Predicting response to hypomethylating agents in myeloid neoplasms: a robust model Authors: Chao Guo, Ya-yue Gao, Qian-qian Ju, Zhen-ling Li Published: 2026 ABSTRACT Current biomarkers and prognostic systems lack sufficient evaluation of their predictive power on HMA efficacy for myeloid neoplasms. Weighted Gene Co-expression Network Analysis (WGCNA) was employed to uncover the target module correlated with HMA response. Based on this module, Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to establish a predictive model with non-zero coefficients, termed the HMA-29 model. The magenta module was identified as the target module correlating with HMA response (R² = 0.37, p = 2 × 10⁻⁴). LASSO regression established the HMA-29 model, comprising the expression signatures of 29 genes, which stratified patients into responsive-like and resistant-like groups. HMA-29 risk scores significantly differed between HMA responders and non-responders and demonstrated superior predictive performance (AUC 0.9982, p < 0.0001). The predictive capability was validated across independent cohorts/disease subtypes/HMA drugs/sampling conditions. Notably, HMA-29 risk scores did not significantly differ between pre- and post-treatment samples. Moreover, Kaplan-Meier analysis demonstrated significant differences of disease-free survival (DFS) and overall survival (OS) between the HMA-29-defined groups. GSEA indicated that the G2M checkpoint, MYC targets, and E2F targets were positively associated with HMA-29 risk. We developed a novel predictive model for HMA response in myeloid neoplasms, demonstrating superior predictive and prognostic value compared to existing systems. Our findings provide insights into potential therapeutic strategies, including the integration of targeted therapies with HMA treatment.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s00277-026-07137-3. Introduction Hypomethylating agents (HMAs), including decitabine and azacitidine, represent the current standard of care for several myeloid neoplasms, such as intermediate- to high-risk myelodysplastic syndromes (MDS) [1], chronic myelomonocytic leukemia (CMML) [2, 3], elderly or unfit acute myeloid leukemia (AML) [4], and blast-phase myeloproliferative neoplasms (MPN) [5]. However, fewer than 50% of patients with MDS or elderly AML respond to HMA treatment, with the majority achieving hematological response rather than complete remission. Therefore, predicting HMA response in myeloid patients remains a critical challenge. Several predictive models and biomarkers have been developed based on molecular variations [6–10], methylation profiles [11], or gene expression signatures [12, 13]. The presence of mutant ASXL1 has been independently associated with adverse responses to HMA therapy in both CMML [10] and MDS [9], whereas the TET2mut/ASXL1wt genotype correlates with higher complete remission rates [10]. Genomic models, primarily based on ASXL1, have been developed using artificial intelligence to predict resistance to HMA therapy in MDS with high accuracy [8]. Nevertheless, despite strong evidence supporting the predictive value of ASXL1 mutations, their clinical utility remains limited due to their low frequency in myeloid neoplasms [14]. Mutant DNMT3A was also demonstrated to be associated with HMA response in MDS patients [15]. In AML, a risk model approved by the Mayo Clinic has been established to predict HMA response based on the mutation status of six genes (NPM1, IDH2, DDX41, TP53, FLT3-ITD, and RUNX1) [6]. Mutant DNMT3A [16, 17], IDH1/2 [18], and TET2 [19] were also reported to be associated with HMA response. Given that HMA drugs function by modifying the methylation landscape, leading to the re-expression of silenced tumor suppressor genes and cellular reprogramming [20], additional efforts have been made to develop predictive models based on transcriptomic and methylation profiles [11, 12, 20–23]. However, the external validity of these models has not been broadly tested across different cohorts within myeloid neoplasms. Another key challenge in assessing HMA response prediction is the heterogeneity of omics data, including inclusion of diverse disease subtypes and/or drugs, variations in methodology and sampling conditions. Emerging methodologies have identified key genes and pathways using RNA-seq data, with Weighted Gene Co-expression Network Analysis (WGCNA) serving as a powerful tool to analyze co-expressed gene clusters based on scale-free network theory. This approach partitions the entire genome into a limited number of modules, each containing genes co-expressed within individual samples. Correlation analysis is then performed between clinical traits and Module Eigengenes (MEs), which are defined as the primary principal components of gene modules [24, 25]. This analysis facilitates the identification of key genes associated with HMA response. In this study, we attempted to develop a robust, transcriptome-based HMA response prediction model and validate its performance across independent cohorts encompassing common myeloid neoplasms, including AML, MDS, and CMML. The overall study design is illustrated in Fig. 1.Fig. 1The flowchart of the present study The flowchart of the present study Methods Data download From the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/), clinical and transcriptomic data were retrieved for the following datasets: GSE76203, GSE61162, GSE272173, GSE77750, GSE40442, and GSE138696. Additionally, clinical and expression data of BEATAML cohort, who were inducted by HMA therapy, were obtained from the online database (http://www.vizome.org/aml/). The TCGA-LAML dataset was accessed via the Genomic Data Commons (GDC) database