PO.BCS02.06 · 生物信息与计算
AI 驱动的表观基因组分析揭示化脓性汗腺炎中皮肤鳞状细胞癌风险的早期预测因子
AI-driven epigenomic profiling reveals early predictors of cutaneous squamous cell carcinoma risk in hidradenitis suppurativa
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景：化脓性汗腺炎（HS）是一种慢性炎症性皮肤病，皮肤鳞状细胞癌（cSCC）风险显著升高，包括侵袭性且常致命的变异型。尽管 cSCC 是通过紫外线损伤、慢性炎症、伤口修复受损和吸烟而发生，但其分子驱动因素仍未完全明确。新出现的证据表明，表观遗传失调，特别是 DNA 甲基化，是致癌潜能的早期指标。我们的 HS 队列在招募时不含任何 cSCC 病例，这提供了识别恶性转化之前早期表观基因组改变的独特机会。方法：使用 Illumina MethylationEPIC 芯片对 24 例 HS 患者及匹配对照的血液进行全基因组 DNA 甲基化分析。通过严格的生物信息学流程识别差异甲基化 CpG，并与 cSCC 数据集交叉参照，以界定其与致癌通路的重叠。采用人工智能方法，包括深度学习、Cox elastic-net 生存建模、随机森林以及整合的 AI/ML 流程，对与 cSCC 风险相关的 CpG 进行优先排序。结果：我们在 HS 中识别出 32 个基因上 32 个 CpG 位点的显著甲基化改变（FDR ≤ 0.05），包括 24 个低甲基化位点和 8 个高甲基化位点。所有位点此前均已在 cSCC 中被提及，并汇聚于经典致癌通路，包括 EGFR/MAPK、p53、TERT、NOTCH 信号以及 DNA 修复/染色质重塑。以 cSCC 肿瘤训练的 Cox 比例风险模型（Cox 模型）表现出强大的预后性能（交叉验证中 C-index 为 0.78-0.84）。将该模型应用于 HS 样本，生成了一个连续的 cSCC 表观遗传预后评分（cSCC-EPS），将 HS 患者分层为低、中、高风险甲基化表型。SHapley Additive exPlanations（SHAP）分析强调 EGFR、DNMT1、TP53、NOTCH3、BRAF 和 TERT 内的 CpG 是 cSCC 预测风险的最强贡献因素。这一整合的 AI-ML 表观遗传流程识别出甲基化特征汇聚于高风险 cSCC 肿瘤生物学的 HS 患者，提示存在通向早期致癌过程的基于血液的分子窗口。结论：HS 患者携带基于血液的甲基化特征，可反映 cSCC 中所见的早期表观基因组改变。这一 AI 整合框架首次提供了证据，表明这些生物标志物可以预测 HS 中的恶性转化，使甲基化分析成为一种有前景的早期 cSCC 风险评估工具。
查看英文原文 English abstract
Background: Hidradenitis Suppurativa (HS) is a chronic inflammatory skin disease with a markedly increased risk of cutaneous squamous cell carcinoma (cSCC), including aggressive and often fatal variants . Although cSCC arises through ultraviolet injury, chronic inflammation, impaired wound repair, and smoking, its molecular drivers remain incompletely defined. Emerging evidence implicates epigenetic dysregulation, particularly DNA methylation, as an early indicator of carcinogenic potential. Our HS cohort contained no cSCC cases at recruitment, providing a unique opportunity to identify early epigenomic alterations preceding malignant transformation. Methods: Genome-wide DNA methylation was profiled in blood from 24 HS cases and matched controls using the Illumina MethylationEPIC array. Differentially methylated CpGs were identified through a rigorous bioinformatic pipeline and cross-referenced with cSCC datasets to define overlap with oncogenic pathways. Artificial intelligence approaches, including deep learning, Cox elastic-net survival modeling, random forest, and integrated AI/ML pipelines, were used to prioritize CpGs associated with cSCC risk. Results: We identified significant methylation alterations (FDR ≤ 0.05) at 32 CpG sites across 32 genes in HS, comprising 24 hypomethylated and 8 hypermethylated loci. All are previously implicated in cSCC and converge on canonical oncogenic pathways, including EGFR/MAPK, p53, TERT, NOTCH signaling, and DNA repair/chromatin remodeling. The cSCC tumor-trained Cox proportional hazards model (Cox model) demonstrated strong prognostic performance (C-index 0.78-0.84 across cross-validation). Applying this model to HS samples generated a continuous cSCC Epigenetic Prognostic Score (cSCC-EPS) that stratified HS patients into low-, intermediate-, and high-risk methylation phenotypes. SHapley Additive exPlanations (SHAP) analyses highlighted CpGs within EGFR, DNMT1, TP53, NOTCH3, BRAF, and TERT as the strongest contributors to cSCC predicted risk. This integrative AI-ML epigenetic pipeline identified HS patients with methylation profiles converging on high-risk cSCC tumor biology, suggesting a blood-based molecular window into early carcinogenic processes. Conclusions: HS patients harbor blood-based methylation signatures that mirror early epigenomic alterations seen in cSCC. This AI-integrated framework provides the first evidence that these biomarkers can predict malignant transformation in HS, positioning methylation profiling as a promising tool for early cSCC risk assessment.
利益披露 Disclosure
M. R. Kuracha, None..
S. V. Kuracha, None..
A. Vedangi, None..
R. Uppala, None..
L. Uppala, None..
V. Duvvuri, None.