Integrated Transcriptomic and Functional Analyses Reveal lncRNA‐miRNA‐mRNA Axis in AML Relapse After Allo‐HSCT

ABSTRACT

Relapse after allogeneic hematopoietic stem cell transplantation (allo‐HSCT) remains the leading cause of treatment failure in acute myeloid leukemia (AML), yet the non‐coding RNA‐mediated regulatory mechanisms in relapse‐driving tumor clones are not fully understood. In this study, we performed transcriptomic profiling of CD34⁺ cells from the bone marrow of patients who had relapsed after allo‐HSCT and from those in sustained remission, and observed that long non‐coding RNAs (lncRNAs) exhibited the largest magnitude of differential expression relative to other RNA classes. To systematically elucidate their function, we integrated correlation analysis, neighborhood topology, and thermodynamic modeling to construct relapse‐associated competing endogenous RNA (ceRNA) networks. These networks highlighted multiple lncRNA‐miRNA‐mRNA axes potentially promoting leukemic progression. Functional interrogation using CRISPR‐based xenograft assays and a single‐cell CRISPR screening platform (Perturb‐seq‐like strategy) revealed that relapse‐associated lncRNAs regulate AML cell proliferation and survival. Among them, SNHG8 emerged as a representative regulator that promotes relapse through the SNHG8‐miR‐625‐ZC3H13/15 signaling axis. SNHG8 knockdown markedly impaired AML cell growth and triggered apoptosis. Furthermore, treatment with hypomethylating agents (HMAs) such as azacitidine and decitabine differentially modulated lncRNA/ceRNA expression signatures, thereby linking clinical interventions to transcriptomic regulation. Together, our findings identify lncRNA‐ceRNA networks in post‐transplant relapse of AML and identify the SNHG8 axis as a potential therapeutic vulnerability, providing a rationale for further investigation of lncRNA‐targeted strategies in this clinical setting.

Transcriptomic profiling of CD34+ cells from AML patients relapsing after allo‐HSCT reveals lncRNA‐driven ceRNA networks, with the SNHG8‐miR‐625‐ZC3H13/15 axis identified as a key relapse‐promoting module through CRISPR screening and single‐cell perturbation. Hypomethylating agents reverse this axis, linking clinical intervention to ceRNA regulation.

Abbreviations

allo‐HSCT

allogeneic hematopoietic stem cell transplantation

HMAs

hypomethylating agents

lncRNAs

long non‐coding RNAs

1. Introduction

Acute myeloid leukemia (AML) is one of the most common hematologic malignancies, and allogeneic hematopoietic stem cell transplantation (allo‐HSCT) remains the most effective curative therapy. However, relapse after allo‐HSCT (20%–50%) continues to be the leading cause of treatment failure and mortality [1]. Despite advances in measurable residual disease (MRD) monitoring [1, 2, 3] and novel therapeutic strategies [4, 5], the ability to predict who will relapse, when relapse will occur, and through which molecular trajectories remains limited. In particular, the molecular programs within hematopoietic stem and progenitor cells (HSPCs) that drive relapse remain poorly defined [6].

Non‐coding RNAs (ncRNAs), especially long non‐coding RNAs (lncRNAs) and microRNAs (miRNAs), have emerged as critical regulators of hematopoiesis, leukemogenesis, and therapy resistance [7, 8]. lncRNAs can act through diverse mechanisms, including chromatin remodeling, transcriptional regulation, and post‐transcriptional modulation [9, 10]. A widely recognized model is the competing endogenous RNA (ceRNA) axis, in which lncRNAs function as sponges for miRNAs to regulate downstream mRNA stability and translation [11, 12]. However, the strength and biological relevance of ceRNA interactions depend on multiple factors, such as RNA abundance, subcellular localization, and binding accessibility [13]. Indeed, the physiological relevance of many proposed ceRNA interactions has been questioned, necessitating robust, multi‐modal approaches to identify bona fide regulatory axes. Whether such lncRNA‐centered ceRNA networks are systematically rewired in HSPCs during post‐transplant relapse of AML remains unclear.

Here we systematically investigated the role of lncRNAs in AML relapse after allo‐HSCT (Figure 1). Transcriptome profiling of CD34⁺ cells revealed that lncRNA alterations were among the most prominent differences between relapse and remission, suggesting a potential regulatory role. Because lncRNAs can act through diverse mechanisms, we focused on their function as competing endogenous RNAs (ceRNAs) within miRNA‐mRNA networks. We constructed relapse‐associated ceRNA networks by integrating expression correlations, neighborhood topology [14], and thermodynamic modeling [15], and validated functional relevance through in vivo CRISPR screening, Perturb‐seq‐like strategy, and mechanistic assays of representative candidates such as SNHG8. Finally, given that hypomethylating agents (HMAs) are clinically explored for relapse prevention after allo‐HSCT but with mixed results [16, 17], we examined how HMA exposure modulates relapse‐associated lncRNA/ceRNA networks, thereby linking mechanistic insights to potential therapeutic implications.

FIGURE 1.

Transcriptomic profiling of CD34⁺ cells from AML patients relapsing after allo‐HSCT reveals lncRNA‐driven ceRNA networks, with the SNHG8‐miR‐625‐ZC3H13/15 axis identified as a key relapse‐promoting module through CRISPR screening and single‐cell perturbation. Hypomethylating agents reverse this axis, linking clinical intervention to ceRNA regulation.

