PO.BCS01.14 · 生物信息与计算
基于转录组的AI元模型用于肝细胞癌免疫治疗反应分类
A transcriptome-based AI meta-model for immunotherapy response classification in hepatocellular carcinoma
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
免疫检查点抑制剂（ICI）常被选作晚期肝细胞癌（HCC）的一线治疗，然而目前预测治疗反应的生物标志物准确性有限。亟需建立稳健且临床适用的HCC ICI反应预测指标。我们对接受ICI治疗的HCC患者（阿替利珠单抗联合贝伐珠单抗、纳武利尤单抗，或伊匹木单抗联合纳武利尤单抗）进行了组织RNA-seq，并开发了一个基于AI堆叠（stacking）的元模型，以分类反应者（R，n=32）和无反应者（NR，n=57）。为解决类别不平衡问题，将80%的样本分配至三个平衡的训练集，其余20%用作内部测试集。使用六种基于树的算法（AB、ERT、GB、LGB、RF和XGB）进行基因特征选择，将平均AUC从0.58提升至0.73，平均MCC从0.25提升至0.43。整合14种算法（AB、CB、ERT、GB、LGB、RF、XGB、四种SVM核、LR、NB和MLP）的堆叠元模型进一步将平均AUC提升至0.95、MCC提升至0.79。该元模型在内部测试集中达到0.92的AUC和0.62的MCC，在一个外部ICI治疗HCC队列中达到0.84的AUC和0.82的MCC（表1）。模型得出的元评分清晰地分层了临床结局。高评分患者显示出改善的PFS（HR 0.27，95% CI 0.17-0.45；p<0.0001）和OS（HR 0.37，95% CI 0.22-0.62；p<0.0001）。特征选择鉴定出11组AI重要基因（AIG）集（每组10-173个基因），每组通过共享AIG与至少两组其他基因集重叠，反映了趋同的生物学信号。在一个外部单细胞RNA-seq队列的免疫细胞群中对AIG集进行分析，揭示了与已知免疫激活和耐药机制一致的富集模式。这些发现表明，该AI元模型可帮助分类HCC的免疫治疗反应并分层生存结局，同时揭示了具有生物学相关性的AIG特征。
$$table_{D7689B83-971F-4C7E-8E56-BB92EAEC2020}$$
表1. 用于HCC免疫治疗反应分类的最终AI元模型的性能 MCC ACC AUC 敏感性 特异性 训练集 0.88 0.94 0.96 0.96 0.93 内部测试 0.62
0.76 0.92 1.00 0.64 外部验证 0.82 0.90 0.84 1.00 0.80
查看英文原文 English abstract
Immune checkpoint inhibitor (ICI) is frequently selected as first-line therapy in advanced hepatocellular carcinoma (HCC), yet current biomarkers that predict therapeutic response show limited accuracy. Robust and clinically applicable predictors of ICI response in HCC need to be established. We performed tissue RNA-seq on ICI-treated HCC patients (Atezolizumab plus Bevacizumab, Nivolumab, or Ipilimumab plus Nivolumab) and developed an AI stacking-based meta-model to classify responders (R, n=32) and non-responders (NR, n=57). To address class imbalance, 80% of samples were distributed into three balanced training sets, and the remaining 20% were used as an internal test set. Gene-feature selection performed with six tree-based algorithms (AB, ERT, GB, LGB, RF, and XGB) improved mean AUC from 0.58 to 0.73 and mean MCC from 0.25 to 0.43. The stacking meta-model, incorporating 14 algorithms (AB, CB, ERT, GB, LGB, RF, XGB, four SVM kernels, LR, NB, and MLP), further increased mean AUC to 0.95 and MCC to 0.79. The meta-model achieved an AUC of 0.92 and MCC of 0.62 in the internal test set, and an AUC of 0.84 and MCC of 0.82 in an external ICI-treated HCC cohort (Table 1). Meta-scores derived from the model clearly stratified clinical outcomes. Patients with high scores showed improved PFS (HR 0.27, 95% CI 0.17-0.45; p<0.0001) and OS (HR 0.37, 95% CI 0.22-0.62; p<0.0001). Feature selection identified 11 AI-important gene (AIG) sets (10-173 genes each), and each set overlapped with at least two others through shared AIGs, reflecting convergent biological signals. Analysis of AIG sets in immune-cell populations from an external single-cell RNA-seq cohort revealed enrichment patterns consistent with known mechanisms of immune activation and resistance. These findings demonstrate that the AI meta-model can help to classify immunotherapy response in HCC and stratify survival outcomes, while revealing biologically relevant AIG signatures.
$$table_{D7689B83-971F-4C7E-8E56-BB92EAEC2020}$$
Table 1. Performance of the final AI meta-model for immunotherapy response classification in HCC MCC ACC AUC Sensitivity Specificity Training 0.88 0.94 0.96 0.96 0.93 Internal Test 0.62
0.76 0.92 1.00 0.64 External Validation 0.82 0.90 0.84 1.00 0.80
利益披露 Disclosure
J. Seo, None..
N. Kwon, None..
K. Lee, None..
H. Lo, None..
Y. Jeon, None.