Dissecting PCD-driven molecular landscapes in AML: a multi-omic framework for prognostication and therapeutic targeting Authors: Jing Liu, Fangmin Zhong, Xiaozhong Wang, Liuqing Xu, Zanwei Tu, Xiaofang Cheng, Linlin Zhang, Guangyao Kong Published: 2026 ABSTRACT AbstractAcute myeloid leukemia (AML) remains a molecularly heterogeneous malignancy with poor prognosis, necessitating robust biomarkers for risk stratification. By integrating single-cell RNA-seq and multiomics data from 2,680 AML samples across 10 cohorts, we demonstrate that dysregulated programmed cell death (PCD) pathways are significantly elevated in AML cells and correlate with adverse outcomes. Unsupervised clustering identified two PCD-driven subtypes: Subtype A exhibits high PCD activity, an immunosuppressive microenvironment (M2 macrophages, upregulated PD-L1/HAVCR2), increased somatic mutations (NPM1/DNMT3A/FLT3), and poor survival, while Subtype B shows lower PCD activity, immune-active features, and better prognosis. As an exploratory analysis, subtype-specific therapeutic vulnerabilities were revealed: Subtype A displays predicted sensitivity to immune checkpoint inhibitors (anti-PD-1) and tipifarnib, whereas Subtype B responds better to cytarabine/doxorubicin. We developed PCDRScore—a prognostic model incorporating 13 PCD-related genes via machine learning—which outperformed existing models (higher C-index) in 10 validation cohorts and remained independent of clinicopathological factors. A nomogram combining PCDRScore, age, and cytogenetic risk enhanced clinical applicability. Crucially, experimental validation confirmed significant upregulation of key model genes (HIP1, SQLE, VNN1) in AML patient samples and cell lines (P < 0.05), reinforcing the model’s biological relevance. These findings establish PCD dysregulation as a central axis of AML heterogeneity, providing a framework for precision risk stratification and hypothesis-generating immunophenotype-guided therapy.Supplementary InformationThe online version contains supplementary material available at 10.1007/s10238-026-02205-4.Keywords: Acute myeloid leukemia, programmed cell death, molecular subtype, Prognostic model Introduction Acute myeloid leukemia (AML) is a hematologic malignancy caused by the malignant clonal proliferation of myeloid hematopoietic stem cells 1. The highly heterogeneous molecular characteristics of AML contribute to significant variability in clinical outcomes 2. Despite advances in chemotherapy, targeted therapies, and hematopoietic stem cell transplantation, the 5-year overall survival rate for AML patients remains below 30%, with substantial differences in treatment response and relapse risk across molecular subtypes 3. In recent years, programmed cell death (PCD), a critical mechanism regulating cellular homeostasis, has emerged as a research focus in AML pathogenesis and treatment 4. Dysregulation of PCD pathways, including apoptosis, necroptosis, and ferroptosis pathways, not only influences leukemia cell proliferation and drug resistance but also correlates closely with patient prognosis 5–8. Recent breakthroughs in genomics, transcriptomics, and epigenetics have elucidated the associations between key driver mutations (e.g., FLT3, NPM1, and TP53) and clinical outcomes in AML 9,10. However, the role of the dynamic imbalance in PCD pathways remains systematically uncharacterized. Studies have demonstrated that AML cells evade apoptosis through the upregulation of antiapoptotic proteins (e.g., BCL-2 and MCL-1) or the suppression of proapoptotic signals (e.g., the CASPASE family) 11–13. Concurrently, aberrant expression of ferroptosis-related metabolic regulators (e.g., GPX4 and ACSL4) and necroptosis mediators (e.g., RIPK1/RIPK3) may further impact chemotherapy sensitivity and immune microenvironment remodeling in AML 14–17. Nevertheless, current research predominantly focuses on individual PCD pathways, and an integrative analysis of multidimensional regulatory networks is lacking. The crosstalk between distinct programmed cell death modalities and their subtype-specific roles in AML remain poorly understood. This study aims to systematically characterize PCD-related molecular subtypes in AML through bioinformatics approaches and construct a prognostic prediction model on the basis of multiomics data. By integrating the gene expression profiles of apoptosis, autophagy, and other PCD pathways with clinical outcomes, this research not only reveals the biological significance of PCD-related molecular subtypes in AML but also provides a theoretical foundation for the development of personalized therapeutic strategies. These findings are expected to advance the application of precision medicine in AML diagnosis and treatment. To ensure the translational relevance of our prognostic model, we experimentally validated the expression patterns of key signature genes in clinical AML samples and cell lines, bridging bioinformatics predictions with biological validation. Materials and methods Data acquisition and preprocessing A total of 2,680 AML samples with clinical annotations were aggregated from ten independent cohorts. These included seven Gene Expression Omnibus (GEO) datasets (GSE10358-GPL570, GSE12417-GPL96, GSE12417-GPL570, GSE37642-GPL96, GSE37642-GPL570, GSE71014-GPL10558, and GSE106291), 173 TCGA-LAML samples from the UCSC Xena database, the HOVON cohort from ArrayExpress 18, and RNA sequencing data from the Beat AML project (Vizome database) 9. Detailed genomic platform used, sample size, clinical context, and specific application in each analytical step—including risk score