using the TCGAbiolinks package in R (version 4.2.1). From the ArrayExpress database (https://www.ebi.ac.uk/biostudies/arrayexpress), E-MEXP-3737 dataset was retrieved. Among these datasets, GSE76203, BEATAML, GSE61162, GSE272173, E-MEXP-3737, and TCGA-LAML were derived using RNA sequencing (RNA-seq), whereas GSE77750, GSE40442, and GSE138696 were generated using microarray. WGCNA Four datasets were analyzed using WGCNA, incorporating clinical outcome data (response status) and RNA-seq derived expression matrices for individual patients. Cohort 1 (GSE76203): Included 9 CMML and 10 MDS pretreatment samples (13 responders, 6 non-responders) and 14 post-treatment samples after six cycles of azacitidine.Cohort 2 (GSE61162): Consisted of 14 pretreatment CMML samples (8 responders, 6 non-responders) following six cycles of decitabine.Cohort 3 (BEATAML): Comprised 20 AML samples, with 11 complete remission (CR/CRi) and 9 refractory patients treated with azacitidine- or decitabine-based therapy.Cohort 4 (GSE272173): Included 14 pretreatment MDS samples (6 responders, 8 non-responders), 14 matched post-treatment samples, and 3 healthy controls. Cohort 1 (GSE76203): Included 9 CMML and 10 MDS pretreatment samples (13 responders, 6 non-responders) and 14 post-treatment samples after six cycles of azacitidine. Cohort 1 (GSE76203): Included 9 CMML and 10 MDS pretreatment samples (13 responders, 6 non-responders) and 14 post-treatment samples after six cycles of azacitidine. Cohort 2 (GSE61162): Consisted of 14 pretreatment CMML samples (8 responders, 6 non-responders) following six cycles of decitabine. Cohort 2 (GSE61162): Consisted of 14 pretreatment CMML samples (8 responders, 6 non-responders) following six cycles of decitabine. Cohort 3 (BEATAML): Comprised 20 AML samples, with 11 complete remission (CR/CRi) and 9 refractory patients treated with azacitidine- or decitabine-based therapy. Cohort 3 (BEATAML): Comprised 20 AML samples, with 11 complete remission (CR/CRi) and 9 refractory patients treated with azacitidine- or decitabine-based therapy. Cohort 4 (GSE272173): Included 14 pretreatment MDS samples (6 responders, 8 non-responders), 14 matched post-treatment samples, and 3 healthy controls. Cohort 4 (GSE272173): Included 14 pretreatment MDS samples (6 responders, 8 non-responders), 14 matched post-treatment samples, and 3 healthy controls. All four transcriptomic datasets were log₂-normalized in R software. WGCNA was conducted using the WGCNA package (version 1.73) for co-expression analysis [19]. Inter-individual heterogeneity was assessed via hierarchical clustering with average linkage to detect and exclude outliers. No significant outliers were identified and thus no samples were excluded (see Supplementary Fig. 1). The combat function (from sva package, version 3.54.0) was applied to correct inter-cohort batch effects, confirmed by sample dendrograms and trait heatmaps (Supplementary Fig. 1). Intra-cohort heterogeneity was further adjusted using the normalizeBetweenArrays function of the limma package (version 3.62.2). The soft-threshold power was determined as the minimum beta value yielding a scale-free R² > 0.85. Pearson’s correlation coefficients between individual genes were computed to construct the gene adjacency matrix and topological overlap matrix. Using hierarchical clustering with average linkage, gene co-expression modules were identified, requiring > 20% inter-module dissimilarity and with 30 genes at least. The association analysis was performed using Pearson’s correlation coefficients between module eigengenes (MEs) and eight clinical conditions: HMA response, HMA non-response, pre-treatment, post-treatment, AML, CMML, MDS, and healthy controls. The module exhibiting the highest Pearson’s correlation coefficient with HMA response and a statistically significant p-value (< 0.05) was identified as the target module. To characterize enriched pathways within the target module, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) databases were utilized via the enrichGO and enrichKEGG functions in the clusterProfiler package (version 4.14.4) [26–28]. The ReactomePA package (version 1.50.0) was used to generate an interactive network of enriched pathways based on the Reactome database [29]. Pathways showing a local false discovery rate (FDR)-adjusted p-value below 0.05 were deemed statistically significant. LASSO and model development The expression data of the target gene module from 67 pre-treatment samples, derived from combination of four RNA-seq-based datasets (GSE76203, GSE61162, BEATAML, and GSE272173), were used as the training set for LASSO regression analysis. Although RNA-seq is the most widely used method for transcriptomic analysis, microarray technology remains a cost-effective and commonly utilized alternative. To enhance the generalizability of the model, the GSE77750 dataset, a microarray-based cohort, was used for external validation. LASSO regression was performed using the glmnet package (v4.1–8) in R (v4.4.2). The optimal lambda value was determined through 10-fold cross-validation (using the cv.glmnet() function with nfolds = 10), and lambda.min was selected as the final regularization parameter. Model performance was assessed using mean squared error (MSE) and the coefficient of determination (R²). To mitigate overfitting, LASSO regression was iteratively performed 10,000 times, and variables with non-zero coefficients were identified. Subsequently, bootstrap aggregation was applied to develop a robust predictive model, resulting in a 29-gene expression signature, referred to as the HMA-29 model. Individual patient risk scores were calculated by summing the expression values of the