2. Materials and Methods

2.1. Clinical Sample Collection and Processing

Bone marrow aspirates were obtained from acute myeloid leukemia (AML) patients who received allo‐HSCT at our institute. Three cohorts were included: patients with hematologic relapse after allo‐HSCT (RE, n = 5), patients who remained in complete remission within 2 years after allo‐HSCT (CR, n = 5), and healthy donors (NC, n = 4). Baseline characteristics of the patients are listed in Table S1. Approximately 4 mL of bone marrow was collected in EDTA‐anticoagulated tubes. Plasma was separated by centrifugation and stored at −80°C. BMMCs were isolated by Ficoll density gradient centrifugation. CD34⁺ HSPCs were purified using CD34 MicroBeads (Miltenyi#130–046‐702) following the manufacturer's instructions, resuspended in TRIzol (Invitrogen#15596026), snap‐frozen in liquid nitrogen, and stored at −80°C for downstream RNA extraction. Post‐enrichment purity was assessed by flow cytometry (Figure S1).

2.2. CD34 ⁺ Cell Isolation and Culture

CD34⁺ cells were magnetically isolated using CD34 MicroBeads with MS/LS columns on a magnetic separator in PBS containing 0.5% BSA and 2 mM EDTA. Cells were maintained in one of the following media: MyeloCult H5100 (StemCell#05150), StemSpan SFEM II (StemCell#09605), or IMDM (Gibco#12440053), supplemented as indicated with recombinant human cytokines and additives, including SCF (StemCell #78062), FLT3L (Miltenyi#130–096‐474), TPO (StemCell #78210.1), IL‐6 (MCE# HY‐P7044), IL‐3 (MCE# HY‐P70576), G‐CSF (MCE# HY‐P73523), LDL (MCE# HY‐NP012), hydrocortisone sodium hemisuccinate (Sigma#H2270), and FBS (Gibco# A5669701). For LTC‐IC conditions, MyeloCult H5100 was supplemented with 10⁻⁶ M hydrocortisone sodium hemisuccinate, 4 ng/mL IL‐3, 190 ng/mL G‐CSF, and 4 ng/mL SCF, and cells were co‐cultured on irradiated (80 Gy) feeder layers of engineered mouse fibroblasts expressing human IL‐3, G‐CSF, and SCF. For SFEM II conditions, cultures contained 100 ng/mL SCF, 100 ng/mL FLT3L, and 50 ng/mL TPO. For IMDM‐based conditions, IMDM was supplemented with 20% FBS, 20 ng/mL SCF, 20 ng/mL FLT3L, 20 ng/mL TPO, 20 ng/mL IL‐6, and 10 ng/mL IL‐3. For BBMCs sourced CD34⁺ cells, freshly isolated cells (1 × 10⁵/well) were pre‐stimulated for 48 h in IMDM +10% FBS with SCF (100 ng/mL), FLT3L (100 ng/mL), and TPO (50 ng/mL); Retronectin‐coated 24‐well plates were then prepared, cells were resuspended in IMDM +10% FBS with the above cytokines and polybrene (6 μg/mL), and lentiviral transduction was performed at MOI = 1 for 5–6 h, after which cells were washed, re‐cultured, and analyzed for CD34 expression by flow cytometry 48 h later.

2.3. RNA Extraction, Library Preparation, and Sequencing

Total RNA was extracted using TRIzol according to the manufacturer's instructions. Ribosomal RNA was depleted, and strand‐specific libraries were prepared for long RNA sequencing (lncRNA, mRNA, circRNA) using paired‐end 150 bp (PE150) reads. For small RNA sequencing, libraries enriched for short fragments were constructed and sequenced in single‐end 50 bp (SE50) mode. Clean reads were aligned to the human genome (hg38), and transcript quantification was performed for lncRNA, mRNA, circRNA, and miRNA. Differential expression analyses were carried out across groups.

2.4. Thermodynamics and Expression‐Integrated ceRNA Scoring

We evaluated lnc‐mi‐mRNA triplets curated from whole‐transcriptome data. Reference sequences were prepared as FASTA, with 3′UTRs reconstructed from hg38 where needed. Expression levels (TPM), pairwise correlations, and KNN distances were integrated with thermodynamic calculations (RNAup, RNAduplex, RNAhybrid, RNAplfold) to compute binding free energies (ΔG_eff), binding probabilities (P_bind), and ceRNA scores. Details of computational modeling, integration modules (stoichiometry, correlation, KNN), scoring formulas, and significance assessment by permutation tests are described in the Supporting Information and Methods Doc S1.

2.5. sgRNA Library Design and Construction

Based on differentially expressed lncRNAs (top 193 genes), we employed the pLV‐hEF1a‐Puro‐U6/7SK dual‐gRNA vector system, which expresses two sgRNAs per construct. To ensure maximum disruptive impact on lncRNA function, we designed our library based on the following principles: (i) For each lncRNA, we targeted the Transcription Start Site (TSS) and the first exon, as these regions are most critical for transcript initiation and stability. (ii) We utilized the pLV‐hU6‐sgRNA1‐7SK‐sgRNA2 vector system. This specific backbone allows for the simultaneous expression of two distinct sgRNAs (driven by hU6 and 7SK promoters, respectively) from a single plasmid. (iii) For each lncRNA gene, we designed 3 pairs of sgRNAs. Each pair was positioned to flank the target region, aiming for a genomic deletion typically ranging from 500 bp to 1.5 kb. This ensures the removal of essential promoter elements or functional structural motifs of the lncRNA. Oligonucleotides were synthesized (Table S2), cloned into pLV‐hEF1a‐Puro‐U6/7SK dual‐gRNA vectors, and validated by next‐generation sequencing. Lentiviral particles were packaged in HEK293T cells (ATCC# CRL‐3216) and concentrated for use in AML cell transduction. To verify knockout efficiency, Reverse Transcription quantitative PCR (RT‐qPCR) was performed on representative targets (SNHG5, SNHG8, and ARHGAP31‐AS1) in bulk library‐transduced MOLM13 cells using target‐specific primers, with GAPDH as the internal reference gene. Relative expression was calculated by the 2^(−ΔΔCt) method. Genomic PCR was performed using primers flanking the dual‐gRNA deletion sites to confirm target locus disruption, and products were visualized by agarose gel electrophoresis.