derivation, molecular subtyping (clustering), model training, external validation, and therapeutic response prediction, are provided in Supplementary Table S1. To address platform-specific variability, raw CEL files from the GPL96 and GPL570 microarray platforms were normalized via robust multiarray averaging (RMA). For RNA-seq datasets (TCGA-LAML, GSE106291, and Beat AML), raw read counts were converted to transcripts per million (TPM) using the formula: TPM = (reads per gene / gene length in kb) / (sum of all (reads per gene / gene length in kb)) × 1,000,000. Gene lengths were obtained from the respective genome annotations (GENCODE v22 for TCGA-LAML, and corresponding annotations for other datasets). The somatic mutation data, and copy number alteration data were retrieved from the TCGA portal (https://portal.gdc.cancer.gov/). PCD genes were derived from a previous study 19,20 and are listed in full in Supplementary Table S2. TCGA-LAML was selected as the discovery cohort due to its historical role as a benchmark for AML prognostic modeling, its homogeneous treatment background (standard induction chemotherapy), and to enable direct comparison with previously published signatures. The Beat AML cohort, despite its larger size and richer annotation, was reserved as an external validation cohort to test model generalizability across different treatment eras and patient populations, thereby avoiding overfitting and ensuring robust performance assessment. Single-cell RNA sequencing analysis Single-cell RNA-seq data (GSE116256) spanning 21 distinct AML cell populations were obtained from GEO 21. Preprocessing adhered to stringent quality control criteria: cells with a mitochondrial gene content exceeding 20%, UMI counts below 200, or genes detected outside the range of 1,500–6,000 were excluded. Genes not expressed in at least three cells were filtered out. Library size normalization was applied, followed by selection of the top 2,000 highly variable genes on the basis of the coefficient of variation. Dimensionality reduction was performed via principal component analysis (PCA) and uniform manifold approximation and projection (UMAP). Cell type annotations were obtained directly from the original publication 21, which used canonical marker genes and known differentiation trajectories to define 21 distinct cell populations, including normal hematopoietic progenitors and leukemic malignant cells. Functional enrichment and cellular composition estimation Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses were executed via the clusterProfiler package (v4.8.3) 22. Pathway activity scores were quantified via gene set variation analysis (GSVA) 23. PCD scores for 18 programmed cell death modalities were computed using the GSVA package (v1.48.3) with default parameters (method=’gsva’, kcdf=’Gaussian’, min.sz = 1, max.sz = Inf). Expression data were log₂-transformed prior to GSVA analysis. For deconvolution of bulk RNA-seq data, a reference matrix of 21 AML cell types was generated by randomly sampling 200 cells per type from single-cell profiles. The CIBERSORT algorithm, which employs a support vector regression model, was used to estimate the relative proportions of hematopoietic stem cells (HSCs), HSC-like progenitors (Prog), Prog-like, granulocyte–monocyte–progenitor (GMP), GMP-like, promonocyte (ProMono), ProMono-like, monocyte (Mono), Mono-like, conventional dendritic cell (cDC), cDC-like, plasmacytoid dendritic cell (pDC), early erythroid progenitor (earlyEry), late erythroid progenitor (lateEry), progenitor B-cell (proB), mature B-cell (B), plasma cell (plasma), naïve T-cell (T), cytotoxic T lymphocyte (CTL), and natural killer cell (NK) populations in the bulk datasets 24. Molecular subtype identification Unsupervised consensus clustering was performed via the ConsensusClusterPlus package (v1.64.0) with 1,000 resampling iterations (80% item resampling) to ensure robust classification of AML samples into molecular subtypes 25. Clustering was performed on the PCD scores calculated within the TCGA-LAML cohort. We used Euclidean distance, k-means clustering, and set maxK = 9 with resampling proportion pItem = 0.8 and pFeature = 1. The optimal k = 2 was determined by consensus cumulative distribution function (CDF) and delta area plots. Prognostic model development PCD-related genes (PCDRGs) associated with survival (P < 0.05 in ≥ 5 cohorts) were identified through univariate Cox regression. Ten machine learning algorithms—CoxBoost, stepwise Cox, Lasso, Ridge, elastic net (Enet), survival support vector machines (survival-SVMs), generalized boosted regression models (GBMs), supervised principal components (SuperPCs), partial least Cox (plsRcox), and RSF—were systematically combined into 118 distinct models. Model training and validation employed a 10-fold cross-validation framework, with random seed set for reproducibility. Nested cross-validation was employed, where the inner loop performed feature selection and hyperparameter tuning, and the outer loop evaluated model performance. This design ensures that feature selection is performed independently on each training fold, preventing information leakage. Performance was evaluated via the concordance index (C-index) across primary (TCGA-LAML) and external validation cohorts. The final PCDRScore model was constructed using genes selected by CoxBoost with coefficients optimized via Ridge regression on the full training set: Patients were stratified into high- and low-PCDRScore groups using an optimal cutoff determined by survival analysis. Therapeutic sensitivity and immune response prediction Drug