selected genes, each weighted by its corresponding regression coefficient. Patients were stratified into responsive-like and resistant-like groups based on a cutoff value, determined using the coords function in the pROC package (version 1.18.5). The predictive performance of the HMA-29 model was evaluated by calculating the area under the receiver operating characteristic (ROC) curve (AUC) using the roc function in the ROCR package (version 1.0–11). Additionally, batch effects and inter-array heterogeneity were managed based on the median value and variance of the cohort, ensuring comparability across datasets. In clinical practice, these cohort-based normalizations, cannot be applied to RNA-seq data from individual patients. Therefore, the HMA-29 model was evaluated using RNA-seq data subjected to a simple log₂ transformation of raw counts, followed by ROC analysis. Furthermore, the predictive performance of the HMA-29 model was assessed across multiple RNA-seq data subsets, stratified by platform (RNA-seq vs. microarray), disease type (AML, CMML, and MDS), specific therapeutic agents (decitabine vs. azacitidine), and sampling conditions (BMMC, PBMC, and CD34⁺ cells). We conducted a literature review to identify relevant markers/models for use as controls. One such model, developed by Aziz Nazha et al. and published in the Journal of Clinical Oncology (hereafter referred to as the “JCO model”), has demonstrated high accuracy in classifying HMA non-responders among MDS patients [8]. Additionally, several genomic and transcriptomic biomarkers have been reported to be associated with HMA response in MDS, including DNMT3Amut [15], TET2mut [15, 19, 30], ASXL1mut [9] status, expression of ITGA5 [12], RPL28 [31], MEG3 [21], CXCL4 and CXCL7 [22]. These markers were selected for comparative analysis with the HMA-29 model using ROC analysis in MDS cohorts (GSE76203 and GSE272173). For AML, the Mayo Clinic has established a prognostic model incorporating a six-gene mutation panel as a predictor of HMA response [6]. Moreover, mutations in DNMT3Amut [16, 32], IDH1/2mut [18], TET2mut [15] status and SAMHD1 expression [13], have been reported to influence HMA response in AML patients. Accordingly, the Mayo model and above markers were included in the comparison with the HMA-29 model using the BEATAML cohort. Clinical parameter association To illustrate the clinical relevance of HMA-29 risk scores, the GSE272173 cohort (n = 14) was utilized to compare key clinical parameters—including bone marrow blast percentage, WBC count, HGB levels, and PLT count—between HMA-29 defined responsive-like and resistant-like groups. To investigate the impact of HMA treatment on the target module, we analyzed multiple transcriptomic datasets. The GSE40442 dataset includes 67 primary AML samples cultured for seven days, with a three-day treatment period. These samples were divided into four groups: decitabine (n = 17), cytarabine (n = 16), dimethyl sulfoxide (DMSO) vehicle control (n = 17), and untreated control (n = 17) [12]. Comparisons were made between the decitabine-treated group and the other groups. Additionally, the E-MEXP-3737 dataset consists of seven baseline AML samples and their corresponding post-treatment samples following six to seven days of decitabine exposure [33]. The GSE138696 dataset includes technical triplicates (with duplicates for samples 02_d8, 04_d0, 17_d0, and 17_d8) and comprises primary AML blast cells from 14 patients, assessed at day 0 and day 8 of decitabine treatment [34]. Furthermore, the GSE272173 dataset contains 14 pre-treatment MDS samples alongside their corresponding post-treatment samples collected after four to twelve cycles of azacitidine. HMA-29 risk scores were compared between pre- and post-treatment samples to evaluate changes following HMA exposure. Survival analysis A total of 15 AML patients from the TCGA-LAML cohort who received HMA treatment (14 treated with decitabine and one with azacitidine) and had available disease-free survival (DFS) and overall survival (OS) data were included in the survival analysis. Additionally, 18 out of 20 AML patients from cohort 3 (BEATAML) were selected for survival analysis based on the availability of OS data. Kaplan-Meier analysis was performed to evaluate survival differences between high- and low-risk groups as defined by the HMA-29 risk model. The optimal cut-off value for risk stratification was determined using the “surv_cutpoint” function from the “survminer” package (version 0.5.0), effectively dividing patients into HMA-29 high- or low-risk groups. The “survival” package (version 3.8-3) was utilized to conduct the Kaplan-Meier survival analysis. Gene set enrichment analysis (GSEA) Pearson’s correlation coefficients were calculated to assess the association between HMA-29 risk scores and individual gene expression values. Gene Set Enrichment Analysis (GSEA) was conducted to interpret the genome-wide correlation analysis of HMA-29 risk scores, using gene sets from the Molecular Signatures Database (MSigDB) (http://software.broadinstitute.org/gsea/msigdb) [35]. For each predefined gene set, genes were ranked based on their Pearson’s correlation coefficient with HMA-29 risk scores. To account for multiple hypothesis testing, q values were applied for statistical correction. Unlike analyses focusing solely on a few key regulators, GSEA evaluates the expression levels of all genes within a given pathway, providing a more comprehensive representation of signaling pathway activation or suppression associated with HMA-29 risk scores. Signaling pathways with an absolute Normalized Enrichment Score (|NES|) greater than 1 and a q value below 0.05 were considered significantly