DeepMeta：多中心泛癌转录组学数据的恶性肿瘤复发转移风险分层新模型

詹奇; 胡正雨; 潘绍锦; 伍亚舟; 李芳

doi:10.16462/j.cnki.zhjbkz.2025.09.010

DeepMeta：多中心泛癌转录组学数据的恶性肿瘤复发转移风险分层新模型

doi: 10.16462/j.cnki.zhjbkz.2025.09.010

詹奇^{1, 2},
胡正雨^{1, 2},
潘绍锦^{1, 2},
伍亚舟¹,
李芳^1, ,

1.
陆军军医大学军事预防医学系军队卫生统计学教研室，重庆 400038
2.
陆军军医大学基础医学院，重庆 400038

基金项目:

国家自然科学基金 82173621

详细信息

通讯作者:
李芳，E-mail: oucfli@163.com

中图分类号: R730.1;TP391
计量
- 文章访问数: 22
- HTML全文浏览量: 11
- PDF下载量: 2
- 被引次数: 0
出版历程
- 收稿日期: 2024-03-05
- 修回日期: 2025-05-06
- 网络出版日期: 2025-10-10
- 刊出日期: 2025-09-10

DeepMeta: an innovative framework for stratifying cancer recurrence and metastasis risks using multi-center pan-cancer transcriptomic data

ZHAN Qi^{1, 2},
HU Zhengyu^{1, 2},
PAN Shaojin^{1, 2},
WU Yazhou¹,
LI Fang^{1
, ,}

1.
Department of Health Statistics, Department of Military Preventive Medicine, Army Medical University, Chongqing 400038, China
2.
School of Basic Medical Sciences Army Medical University, Chongqing 400038, China

Funds:

National Natural Science Foundation of China 82173621

More Information

Corresponding author: LI Fang, E-mail: oucfli@163.com

摘要

摘要: 目的构建恶性肿瘤复发转移风险分层模型，为临床精准诊疗提供决策支持。方法基于癌症基因组图谱(The Cancer Genome Atlas, TCGA)和基因表达数据库(Gene Expression Omnibus, GEO)多中心数据构建泛癌队列，开发DeepMeta模型，通过自编码器预训练提取组学特征，结合迁移学习实现跨癌种知识迁移，利用高斯混合模型完成无监督亚型聚类，并基于极端梯度提升构建有监督分类器；采用一致性指数(concordance index, C-index)和对数秩(log rank, Log-rank)检验评估模型效能，结合沙普利加性解释(SHapley additive explanations, SHAP)算法与基因富集分析解析生物学机制。结果 DeepMeta模型在TCGA的31种恶性肿瘤中，RNA与微小RNA多组学联合分析可区分28种恶性肿瘤风险亚型，RNA单组学分析可区分26种，C-index范围分别为0.58~0.91和0.52~0.91，模型性能优于未预训练模型和传统基线模型；GEO外部验证中，肝细胞癌和乳腺癌的C-index分别为0.85和0.90，且Log-rank检验结果进一步证实DeepMeta模型的风险分层能力。鉴定出76个泛癌关键基因，形成两类功能模块，基因簇Ⅰ调控力学信号转导和机械刺激响应，基因簇Ⅱ主导胞外基质重塑和细胞黏附通路，从生物学角度验证了模型的可靠性。结论 DeepMeta模型在恶性肿瘤复发转移风险分层中展现良好效能，为恶性肿瘤关键靶向基因的研究提供有力支持。
- DeepMeta模型 /
- 恶性肿瘤复发转移 /
- 风险分层 /
- 迁移学习
Abstract: Objective To develop a risk stratification model for cancer recurrence and metastasis, providing decision support for clinical precision diagnosis and treatment. Methods A pan-cancer cohort was constructed using multi-center data from The Cancer Genome Atlas Program (TCGA) and Gene Expression Omnibus (GEO). The DeepMeta model was developed through a multi-stage computational framework: extracting latent omics features via autoencoder pre-training, enabling cross-cancer knowledge transfer through transfer learning, identifying distinct risk subtypes using unsupervised Gaussian mixture model clustering, and constructing an optimized supervised classifier with extreme gradient boosting. Model performance was evaluated using the concordance index (C-index) and Log-rank test, while biological mechanisms were elucidated through SHAP analysis and gene enrichment analysis. Results The DeepMeta model effectively stratified risk subtypes in 28/31 cancer types (C-index: 0.58-0.91) when using combined RNA+miRNA multi-omics analysis, and in 26 cancers (C-index: 0.52-0.91) with RNA single-omics analysis. The model outperformed both non-pre-trained and traditional baseline models. In external validation using GEO data, the C-index reached 0.85 for hepatocellular carcinoma and 0.90 for breast cancer, with Log-rank test results further confirming the model's risk stratification capability. The study identified 76 pan-cancer key genes that segregated into two distinct functional modules: Cluster Ⅰ (mechanical signal transduction and cellular stress response) and Cluster Ⅱ (extracellular matrix remodeling and cell adhesion pathways). This functional characterization provides biological validation of the model's reliability. Conclusions The DeepMeta model demonstrates strong performance in stratifying cancer recurrence and metastasis risk, offering a powerful tool for identifying key therapeutic targets in cancer progression.
- DeepMeta model /
- Cancer recurrence and metastasis /
- Risk stratification /
- Transfer learning

