文献速递消化系统肿瘤

基于肝转移基因组预测因素的胰腺导管腺癌术前风险分层深度学习模型

A deep learning model for preoperative risk stratification of pancreatic ductal adenocarcinoma based on genomic predictors of liver metastasis.
PubMed 文献 · Eur J Cancer · 2025-07-05

该回顾性研究利用MSK-MET队列(n=1779)的基因组与转移部位注释,以肝转移为二分类目标筛选8个基因组特征,构建深度学习风险模型PanScore,并在MSK-IMPACT队列(n=2181)独立验证。PanScore将患者分为低、中、高风险组,中位总生存期依次缩短,在影像学可切除病例中亦可进一步分层,高风险可切除患者生存与交界性/局部晚期疾病相似。

要点

  • PanScore基于8个基因组特征(TMB、FGA、TP53突变及AKT2、MYC、KRAS、CDKN2A、SMAD4拷贝数变异)构建。
  • MSK-MET队列中低、中、高风险组中位OS分别为21.39、15.34、9.36个月(p<0.001;高风险 vs 低风险 HR=2.07)。
  • 影像学可切除PDAC中低、中、高风险组中位OS分别为35.4、29.0、17.9个月,高风险组生存与交界性/局部晚期疾病相似。

核心结论

  • PanScore是一种基于基因组特征训练并验证的深度学习模型,可将PDAC患者尤其是可切除疾病患者分为生物学和预后不同的风险组,可能为超越传统分期的个体化治疗决策提供信息。

研究设计与样本

  • 回顾性研究,开发并验证基于深度学习的风险模型PanScore,使用两个独立队列。
  • 训练队列为MSK-MET数据集(n=1779),具有基因组和转移部位注释。
  • 以肝转移作为二分类目标指导特征选择,肝转移被确定为不良预后最强的独立预测因素。
  • 根据患病率(>2.5%)和统计学显著性(p<0.05)选择8个基因组特征:TMB、FGA、TP53突变、AKT2、MYC、KRAS、CDKN2A和SMAD4的拷贝数变异。
  • 使用H2O AutoML平台(v3.42.0.1)进行模型构建和特征评估,采用五折交叉验证、基于AUC的排序和超参数优化,最终模型架构为6层深度神经网络。

主要发现

  • 在MSK-MET队列中,PanScore将患者分为低风险(PanScore<50,n=566)、中风险(50≤PanScore<70,n=692)和高风险(PanScore≥70,n=392)组,中位总生存期分别为21.39、15.34和9.36个月(p<0.001;高风险 vs 低风险 HR=2.07)。
  • 在MSK-IMPACT队列的影像学可切除PDAC病例(n=680)中,PanScore进一步将患者分为低风险(n=312,中位OS:35.4个月)、中风险(n=306,29.0个月)和高风险(n=62,17.9个月)组。
  • PanScore高的可切除患者生存与交界性/局部晚期疾病相似,提示存在隐匿性转移扩散。

关键数值与推荐

  • PanScore风险分组阈值:低风险为PanScore<50,中风险为50≤PanScore<70,高风险为PanScore≥70。
  • MSK-MET队列低、中、高风险组中位OS分别为21.39、15.34、9.36个月(p<0.001;高风险 vs 低风险 HR=2.07)。
  • MSK-IMPACT队列影像学可切除病例低、中、高风险组中位OS分别为35.4、29.0、17.9个月。

临床怎么用

  • 原文未据此给出推荐。

局限性

  • 原文未报告局限性。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

BACKGROUND: Pancreatic ductal adenocarcinoma (PDAC) frequently presents with occult metastatic disease which can undermine the benefit of local treatment. Improved preoperative tools may enhance risk stratification. METHODS: We conducted a retrospective study to develop and validate a deep learning-based risk model, PanScore, using two independent cohorts. The training cohort (n = 1779, MSK-MET dataset) has genomic and metastatic site annotations. Liver metastasis-identified as the strongest independent predictor of poor prognosis-was used as the binary target to guide feature selection. Eight genomic features (TMB, FGA, TP53 mutation, CNVs of AKT2, MYC, KRAS, CDKN2A, and SMAD4) were selected based on prevalence (>2.5 %) and statistical significance (p < 0.05). Model construction and feature evaluation were performed using the H2O AutoML platform (v3.42.0.1) with five-fold cross-validation, AUC-based ranking, and hyperparameter optimization. The final model architecture was refined as a 6-layer deep neural network. RESULTS: PanScore classified patients in the MSK-MET cohort into low (PanScore<50, n = 566), intermediate (50 ≤PanScore<70, n = 692), and high (PanScore≥70, n = 392) risk groups with median overall survival (OS) of 21.39, 15.34, and 9.36 months, respectively (p < 0.001; HR=2.07 for high vs. low). The model was independently validated in the MSK-IMPACT cohort (n = 2181), which includes staging and survival data. Among radiographically resectable PDAC cases (n = 680), PanScore further stratified patients into low (n = 312, median OS: 35.4 months), intermediate (n = 306, 29.0 months), and high (n = 62, 17.9 months) risk groups. Resectable patients with high PanScore had survival similar to borderline/locally advanced disease, suggesting occult metastatic spread. CONCLUSION: PanScore is a validated deep learning model trained on genomic features that stratifies PDAC patients-particularly those with resectable disease- into biologically and prognostically distinct risk groups. It may inform personalized treatment decision-making beyond conventional staging.

来源:Eur J Cancer,查看原文。版权归原作者所有。

引用

  1. Zheng S, Liu Y, Troy Teo P, et al. A deep learning model for preoperative risk stratification of pancreatic ductal adenocarcinoma based on genomic predictors of liver metastasis. Eur J Cancer. 2025. PMID: 40628176. doi:10.1016/j.ejca.2025.115608

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。