2.6. AML Xenograft Mouse Model

Nude mice (6–8 weeks, female, Hfkbio. Inc) were injected subcutaneously with 1 × 10⁷ MOLM13 cells (DSMZ#ACC554) transduced with either Cas9 alone or Cas9 plus the sgRNA library. Tumor growth was monitored, and mice were sacrificed when tumor diameter reached 1–2 cm. Tumor DNA was extracted, and sgRNA representation was quantified by sequencing.

2.7. Single‐Cell Perturb‐Seq‐Like Strategy

AML cells with integrated Cas9 were transduced with the sgRNA library at low MOI (0.1) and selected with puromycin. Single cells were captured by FACS into 8‐tube strips, lysed, and subjected to Smart‐seq2‐based library construction (TaKaRa# R400751) with sgRNA barcoding. Libraries were purified, size‐validated (300–500 bp and 1.8 kb peaks), and sequenced on an Illumina NovaSeq 6000 platform (PE150).

2.8. HMA Treatment Assays

MOLM13 or CD34⁺ cells from NC, CR, and RE cohorts (1 × 10⁵/well, 12‐well plates) were treated with AZA (MCE#HY‐10586) or DAC (MCE#HY‐A0004) for 72 h, with medium refreshed every 24 h. Control and treated cells were harvested for downstream RNA extraction and sequencing. For dose–response validation of the SNHG8 ceRNA axis, MOLM13 cells and RE‐derived CD34⁺ cells were treated with AZA at 250, 500, or 750 nM, or DAC at 50, 75, or 100 nM, for 72 h with daily medium changes. DMSO served as vehicle control. Total RNA was extracted and subjected to quantitative RT‐PCR for SNHG8, miR‐625‐5p, and ZC3H13 as described above.

2.9. shRNA Design

shRNAs targeting lncRNAs were designed using public algorithms. Two independent shRNA sequences per target were selected to minimize off‐target effects. Oligonucleotides were cloned into a lentiviral shRNA expression vector (pLKO.1 puromycin). After lentivirus was produced, AML cell lines (MOLM13 and THP‐1) were infected with lentivirus at MOI = 1 in the presence of 6 μg/mL polybrene, followed by puromycin selection (1–2 μg/mL) for 72 h. Knockdown efficiency was verified by qPCR, and functional assays (cell proliferation, cell cycle, apoptosis) were performed as described above.

2.10. Statistical Analysis

Comparisons between two groups were assessed using unpaired two‐tailed Student's t‐test. For transcriptomic analyses, differential expression was determined using DESeq2 with Benjamini‐Hochberg correction for false discovery rate (FDR). GSEA was performed with MSigDB hallmark, Reactome, and GO‐BP collections. A p/q‐value < 0.05 or FDR < 0.05 was considered statistically significant.

3. Results

3.1. Altered Non‐Coding and Coding RNA Profiles in CD34 ⁺ Cells From Relapsed AML‐HSCT Patients

To characterize transcriptional changes underlying relapse after allogeneic HSCT, we analyzed bone marrow CD34⁺ cells from relapsed patients (RE, n = 5), patients in complete remission (CR, n = 5), and healthy controls (NC, n = 4). Using magnetic‐activated cell sorting and whole‐transcriptome sequencing, we profiled both long (lncRNA/circRNA/mRNA) and short (miRNA) transcripts (Figure 2A). Principal component analysis (PCA) showed that expression profiles of lncRNAs, miRNAs, and mRNAs clearly segregated RE samples from CR and NC groups, while CR and NC clustered closely (Figure 2B), indicating that relapse CD34⁺ cells harbor distinct transcriptional states. Differential expression analyses across groups showed that lncRNAs accounted for the greatest number of differentially expressed transcripts (Figure 2C, Figure S2), suggesting that lncRNA dysregulation may be involved in relapse biology.

FIGURE 2.

Transcriptomic alterations in CD34+ cells from AML‐HSCT relapse patients. (A) Bone marrow samples from relapsed (RE, n = 5), complete remission (CR, n = 5), and normal controls (NC, n = 4) were subjected to CD34+ sorting and whole‐transcriptome sequencing of long and short RNAs. (B) PCA of lncRNA, miRNA, and mRNA expression showing clear separation of RE from CR/NC groups. (C) Volcano plots highlighting differential expression of lncRNAs, miRNAs, and mRNAs across groups. (lncRNA 1664 [up 577, down 1087], miRNA 1206, mRNA 3173) (D) GO‐BP enrichment of mRNA differences between RE and CR. (E) Heatmap of the top 40 downregulated protein‐coding genes in RE relative to CR. (F) Hallmark GSEA showing the enriched Hallmark gene sets (CR vs. RE). (G) Differential pathway enrichment in RE, CR, and NC (Hallmark).

To infer functional consequences, we examined mRNA expression changes between RE and CR. Gene ontology enrichment highlighted alterations in cell cycle regulation and immune activation (Figure 2D). The top 40 downregulated protein‐coding genes in RE were visualized by z‐score ranking (Figure 2E). Hallmark Gene set enrichment analysis (GSEA) revealed that RE samples were negatively associated with metabolism, IFN‐γ response, KRAS signaling, G2M checkpoint, graft‐versus‐host disease, and hypoxia pathways, but positively correlated with TNF‐NFκB signaling (Figure 2F). Comparison of CR with NC showed persistent abnormalities, including enhanced graft rejection, metabolic activation, KRAS signaling, and impaired DNA repair (Figure 2G, Figure S3), suggesting directions for clinical consolidation strategies.