HTML全文

图 1 TCGA泛癌数据的分布情况

外环数字：代表总病例数；内环数字：代表复发转移数；UCEC：子宫内膜癌；HNSC：头颈鳞状细胞癌；KIRC：肾透明细胞癌；BLCA：膀胱尿路上皮癌；LUSC：肺鳞癌；LIHC：肝细胞癌；KIRP：肾乳头状细胞癌；CESC：宫颈鳞癌和腺癌；PCPG：嗜铬细胞瘤和副神经瘤；TGCT：睾丸生殖细胞肿瘤；DLBC：弥漫性大B细胞淋巴瘤; TCGA: 癌症基因组图谱。

Figure 1. The distribution of pan-cancer data from TCGA

Outer ring number: represents the total number of cases; Inner ring number: represents the number of recurrences and metastases; UCEC: uterine corpus endometrial carcinoma; HNSC: head and neck squamous cell carcinoma; KIRC: kidney renal clear cell carcinoma; BLCA: bladder urothelial carcinoma; LUSC: lung squamous cell carcinoma; LIHC: liver hepatocellular carcinoma; KIRP: kidney renal papillary cell carcinoma; CESC: cervical squamous cell carcinoma and endocervical adenocarcinoma; PCPG: pheochromocytoma and paraganglioma; TGCT: testicular germ cell tumor; DLBC: diffuse large B-cell lymphoma; TCGA: The Cancer Cenome Atlas.

下载: 全尺寸图片幻灯片

图 2 技术路线图

TCGA：癌症基因组图谱；miRNA：微小RNA；GMM：高斯混合模型；XGBoost：极端梯度提升；GO: 基因本体论。

Figure 2. Technology roadmap

TCGA: The Cancer Genome Atlas; miRNA: microRNA; GMM: Gaussian mixture model; XGBoost: extreme gradient boosting; GO: gene ontology.

下载: 全尺寸图片幻灯片

图 3 DeepMeta模型框架图

miRNA：微小RNA；XGBoost：极端梯度提升；GMM：高斯混合模型。

Figure 3. DeepMeta framework structure

miRNA: microRNA; XGBoost: extreme gradient boosting; GMM: Gaussian mixture model.

下载: 全尺寸图片幻灯片

图 4 模型风险分层结果

A：测试集在RNA+miRNA联合分析风险分层结果；B~C：分别为训练集、测试集风险分层结果比较；DLBC：弥漫性大B细胞淋巴瘤；HNSC：头颈鳞状细胞癌；KIRC：肾透明细胞癌；miRNA：微小RNA；C-index：一致性指数。

Figure 4. Model risk stratification results

A: Testing sets demonstrate risk stratification outcomes in integrated RNA-miRNA analysis; B-C: Comparative analysis of risk stratification between training and testing cohorts; DLBC: diffuse large B-cell lymphoma; HNSC: head and neck squamous cell carcinoma; KIRC: kidney renal clear cell carcinoma; miRNA: microRNA; C-index: concordance index.

下载: 全尺寸图片幻灯片

图 5 模型效果比较和外部验证

A: 模型风险分层效果比较；B: DeepMeta外部验证；KIRC: 肾透明细胞癌；C-index：一致性指数；GEO：基因表达数据库。

Figure 5. Model effect comparison and external validation

A: Model performance evaluation in risk stratification; B: External validation of DeepMeta; KIRC: kidney renal clear cell carcinoma; C-index: concordance index; GEO: Gene Expression Omnibus.

下载: 全尺寸图片幻灯片

图 6 关键基因分析结果

A~B：乳腺癌；C：泛癌关键基因的聚类热图；miRNA：微小RNA；SHAP：沙普利加性解释。

Figure 6. Key genetic analysis results

A-B: Critical genes underlying metastatic recurrence in breast cancer; C: Hierarchically clustered heatmap of pan-cancer critical gene signatures; miRNA: microRNA; SHAP: SHapley additive explanations.

下载: 全尺寸图片幻灯片

图 7 基因本体论富集分析结果

A：基因簇Ⅰ的基因本体论富集分析结果；B：基因簇Ⅱ的基因本体论富集分析结果。

Figure 7. Gene ontology enrichment analysis results

A: gene ontology enrichment profiles of gene cluster Ⅰ; B: gene ontology enrichment profiles of gene cluster Ⅱ.