associated with HMA-29 risk scores. Statistical analysis For normally distributed continuous variables, unpaired t-tests were used; otherwise, Mann-Whitney tests were applied. One-way ANOVA was used for multiple group comparisons. For categorical variables, Fisher’s exact test was employed. Data sharing statement All datasets supporting the present study were publicly available, which can be obtained from GEO database (https://www.ncbi.nlm.nih.gov/geo/), BEATAML database (http://www.vizome.org/aml/), GDC database (https://portal.gdc.cancer.gov/), and ArrayExpress database (https://www.ebi.ac.uk/biostudies/arrayexpress). Results The HMA response-related gene co-expression module The module-trait relationship is illustrated in Fig. 2. HMA treatment significantly altered the expression profiles of more than half of the gene modules (7 out of 12, p < 0.05), aligning with the known mechanisms of HMA. The magenta module demonstrated the highest correlation with HMA response (R2 = 0.37, p = 2e-4) and was designated as the target module.Fig. 2The relationship of gene co-expression clusters and clinical traits. Each box contains Pearson’s coefficients (by gradient of color, red = 1, blue = − 1) and p values, for correlation between MEs (X-axis) and clinical/genetic variables (Y-axis) The relationship of gene co-expression clusters and clinical traits. Each box contains Pearson’s coefficients (by gradient of color, red = 1, blue = − 1) and p values, for correlation between MEs (X-axis) and clinical/genetic variables (Y-axis) Notably, the target module was not affected by HMA treatment, as no significant correlation was observed between the ME and pre/post-HMA treatment states (p > 0.05, Fig. 2). This suggests that the uninfluenced portion of HMA administration plays a key role in determining the fate of malignant cells. The association between HMA response and the magenta module was independent of disease type, as no significant correlation was found between its expression profile (ME) and AML, CMML, or MDS (p > 0.05, Fig. 2). Gene enrichment analysis (GO, KEGG, and Reactome databases) revealed that genes within the target module were primarily associated with pathways related to cellular mitosis and the TP53 pathway (Fig. 3, Supplementary Fig. 2). The result was in line with previous investigation by John E. Pimanda et al. [12], which demonstrated that AZA responders exhibited increased hematopoietic progenitor cell cycling.Fig. 3The emplot demonstrating gene enrichment results based on REActome database The emplot demonstrating gene enrichment results based on REActome database Establishment of the HMA-29 model The variables with non-zero coefficients derived from LASSO analysis are listed in Table 1, encompassing expression signatures from 29 genes. The HMA-29 risk scores were significantly higher in the HMA-response group than that of the non-response group in both the training set (GSE76203/GSE61162/BEATAML/GSE272173, Fig. 4A) and validation set (GSE77750, Fig. 4B). No significant correlation was observed between HMA-29 risk scores and disease type.Table 1The details of HMA-29 risk modelGene symbolCoefficientDDX110.011660188TNPO3−0.066005729ABCA70.086742025MCM60.038041418CTTN−0.054459855HIRA0.039359242POLE20.003811435MTMR4−0.098772592CNRIP10.008533114TNFAIP60.020250119DEK0.010190672MIS18BP1−0.006510425DSCC10.028731982BMP2K0.001874367GUCD1−0.032661179TAF50.001210664CENPH−0.085657393SLFN130.053665369PAQR40.015677469FCGR3B0.029809068KIF150.065335393RAD210.06532429LOXHD1−0.035869945FEN10.01083255RNF26−0.159675355NFATC2IP0.037043701CENPP0.069793713EPHB40.013549046BLM0.010723163Fig. 4The comparison of HMA-29 risk scores between HMA responders and non-responders in combined RNAseq datasets (A), and microarray datasets (B). (ns, p>0.05, * p<0.05, ** p < 0.01, *** p < 0.001) Table 1The details of HMA-29 risk modelGene symbolCoefficientDDX110.011660188TNPO3−0.066005729ABCA70.086742025MCM60.038041418CTTN−0.054459855HIRA0.039359242POLE20.003811435MTMR4−0.098772592CNRIP10.008533114TNFAIP60.020250119DEK0.010190672MIS18BP1−0.006510425DSCC10.028731982BMP2K0.001874367GUCD1−0.032661179TAF50.001210664CENPH−0.085657393SLFN130.053665369PAQR40.015677469FCGR3B0.029809068KIF150.065335393RAD210.06532429LOXHD1−0.035869945FEN10.01083255RNF26−0.159675355NFATC2IP0.037043701CENPP0.069793713EPHB40.013549046BLM0.010723163 The details of HMA-29 risk model The comparison of HMA-29 risk scores between HMA responders and non-responders in combined RNAseq datasets (A), and microarray datasets (B). (ns, p>0.05, * p<0.05, ** p < 0.01, *** p < 0.001) When cutoff values were set at 0.6338387 for RNAseq datasets and 0.5337593 for microarray datasets, ROC analysis (Fig. 5) demonstrated high predictive power, with the corresponding AUC values as follows: 0.9982 (p<0.0001, sensitivity 100%, specificity 96.552%) for the whole training set, 1.0 (p = 0.0006, sensitivity 100%, specificity 100%) for GSE76203, 1.0 (p=0.0019, sensitivity 100%, specificity 100%) for GSE61162, 1.0 for BEATAML (p = 0.0002, sensitivity 100%, specificity 100%), 1.0 (p = 0.0019 sensitivity 100%, specificity 100%) for GSE272173, and 0.7045 (p = 0.0527 sensitivity 89.47%, specificity 53.846%) for GSE77750.Fig. 5The ROC analysis of HMA-29 risk on the training set, GSE76203, GSE61162, BEATAML, GSE7750 and GSE272173 The ROC analysis of HMA-29 risk on the training set, GSE76203, GSE61162, BEATAML, GSE7750 and GSE272173 Given that LASSO analysis was performed on cohort-adjusted data, the raw count data from four RNAseq datasets (GSE76203/BEATAML/GSE61162/GSE272173) were log2-transformed and analyzed by ROC (Supplementary Fig. 3). The AUC for the entire RNAseq dataset was 0.8757 (p < 0.0001), and individual