Table 5_Local Promoter Methylation Disorder algorithm reveals bidirectional epigenetic disruption in DNMT3A-mutated AML and predicts azacitidine treatment response.xlsx

BackgroundDNMT3A mutations occur in 20-25% of acute myeloid leukemia (AML) cases and are associated with poor prognosis, yet the epigenetic mechanisms underlying treatment response remain poorly understood. Traditional methylation analyses focus on mean levels, overlooking the heterogeneity that may be central to therapeutic vulnerability. We developed a Local Promoter Methylation Disorder (LPMD) algorithm to quantify methylation instability and evaluate its clinical utility in predicting azacitidine response.
MethodsThe LPMD algorithm was developed using the GSE62298 dataset (68 AML patients: 15 DNMT3A-mutant, 53 wild-type) to quantify local methylation heterogeneity through a 1-kb sliding window across CpG sites. Algorithm performance was validated in an independent WGBS cohort (20 AML samples) and further confirmed via R882-stratified analysis in the TCGA-LAML cohort (n = 194). Clinical predictive value was assessed in GSE152710 (63 high-risk MDS/secondary AML patients receiving azacitidine), where differentially methylated disorder regions (DMDRs) were identified and a consensus feature selection strategy was employed to construct a predictive panel. Longitudinal samples (n = 153) enabled treatment dynamics analysis.
ResultsThe LPMD algorithm effectively captured DNMT3A mutation-associated epigenetic instability (Cohen’s d = 0.8, p < 0.001), with the strongest effects observed in the 5’UTR-Exon1 region (d = 0.74) and a gradient pattern from CpG islands to shores (d: 0.59→0.54→0.43). Genome-wide scanning identified 7,097 DMDRs exhibiting a striking bidirectional pattern: 85.3% showed decreased disorder (aberrant stabilization) while 14.7% showed increased disorder (maintenance failure), with the latter enriched in promoters (91% of high-priority DMDRs). Although genome-wide LPMD failed to predict azacitidine response, a 5-DMDR panel derived from multi-algorithm consensus achieved AUC = 0.777, 81% sensitivity, and 73% specificity. Treatment monitoring revealed a significant LPMD decrease at 3–5 months (-3.7%, p < 0.001), defining a critical window for efficacy assessment.
ConclusionsThe LPMD algorithm reframes DNMT3A-mutant AML from a hypomethylation paradigm to a methylation disorder paradigm, revealing dual mechanisms of aberrant stabilization and maintenance failure at distinct genomic regions. The 5-DMDR panel offers a practical tool for azacitidine response prediction, while dynamic LPMD monitoring provides a potential biomarker for therapeutic guidance. These findings establish methylation disorder as a clinically actionable dimension of epigenetic dysregulation in myeloid malignancies.

背景 DNMT3A突变在20%~25%的急性髓系白血病(acute myeloid leukemia, AML)病例中发生，且与不良预后相关，但目前对治疗应答背后的表观遗传机制仍知之甚少。传统甲基化分析多聚焦于平均甲基化水平，忽略了可能是治疗脆弱性核心的异质性特征。本研究开发了局部启动子甲基化紊乱(Local Promoter Methylation Disorder, LPMD)算法，以量化甲基化不稳定性，并评估其在预测阿扎胞苷(azacitidine)应答中的临床应用价值。
方法 本研究利用GSE62298数据集（包含68例AML患者：15例DNMT3A突变型、53例野生型）开发LPMD算法，通过对CpG位点采用1kb滑动窗口来量化局部甲基化异质性。算法性能在独立的全基因组亚硫酸氢盐测序(whole-genome bisulfite sequencing, WGBS)队列（20例AML样本）中得到验证，并通过TCGA-LAML队列（n=194）中按R882分层的分析进一步确认。临床预测价值在GSE152710队列（63例接受阿扎胞苷治疗的高危骨髓增生异常综合征(myelodysplastic syndrome, MDS)/继发AML患者）中进行评估，该队列中鉴定出差异甲基化紊乱区域(differentially methylated disorder regions, DMDRs)，并采用共识特征选择策略构建预测模型。纵向样本（n=153）则支持治疗动态学分析。
结果 LPMD算法可有效捕捉DNMT3A突变相关的表观遗传不稳定性（Cohen’s d=0.8，p<0.001），其中以5’UTR-外显子1区域效应最强（d=0.74），且呈现从CpG岛到CpG岸的梯度变化模式（d值：0.59→0.54→0.43）。全基因组扫描共鉴定出7097个DMDRs，呈现显著的双向特征：85.3%的区域表现为紊乱程度降低（异常稳定化），14.7%的区域表现为紊乱程度升高（维持失败），后者在启动子区域显著富集（占高优先级DMDRs的91%）。尽管全基因组LPMD评分无法预测阿扎胞苷应答，但基于多算法共识得到的5个DMDRs构建的预测模型的曲线下面积(Area Under the Curve, AUC)达0.777，灵敏度为81%，特异度为73%。治疗监测显示，在治疗3~5个月时LPMD评分显著降低（-3.7%，p<0.001），此为疗效评估的关键窗口。
结论 LPMD算法将DNMT3A突变型AML从低甲基化范式重构为甲基化紊乱范式，揭示了不同基因组区域存在异常稳定化与维持失败的双重机制。该5-DMDR预测模型可为阿扎胞苷应答预测提供实用工具，而动态LPMD监测则为治疗指导提供了潜在的生物标志物。本研究结果确立了甲基化紊乱作为髓系恶性肿瘤表观遗传失调的临床可干预维度。