sensitivity (IC50 values) for conventional AML therapies was predicted via the pRRophetic package (v0.5). The drug panel and source are listed in Supplementary Table S3. The underlying predictive model employs ridge regression to estimate IC50 values from bulk tumor gene expression profiles, leveraging a pan-cancer training framework built on transcriptomic and pharmacologic data from over 1,000 cancer cell lines and more than 200 bioactive compounds. Prior to prediction, input expression data were harmonized to the reference distribution of the training cohort via quantile normalization, as implemented in pRRophetic. Importantly, all predictions are hypothesis-generating and should be interpreted with caution, as they are derived from models trained exclusively on in vitro cell-line data and not validated in primary patient samples. SubMap analysis was used to assess the likelihood of a response to anti-PD-1 and anti-CTLA4 immune checkpoint inhibitors. Tumor immune evasion potential was quantified via the Tumor Immune Dysfunction and Exclusion (TIDE) algorithm. Clinical sample collection, cell culture and gene expression detection This study was approved by the Ethics Committee of the Second Affiliated Hospital of Nanchang University. The remaining samples after patient examinations were collected, and the sample processing conformed to the standards. We collected peripheral blood samples from 12 AML patients (blast percentage > 70%) and 12 healthy controls. While peripheral blood may not fully recapitulate bone marrow disease biology, the high blast fraction minimizes confounding from non-leukocytic cells, and the findings provide orthogonal support for the bioinformatics predictions. AML cell lines MV4-11 and Molm-14 were cultured in RPMI1640 medium supplemented with 10% fetal bovine serum and 1% penicillin and streptomycin. HS-5 cells were used as a stromal niche reference to compare gene expression profiles between AML cells and the bone marrow microenvironment cell type, and cultured as adherent cells in DMEM complete medium containing 10% fetal bovine serum and 1% penicillin and streptomycin. This comparison does not represent a lineage-matched normal hematopoietic control but rather addresses leukemia-stroma biology. The culture conditions for all were 37 °C and 5% carbon dioxide in a humidified incubator. The expression of key genes in the model was detected by quantitative real-time polymerase chain reaction (RT-qPCR). RT-qPCR was performed using a TAKARA kit on an ABI7500 instrument, with GAPDH as the internal reference gene, and normalization via the 2 − ΔΔCt method. Primers are shown in Supplementary Table S4. Statistical analysis All analyses were conducted in R (v4.3.0). Differences between groups were evaluated via Wilcoxon rank-sum (two groups) or Kruskal–Wallis tests (multiple groups), with significance thresholds set at *P < 0.05, **P < 0.01, and ***P < 0.001. Results Evaluation of the PCD activity characteristics of AML cells via single-cell data We first evaluated the PCD activity characteristics across different cell types in the AML single-cell dataset GSE116256. The GSE116256 cohort comprises 21 cell types (Fig. 1A), including six leukemic malignant cell populations (cDC-like, GMP-like, HSC-like, Mono-like, Prog-like, and ProMono-like) (Fig. 1B). Compared with normal cells, AML malignant cells presented higher scores for all PCD types except for alkaliptosis, reflecting their overall more active state (Fig. 1C and D). Notably, HSC-like cells displayed low PCD scores across all categories, which was consistent with their quiescent status (Fig. 1C). We subsequently calculated PCD scores for each AML sample in the TCGA-LAML cohort. Univariate Cox regression analysis revealed that most PCD scores were significantly associated with poor prognosis in AML patients (Fig. 1E). Furthermore, significant positive correlations were observed among the PCD scores (Fig. 1F), indicating PCD dysregulation in AML cells, which may synergistically drive disease progression. Characteristics of the programmed cell death (PCD) scores of AML cells. (A) Cell classification features of the AML single-cell dataset. (B) Distribution of normal and malignant cells in the AML single-cell dataset. (C) Differences in PCD scores among different cell types. (D) Differences in the PCD scores between AML malignant cells and normal cells. (E-F) Univariate prognostic analysis (E) and correlation analysis (F) of 18 PCD scores in the TCGA-LAML cohort Identification and clinical characterization of PCD-related molecular subtypes To elucidate the relationships between PCD pathways and AML clinical features, we applied an unsupervised clustering algorithm based on 18 PCD scores to identify distinct PCD-related subtypes. The area under the CDF curve plateaued at k = 2, indicating that increasing k beyond 2 did not substantially improve consensus (Figure S1A). The relative change in CDF area was maximal at k = 2 and decreased thereafter, supporting two-cluster solution (Figure S1B). Heatmap of consensus matrices showed clear diagonal blocks with very small off-diagonal cross-assignment for k = 2, indicating high stability (Fig. 2A). AML patients were classified into two distinct subtypes (Subtypes A and B), with PCA validating the clustering reliability (Fig. 2B). Survival analysis revealed that Subtype A patients had significantly worse prognoses than did Subtype B patients (Fig. 2C). Strikingly, nearly all PCD scores were higher in Subtype A (Fig. 2D). Clinical feature analysis revealed significant differences in FAB classification and peripheral blood white blood