Coagulation-associated genes stratify acute myeloid leukemia and spotlight SPINT2

Acute myeloid leukemia (AML) contains distinct transcriptional states associated with coagulation-related genes, and a seven-gene signature built from this activity separated patients with different survival outcomes in retrospective public cohorts, according to an integrated study combining single-cell RNA sequencing, bulk transcriptomics, machine learning, computational perturbation and laboratory experiments. The analysis identified three molecular subtypes: Cluster A had the most favorable prognosis, Cluster B an intermediate outcome and Cluster C the poorest. The model was developed in 547 patients from the TCGA-LAML and GSE37642 cohorts and evaluated in an independent cohort of 104 patients from GSE12417. High signature scores were associated with shorter overall survival in both cohorts, although discrimination was stronger in the development data than in external validation. SPINT2, a serine peptidase inhibitor, was the model’s most influential feature, and silencing it reduced proliferation and increased apoptosis in two AML cell lines.

At single-cell resolution, the researchers analyzed 10x Genomics data from 12 primary AML samples in GEO dataset GSE289435, identifying myeloid progenitors, monocytes, cycling erythroid cells, erythroid cells, B cells, T cells and dendritic cells. The relative activity of 39 predefined coagulation-associated genes, curated from Molecular Signatures Database gene sets and summarized with the UCell algorithm, was not uniform across the marrow landscape. Cycling erythroid and erythroid cells showed the highest relative activity, while myeloid progenitors, monocytes and subsets of immune cells displayed heterogeneous signals. Cell proportions also varied substantially between patients, with myeloid progenitors and monocytes predominating in most samples. Because UCell measures enrichment of a transcriptional gene set rather than thrombin generation, fibrin formation or another direct coagulation function, the result shows cell-state-specific expression of coagulation-associated programs, not proof that erythroid cells directly activate coagulation.

Consensus clustering of the bulk profiles identified three CARG-related subtypes in the 547-patient development cohort: Cluster A included 195 patients, Cluster B 217 and Cluster C 135. Their overall survival differed significantly in Kaplan–Meier analysis, with a log-rank P value below 0.001, and principal-component analysis separated Cluster C most clearly from the other groups. Gene set variation analysis revealed distinct biological programs: compared with Cluster B, Cluster A showed greater enrichment of glycolysis, fatty acid metabolism, peroxisomal, oxidative-phosphorylation, PI3K–AKT–mTOR, reactive-oxygen-species, apoptosis and MYC-target pathways. Cluster B was relatively enriched for heme metabolism, Hedgehog signaling and differentiation-related processes, whereas Cluster C combined heme and lipid metabolism with proliferation, cellular-stress and immune-regulatory programs. These are retrospective associations and do not establish that a particular metabolic pathway causes the observed survival differences.

Immune and stromal profiling further distinguished the subtypes. Cluster B had lower Immune, Stromal and ESTIMATE scores and higher estimated tumor purity than Clusters A and C. Single-sample gene set enrichment analysis detected differences across 23 immune-cell signatures, including activated CD4-positive and CD8-positive T cells, natural killer cells and regulatory T cells. The 243 genes shared in pairwise subtype comparisons were enriched for blood coagulation, hemostasis, platelet alpha-granules, fibrillar collagen and proteolytic regulation; KEGG analysis highlighted platelet activation, NF-kappa B signaling and the pentose phosphate pathway. These immune scores reflect relative transcriptional enrichment rather than direct measurements of cell abundance or function, while the shared-gene results indicate that coagulation-related expression accompanies broader differences among the subtypes.

For prognostic modeling, the investigators screened the 243 shared differentially expressed genes with univariable Cox regression, retaining 30 survival-associated candidates. They then compared 68 configurations drawn from nine survival-machine-learning classes, including LASSO-Cox, Ridge-Cox, elastic net, stepwise Cox, CoxBoost, survival support vector machine, supervised principal components, partial least-squares Cox regression and gradient boosting. Model selection relied on the development-cohort concordance index, and the external cohort was kept separate from feature selection, tuning and coefficient estimation. The locked final model combined seven genes: AGTPBP1, TMEM38B, SPINT2, SH3BP5, LAPTM4B, HIST1H4H and SLC35C2. Its score was the weighted sum of expression values, with coefficients of −0.0074, 1.0348, 2.0164, 0.7373, 0.5050, 0.2137 and −1.0218, respectively.

Patients divided at each cohort’s median score had significantly different survival: the development-cohort log-rank P value was below 0.001, and the external-validation value was 0.035. Time-dependent areas under the curve for predicting three-, five- and seven-year overall survival were 0.751, 0.719 and 0.739 in development, compared with more modest values of 0.644, 0.681 and 0.647 in validation. The score increased progressively from Cluster A through Cluster C, matching the subtype survival gradient. Post hoc SHapley Additive Explanations identified SPINT2 as the largest contributor, with a mean absolute SHAP value of 0.125, followed by SH3BP5, SLC35C2, LAPTM4B, TMEM38B, AGTPBP1 and HIST1H4H. Higher SPINT2 expression generally shifted the model’s linear predictor toward greater estimated risk.