dataset AUC values were: 1.0 (p = 0.0006) for GSE76203, 0.9792 (p = 0.003) for BEATAML, 0.9091 (p = 0.0021) for GSE61162, and 0.9792 (p = 0.003) for GSE272173. This indicated the high predictive capacity was preserved in non-adjusted datasets. Further validation assessed the impact of disease type, drug type, and sample type on HMA-29 model performance (Supplementary Fig. 4). The AUCs and corresponding p-values were: 0.9982 (p < 0.0001) and 0.7045 (p = 0.0527) for RNAseq and microarray datasets, respectively. By disease type: AML (AUC = 0.9091, p = 0.0021), CMML (AUC = 0.9802, p < 0.0001), and MDS (AUC = 0.9792, p = 0.003). By drug type: azacitidine (AUC = 0.8740, p < 0.0001) and decitabine (AUC = 0.9402, p = 0.0006). By sample conditions: BMMC (AUC = 0.9533, p = 0.0002), PBMC (AUC = 0.85, p = 0.0864), and CD34 + cells (AUC = 0.9925, p < 0.0001). In the BEATAML cohort (Fig. 6A), the AUC of HMA-29 (AUC = 1) surpassed that of the Mayo model [6] (AUC = 0.5909, p = 0.4941) and SAMH1 expression (AUC = 0.5657, p = 0.6214). Similarly, in MDS cohorts (GSE76203, Fig. 6B), HMA-29 outperformed other markers/models, with only MEG3 achieving significance (AUC = 0.8333, p = 0.0226), while CXCL7 approached significance (AUC = 0.7821, p = 0.0537). In GSE272173, no markers/models, except HMA-29 risk, showed significant classification power (data not shown).Fig. 6The ROC analysis of HMA-29 risk and other markers (with p value < 0.05), on BEATAML (A), and GSE76203 (B) The ROC analysis of HMA-29 risk and other markers (with p value < 0.05), on BEATAML (A), and GSE76203 (B) To evaluate whether an alternative machine-learning approach could generate a distinct predictive model, we additionally performed Random Forest analysis using the same response-associated candidate gene set. The Random Forest model was constructed in the training cohort and tested in the validation cohort. However, its predictive performance in the validation set was limited, with an AUC of 0.555, indicating poor discriminatory ability. Therefore, the Random Forest model did not provide a viable adjunctive model to the LASSO-derived HMA-29 signature. These findings further support the superior predictive performance and generalizability of the HMA-29 model in the present analysis. Clinical relevance of HMA-29 defined risk With cutoff values of 0.6338387 (RNAseq) and 0.7045 (microarray), patients were dichotomized into responsive-like and resistant-like groups. No significant differences were observed in blast percentage, WBC count, HGB level, or PLT count between these groups (Supplementary Fig. 5). Consistent with WGCNA results, HMA-29 risk scores remained stable pre- and post-AZA treatment across E-MEXP-3737, GSE138696, and GSE272173 (Fig. 7). Likewise, no significant differences were found between decitabine and cytarabine groups in GSE40442 (Fig. 7). The stability of HMA-29 scores before and after treatment underpins its predictive power based on baseline transcriptomic data. This characteristic was consistently observed across myeloid malignancies, including AML (GSE40442/E-MEXP-3737/GSE138696) and MDS (GSE272173).Fig. 7The comparison of HMA-29 risk scores between pre- and post-treatment, in GSE40442 (A), E-MEXP-3737 (B), GSE138696 (C), and GSE272173 (D). (ns, p > 0.05, * p < 0.05, ** p < 0.01, *** p < 0.001) The comparison of HMA-29 risk scores between pre- and post-treatment, in GSE40442 (A), E-MEXP-3737 (B), GSE138696 (C), and GSE272173 (D). (ns, p > 0.05, * p < 0.05, ** p < 0.01, *** p < 0.001) Prognostic value of HMA-29 risk in myeloid neoplasms In the TCGA-LAML cohort, HMA-29-defined high-risk and low-risk groups exhibited significant differences in both DFS and OS (Fig. 8A&B). In the BEATAML cohort, Kaplan-Meier curves showed separation between high- and low-risk groups, although the log-rank test was not significant (Fig. 8C).Fig. 8The Kaplan-Meier illustrating survival difference between HMA-29 high- and low-risk groups, by DFS in TCGA cohort (A), OS in TCGA cohort (B), and OS in BEATAML cohort (C) The Kaplan-Meier illustrating survival difference between HMA-29 high- and low-risk groups, by DFS in TCGA cohort (A), OS in TCGA cohort (B), and OS in BEATAML cohort (C) Results of GSEA Pathway analysis (Fig. 9) revealed that cell proliferation-related pathways were activated in association with HMA-29 risk, including G2M checkpoint, DNA repair, MYC targets, and E2F targets. Conversely, TNFa signaling via NF-kappaB was suppressed in the resistant-like group. An intriguing finding was the increased activation of heme metabolism in the resistant-like group, aligning with the observed (though non-significant) trend toward higher HGB levels in this group (8.6 g/L vs. 9.2 g/L).Fig. 9The results of GSEA showing the activated/suppressed cell signaling pathways correlating with HMA-29 risk The results of GSEA showing the activated/suppressed cell signaling pathways correlating with HMA-29 risk Discussion Considering the biological rationale of HMA therapy, the significant association between genetic/epigenetic predictors and treatment response has been reported [20, 36]. However, before translating these findings into clinical practice, three critical issues must be addressed: Intrinsic Predictive Power: The sensitivity, specificity, and AUC of predictive models should be thoroughly evaluated and compared, rather than solely relying on significant differences in treatment response rates among predictor-defined groups, as done in previous studies [20]. Our analysis revealed that genetic markers generally had inferior AUC values and lacked strong classification performance (most with p > 0.05).Limited Applicability of Genetic Markers: Only a fraction of patients exhibited