cell (WBC) counts between the two subtypes (P < 0.05) (Fig. 2E). Subtype A included a greater proportion of M4 and M5 FAB subtypes, whereas Subtype B had a significantly greater proportion of M0 and M3 subtypes (Fig. 2F). Additionally, Subtype A patients had higher WBC counts and mortality rates (Fig. 2G and H). Given the enrichment of M3 (promyelocytic leukemia) subtype in Subtype B, which has excellent prognosis with modern ATRA/ATO-based therapy, we performed additional analyses to exclude this potential confounder. After excluding M3 patients, the survival difference between Subtype A and Subtype B remained significant (HR = 1.55, 95% CI = 1.05–2.31, P = 0.0025; Figure S2). These analyses demonstrate that the prognostic value of the PCD subtypes is not solely attributable to the favorable outcomes of M3 patients. Identification of PCD-related subtypes. (A) Consensus clustering algorithm identified PCD-related subtypes on the basis of key PCD scores. (B) The PCA algorithm verified the reliability of the clustering. (C-I) Differences in patient prognosis (C), PCD scores (D), and clinicopathological features (E-I) among the PCD-related subtypes Immune microenvironment and genomic differences between PCD-related subtypes We further analyzed immune- and pathway-related differences between the subtypes. Immune infiltration analysis revealed that Subtype B cells presented increased proportions of naïve B cells, plasma cells, CD8 + T cells, resting memory CD4 + T cells, follicular helper T cells, and resting/activated mast cells, whereas Subtype A cells were enriched with monocytes and M2 macrophages (Fig. 3A). GSVA enrichment analysis revealed that most cancer-related hallmark pathways were significantly upregulated in Subtype A (Fig. 3B). Immune checkpoint gene expression analysis revealed significant upregulation of PD-L1, HAVCR2, PD-L2, CD80, CD86, and TNFRSF9 in Subtype A (Fig. 3C). To further characterize the immune phenotype of the two PCD subtypes, we compared the expression of MHC class I genes (HLA-A, HLA-B, HLA-C) and MHC class II genes (HLA-DRA, HLA-DRB1, HLA-DPA1, HLA-DPB1, HLA-DQA1, HLA-DQB1) between Subtype A and Subtype B. Subtype A showed significantly higher expression of all examined MHC class I and class II genes compared to Subtype B (all P < 0.001) (Figure S3). Somatic mutation analysis higher mutation frequencies in Subtype A, including genes such as NPM1, DNMT3A, FLT3, and KIT (Fig. 3D). These results collectively suggest that Subtype A has a more immunosuppressive microenvironment, with high immune checkpoint expression promoting immune evasion, alongside elevated oncogenic pathway activity and somatic mutation frequencies, contributing to its poor prognosis. Furthermore, the elevated MHC expression in subtype A may reflect a state of chronic immune activation or inflammatory stress, which could lead to T-cell exhaustion rather than effective antigen presentation. Analysis of differences in biological characteristics among PCD subtypes. (A-D) Differences in immune cell infiltration (A), tumor marker pathway activity (B), immune checkpoint expression (C), and somatic mutations (D) among subtypes Predicted therapeutic implications for PCD-related subtypes As an exploratory analysis, we predicted treatment sensitivity to immunotherapy and chemotherapy. According to the TIDE algorithm, a greater proportion of patients in Subtype A patients were likely to respond to immunotherapy (Fig. 4A). SubMap analysis further suggested that Subtype A patients might benefit from anti-PD-1 therapy (Fig. 4B). Survival analysis confirmed that Subtype A patients with high PD-1 and PD-L1 expression had the worst survival outcomes (Fig. 4C and D). In terms of predicted chemotherapy sensitivity, tipifarnib, bexarotene, and AICAR presented significantly higher IC50 values in Subtype B, indicating greater sensitivity in Subtype A (Fig. 4E and F). Conversely, Subtype B was more sensitive to GW843682X, and bleomycin (Fig. 4E and F). Among the commonly used AML chemotherapies, Subtype B exhibited increased sensitivity to cytarabine and doxorubicin, whereas midostaurin sensitivity did not differ significantly between the subtypes (Fig. 4G). Prediction of the response to chemotherapy and immunotherapy in different PCD subtypes. (A) Predicted differences in the proportions of patients who responded to immunotherapy among subtypes. (B) Prediction of the response to anti-PD-1 and anti-CTLA4 immunotherapy in patients of different subtypes. (C-D) Prognostic analysis of the combination of PD-1/PD-L1 high- and low-PD-L1 expression groups with molecular subtypes. H-PD-1 + Subtype A (n = 64); H-PD-1 + Subtype B (n = 54); L-PD-1 + Subtype A (n = 19); L-PD-1 + Subtype B (n = 24). H-PD-L1 + Subtype A (n = 41); H-PD-L1 + Subtype B (n = 29); L-PD-L1 + Subtype A (n = 43); L-PD-L1 + Subtype B (n = 48). (E-G) Prediction of the sensitivity of different subtypes to chemotherapy drugs Prognostic value analysis of PCD-related genes Given the significant association between PCD pathways and AML prognosis, we investigated the prognostic value of PCD-related genes (PCDRGs). Univariate Cox regression revealed that 30 PCDRGs were significantly associated with AML prognosis. Using 10 machine learning algorithms, the CoxBoost and ridge combination achieved the highest average C-index across cohorts (Fig. 5A). CoxBoost selects 13 key PCDRGs and ridge regression optimized coefficients on the basis of minimal misclassification error (logλ) (Fig. 5B and C). The final PCD-related score (PCDRScore) model included ATP13A2, HIP1, IL7, ATP6V0E2, DDIT4, SQLE, DOCK1, IL1R1, VNN1, MPO, HINT1, ITGA4, and