The research team, led by Pei Zhu, Wenjuan Zhang, Zhihua Wu and senior author Lin Fu of the Second Affiliated Hospital of Guangzhou Medical University, set out to chart the polarization landscape of macrophages in bone marrow samples taken from AML patients after standard chemotherapy. Macrophages, the tissue-resident scavengers of the immune system, are famously plastic. They can adopt broadly inflammatory, microbicidal states or wound-healing, immunosuppressive ones depending on the signals they receive, and in tumors they often end up shielding malignant cells from both drugs and immune attack. Which of these states dominate the marrow after chemotherapy, and whether their abundance tracks with clinical response, had never been systematically resolved in AML.

To answer this, the investigators applied single-cell RNA sequencing to primary bone marrow samples from five patients in the GSE198681 dataset, three of whom had achieved complete remission after the 7+3 regimen and two of whom had not. Rather than forcing cells into the classical M1-versus-M2 dichotomy, which many immunologists now consider too crude to capture real biology, they used a computational framework called Scupa to classify macrophage polarization states in a data-driven manner. The analysis resolved five distinct polarization states, labeled Mac.a through Mac.e, each carrying its own transcriptional fingerprint and, as it turned out, its own functional program.

The first striking finding emerged when the researchers compared polarization scores at the level of individual cells. Mean Scupa scores for three of the five states, Mac.a, Mac.b and Mac.c, were significantly higher in the marrow of patients who had failed to achieve complete remission than in those who had responded. In other words, the macrophages populating the chemotherapy-resistant microenvironment were not a random mixture; they were skewed toward specific polarization programs. This cell-level shift provides some of the most direct evidence yet that the immune milieu of the marrow is reorganized in association with treatment failure, although the authors are careful to frame these as exploratory findings given the small sample size.

Functional annotation through Gene Ontology and KEGG pathway enrichment gave each state a biological identity. Mac.a macrophages were characterized by FoxO and PI3K-Akt signaling, pathways intimately involved in cell survival, metabolism and stress resistance. Mac.b macrophages lit up with NF-κB activity and inflammatory programs, and indeed showed the highest pro-inflammatory Hallmark activity of any state, suggesting a state of chronic, unresolved inflammation rather than productive immune activation. Mac.c macrophages stood out for oxidative phosphorylation, pointing to an energetically active, mitochondria-driven phenotype. The remaining states, Mac.d and Mac.e, were distinguished less by chemotherapy response and more by their correlations with immune-checkpoint biology and mutation burden in larger datasets, a point that would prove consequential later in the study.

To move from descriptive cell biology toward clinically usable tools, the team turned to high-dimensional weighted gene co-expression network analysis, or hdWGCNA. This method groups genes into co-expression modules that behave in concert across cells, and it identified ten such modules corresponding to the Mac.a through Mac.e polarization axis. Within each module, the researchers pinpointed hub genes, the most connected and therefore most representative members of each program. These hub genes became the raw material for a prognostic signature: a 32-gene panel designed to capture the macrophage polarization landscape of an AML sample and translate it into risk information.

The validation strategy is where the study gains much of its weight. A signature built from five patients is, on its own, little more than a hypothesis, so the authors tested the 32-gene panel in independent bulk AML cohorts. In the GEO-AML training cohort of 943 patients, the signature stratified patients with a hazard ratio of 2.51, meaning high-risk patients carried more than two and a half times the hazard of adverse outcome, with a p-value below 0.001. The result held up in the TCGA-AML validation cohort of 132 patients, where the hazard ratio was 1.66 and the p-value fell below 0.05. Replication across two independent cohorts with hundreds of patients transforms the signature from an interesting observation into a candidate biomarker worthy of prospective evaluation.

Perhaps the most therapeutically tantalizing results came from deconvolution of bulk tumor data. Using CIBERSORTx with a custom signature matrix derived from the single-cell data, the researchers estimated the proportions of each macrophage state in the TCGA-LAML bulk cohort and asked which immune features those proportions tracked with. Mac.d proportions showed the strongest association with expression of LAG-3, an immune checkpoint receptor that has emerged as a drug target in its own right, with a correlation coefficient of 0.581 and a p-value below 0.001. Meanwhile, Mac.e correlated strikingly with hypermutator signatures quantified by mutSig2, with a correlation of 0.86. These are the kinds of correlations that suggest combination strategies: if a particular macrophage state co-occurs with checkpoint expression, patients enriched for that state might be candidates for macrophage-directed therapy paired with anti-LAG-3 treatment.

The clinical context makes these findings timely. Immune checkpoint blockade has transformed solid oncology but has delivered only modest results in AML so far, and one leading explanation is that the bone marrow microenvironment is dominated by suppressive myeloid cells rather than the T-cell-inflamed states that predict response to checkpoint inhibitors. By showing that specific macrophage polarization states associate with checkpoint receptor expression and with clinical response to chemotherapy, the study offers a roadmap for selecting AML patients who might benefit from myeloid-targeted combinations rather than T-cell-centric approaches alone. It also provides the analytical machinery, the Scupa classification and the CIBERSORTx signature matrix, that other groups can apply to their own datasets to test these associations independently.

Caveats remain, and the authors acknowledge them. The single-cell discovery cohort comprised only five patients, so the links between Mac.a, Mac.b, Mac.c and chemotherapy response are explicitly exploratory and will need confirmation in larger, prospectively collected sample sets. Bulk deconvolution infers cell proportions rather than measuring them directly, and correlations between macrophage states and LAG-3 expression or mutational burden do not establish causation. Still, the study delivers a coherent arc from single-cell discovery to independent prognostic validation to therapeutic hypothesis, and it reframes chemotherapy resistance in AML as a property of an ecosystem rather than of malignant cells alone. If the Mac.a through Mac.e framework holds up in larger cohorts, the bone marrow macrophage may move from background player to actionable target in the fight against relapsed leukemia.

Subject of Research: Macrophage polarization heterogeneity in the bone marrow microenvironment of chemotherapy-treated acute myeloid leukemia

Article Title: Decoding macrophage polarization heterogeneity in post-chemotherapy AML microenvironments

Article References: Zhu, P., Zhang, W., Wu, Z., Deng, Y., Huang, X., & Fu, L. (2026). Decoding macrophage polarization heterogeneity in post-chemotherapy AML microenvironments. Cancer Immunology, Immunotherapy. https://doi.org/10.1007/s00262-026-04569-8

Image Credits: AI Generated

DOI: 10.1007/s00262-026-04569-8

Keywords: acute myeloid leukemia, macrophage polarization, single-cell RNA sequencing, bone marrow microenvironment, chemotherapy resistance, LAG-3, hdWGCNA, CIBERSORTx, prognostic signature, tumor immunology, NF-kB signaling, oxidative phosphorylation