mutations investigated in prior studies, limiting their feasibility due to sparsity.Complex Genetic Variations: Patients often harbor multiple genetic alterations rather than single mutations, making individual risk assessment challenging and inconsistent. Intrinsic Predictive Power: The sensitivity, specificity, and AUC of predictive models should be thoroughly evaluated and compared, rather than solely relying on significant differences in treatment response rates among predictor-defined groups, as done in previous studies [20]. Our analysis revealed that genetic markers generally had inferior AUC values and lacked strong classification performance (most with p > 0.05). Intrinsic Predictive Power: The sensitivity, specificity, and AUC of predictive models should be thoroughly evaluated and compared, rather than solely relying on significant differences in treatment response rates among predictor-defined groups, as done in previous studies [20]. Our analysis revealed that genetic markers generally had inferior AUC values and lacked strong classification performance (most with p > 0.05). Limited Applicability of Genetic Markers: Only a fraction of patients exhibited mutations investigated in prior studies, limiting their feasibility due to sparsity. Limited Applicability of Genetic Markers: Only a fraction of patients exhibited mutations investigated in prior studies, limiting their feasibility due to sparsity. Complex Genetic Variations: Patients often harbor multiple genetic alterations rather than single mutations, making individual risk assessment challenging and inconsistent. Complex Genetic Variations: Patients often harbor multiple genetic alterations rather than single mutations, making individual risk assessment challenging and inconsistent. To overcome these limitations, a predictive system must integrate consistent, multi-dimensional parameters rather than relying on discrete and sparse genetic/cytogenetic variables. Furthermore, this system must demonstrate superior predictive power compared to existing markers, validated through independent cohorts. To address these challenges, we focused on transcriptomic data derived from high-throughput methods, which inherently provide consistency and multidimensionality. Unlike previous studies that analyzed the expression of single genes, WGCNA identified co-expressed gene modules, leading to the discovery of the HMA response-specific module. Subsequently, LASSO analysis generated the HMA-29 model, which outperformed conventional genetic/expression markers, as validated by independent cohorts through ROC analysis (Fig. 6A&B). Notably, since the HMA-29 risk score can be applied to log2-transformed raw RNA-seq count data, it enhances feasibility in clinical settings. The module-trait relationship plot demonstrated that the target module was not disease-specific (Fig. 2), encompassing AML, MDS, and CMML. Derived from the target module, the HMA-29 risk score retained this property (Supplementary Fig. 4). Its discriminative power remained unaffected by experimental platforms (RNA-seq vs. microarray), disease type (AML vs. MDS vs. CMML), specific drugs (azacitidine vs. decitabine), or sampling condition (BMMC vs. PBMC vs. CD34 + cells). This suggests that the HMA-29 risk score may represent an intrinsic characteristic of myeloid neoplasms. The expression signature associated with HMA response, including both the target module and HMA-29 risk, exhibits three fundamental characteristics: Consistency with Prior Findings: The signature aligns with existing evidence that HMA response correlates with activation of cell mitosis-related pathways (Fig. 3). HMAs function by incorporating into DNA/RNA and degrading DNMT1, leading to DNA demethylation and reactivation of silenced genes involved in apoptosis, DNA repair, and differentiation [37, 38]. Consistent with this mechanism, HMA responders were shown to have more hematopoietic progenitor cells in the cell cycle, as demonstrated by integrative genomics studies [12].Predictive and Prognostic Value: The signature exhibits strong predictive capability for treatment response and prognostic significance, as demonstrated in our analyses.Temporal Stability: The signature remained stable before and after HMA treatment, with no statistically significant alterations observed (Figs. 2 and 7). Consistency with Prior Findings: The signature aligns with existing evidence that HMA response correlates with activation of cell mitosis-related pathways (Fig. 3). HMAs function by incorporating into DNA/RNA and degrading DNMT1, leading to DNA demethylation and reactivation of silenced genes involved in apoptosis, DNA repair, and differentiation [37, 38]. Consistent with this mechanism, HMA responders were shown to have more hematopoietic progenitor cells in the cell cycle, as demonstrated by integrative genomics studies [12]. Consistency with Prior Findings: The signature aligns with existing evidence that HMA response correlates with activation of cell mitosis-related pathways (Fig. 3). HMAs function by incorporating into DNA/RNA and degrading DNMT1, leading to DNA demethylation and reactivation of silenced genes involved in apoptosis, DNA repair, and differentiation [37, 38]. Consistent with this mechanism, HMA responders were shown to have more hematopoietic progenitor cells in the cell cycle, as demonstrated by integrative genomics studies [12]. Predictive and Prognostic Value: The signature exhibits strong predictive capability for treatment response and prognostic significance, as demonstrated in our analyses. Predictive and Prognostic Value: The signature exhibits strong predictive capability for treatment response and prognostic significance, as