MANBA (Fig. 5D). Survival analysis in the TCGA cohort revealed significantly worse outcomes for patients with high PCDRScores (Fig. 5E). The AUC values for predicting 1-, 3-, and 5-year overall survival (OS) were 0.809, 0.799, and 0.862, respectively, demonstrating the model’s accuracy (Fig. 5F). Construction of the prognostic model via machine learning. (A) A total of 118 combination models were generated by integrating 10 machine learning algorithms, and the C-index of each model was calculated in 10 cohorts and ranked according to the average C-index of all cohorts. (B) The penalty coefficient of the minimum 10-fold cross-validation error point was calculated to identify the corresponding model genes. (C-D) Estimation of the model gene coefficients. (E) Patients in the TCGA-LAML cohort were divided into high- and low-PCDRScore groups on the basis of the optimal cutoff value for survival analysis. (F) ROC curve analysis of the 1-year, 3-year, and 5-year prognosis prediction ability of the PCDRScore model in the TCGA-LAML cohort. Validation of the prognostic power of the PCDRScore model In nine independent validation cohorts, patients with high PCDRScores consistently had worse prognoses (Fig. 6A), confirming the clinical predictive value of the PCDRScore model. ROC curves further validated the accuracy of the OS prediction, with 5-year AUCs of 0.834 (GSE10358), 0.760 (GSE71014), and 0.735 (Beat AML). Univariate Cox regression confirmed significant prognostic power across all 10 cohorts (P < 0.05) (Fig. 6B). Multivariate Cox analysis of the TCGA, Beat AML, and HOVON cohorts revealed the independent predictive value of the PCDRScore (P < 0.001) (Fig. 6C). Validation of the prognostic ability of the PCDRScore model. (A) Survival analysis and ROC curve analysis of patients in the high- and low-PCDRScore groups in the 9 validation cohorts. (B) Univariate Cox regression analysis of the PCDRScore model in 10 AML cohorts. (C) Multivariate Cox regression analysis of clinicopathological factors and the PCDRScore model within each cohort (TCGA-LAML, HOVON, and Beat AML). Each column represents a separate analysis performed on the indicated cohort To assess whether the prognostic value of PCDRScore is consistent across clinically relevant subgroups, we performed stratified analyses within the TCGA-LAML cohort after excluding APL (M3) patients. Patients were divided into three age groups (< 40 years, 40–59 years, and ≥ 60 years). In each age group, high PCDRScore was significantly associated with worse overall survival compared to low PCDRScore: <40 years (HR = 4.39, 95% CI = 1.28–14.98, P = 0.003), 40–59 years (HR = 3.71, 95% CI = 1.67–8.23, P = 0.001), and ≥ 60 years (HR = 2.56, 95% CI = 1.49–4.42, P = 0.004) (Figure S4A). These results demonstrate that the prognostic performance of PCDRScore is consistent across age groups and is not driven solely by age effects. Within non-APL patients with available cytogenetic risk annotations, PCDRScore significantly stratified prognosis in all three risk categories: favorable risk (HR = 2.34, 95% CI = 1.11–4.93, P = 0.025), intermediate risk (HR = 1.81, 95% CI = 1.02–3.21, P = 0.043), and poor risk (HR = 9.43, 95% CI = 3.98–22.32, P = 0.003) (Figure S4B). Notably, the hazard ratio was largest in the poor-risk group, suggesting that PCDRScore may provide particularly strong prognostic information in high-risk patients. Taken together, these findings indicate that PCDRScore offers independent and additive prognostic value beyond established clinical and genetic risk factors. Clinical application assessment of the PCDRScore model We compared the PCDRScore with published prognostic models based on CRISPR-Cas9 screening genes (AFG16) 26, stem cell phenotype genes (LSC17) 27, AML-vs-healthy DEGs (CODEG22) 28, and tumor heterogeneity genes (GENE4) 29. In seven AML cohorts, the PCDRScore presented slightly higher C-index (Fig. 7A), indicating superior predictive accuracy. PCDRScore showed moderate to strong positive correlations with LSC17 (r = 0.67, P < 2.2 × 10⁻¹⁶), CODEG22 (r = 0.75, P < 2.2 × 10⁻¹⁶), and GENE4 (r = 0.53, P < 2.2 × 10⁻¹⁶), and a weak but significant correlation with AFG16 (r = 0.21, P = 0.0064) (Figure S5A). These moderate-to-strong correlations indicate that PCDRScore captures related but not identical biological information compared to existing scores. We next examined whether PCDRScore could further stratify patients within risk groups defined by each existing signature. Using the LSC17 score as an example, among patients classified as high-risk by LSC17, PCDRScore still significantly distinguished high- versus low-risk subgroups (HR = 3.38, P < 0.001; Figure S5B). Similarly, within LSC17 low-risk patients, PCDRScore provided additional risk stratification (HR = 3.87, P = 0.001). Comparable patterns were observed for CODEG22, AFG16, and GENE4 (Figure S5C-S5E). These results suggest that PCDRScore captures prognostic information not fully accounted for by existing signatures. To formally test the independent prognostic value of PCDRScore, we constructed a multivariable Cox model that included age, cytogenetic risk, and all four existing scores (LSC17, CODEG22, AFG16, GENE4) together with PCDRScore. After adjusting for these established factors, PCDRScore remained a significant independent prognostic factor (HR = 2.44, 95% CI = 1.41–4.23, P = 0.002; Figure S6). This finding demonstrates that PCDRScore provides incremental prognostic value beyond currently available AML risk signatures. Moreover, to investigate whether PCDRScore is associated with specific AML-defining fusion events, we