Donor lymphocyte infusion (DLI) is an established immunotherapy for leukemia that has relapsed after an allogeneic stem cell transplant (allo-SCT) - a procedure in which a patient receives blood stem cells from a healthy donor. In DLI, additional immune cells (lymphocytes) from the original donor are infused into the patient, with the goal of reactivating an immune attack on the remaining leukemia cells.
The therapeutic principle underlying both allo-SCT and DLI is the graft-versus-leukemia (GvL) effect: donor-derived T cells can recognize and kill the patient's leukemia cells, which carry proteins foreign to the donor's immune system. DLI induces durable molecular remissions in approximately 75% of patients with relapsed chronic myelogenous leukemia (CML) after allo-SCT - a remarkable success rate for a cancer that has already come back after transplant.
Despite this clinical effectiveness, the cellular mechanisms of DLI response and resistance remain poorly understood. Why does DLI work for some patients and fail for others? What T cell states are present in responding versus non-responding patients? Answering these questions requires detailed, high-resolution profiling of immune cells in the bone marrow before and after treatment.
This study leveraged single-cell RNA sequencing, chromatin accessibility profiling, and T cell receptor sequencing from serial bone marrow biopsies of CML patients treated with DLI - creating one of the most detailed immunological maps of any human cancer immunotherapy response ever assembled.
The study analyzed 12 patients treated with CD8-depleted DLI for relapsed CML - six long-term responders (who achieved molecular remission) and six non-responders (who showed no measurable response). Serial bone marrow biopsies were collected at a median of three time points per patient - before and at multiple intervals after DLI treatment - spanning up to 4.5 years of follow-up.
From each biopsy, the researchers performed single-cell RNA sequencing (scRNA-seq), which measures the gene expression profile of each individual cell rather than averaging across millions of cells. This reveals the full spectrum of cell types and functional states present in the sample. In total, 381,462 cells passed quality filters across all samples.
For T cells specifically, 87,939 T cells were clustered into 43 distinct states, each defined by a unique gene expression signature. Alongside RNA sequencing, ATAC-seq (a method that maps open regions of chromatin, indicating which genes are accessible for transcription) was performed on sorted T cell populations, providing a complementary view of the epigenetic regulatory landscape.
The team developed a novel computational tool called Symphony, a Bayesian model that integrates data across patients and time points, accounting for patient-to-patient variability and batch effects to reveal the shared biological dynamics underlying treatment response. A hierarchical Gaussian Process (GP) regression model was used to model how T cell populations changed over time in responders and non-responders.
A striking finding was that patients who would go on to respond to DLI had significantly higher phenotypic diversity in their T cells before treatment began - meaning a broader range of different T cell states was present in their bone marrow at baseline. This diversity was quantified using a measure called phenotypic volume, which captures how many distinct and independent gene expression programs co-exist in the T cell population.
Within this pre-treatment diversity, four specific T cell clusters (numbered 4, 14, 21, and 27) were consistently enriched in future responders before DLI. These clusters were predominantly CD8+ T cells expressing genes associated with T cell activation (CD160, HAVCR2, CD38) and cytotoxicity (CRTAM, GNLY, GZMK, GZMB) - exactly the functional proteins needed to kill leukemia cells.
Importantly, no single T cell cluster was consistently enriched in non-responders before treatment. Instead, non-responders showed a heterogeneous pattern - different patients had different types of dysfunctional T cells. This suggests that DLI resistance has multiple distinct causes rather than a single shared mechanism, which has important implications for developing therapies to overcome it.
The pre-treatment enrichment of activated, cytotoxic T cells in responders suggests these patients already possessed a primed immune environment in the bone marrow before DLI was given. DLI may then have amplified and expanded this pre-existing response, whereas in non-responders, the immune environment was too dysfunctional to support such expansion.
After DLI treatment, responding patients showed a consistent pattern: certain 'early-differentiated' T cell populations expanded dramatically, while certain 'late-differentiated' T cell populations contracted. The expanding populations expressed genes associated with self-renewal (TCF7, IL7R, SATB1), lymph node homing (SELL, CCR7), and long-term survival - all hallmarks of T cells with the capacity for sustained, durable immune responses.
The expanding T cells closely resembled a subset called precursor exhausted T cells (TPEX) - a population originally identified in studies of chronic viral infection and cancer in mice. TPEX cells are characterized by expression of TCF7 (a key transcription factor for stemness), intermediate levels of exhaustion markers, and the ability to proliferate and give rise to effector cells. In contrast, the contracting pre-DLI clusters resembled terminally exhausted T cells (TEX), which express high levels of co-inhibitory receptors and are less capable of sustained responses.
This finding - that resolution of T cell exhaustion in responding patients was driven not by changing gene expression within cells, but by shifts in which cell types were present (expanding TPEX-like cells replacing TEX-like cells) - is a key conceptual advance. It reframes how immunotherapy works: successful treatment enriches for a stem-like T cell population that can continuously regenerate effector cells.