In clinical analyses, the signature remained associated with overall survival after adjustment for age and FAB M3 status. For each one-unit increase in the score, the multivariable hazard ratio for death was 2.085, with a 95% confidence interval of 1.716 to 2.533 and P value below 0.001. Age remained significant, but FAB M3 status did not after adjustment. A nomogram combining age and the signature estimated three-, five- and seven-year survival, with Brier scores of 0.182, 0.172 and 0.160 and expected-to-observed ratios near 1. Decision-curve analysis indicated greater net benefit from the combined model than from age alone over clinically relevant thresholds. However, the retrospective cohorts lacked consistently complete cytogenetic and molecular information for reliable comparison with contemporary ELN 2022 risk classification, leaving incremental value over established genetic stratification unresolved.

The risk groups also had different immune-expression patterns and ex vivo drug responses in Beat AML 2.0 waves 1 through 4. High-score samples showed greater relative enrichment of activated CD4-positive T cells, activated CD8-positive T cells, CD56-bright natural killer cells and plasmacytoid dendritic-cell signatures, alongside reduced macrophage, myeloid-derived suppressor-cell and regulatory T-cell signatures. Because these were expression-based signatures, they do not demonstrate that the apparent immune differences caused survival or treatment sensitivity. Lower probit area-under-the-curve values mean greater ex vivo sensitivity. Higher CARG scores correlated with higher response-AUC values, and therefore lower sensitivity, for venetoclax, GSK690693, palbociclib, sorafenib and SB-431542. Panobinostat and selumetinib showed the opposite pattern, with higher scores associated with greater sensitivity. These retrospective associations were not presented as evidence for selecting drugs in patients.

SPINT2 was prominent in the cell atlas, with relatively high expression in cycling erythroid cells, B cells and dendritic cells. Two independent whole-atlas virtual-knockout algorithms, scTenifoldKnk and GenKI, produced only partial agreement at the individual-gene level but converged on pathway-level associations involving erythroid and heme biology, fatty acid metabolism, oxidative phosphorylation, mitochondrial function and cellular stress responses. Genes consistently prioritized by both methods included SLC4A1, GYPC, BLVRB, SLC25A37, HBD, PRDX2 and TUBA1B. These are computationally predicted transcriptional associations; because the perturbations were performed across the integrated atlas, they do not establish lineage-specific, cell-autonomous or direct downstream effects. An independent CellChat analysis found MIF signaling to be the largest inferred communication pathway, with interactions involving CD74 paired with CXCR4 or CD44, but that network was not generated as a downstream SPINT2 mechanism.

Laboratory experiments provided the study’s direct functional evidence. SPINT2 messenger RNA and protein were detectable across six AML cell lines, with the highest messenger-RNA level in MOLM-13 cells; THP-1 and MOLM-13 were selected for knockdown using two independent short hairpin RNAs. In both lines, silencing SPINT2 lowered CCK-8 proliferation signals over time, reduced EdU incorporation into DNA and increased Annexin V-positive apoptosis relative to non-targeting controls, with at least three independent biological experiments. The results support a role for SPINT2 in maintaining AML-cell growth and survival in vitro. They do not show that the computationally inferred erythroid, heme or oxidative pathways mediate the phenotype, nor do they establish therapeutic effectiveness in patients. No erythroid leukemia models, in vivo experiments or lineage-resolved mechanistic tests were included, leaving the biology of SPINT2 and the clinical value of the seven-gene score to be evaluated in future prospective studies.

Subject of Research: Coagulation-associated transcriptomic heterogeneity, prognosis and SPINT2 function in acute myeloid leukemia

Article Title: Integrated single-cell and bulk transcriptomic profiling reveals coagulation-associated heterogeneity and prognostic SPINT2 in acute myeloid leukemia

Article References: Wang, Z., Han, M., Zhang, D., Cui, L., Zhang, L., & Wang, B. (2026). Integrated single-cell and bulk transcriptomic profiling reveals coagulation-associated heterogeneity and prognostic SPINT2 in acute myeloid leukemia. Frontiers in Cell and Developmental Biology, 14, Article 1971923. https://doi.org/10.3389/fcell.2026.1971923

Image Credits: AI Generated

DOI: 10.3389/fcell.2026.1971923

Keywords: acute myeloid leukemia, coagulation-associated genes, SPINT2, single-cell RNA sequencing, bulk transcriptomics, prognostic scoring, tumor immunity, AML metabolism, drug sensitivity, machine learning, cell proliferation, apoptosis