analyzed its distribution across major fusion gene statuses in the TCGA-LAML cohort, including BCR-ABL, PML-RARA, MLL rearrangements, CBFB-MYH11, RUNX1-RUNX1T1 (AML1-ETO), and TEL-AML1 (ETV6-RUNX1). For each fusion gene, we compared PCDRScore values between fusion-positive and fusion-negative patients. No significant differences were observed for any of the tested fusions: BCR-ABL (P = 0.61), PML-RARA (P = 0.11), MLL (P = 0.975), CBFB-MYH11 (P = 0.428), RUNX1-RUNX1T1 (P = 0.573), and TEL-AML1 (P = 0.554) (Figure S7). These results suggest that PCDRScore captures a broader transcriptional program related to programmed cell death dysregulation rather than being a surrogate marker for any specific recurrent fusion event in AML. Finally, a nomogram integrating the PCDRScore with clinical factors (age and cytogenetic risk) was developed (Fig. 7B). Calibration curves in the TCGA, HOVON, and Beat AML cohorts confirmed its precision for OS prediction (Fig. 7C). Analysis of the clinical application value of the PCDRScore model. (A) Comparison of the C-index of the PCDRScore model with existing prognostic models in 7 AML cohorts. (B) Construction of a nomogram combining clinicopathological factors and the PCDRScore to predict the overall survival (OS) of AML patients. (C) Calibration curve to evaluate the accuracy of the nomogram in predicting OS To explore the potential clinical relevance of PCDRScore for targeted therapy, we leveraged ex vivo drug sensitivity data available in the Beat AML cohort. Venetoclax (VEN) sensitivity, measured as area under the dose-response curve (AUC), was compared between high- and low-PCDRScore patients. High PCDRScore patients showed significantly lower sensitivity to venetoclax (higher AUC values, indicating greater resistance) compared to low PCDRScore patients (P = 0.00062) (Figure S8). This finding suggests that patients with high PCDRScore may be less responsive to venetoclax‑based therapy, highlighting the potential utility of PCDRScore for guiding treatment selection. Identification and expression validation of key genes in the prognostic model We analyzed the expression of 13 model genes in various cell types in the AML single-cell dataset. We noticed that the expression of HIP1, SQLE, and VNN1 showed significant differences between several normal cells and their corresponding malignant cells (Fig. 8A). For instance, the expression of HIP1 was significantly higher in ProMono_like, Mono_like, GMP-like, and cDC-like cells than in ProMono, Mono, GMP, and cDC cells. The expression of SQLE was higher in ProMono_like, Prog_like, HSC_like, and GMP_like cells than in ProMono, Prog, HSC, and GMP cells. The expression of VNN1 was significantly higher in Prog_like and HSC_like cells than in Prog and HSC cells, indicating that the expression levels of these three genes were significantly elevated in AML malignant cells. Subsequently, we verified that the expression of HIP1, SQLE, and VNN1 was significantly higher in 12 AML patient samples than in normal samples (Fig. 8B), and their expression was also significantly higher in two AML cell lines, MV4-11 and Molm-14, than in bone marrow stromal cells HS-5 used as a stromal niche comparator (Fig. 8C). This supports the upregulated expression characteristics of the three model genes in AML and provides preliminary validation of our bioinformatics analysis results. Identification and validation of key genes in the prognostic model. (A) Expression levels of 13 model genes in various cell types in single-cell data. (B) Expression differences of SQLE, VNN1 and HIP1 between 12 AML samples and 12 normal samples. (C) Expression differences of SQLE, VNN1 and HIP1 between the normal cell line HS-5 and the AML cell lines MV4-11 and Molm-14. All data are presented as mean ± SD Discussion The intricate interplay between PCD pathways and the molecular heterogeneity of AML has long been hypothesized but remains systematically unexplored. In this study, we integrated multiomics data to delineate PCD-related molecular subtypes and constructed a robust prognostic model, PCDRScore, which offers novel insights into AML pathogenesis and clinical management. Our findings highlight the critical role of dynamic PCD pathway dysregulation in shaping AML progression, immune evasion, and therapeutic responses, bridging a significant gap in current research. The identification of two PCD-related subtypes (Subtypes A and B) underscores the heterogeneity of PCD activation patterns in AML. Subtype A, characterized by elevated PCD scores, an immunosuppressive microenvironment, and a high somatic mutation burden, aligns with prior studies linking hyperactivation of PCD pathways to chemoresistance and immune evasion. For example, the upregulation of PD-L1 and M2 macrophages in Subtype A mirrors observations in solid tumors, where PCD-driven inflammation fosters immunosuppression 30,31. Conversely, Subtype B’s lower PCD activity, enriched naïve immune cell populations, and better prognosis suggest a less aggressive disease phenotype. These findings extend previous reports on individual PCD pathways by demonstrating their collective impact on AML biology. Notably, the associations of Subtype A with M4/M5 FAB classifications and high WBC counts echo clinical observations that monocytic differentiation and leukocytosis correlate with adverse outcomes, further validating our subtyping approach. The divergent immune microenvironments between subtypes provide hypothesis-generating insights for personalized therapy. The immunosuppressive features of subtype A patients, including elevated expression of checkpoint molecules (PD-L1 and HAVCR2) and oncogenic pathway activation, suggest potential responsiveness to immunotherapy. This finding aligns with emerging evidence that high PD-L1 expression in AML may predict the response to immune checkpoint inhibitors. Furthermore, when patients were reclassified based on molecular subtypes and high and low PD-1/PD-L1 expression, patients with subtype A disease and high PD-1/PD-L1 expression had the poorest prognosis, indicating that high expression of immune checkpoints is very unfavorable for the prognosis of AML patients. This finding also suggests that the use of immune checkpoint inhibitors may potentially help improve the prognosis of subtype A patients. Moreover, the predicted sensitivity of patients with type B AML to cytarabine and doxorubicin—the cornerstones of AML chemotherapy—highlights the need for further validation of treatment tailoring based on PCD subtypes. The higher mutation burden in Subtype A, particularly in FLT3 and NPM1, further supports the integration of targeted therapies (e.g., FLT3 inhibitors) with immunotherapy in this subgroup. Compared with existing models (LSC17, CODEG22), the PCDRScore model, which is derived from 13 key PCD-related genes, has superior prognostic accuracy. Its robustness across nine validation cohorts highlights the universality of PCD-related signatures in AML prognosis. Genes such as IL7 and DDIT4 within the model have established roles in leukemogenesis; IL7 promotes the survival of leukemic stem cells 32, whereas MPO is linked to cell cycle transition/proliferation 33. The inclusion of less-characterized genes such as ATP6V0E2 and MANBA opens new avenues for mechanistic studies. Beyond basic comparison of C-index values, we further evaluated whether PCDRScore provides independent prognostic information beyond established AML signatures (LSC17, CODEG22, AFG16, GENE4). Correlation analysis showed modest to strong correlations (ranging from 0.21 to 0.75), indicating that PCDRScore captures overlapping yet distinct biological features. Strikingly, within patients classified as high-risk by LSC17, PCDRScore still further stratified them into high- and low-risk subgroups. In a multivariable model containing age, cytogenetic risk, and all four existing scores, PCDRScore remained an independent prognostic factor. These analyses confirm that PCDRScore provides incremental prognostic value beyond currently available clinical and transcriptomic signatures, suggesting its potential as a complementary tool for risk refinement. Notably, the prognostic performance of PCDRScore remained robust across clinically relevant subgroups. In TCGA-LAML non-APL patients stratified by age (< 40, 40–59, and ≥ 60 years), high PCDRScore consistently predicted worse overall survival, with hazard ratios of 4.39, 3.71, and 2.56 respectively (all P < 0.01). Furthermore, within each ELN 2017 risk category (favorable, intermediate, and adverse), PCDRScore significantly stratified patients, with the largest effect observed in the adverse-risk group. These findings demonstrate that the prognostic value of PCDRScore is not driven by age or pre-existing genetic risk stratification, and it may provide particularly strong prognostic discrimination in high-risk patients who currently have limited therapeutic options. Importantly, the nomogram integrating the PCDRScore with age and cytogenetic risk enhances clinical applicability, offering a user-friendly tool for risk stratification in diverse settings. To explore whether PCDRScore is merely a surrogate for specific recurrent fusion events, we examined its distribution across major AML-associated fusions (PML-RARA, RUNX1-RUNX1T1, CBFB-MYH11, BCR-ABL, MLL rearrangements, and TEL-AML1) in the TCGA-LAML cohort. No significant differences in PCDRScore were observed between fusion-positive and fusion-negative patients for any of these fusions (all P > 0.05). This indicates that PCDRScore captures broader transcriptional programs related to programmed cell death dysregulation rather than serving as a proxy for a single genetic event, reinforcing its potential as a generalizable risk stratification tool across diverse genetic backgrounds. Moreover, using ex vivo drug sensitivity data from the Beat AML cohort, we found that patients with high PCDRScore exhibited significantly lower sensitivity to venetoclax (higher AUC), compared to low PCDRScore patients. This finding aligns with the observation that Subtype A, characterized by high PCDRScore, upregulates anti-apoptotic pathways (e.g., BCL-2 family members) and exhibits an immunosuppressive microenvironment—both known contributors to venetoclax resistance 34. If validated in prospective clinical cohorts receiving venetoclax-based therapy, PCDRScore could potentially help identify patients less likely to benefit from venetoclax and guide alternative treatment strategies. The PCDRScore model’s robustness was further reinforced by our experimental validation of key constituent genes. Single-cell analysis revealed subtype-specific dysregulation of HIP1, SQLE, and VNN1 in malignant AML populations. Critically, these findings were corroborated in clinical samples: HIP1, SQLE, and VNN1 expression was significantly elevated in AML patient peripheral blood compared to healthy controls. Consistent upregulation was also observed in AML cell lines (MV4-11, Molm-14) versus stromal controls (HS-5). HIP1 (huntingtin-interacting protein 1) promotes leukemia cell drug resistance via receptor tyrosine kinase signaling 35, aligning with its association with poor prognosis 36. SQLE (squalene epoxidase), a cholesterol biosynthesis