Myeloid-Like Bone Marrow Programs Marked Poor CD19 CAR-T Response in R/R B-ALL
Pre-treatment bone marrow biology appears to matter in relapsed or refractory B-cell acute lymphoblastic leukemia far more than surface phenotype alone. In a small phase II cohort of patients treated with CD19 chimeric antigen receptor T-cell therapy, investigators found that those who later responded poorly carried a distinct transcriptomic pattern before infusion: a myeloid-like signature with higher expression of genes tied to monocytes, macrophages, and other innate immune programs.
The signal emerged from a clinically simple question with difficult implications. Why do some patients with the same target antigen, treated on the same platform, move toward remission while others do not? Here, the answer did not lie only in the CAR-T product or the leukemia blasts after treatment. It was present in the marrow before therapy began.
Response Divergence at the Three-Month Landmark
Fifteen patients with relapsed or refractory B-ALL entered a phase II, single-arm clinical trial and were split at a three-month landmark into 8 good responders and 7 poor responders. That separation defined the analysis. Clinical characteristics, genetic alterations, and pre-treatment marrow transcriptomes were then compared between groups, and an exploratory gene-expression score was built from the data.
The cohort was small, but the contrast was biologically coherent. Differential expression and enrichment analyses pointed in the same direction: poor responders showed upregulation of myeloid-related genes and a broader myeloid-like transcriptional state. In a disease defined by lymphoid blasts, the marrow appeared to be carrying a parallel innate-immune program that may have shaped how effectively CD19-directed cellular therapy could function.
A Marrow Environment That May Resist Immune Clearance
A myeloid-skewed transcriptional state in leukemia raises several plausible mechanisms. Monocytes and macrophages can alter cytokine tone, present antigen differently, and shape local immune suppression. They may also influence T-cell trafficking, activation, or persistence in the marrow niche. If the pre-treatment environment is already dominated by innate immune programs, CAR-T cells may face a less permissive setting even when the target antigen remains available on leukemic blasts.
Seven myeloid-related genes were tested across independent patient cohorts, B-ALL cell lines, and single-cell data from one patient. Three of those genes were expressed by the leukemic cells themselves, while four were mainly expressed by monocytes and macrophages. That split matters because it suggests the signature was not a single-cell artifact or a purely stromal phenomenon. It likely reflected interaction across compartments inside the marrow microenvironment.
The distinction is clinically important. A leukemic clone that expresses myeloid-associated genes may behave differently from one that is immunophenotypically classic but embedded in a myeloid-rich niche. Either route can weaken the practical effect of CD19 CAR-T therapy, but by different biological paths.
The Seven-Gene Score Separated Groups, Then Lost Precision Outside the Discovery Set
Investigators also derived a separate seven-gene least absolute shrinkage and selection operator, or LASSO, score from variance-based modeling. In the discovery cohort, the score cleanly separated good and poor responders, with an apparent area under the curve of 1.00. That result, however, did not hold at the same level once corrected for bias and tested externally. The bias-corrected and external estimates fell to roughly 0.57–0.59, and the external cohort produced an AUC of 0.592 with a 95% confidence interval of 0.356 to 0.828.
Those numbers tell a familiar story in translational oncology. A model can fit a small training set almost perfectly and still underperform when moved beyond the original patients. Here, the drop in discrimination does not erase the signal; it places it in context. The biology looks real enough to merit follow-up, but the score is not yet a clinically reliable selector for treatment decisions.
One sentence captures the limitation plainly: the discovery model was stronger than the validation performance.
Why This Signature May Matter Before Infusion Begins
The value of this study lies in timing. Most CAR-T analyses focus on infused cells, post-infusion expansion, or the leukemia itself after therapy pressure has already begun. This analysis looked earlier, before treatment, and found a host environment that may have been primed for resistance. In practical terms, that opens the possibility that baseline marrow transcriptomics could help identify patients who need closer monitoring, alternative bridging strategies, or future combination approaches designed to blunt myeloid-driven resistance.
Biologically, the finding fits with the broader concept that effective cellular immunotherapy depends on more than target recognition. CAR-T cells need access to the marrow, persistence after infusion, and enough functional space to expand. Myeloid cells can influence all three. They can shape suppressive cytokines, interfere with T-cell activity, and mark a tissue state less favorable to durable immune clearance.
The study did not establish a treatment-changing biomarker. It did define a candidate correlate of outcome rooted in the patient’s own marrow program before therapy started.
What the Cohort Can and Cannot Support
The evidence remains preliminary because the cohort included only 15 patients, and the response split was measured at a three-month landmark. Small numbers limit the stability of any expression signature, particularly one built with variable selection methods such as LASSO. Single-cell analysis was also limited to one patient, which makes the compartment-level interpretation informative but not definitive.
Even so, the convergence across analytic layers is notable. Differential expression, enrichment analysis, independent cohort testing, and single-cell localization all pointed toward the same broad theme: poor responders had a myeloid-like signature before CD19 CAR-T infusion. The signal was not confined to one dataset or one analytic approach.
For now, the most defensible reading is cautious. Pre-treatment transcriptomics may help explain why some patients with R/R B-ALL respond poorly to CD19 CAR-T therapy, and myeloid biology appears to be part of that explanation. Whether this signature can be refined into a usable biomarker, or whether it can guide intervention before infusion, will require larger cohorts and prospective validation.