demonstrated in our analyses. Temporal Stability: The signature remained stable before and after HMA treatment, with no statistically significant alterations observed (Figs. 2 and 7). Temporal Stability: The signature remained stable before and after HMA treatment, with no statistically significant alterations observed (Figs. 2 and 7). Analyzing the activated/inhibited pathways associated with the expression signature may help identify targeted therapies that could synergize with HMAs. Among the GSEA results, three signaling pathways warrant particular attention: G2M checkpoint, E2F targets, and MYC targets. G2M Checkpoint: Re-establishing the G2/M checkpoint has been linked to azacitidine resistance [39]. Our analysis demonstrated a significant correlation between WEE1 expression, a crucial G2M checkpoint regulator, and HMA-29 risk (p = 2.13e-4, Supplementary Fig. 6). Notably, WEE1 inhibition has been explored as a potential therapeutic strategy in MDS and AML ex vivo, particularly in combination with PARP-1 inhibitors [40]. Additionally, WEE1 inhibition can eliminate AML cells when synergizing with CHK1 inhibitor [41, 42], and there were ongoing clinical trials for the WEE1 inhibitor adavosertib in solid malignancies currently [43].E2F Targets: As a master regulator of cell-cycle progression, E2F1 contributes to dysregulated granulocyte differentiation and AML initiation/maintenance, in conjunction with C/EBPα [44, 45]. E2F7 has also been implicated in promoting AML cell cycle progression while inhibiting monocytic differentiation [46]. Both E2F1 and E2F7 were included in the target module and correlated significantly with HMA-29 risk (Supplementary Figs. 7 & 8).MYC Targets: accumulated evidence indicates that c-Myc serves as a key role in apoptosis prevention, drug resistance development, and leukemogenesis progression [47, 48]. Several inhibitors, including HDAC inhibitors, BET inhibitors, and CDK9 inhibitors, have been found to enhance the anti-leukemic activity of HMAs by suppressing c-MYC activity or expression, leading to ongoing clinical trials [43]. Additionally, the XPO1 inhibitor Selinexor demonstrated synergy with azacitidine by suppressing the XPO1/eIF4E/c-MYC signaling pathway [49]. Selinexor has been shown to enhance the effects of HMAs in eliminating AML/MDS cells [50, 51] and has exhibited promising efficacy in clinical trials [52, 53]. G2M Checkpoint: Re-establishing the G2/M checkpoint has been linked to azacitidine resistance [39]. Our analysis demonstrated a significant correlation between WEE1 expression, a crucial G2M checkpoint regulator, and HMA-29 risk (p = 2.13e-4, Supplementary Fig. 6). Notably, WEE1 inhibition has been explored as a potential therapeutic strategy in MDS and AML ex vivo, particularly in combination with PARP-1 inhibitors [40]. Additionally, WEE1 inhibition can eliminate AML cells when synergizing with CHK1 inhibitor [41, 42], and there were ongoing clinical trials for the WEE1 inhibitor adavosertib in solid malignancies currently [43]. G2M Checkpoint: Re-establishing the G2/M checkpoint has been linked to azacitidine resistance [39]. Our analysis demonstrated a significant correlation between WEE1 expression, a crucial G2M checkpoint regulator, and HMA-29 risk (p = 2.13e-4, Supplementary Fig. 6). Notably, WEE1 inhibition has been explored as a potential therapeutic strategy in MDS and AML ex vivo, particularly in combination with PARP-1 inhibitors [40]. Additionally, WEE1 inhibition can eliminate AML cells when synergizing with CHK1 inhibitor [41, 42], and there were ongoing clinical trials for the WEE1 inhibitor adavosertib in solid malignancies currently [43]. E2F Targets: As a master regulator of cell-cycle progression, E2F1 contributes to dysregulated granulocyte differentiation and AML initiation/maintenance, in conjunction with C/EBPα [44, 45]. E2F7 has also been implicated in promoting AML cell cycle progression while inhibiting monocytic differentiation [46]. Both E2F1 and E2F7 were included in the target module and correlated significantly with HMA-29 risk (Supplementary Figs. 7 & 8). E2F Targets: As a master regulator of cell-cycle progression, E2F1 contributes to dysregulated granulocyte differentiation and AML initiation/maintenance, in conjunction with C/EBPα [44, 45]. E2F7 has also been implicated in promoting AML cell cycle progression while inhibiting monocytic differentiation [46]. Both E2F1 and E2F7 were included in the target module and correlated significantly with HMA-29 risk (Supplementary Figs. 7 & 8). MYC Targets: accumulated evidence indicates that c-Myc serves as a key role in apoptosis prevention, drug resistance development, and leukemogenesis progression [47, 48]. Several inhibitors, including HDAC inhibitors, BET inhibitors, and CDK9 inhibitors, have been found to enhance the anti-leukemic activity of HMAs by suppressing c-MYC activity or expression, leading to ongoing clinical trials [43]. Additionally, the XPO1 inhibitor Selinexor demonstrated synergy with azacitidine by suppressing the XPO1/eIF4E/c-MYC signaling pathway [49]. Selinexor has been shown to enhance the effects of HMAs in eliminating AML/MDS cells [50, 51] and has exhibited promising efficacy in clinical trials [52, 53]. MYC Targets: accumulated evidence indicates that c-Myc serves as a key role in apoptosis prevention, drug resistance development, and leukemogenesis progression [47, 48]. Several inhibitors, including HDAC inhibitors, BET inhibitors, and CDK9 inhibitors, have been found to enhance the anti-leukemic activity of HMAs by suppressing c-MYC activity or expression, leading to ongoing clinical trials [43]. Additionally, the XPO1 inhibitor Selinexor demonstrated synergy