Fisher's exact p = 1.000). (G) Heatmap showing the results of unsupervised clustering of 15 patients based on the top 3000 variable genes. Good and poor responders did not form separate clusters.

Figure S2: Enrichment analysis and immune infiltration. (A–F) Gene set enrichment analysis (GSEA) of specific gene sets: negative regulation of innate immune response (A), regulation of adaptive immune response (B), regulation of immune effector processes (C), myeloid cell activation involved in immune response (D), myeloid dendritic cell activation (E), and positive regulation of leukocyte proliferation (F). Nominal p‐value, FDR q‐value, enrichment score (ES), and normalized enrichment score (NES) were labeled. FDR q < 0.25 was considered significant. (G) Box plot comparing estimated proportions of immune cell types between good and poor responders (CIBERSORT with the LM22 signature matrix). ns, not significant.

Figure S3: Cell–cell communications analysis and key signaling pathway identification in the R/R B‐ALL patient (P03). (A,B) Cell–cell interactions between major cell clusters. Edge width represents interaction intensity. (A) Number of interactions. (B) Interaction weights or strength. (C) Heatmap illustrating communication intensity between different signal pathways ligand‐receptor pairs across different cell clusters. (D) Heatmap illustrating MIF signaling pathway interaction intensity between sender clusters and receiver clusters. (E) Violin plot showing the expression of MIF signaling pathway‐related genes among cell clusters.

Table S1: Clinical profiles of R/R B‐ALL patients receiving CNCT19 in IHCAMS.

Table S2: Non‐silent mutations of R/R B‐ALL patients receiving CNCT19 in IHCAMS.

Table S3: Differential expression analysis result between good and poor responders.

Table S4: Gene Ontology (GO) enrichment analysis results.

Table S5: Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis results.

Table S6: Discrimination of the seven‐gene risk score and clinical variables for CAR‐T response in the discovery cohort.

Table S7: Differential expression of the seven myeloid genes under 3‐month and 3‐month response definitions.

Table S8: Differential expression of the seven myeloid genes with and without covariate adjustment.

Data Availability Statement

RNA‐seq data of R/R B‐NHL patients prior to CD19 CAR‐T therapy were obtained from Stahl et al. (https://doi.org/10.5281/zenodo.15280550). The expression data of B‐ALL cell lines were retrieved from the Cancer Cell Line Encyclopedia (CCLE) database (https://portals.broadinstitute.org/ccle). Expression data of B‐ALL patients were retrieved from GSE270185 and GSE34861. RNA‐seq data of R/R B‐ALL patients prior to CD19 CAR‐T infusion were retrieved from GSE153670. The scRNA‐seq data of the B‐ALL patient P03 can be obtained from the Genome Sequence Archive (https://ngdc.cncb.ac.cn/gsa‐human/browse/HRA005240). The RNA‐seq data of our in‐house B‐ALL patients are available from the corresponding author upon reasonable request.