rate-limiter, may fuel AML proliferation through metabolic reprogramming—a mechanism observed in solid tumors 37. VNN1 (pantetheinase) modulates oxidative stress response and has been linked to chemoresistance 38. Their coordinated overexpression in malignant cells underscores their functional synergy within PCD networks and validates the biological foundation of the PCDRScore. Regarding clinical translation, the PCDRScore model could be implemented as a quantitative RT-qPCR panel targeting the 13 signature genes, as all are readily detectable using standard qPCR assays. Such a panel could be applied to RNA extracted from diagnostic bone marrow aspirates or peripheral blood samples, with results available within days of sample collection. Before clinical implementation, prospective validation in a uniformly treated cohort and standardization of the detection protocol are required. While this study provides a comprehensive framework for PCD-related subtyping, several limitations warrant consideration. First, the retrospective nature of public datasets may introduce selection bias, and prospective validation in uniformly treated cohorts is essential. Second, bulk transcriptomic data may obscure cell type-specific PCD dynamics, necessitating single-cell functional validation. Third, our experimental validation used peripheral blood samples and a stromal cell line (HS-5) as a comparator. Peripheral blood samples, while accessible and clinically relevant, may not fully recapitulate the bone marrow microenvironment, and the HS-5 comparison reflects leukemia-stroma differences rather than leukemia-normal hematopoietic differences. Future studies using sorted CD34 + hematopoietic progenitor cells and bone marrow aspirates would provide more definitive validation of these expression differences in the appropriate biological context. Fourth, the mechanistic links between PCD pathways and immune modulation remain speculative; experimental studies using AML cell lines or patient-derived xenografts are needed to dissect causal relationships. Fifth, none of the public cohorts used in this study contained information on hematopoietic stem cell transplantation (HSCT). Consequently, we could not evaluate whether the PCDRScore predicts clinical benefit from HSCT or could inform transplant decision-making. Future prospective studies with HSCT data are needed to address this question. Sixth, the risk stratification cutoffs for PCDRScore were determined independently within each cohort using optimal threshold selection. While this approach is standard for discovery-phase prognostic studies, clinical implementation would require a fixed cutoff validated on a standardized assay platform. Cross-cohort normalization and prospective studies are needed before the PCDRScore can be translated into a clinically actionable threshold. Finally, the immunotherapy and drug sensitivity predictions are computationally derived and require validation in prospective AML cohorts treated with checkpoint inhibitors or targeted therapies. The clinical utility of PCDRScore-guided therapy requires testing in randomized trials, particularly for immunotherapy responses predicted in Subtype A patients. Conclusion In summary, through integrated computational and experimental approaches, our study establishes PCD dysregulation as a central axis of AML heterogeneity, with profound implications for prognosis and treatment. The PCDRScore model not only refines risk stratification but also illuminates potential therapeutic vulnerabilities tied to PCD activation. By transcending the traditional focus on individual pathways, this work paves the way for a systems-level understanding of PCD in AML, ultimately advancing the paradigm of precision oncology. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Acknowledgements Not applicable. Author contributions L.Z. and G.K.: Conceptualization, Supervision, and Project administration; J.L., X.W. F.Z. and G.K.: Funding acquisition; J.L.: Resources; J.L., F.Z., X.W., L.X., X.C. and Z.T.: Validation and Visualization; J.L.: Methodology; J.L.: Data Curation; J.L.: Formal analysis; J.L.: Software; J.L.: Writing - original draft; L.Z. and G.K.: Writing - review & editing. All authors edited and approved the final version of the submitted manuscript. Funding The study was funded by the National Natural Science Foundation of China (82260035), and the Natural Science Foundation of Jiangxi Province (20252BAC240524). Data availability All data used in this work can be acquired from the Gene-Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/geo/) and The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov/). The code used for statistical analysis and figure generation in this study can be accessed by the corresponding author upon reasonable request. Declarations Competing interests The authors declare no competing interests. Ethical approval This study was approved by the Ethics Committee of the Second Affiliated Hospital of Nanchang University. Consent for publication Not applicable. Consent to participate Not applicable. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Linlin Zhang, Email: zhanglinlin_fly@163.com. Guangyao Kong, Email: konggy@xjtu.edu.cn. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Data Availability Statement All data used in this work can be acquired from the Gene-Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/geo/) and The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov/). The code used for statistical analysis and figure generation in this study can be accessed by the corresponding author upon reasonable request.