with azacitidine by suppressing the XPO1/eIF4E/c-MYC signaling pathway [49]. Selinexor has been shown to enhance the effects of HMAs in eliminating AML/MDS cells [50, 51] and has exhibited promising efficacy in clinical trials [52, 53]. Targeting these pathways may provide strategies to overcome HMA resistance in myeloid neoplasms. Although HMA-29 showed robust predictive performance across multiple transcriptomic cohorts, this study was based primarily on bulk transcriptomic datasets and therefore could not resolve the spatial distribution of HMA-response-associated gene expression within the bone marrow microenvironment. Future studies incorporating spatial transcriptomics may help determine whether HMA-29-related genes are preferentially expressed in leukemic blasts, stromal components, immune cell populations, or specific hematopoietic niches, and may further clarify the spatial mechanisms underlying HMA sensitivity or resistance. The scRNA-seq may help identify the cellular sources of HMA-29-related genes, this study did not include single-cell validation. HMA-29 was constructed using bulk transcriptomic datasets as a clinically accessible response-prediction model. Because scRNA-seq and bulk RNA-seq differ in data structure, sparsity, dropout effects, normalization, and cell-type composition, direct application of a bulk-derived signature to single-cell data requires careful methodological validation. Moreover, scRNA-seq alone does not retain spatial information, and publicly available single-cell datasets with clearly annotated HMA response in MDS remain limited. Future studies based on paired bulk RNA-seq, scRNA-seq, spatial transcriptomics, and clinical response data may further clarify the cellular and spatial context of HMA-29. Several limitations of this study should be acknowledged. First, although HMA-29 was evaluated across multiple publicly available cohorts, the overall sample size remained relatively limited. This limitation may affect the stability and generalizability of the predictive model, and larger prospective multicenter cohorts are needed to further validate its clinical utility. Second, the analyzed cohorts included different types of myeloid neoplasms, including MDS, AML, CMML, and other related entities. Although these diseases share certain biological features and are frequently treated with hypomethylating agents, they differ in clinical behavior, disease biology, mutational landscape, blast burden, and treatment context. Therefore, incorporating different myeloid malignancies into a combined analysis may introduce heterogeneity and potentially influence model performance. Future studies should validate HMA-29 in larger disease-specific cohorts and determine whether disease-adapted versions of the model are required for optimal clinical application. Conclusion In this study, we developed the HMA-29 model, a novel predictive tool for assessing HMA response in patients with myeloid neoplasms. The HMA-29 model demonstrated superior predictive performance compared to existing genetic and transcriptomic markers and also exhibited prognostic significance. Its stability across treatment phases and strong association with cellular mitosis suggest an intrinsic link to myeloid neoplasms. Furthermore, the implicated signaling pathways offer valuable insights into potential targets for future HMA-based combination therapies. Supplementary Information Below is the link to the electronic supplementary material. Supplementary File 1 The sample dendrogram and trait heatmap (PDF 14.1 KB) Supplementary File 1 The sample dendrogram and trait heatmap (PDF 14.1 KB) Supplementary File 2 The results of gene enrichment analysis of target module based on GO (A), KEGG (B) and REActome (C) databases (PDF 14.0 KB) Supplementary File 2 The results of gene enrichment analysis of target module based on GO (A), KEGG (B) and REActome (C) databases (PDF 14.0 KB) Supplementary File 3 The ROC analysis of HMA-29 risk on the training set, GSE76203, GSE61162, BEATAML, and GSE272173, based on non-adjusted RNA-seq data (PDF 8.04 KB) Supplementary File 3 The ROC analysis of HMA-29 risk on the training set, GSE76203, GSE61162, BEATAML, and GSE272173, based on non-adjusted RNA-seq data (PDF 8.04 KB) Supplementary File 4 The ROC analysis of HMA-29 risk on subsets, defined by methods of high-throughput (A, RNAseq vs. microarray), disease types (B, MDS vs. CMML vs. AML), detail drugs (C, decitabine vs. azacitidine) and sampling origins (D, BMMC vs. PBMC vs. CD34 + cells) (PDF 14.0 KB) Supplementary File 4 The ROC analysis of HMA-29 risk on subsets, defined by methods of high-throughput (A, RNAseq vs. microarray), disease types (B, MDS vs. CMML vs. AML), detail drugs (C, decitabine vs. azacitidine) and sampling origins (D, BMMC vs. PBMC vs. CD34 + cells) (PDF 14.0 KB) Supplementary File 5 Comparison of clinical variables between HMA-29 defined response-like and resistant-like groups (A. blast percentage; B. WBC count; C HGB level; D PLT count) (PDF 247 KB) Supplementary File 5 Comparison of clinical variables between HMA-29 defined response-like and resistant-like groups (A. blast percentage; B. WBC count; C HGB level; D PLT count) (PDF 247 KB) Supplementary File 6 Correlation of WEE1 expression and HMA-29 scores (PDF 253 KB) Supplementary File 6 Correlation of WEE1 expression and HMA-29 scores (PDF 253 KB) Supplementary File 7 Correlation of E2F1 expression and HMA-29 scores (PDF 491 KB) Supplementary File 7 Correlation of E2F1 expression and HMA-29 scores (PDF 491 KB) Supplementary File 8 Correlation of E2F7 expression and HMA-29 scores (PDF 1.28 MB) Supplementary File 8 Correlation of E2F7 expression and HMA-29 scores (PDF 1.28 MB)