3.2. Integrated Correlation‐KNN and Energy‐Based Analyses Identify ceRNA Regulatory Axis in Relapse CD34 ⁺ Cells

To explore ceRNA networks potentially promoting relapse, we constructed expression matrices of lncRNA, miRNA, and mRNA from 14 patients encompassing matched RE, CR, and NC samples. Following normalization, we performed KNN distance analysis across hierarchical layers (all RNA, lnc‐miRNA pairs, and functionally annotated lncRNA–miRNA–mRNA triplets) (Figure 3A,B, Figure S4). UMAP visualization demonstrated distinct clustering of ceRNA modules, from which the top 30 lnc‐mi‐mRNA triplets were selected based on their neighborhood distributions (Figure 3C,D).

FIGURE 3.

ceRNA regulatory networks inferred from correlation, KNN, and thermodynamic analyses. (A) Analytical pipeline integrating expression normalization, KNN distance computation, and ceRNA triplet assembly. (B) KNN‐based clustering across hierarchical layers: All RNAs, lnc‐miRNA, and lnc‐mi‐mRNA triplets. (C) UMAP visualization of lnc‐mi‐mRNA associations; top 30 triplets highlighted. (D) Sankey plot showing representative lnc‐mi‐mRNA regulatory relationships. (E) Thermodynamic modeling (RNA folding/binding ΔG_eff) supporting binding feasibility of miRNA‐target pairs. (F‐G) Ranked top 30 lnc‐miRNA (F) and mi‐mRNA (G) axes supported by correlation and energy integration. (H) Chord diagram illustrating the global ceRNA regulatory network in relapse CD34+ cells.

Parallel correlation analyses identified high‐confidence lnc‐mi‐mRNA associations (|Spearman's ρ| > 0.6). Structural energy modeling (RNA folding and binding free energy ΔG_eff) was then applied to assess binding feasibility (Figure 3E). Integration of correlation and thermodynamic evidence yielded ranked ceRNA regulatory axis, highlighting the top 30 lncRNA‐miRNA (Figure 3F) and miRNA‐mRNA (Figure 3G) interactions. A global ceRNA regulatory map was visualized using chord diagrams (Figure 3H).

3.3. AML‐HSCT RE‐Associated lncRNAs Exert Distinct Effects on AML Cell Survival and Proliferation in Xenograft Models

To obtain preliminary evidence whether AML‐HSCT RE‐associated lncRNAs directly influence the survival and proliferation of AML tumor cells, a CRISPR‐Cas9 sgRNA library targeting these lncRNAs was constructed and introduced into MOLM13 cells via lentiviral transduction. To confirm that the dual‐gRNA strategy effectively disrupted lncRNA transcription, we performed qPCR and genomic PCR validation on representative targets. qPCR analysis of bulk library‐transduced MOLM13 cells demonstrated significant reduction in SNHG5, SNHG8, and ARHGAP31‐AS1 expression compared to non‐targeting controls, confirming RNA‐level disruption. Genomic PCR with primers flanking the deletion sites further confirmed the expected genomic deletions at the target loci (Figure S5). Following subcutaneous inoculation into nude mice, the frequency dynamics of sgRNAs within tumor cells were monitored (Figure 4A). After 30 days, tumor diameters exceeded 1 cm (Figure 4B). Genomic DNA from tumor cells harvested at day 0 and day 30 was used for amplification of sgRNA/GAPDH fragments (600 bp) with template‐switching oligo (TSO) to generate sequencing libraries (Figure 4C). Deep sequencing in NSG mice revealed depletion of multiple lncRNA‐targeting sgRNAs, suggesting loss of growth advantage for cells carrying these perturbations (Figure 4D).

FIGURE 4.

Functional relevance of relapse‐associated lncRNAs in AML xenograft models. (A) Construction of a CRISPR‐Cas9 sgRNA library targeting relapse‐associated lncRNAs and transduction into MOLM13 cells. (B) Schematic of NSG xenograft (n = 2) assays; tumors reached > 1 cm by day 30. (C) Library amplification using TSO primers from day 0 and day 30 tumors. (D) Relative sgRNA abundance showing depletion of multiple sgRNAs. SgRNA abundance was quantified by NGS of tumor genomic DNA; log2 fold change (Day 30/Day 0) is shown. Non‐targeting sgRNAs served as normalization controls. Error bars represent ± SEM of n biological replicates. (E) Classification of lncRNAs into depletion, ambiguous, or enrichment groups based on functional outcomes (F‐G) Cumulative correlation scores ranking miRNA (F) and mRNA (G) partners of depletion‐associated lncRNAs, indicating their network consistency.

We next used correlation‐based integration of lncRNA–miRNA and lncRNA–mRNA interactions to classify lncRNAs according to their functional impact in vivo: reduced growth advantage (depletion), unchanged (ambiguous), or enhanced growth advantage (enrichment). For each category, the total number of lnc‐mi‐mRNA mappings, as well as the numbers of unique miRNAs and mRNAs were summarized (Figure 4E). Within the depletion group, cumulative correlation scores were computed separately for miRNAs (Figure 4F) and mRNAs (Figure 4G), where higher absolute cumulative values reflected stronger concordance between depletion‐associated lncRNAs and their mapped partners. The same analyses were also performed for the ambiguous and enrichment groups (Figure S6).