Using paired single-cell protein measurements (CITE-seq), the researchers confirmed that pre-DLI T cells in responders co-expressed multiple co-inhibitory receptors (CTLA4, LAG3, TIGIT, TIM3, PD1, 2B4) - a classic exhausted signature. In contrast, post-DLI expanding T cells expressed co-stimulatory markers (OX40, CD28) and memory markers (CD62L, IL7RA) consistent with TPEX cells that can sustain long-term anti-leukemia immunity.
To understand what drives the difference between TPEX and TEX cells, the team used ATAC-seq chromatin accessibility data to infer gene regulatory networks - the set of transcription factors and their target genes that define each T cell state. The computational tool Symphony was specifically designed to infer these regulatory circuits from sparse single-cell data.
TEX cells were governed by transcription factors including TOX, PRDM1, and ID2, which are known to drive terminal exhaustion by locking cells into a dysfunctional state. TPEX cells were governed by TCF7, ID3, and LEF1, transcription factors associated with stemness and self-renewal capacity. These regulatory programs are epigenetically stable, meaning they are maintained not just by gene expression but by the physical accessibility of the relevant DNA regions.
Tracking the T cell receptor (TCR) clonotypes - unique sequences that identify individual T cell lineages - revealed that the TPEX-like cells expanding after DLI were not primarily coming from the infused donor T cells. Instead, they were pre-existing clones from the patient's bone marrow and newly recruited clones, suggesting that DLI works by unleashing and expanding T cells already present in the leukemic microenvironment rather than simply replacing them with donor cells.
A case study of a patient with chronic lymphocytic leukemia (CLL) who received DLI showed the same kinetics of expanding TPEX-like and contracting TEX-like cells observed in CML responders. When this patient's disease relapsed 11 years later, the T cell states reverted back to the pre-DLI pattern - directly linking the T cell composition to disease control. This case extends the model beyond CML to other leukemias.
While response to DLI was associated with a common molecular program (expansion of TPEX-like cells), resistance was heterogeneous. Statistical analysis identified distinct gene programs underlying non-response, including hypoxia (oxygen deprivation), anergy (a state of T cell unresponsiveness), and peripheral and deletional tolerance (processes by which the immune system normally avoids attacking its own tissues).
This diversity of resistance mechanisms has important therapeutic implications: there is no single 'DLI resistance pathway' that one drug could overcome. Instead, different patients may require different interventions to restore T cell function - for example, targeting hypoxia in the tumor microenvironment, providing co-stimulatory signals to break anergy, or blocking tolerance mechanisms.
The finding that even non-responders showed significant T cell remodeling after DLI - with much larger increases in phenotypic diversity than responders - suggests their immune systems are not simply unresponsive. Rather, something prevents this remodeling from translating into effective leukemia killing. Identifying what prevents this conversion from immune activation to tumor clearance is a key question for future research.
The TPEX/TEX framework identified here extends the relevance of exhausted T cell biology beyond checkpoint blockade - the type of immunotherapy that blocks inhibitory receptors like PD-1 - to adoptive cellular therapies such as DLI and CAR-T cells. This suggests that the fundamental principles of effective T cell responses may be shared across different immunotherapy modalities.
The Bayesian computational framework introduced in this study - particularly the Symphony tool for integrating data across time points and patients - is broadly applicable beyond leukemia and DLI. It provides a general approach for analyzing longitudinal single-cell data in any disease context where immune responses evolve over time in response to therapy.
Practically, this research suggests that measuring T cell composition before DLI could help predict who will respond. Patients with a pre-existing reservoir of activated, cytotoxic T cells in the bone marrow may be more likely to benefit from DLI, while those with uniformly dysfunctional T cell landscapes might benefit from pre-treatment interventions that restore immune competence before receiving DLI.
The observation that TPEX-like cells expand primarily from pre-existing bone marrow clones rather than from infused donor cells suggests that strategies to enhance endogenous T cell function - such as PD-1 blockade or cytokine support - might be more effective than simply increasing the dose of donor cells in patients who fail initial DLI.
By integrating single-cell transcriptomics, chromatin accessibility, T cell receptor clonality, and protein measurements across 4.5 years of treatment in the same patients, this study provides an unprecedented view of how the immune system evolves during cancer therapy. The scale and longitudinal depth of the dataset sets a new benchmark for mechanistic immunotherapy research in humans.
The central finding - that effective anti-leukemia immunity depends on a stem-like T cell subset with defined transcription factor circuitry - provides a molecular explanation for durable remissions and a framework for understanding why some patients fail. It also directly connects insights from preclinical models of viral infection and mouse cancer to human clinical outcomes.
Looking forward, this work supports developing biomarkers that measure TPEX:TEX ratios in bone marrow biopsies as predictors of DLI response, and designing combination strategies that convert dysfunctional T cells into TPEX-like states before or during immunotherapy. The ultimate goal is to identify which patients will benefit from DLI and tailor treatment to maximize the graft-versus-leukemia effect while minimizing harmful graft-versus-host disease.