下载: 全尺寸图片幻灯片

参考文献(19)

[1]	International Agency for Research on Cancer. Global cancer burden growing, amidst mounting need for services[EB/OL]. (2024-02-01)[2024-04-10]. https://www.who.int/news/item/01-02-2024-global-cancer-burden-growing--amidst-mounting-need-for-services.
[2]	Han BF, Zheng RS, Zeng HM, et al. Cancer incidence and mortality in China, 2022[J]. J Natl Cancer Cent, 2024, 4(1): 47-53. DOI: 10.1016/j.jncc.2024.01.006.
[3]	郑荣寿, 陈茹, 韩冰峰, 等. 2022年中国恶性肿瘤流行情况分析[J]. 中华肿瘤杂志, 2024, 46(3): 221-231. DOI: 10.3760/cma.j.cn112152-20240119-00035. Zheng RS, Chen R, Han BF, et al. Cancer incidence and mortality in China, 2022[J]. Chin J Oncol, 2024, 46(3): 221-231. DOI: 10.3760/cma.j.cn112152-20240119-00035.
[4]	Hu Z, Li Z, Ma ZC, et al. Multi-cancer analysis of clonality and the timing of systemic spread in paired primary tumors and metastases[J]. Nat Genet, 2020, 52(7): 701-708. DOI: 10.1038/s41588-020-0628-z.
[5]	Nguyen B, Fong C, Luthra A, et al. Genomic characterization of metastatic patterns from prospective clinical sequencing of 25, 000 patients[J]. Cell, 2022, 185(3): 563-575.e11. DOI: 10.1016/j.cell.2022.01.003.
[6]	林跃玮, 黄永明, 李文乐, 等. 机器学习和传统列线图预测软骨肉瘤肺转移风险的预测模型的建立与评估[J]. 中国骨与关节杂志, 2022, 11(1): 19-26. DOI: 10.3969/j.issn.2095-252X.2022.01.004. Lin YW, Huang YM, Li WL, et al. Development and evaluation of predictive models for pulmonary metastasis risk in chondrosarcoma using machine learning and traditional nomograms[J]. Chinese Journal of Bone and Joint, 2022, 11(1): 19-26. DOI: 10.3969/j.issn.2095-252X.2022.01.004.
[7]	Jiang BB, Mu QH, Qiu FF, et al. Machine learning of genomic features in organotropic metastases stratifies progression risk of primary tumors[J]. Nat Commun, 2021, 12(1): 6692. DOI: 10.1038/s41467-021-27017-w.
[8]	Xu YN, Cui XR, Wang YD. Pan-cancer metastasis prediction based on graph deep learning method[J]. Front Cell Dev Biol, 2021, 9: 675978. DOI: 10.3389/fcell.2021.675978.
[9]	National Cancer Institute, National Human Genome Research Institute. The cancer genome atlas program (TCGA)[EB/OL]. [2024-04-10]. https://tcga-data.nci.nih.gov/tcga/.
[10]	National Center for Biotechnology Information. Gene expression omnibus (GEO)[EB/OL]. [2024-04-10]. https://www.ncbi.nlm.nih.gov/geo/.
[11]	National Cancer Institute. Genomic data commons data portal[EB/OL]. [2024-04-10]. https://portal.gdc.cancer.gov/.
[12]	Umberto M. An introduction to autoencoders[EB/OL]. (2022-01-11)[2025-05-06]. https://doi.org/10.48550/arXiv.2201.03898.
[13]	D'Amico S, Dall'Olio L, Rollo C, et al. MOSAIC: an artificial intelligence-based framework for multimodal analysis, classification, and personalized prognostic assessment in rare cancers[J]. JCO Clin Cancer Inform, 2024, 8: e2400008. DOI: 10.1200/CCI.24.00008.
[14]	Wang MX, Wu YS, Li XJ, et al. IGJ suppresses breast cancer growth and metastasis by inhibiting EMT via the NF-κB signaling pathway[J]. Int J Oncol, 2023, 63(3): 105. DOI: 10.3892/ijo.2023.5553.
[15]	Wen RM, Qiu ZY, Marti GEW, et al. AZGP1 deficiency promotes angiogenesis in prostate cancer[J]. J Transl Med, 2024, 22(1): 383. DOI: 10.1186/s12967-024-05183-x.
[16]	Xu ZP, Chen SQ, Liu RJ, et al. Circular RNA circPOLR2A promotes clear cell renal cell carcinoma progression by facilitating the UBE3C-induced ubiquitination of PEBP1 and, thereby, activating the ERK signaling pathway[J]. Mol Cancer, 2022, 21(1): 146. DOI: 10.1186/s12943-022-01607-8.
[17]	Gong BC, Qu TY, Zhang JJ, et al. Downregulation of ABLIM3 confers to the metastasis of neuroblastoma via regulating the cell adhesion molecules pathway[J]. Comput Struct Biotechnol J, 2024, 23: 1547-1561. DOI: 10.1016/j.csbj.2024.04.024.
[18]	Koning T, Cordova F, Aguilar G, et al. S-Nitrosylation in endothelial cells contributes to tumor cell adhesion and extravasation during breast cancer metastasis[J]. Biol Res, 2023, 56(1): 51. DOI: 10.1186/s40659-023-00461-2.
[19]	Mao L, Chen J, Lu X, et al. Proteomic analysis of lung cancer cells reveals a critical role of BCAT1 in cancer cell metastasis[J]. Theranostics, 2021, 11(19): 9705-9720. DOI: 10.7150/thno.61731.