3.4. Perturb‐Seq‐Like Strategy Enables High‐Throughput Dissection of AML‐HSCT RE‐Associated lncRNA Effects on Downstream mRNA Programs and Cellular States

To systematically evaluate the direct effects of lncRNAs on AML cells, a Perturb‐seq‐like strategy was performed on sgRNA library‐transduced cells (Figure 5A). Three AML‐derived cell lines were generated through lentiviral infection, blasticidin selection, and stable Cas9 integration (Figure 5B). A second round of infection with sgRNAs (MOI = 0.1) followed by puromycin selection was then applied (Figure 5C), with MOLM13 exhibiting the most robust Cas9/sgRNA integration and stable expression. PCR with specific primers confirmed an increase in sgRNA copy numbers after two rounds of selection (Figure 5D). Smart‐seq combined with TSO‐based library preparation yielded transcriptomes from 980 single cells, after cells harboring multiple sgRNAs were excluded, including 596 lncRNA‐sgRNA‐captured cells and 384 cells carrying either non‐targeting sgRNAs or no detectable sgRNAs. Sixteen distinct lncRNAs were represented by ≥ 4 cells each (Figure 5E).

FIGURE 5.

Perturb‐seq‐like strategy analysis of relapse‐associated lncRNAs in AML cells. (A) Perturb‐seq‐like strategy workflow combining CRISPR‐Cas9 sgRNA library with single‐cell RNA sequencing. (B&C) Establishment of AML cell lines with stable Cas9 expression and sgRNA integration. (D) Schematic of two‐step infection and selection strategy. (E) Single‐cell sequencing results: 980 cells profiled, with 596 capturing lncRNA‐targeting sgRNAs covering 16 lncRNAs. (F‐G) UMAP plots showing transcriptional divergence between sgTarget and sgNC cells, and the targeted gene label is shown in Figure S5. (H‐J) Gene expression changes upon knockout of representative lncRNAs (SNHG5, SNHG8, ARHGAP31‐AS1). (K‐M) Hallmark pathway alterations: SNHG8 KO suppressed PI3K‐AKT and oxidative phosphorylation, SNHG5 KO inhibited TNF‐NFκB, WNT, and PI3K‐AKT, while ARHGAP31‐AS1 KO suppressed IL6‐JAK‐STAT3.

UMAP dimensionality reduction based on single‐cell transcriptomes revealed distributions of sgTarget versus sgNC cells (Figure 5E) and heterogeneous clustering across different lncRNA perturbations (Figure 5F, Figure S7). Consistent with the bulk‐level knockout validation (Figure S5A–B), single‐cell expression profiles also showed trends toward reduced target lncRNA expression in sgTarget cells compared to sgNC cells (Figure S5D), further supporting functional perturbation at the single‐cell level despite the inherent dropout noise of scRNA‐seq. Integrating lnc‐mi‐mRNA correlation analyses, differential expression and pathway perturbations were examined for SNHG5, SNHG8, and ARHGAP31‐AS1 knockouts (Figure 5H–J). Hallmark pathway analysis suggested that SNHG8 knockout was associated with enhancement of KRAS downregulation signaling, suppression of PI3K‐AKT and oxidative phosphorylation pathways; SNHG5 knockout was associated with downregulation of TNF‐NFκB, WNT, and PI3K‐AKT signaling pathways, whereas ARHGAP31‐AS1 knockout was linked to reduced IL6‐JAK‐STAT3 signaling signatures in AML cells (Figure 5K–M).

3.5. The SNHG8‐miR‐625‐5p‐ZC3H13/15 Axis Contributes to Relapse After AML‐HSCT

Our integrative analyses suggested that relapse‐associated lncRNAs, particularly SNHG5 and SNHG8, positively correlate with pathways upregulated in relapse, and that HMA treatment suppresses these lncRNAs and their associated signaling. To further validate functional relevance, we focused on the SNHG8‐miR‐625‐5p‐ZC3H13/15 axis.

We collected BMMCs from AML patients at diagnosis (DX), at post‐HSCT relapse (RE), and from matched donors (NC). CD34⁺ cells were isolated, and qPCR revealed step‐wise up‐regulation of SNHG8, ZC3H13, ZC3H15, and TSC2, and concomitant down‐regulation of miR‐625 along the NC → DX → RE trajectory (Figure 6A).

FIGURE 6.

The SNHG8‐miR‐625‐5p‐ZC3H13/15 axis promotes AML relapse after HSCT. (A) qPCR analysis of CD34+ cells from respective donors (NC), AML at diagnosis (DX), and respective relapsed post‐HSCT patients (RE) with specific primers (B) Lentiviral shRNA‐mediated knockdown of SNHG8 in MOLM13 and THP‐1 cells reduced SNHG8 expression, increased miR‐625/7705, and decreased ZC3H13/15 and TSC2 mRNA. (C) CCK‐8 assays showing impaired proliferation after SNHG8 knockdown. (D) Cell cycle analysis demonstrating increased G0 and decreased G1/2 phase fractions. (E‐F) Apoptosis assays indicating increased apoptotic cell populations upon SNHG8 knockdown.

To directly probe SNHG8 function, we designed shRNAs and delivered them into MOLM13 and THP‐1 cells using lentiviral vectors. Both shRNAs efficiently reduced SNHG8 expression after 72 h, which was accompanied by increased expression of miR‐625 and decreased expression of ZC3H13, ZC3H15, and TSC2 (Figure 6B). Functionally, SNHG8 knockdown significantly inhibited proliferation as measured by CCK‐8 assays (Figure 6C). Cell cycle analysis revealed accumulation of cells in G0 phase and reduction in G1/2 phases (Figure 6D). Moreover, apoptosis assays demonstrated a significant increase in apoptotic fractions following SNHG8 knockdown (Figure 6E,F).

Together, these findings implicate the SNHG8‐miR‐625‐ZC3H13/15 axis as a potential contributor to relapse after AML‐HSCT.

3.6. Hypomethylating Agents Reverse Relapse‐Associated ceRNA Networks, Including the SNHG8‐miR‐625‐ZC3H13/15 Axis

Given the clinical application of HMA in post‐transplant MRD prevention, we next examined their impact on RNA profiles. Relapse‐derived CD34⁺ cells (Figure S8) and the AML cell line MOLM13 were treated with azacitidine (AZA) or decitabine (DAC), followed by RNA sequencing (Figure 7A). Both lncRNA and mRNA were more frequently upregulated by DAC than AZA (Figure 7B). Integrating with relapse‐specific expression changes, we observed that AZA induced limited mRNA changes in RE‐CD34⁺ cells but suppressed expression in MOLM13, whereas DAC broadly elevated expression in both cell types (Figure 7C). For lncRNAs, AZA suppressed expression in RE‐CD34⁺ cells but promoted it in MOLM13, while DAC generally enhanced expression in both (Figure 7D).

FIGURE 7.

Effects of hypomethylating agents (HMAs) on non‐coding and coding RNAs. (A) Experimental design: RE‐CD34+ and MOLM13 cells treated with AZA or DAC, followed by RNA‐seq. (B) Proportion of upregulated lncRNAs/mRNAs post‐treatment induced by DAC than AZA (C‐D) Scatter plots comparing baseline RE/CR differences (x‐axis) with AZA/DAC‐induced changes (y‐axis) in mRNA (C) and lncRNA (D). (E) Summary of lncRNAs/mRNAs reversed by AZA/DAC relative to RE/CR differences (F) Hallmark, Reactome, and GO‐BP enrichment showing different profiles of AZA and DAC effects on AML cells. (G) Pathway‐specific effects of AZA or DAC on MOLM13 or RE‐CD34+ cells Expression values were normalized to TPM/CPM. Comparisons between treatment groups were performed using an unpaired t‐test. Significance thresholds: *p < 0.05, **p < 0.01, ***p < 0.001.

We further summarized lncRNA/mRNA changes reversed by AZA/DAC relative to RE‐CR differences (Figure 7E, Figure S9). Functional enrichment analysis revealed distinct effects: AZA primarily suppressed E2F targets and epigenetic regulation, while DAC enhanced TNF‐NFκB signaling, graft rejection, and immune activation (Figure 7F). Pathway‐specific analyses demonstrated non‐overlapping actions: AZA downregulated TNF‐NFκB signaling in RE‐CD34⁺ but not MOLM13 cells, while DAC upregulated this pathway in MOLM13 but not RE‐CD34⁺ cells. Conversely, AZA suppressed graft rejection pathways in both cell types, whereas DAC enhanced them (Figure 7G, Figure S10). These findings suggest that AZA and DAC exert distinct and potentially complementary effects with potential implications for relapse prevention.

To directly assess whether HMA exposure modulates the SNHG8 ceRNA axis identified in earlier sections, we treated MOLM13 cells and relapse‐derived CD34⁺ (RE‐CD34⁺) cells with increasing concentrations of AZA (250, 500, and 750 nM) or DAC (50, 75, and 100 nM). Quantitative RT‐PCR revealed that both agents dose‐dependently suppressed SNHG8 expression and concurrently upregulated miR‐625‐5p, while the downstream targets ZC3H13 were correspondingly reduced (Figure 8A–D). These expression changes precisely recapitulated the effects observed upon shRNA‐mediated SNHG8 knockdown (Figure 6B), indicating that HMA treatment functionally phenocopies genetic perturbation of the SNHG8 axis. RNA‐seq data further confirmed concordant changes across the entire SNHG8 ceRNA module, including SNHG8, miR‐625‐5p, ZC3H13, ZC3H15, and TSC2, in both cell types treated with either agent (Figure 9A). Cross‐validation between RNA‐seq and qPCR measurements showed strong agreement (R² = 0.782, p < 0.01; Figure 9B), supporting the robustness of these findings.

FIGURE 8.

HMA treatment dose‐dependently modulates the SNHG8‐miR‐625‐ZC3H13 ceRNA axis. (A) Quantitative RT‐PCR showing dose‐dependent effects of azacitidine (AZA; 250, 500, and 750 nM) on SNHG8, miR‐625‐5p, and ZC3H13 expression in MOLM13 cells. DMSO served as vehicle control. (B) Same as (A) but in relapse‐derived CD34+ (RE‐CD34+) cells (C) Dose‐dependent effects of decitabine (DAC; 50, 75, and 100 nM) on SNHG8, miR‐625‐5p, and ZC3H13 expression in MOLM13 cells. (D) Same as (C) but in RE‐CD34+ cells. Data are presented as mean ± SD; ns, not significant; **p < 0.01; ***p < 0.001, by unpaired two‐tailed Student's t‐test.

FIGURE 9.

HMA treatment reverses relapse‐associated ceRNA networks. (A) Heatmap of RNA‐seq log2FC values for the SNHG8 ceRNA axis members (SNHG8, miR‐625‐5p, ZC3H13, ZC3H15, TSC2) in MOLM13 and RE‐CD34+ cells treated with vehicle (Veh), AZA, or DAC. (B) Scatter plot of RNA‐seq versus qPCR log2FC values for SNHG8 axis genes across both cell types and treatment conditions (R² = 0.782, p < 0.01). Circle and square symbols denote MOLM13 and RE‐CD34+ cells, respectively; colors indicate AZA (pink) or DAC (yellow) treatment. (C) Reversal Index (RI) heatmap for the top 30 ceRNA triplets under AZA and DAC treatment. Yellow‐highlighted rows indicate SNHG8‐associated axes (SNHG8‐miR625‐ZC3H13 and SNHG8‐miR625‐ZC3H15). (D) Stacked bar plot summarizing the proportion of ceRNA triplets classified as reversed, unchanged, or strengthened by AZA and DAC treatment. (E) Scatter plot comparing Reversal Index values of AZA (x‐axis) versus DAC (y‐axis) for individual ceRNA triplets. Circle size reflects network centrality; colors denote pathway annotations (Cell Cycle, E2F/Epigenetic, Metabolism, PI3K‐AKT, TNF‐NFkB/Immune, and Other). The SNHG8 axis (yellow circle with border) falls in the “both reversed” quadrant. Wilcoxon signed‐rank test p = 0.0004.

We next asked whether HMA treatment could reverse the relapse‐associated ceRNA regulatory networks. We computed a Reversal Index (RI) for each ceRNA triplet by comparing HMA‐induced expression changes against the RE‐versus‐CR baseline. Among the top 30 ceRNA triplets, AZA reversed 60% and DAC reversed 73% of relapse‐associated expression patterns, whereas only 27% (AZA) and 10% (DAC) were further strengthened (Figure 9C,D). Notably, the SNHG8‐miR‐625‐ZC3H13 and SNHG8‐miR‐625‐ZC3H15 triplets ranked among the most strongly reversed axes by both agents, as highlighted in the RI heatmap (Figure 9C). Direct comparison of AZA and DAC reversal capacities revealed that DAC produced significantly greater ceRNA network reversal than AZA (Figure 9E), with the SNHG8 axis falling into the “both reversed” region. Pathway‐level annotation further showed that reversed triplets were preferentially enriched in cell cycle, E2F/epigenetic, and PI3K‐AKT pathways, whereas metabolism‐ and immune‐related axes showed more heterogeneous responses between the two agents (Figure 9E). These results establish a direct mechanistic link between HMA treatment and the reversal of relapse‐associated ceRNA dysregulation and reveal that DAC exhibits broader and stronger ceRNA‐reversal activity than AZA.

4. Discussion

Our findings suggest that lncRNA signatures and ceRNA networks may serve as relapse risk after HSCT, and their integration with MRD monitoring could improve predictive power. Relapse‐associated lncRNAs, such as SNHG8, also represent potential therapeutic targets, with RNA‐based interventions (ASOs, siRNA, CRISPRi/a) offering translational promise [18, 19, 20, 21]. Importantly, our data demonstrate that HMA treatment directly impinges on the ceRNA networks identified in this study, rather than merely exerting broad transcriptomic effects. Both AZA and DAC dose‐dependently suppressed SNHG8, de‐repressed miR‐625‐5p, and reduced downstream ZC3H13/15 and TSC2, effectively reversing the relapse‐associated ceRNA axis. This convergence between genetic perturbation and pharmacologic intervention (HMA treatment) on the same molecular axis strengthens the biological plausibility of the SNHG8‐miR‐625‐ZC3H13/15 regulatory module in relapse pathogenesis. Moreover, the ceRNA reversal analysis revealed that DAC reversed a significantly larger proportion of relapse‐associated ceRNA triplets than AZA, with pathway‐level divergence: AZA preferentially targeted E2F/epigenetic axes, whereas DAC more broadly reversed immune‐related and PI3K‐AKT modules. These mechanistic distinctions may help explain the heterogeneous clinical outcomes observed with different HMA regimens and support the rationale for personalized selection or sequential combination of these agents in post‐transplant maintenance. Furthermore, the CRISPR and Perturb‐seq‐like platforms employed in this study provide a generalizable framework for functional interrogation of non‐coding RNAs, with applicability to other hematologic malignancies.

This study has several limitations. First, the patient cohort was small (RE, n = 5; CR, n = 5; NC, n = 4), limiting our ability to fully capture the heterogeneity of relapse. Second, the computational framework did not incorporate RNA‐binding proteins, RNA modifications, or subcellular localization, all of which may substantially affect ceRNA interactions. Third, functional validation primarily focused on SNHG8 as a representative lncRNA, leaving other candidate molecules less explored. Fourth, while we confirmed CRISPR‐mediated lncRNA disruption by qPCR and genomic PCR at the bulk population level, single‐cell‐level quantification of knockout efficiency was not performed owing to the technical constraints of the Smart‐seq2‐based Perturb‐seq workflow, where the entire single‐cell lysate is consumed for transcriptomic library construction. Future studies employing higher‐throughput droplet‐based perturbation platforms (e.g., CROP‐seq or Perturb‐CITE‐seq) may enable direct assessment of knockout efficiency at single‐cell resolution.

Future studies should expand patient cohorts and perform longitudinal sampling to capture dynamic transitions before and after relapse. Multi‐omic integration, including epigenomics, proteomics, and spatial transcriptomics, will be essential to dissect the coupling between lncRNA networks, chromatin states, and microenvironmental cues. Functional validation should be extended to additional lncRNA candidates using patient‐derived xenografts (PDX) [22], organoids [23], and single‐cell perturbation platforms [24] for more precise characterization. Mechanistically, non‐canonical ceRNA modalities, such as RBP‐mediated or RNA modification‐dependent interactions [25], merit further investigation to refine our understanding of lncRNA‐driven relapse biology.

Randomized trials of HMA maintenance after allo‐HSCT have yielded conflicting results. A phase III trial showed that subcutaneous azacitidine maintenance did not improve relapse‐free survival in high‐risk AML/MDS [26], whereas a multicenter trial demonstrated that prophylactic low‐dose decitabine plus rhG‐CSF significantly reduced relapse in selected MRD‐negative patients [27]. MRD‐guided pre‐emptive azacitidine, as in the RELAZA‐2 study, effectively delayed overt relaps [28], and oral CC‐486 achieved sustained exposure with encouraging relapse rates compared with historical controls [29]. Together, these findings suggest that the efficacy of HMAs depends on drug, regimen, timing, and patient selection. Our work extends these clinical observations by providing mechanistic evidence that HMA exposure alters relapse‐associated lncRNA and ceRNA networks, thereby linking therapeutic benefit to transcriptomic regulation.

Author Contributions

Fei Zhao: writing – review and editing, investigation, validation, methodology, conceptualization, data curation, writing – original draft. Wenhao Wang: writing – original draft, validation, data curation. Hongye Gao: software, formal analysis. Ping Zhu: supervision, software, methodology. Weihua Zhai: supervision, visualization, project administration. Yi He: formal analysis, supervision. Aiming Pang: formal analysis, investigation. Sizhou Feng: supervision. Mingzhe Han: supervision. Lina Wu: software, methodology. Shukai Qiao: methodology, software. Lei Zhang: funding acquisition, visualization, project administration. Erlie Jiang: supervision. Xiaolei Pei: writing – original draft, writing – review and editing, supervision, software, methodology, conceptualization, funding acquisition. Donglin Yang: resources, writing – review and editing, supervision, conceptualization, investigation, validation, project administration, data curation.

Funding

This work was supported by the Natural Science Foundation of Tianjin Municipality, 23JCZXJC00040, 23JCZXJC00220. Natural Science Foundation of Hebei Province, H2023206910 the Natural Science Foundation of Beijing, China, J230004. Tianjin Municipal Science and Technology Commission, 25ZXZSSS00510. National Key R&D Program of China, 2023YFC2508902 the National Natural Science Foundation of China, 82270152. Chinese Academy of Medical Sciences Innovation Fund for Medical Sciences (CIFMS), 2025‐I2M‐XHZY‐034.

Ethics Statement

The collection and treatment of clinical samples from the patients and the animal model are approved by the ethical committee of the Institute of Hematology and Blood Diseases Hospital, Chinese Academy of Medical Sciences (TJ2023009‐EC‐2).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Doc S1 Thermodynamics‐ and expression‐integrated scoring of lnc–mi–mRNA triplets.

Figure S1: Representative flow cytometry gating strategy for fluorescence‐activated cell sorting (FACS) of CD34+ cells. (A) Negative control (isotype‐matched antibody). (B) Representative patient sample. Bone marrow mononuclear cells were initially gated on forward scatter area (FSC‐A) versus side scatter area (SSC‐A) to identify the lymphocyte population (left panels). Viable cells were subsequently selected by excluding 4, 6‐diamidino‐2‐phenylindole (DAPI)‐positive dead cells using the Hoechst Blue channel (middle panels). CD34+ cells were then identified and sorted using a fluorescein isothiocyanate (FITC)‐conjugated anti‐CD34 antibody (right panels). The percentage of each gated population is indicated within the respective plot.

Figure S2: Pairwise comparisons among RE, CR, and NC groups showing changes in total RNA, lncRNA, miRNA, and circRNA. Differential expression was visualized by volcano plots.

Figure S3: Pathway and functional enrichment of RE‐associated transcriptional programs. (A) GSEA of Hallmark gene sets highlighting major pathway differences among RE, CR, and NC groups. (B) Enrichment analysis of significantly altered genes in RE versus CR within oncogenic signature gene sets and transcription factor–targeting gene sets.

Figure S4: GSEA‐Hallmark analyses of mRNAs from RE, CR, and NC groups. NES values across all gene sets were subjected to t‐SNE–based unsupervised dimensionality reduction, displaying sample distributions and clustering patterns.

Figure S5: Validation of CRISPR‐mediated lncRNA knockout efficiency. (A) qPCR analysis of SNHG5, SNHG8, and ARHGAP31‐AS1 expression in MOLM13 cells transduced with lncRNA‐targeting sgRNAs (sgTarget) or non‐targeting controls (sgNC), confirming significant reduction of all three lncRNAs upon CRISPR‐mediated disruption. (B) Agarose gel electrophoresis of genomic PCR products validating successful deletion of target loci. The 4.5 kb band represents the wild‐type allele, and the 1.2 kb band corresponds to the deletion product generated by the dual‐sgRNA strategy. (C) Schematic of the dual‐sgRNA deletion strategy for lncRNA knockout. sgRNA‐1 targets a region near the promoter/first exon (E1), and sgRNA‐2 targets the last exon (E4), resulting in excision of the intervening genomic segment between breakpoints BP1 and BP2. The dashed line indicates the deleted region. (D) Box plot of lncRNA expression levels [log2 (TPM + 1)] from single‐cell Perturb‐seq data for cells harboring SNHG5‐sg, SNHG8‐sg, or ARHGAP31‐AS1‐sg compared with non‐targeting controls (sgNC), demonstrating effective transcriptional suppression at the single‐cell level. Statistical significance was determined by unpaired two‐tailed Student's t‐test. **p < 0.01; ***p < 0.001.

Figure S6: Cumulative correlation analyses of lncRNA–miRNA and lncRNA–mRNA mappings (A–B). Cumulative correlation scores for enrichment‐group lncRNAs (C–D). Cumulative correlation scores for ambiguous‐group lncRNAs (E). Cumulative mRNA correlation analysis of depletion‐group lncRNAs with GO‐BP enrichment of the top 100 mapped genes.

Figure S7: Supplementary annotation corresponding to Figure 5G, showing sgRNA‐targeted lncRNAs.

Figure S8: CD34+ cells isolated from BMMCs by magnetic sorting and cultured in vitro for 3 days. Flow cytometry analysis showing the distribution of surface CD34 and CD38 expression.

Figure S9: Differentially expressed mRNAs identified in HMA‐treated AML cells (RE‐CD34 cells and MOLM13) before and after treatment, using a relaxed log2FC threshold (> 0.7 or < 0.7).

Figure S10: GSEA of significantly altered Hallmark pathways in RE versus CR before and after HMA treatment in AML cells (RE‐CD34 cells and MOLM13), shown separately for each condition.

Table S1: Clinical features of the individuals in the transcriptome sequence.

Table S2: sgRNA library.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.