PO.BCS01.09 · 生物信息与计算
基于表达的免疫表型ML模型可预测肺腺癌的ICI应答和长期临床获益
Expression-based immune-phenotyping ML model predict ICI response and long-term clinic benefit in lung adenocarcinoma
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
免疫检查点抑制剂（ICI）已改变了肺癌的治疗格局，但仅有一部分患者能获得持久的临床获益，而可靠的预测性生物标志物仍然有限。为应对这一挑战，我们开发了一个基于转录组的人工智能（AI）框架，通过利用肺腺癌（LUAD）中的免疫表型（IP）来预测ICI应答和长期临床获益。使用359个TCGA-LUAD样本的转录组谱（这些样本以来自全切片图像（WSI）的IP分类进行注释），通过基于树的分类器鉴定与免疫浸润相关的AI指导基因（AIG）。随后将这些特征在14种机器学习算法中进行训练，以区分免疫浸润（IF）与非浸润（non-IF）肿瘤，之后进行集成优化。随后将所得AIG应用于接受ICI治疗的LUAD队列（N=300），以构建反映长期治疗获益的无进展生存期（PFS）预测模型。该免疫表型模型取得了强劲的预测性能，训练集AUC为0.907，独立测试集（N=90）为0.810，外部验证（N=76）为0.842。值得注意的是，基于该模型的免疫表型分型优于Lunit-SCOPE和PD-L1肿瘤比例评分（TPS）等基于图像的预测方法，在1%<TPS<50%亚组中AUC达到0.933，在TPS>50%中达到0.809，而后者分别为0.733和0.559。PFS预测模型在预测PFS与观测PFS之间显示出高相关性（R=0.94），由该模型得出的风险评分对ICI应答展现出优异的预测准确性（训练集AUC为0.964，两个外部验证分别为0.887和0.849）。单细胞RNA-seq分析进一步支持了生物学可解释性，该分析揭示模型衍生的基因富集于T细胞活化和耗竭区室，反映了与治疗应答相关的免疫活化。这一整合框架展现了对短期ICI应答和长期临床获益的双重预测能力，提供了一个具有生物学可解释性和临床可扩展性的基于转录组的平台，在精准免疫肿瘤学中具有强大的转化潜力。
查看英文原文 English abstract
Immune checkpoint inhibitors (ICIs) have transformed the treatment landscape of lung cancer, yet only a subset of patients derive durable clinical benefit, and reliable predictive biomarkers remain limited. To address this challenge, we developed a transcriptome-based artificial intelligence (AI) framework that predicts both ICI response and long-term clinical benefit by leveraging immune phenotype (IP) in lung adenocarcinoma (LUAD). Transcriptomic profiles of 359 TCGA-LUAD samples annotated with IP classes derived from whole-slide images (WSIs) were used to identify immune infiltration-associated AI-informed genes (AIGs) through tree-based classifiers. These features were subsequently trained across 14 machine learning algorithms to classify immune-infiltrated (IF) versus non-infiltrated (non-IF) tumors, followed by ensemble refinement. The resulting AIGs were then applied to ICI-treated LUAD cohorts (N=300) to construct a progression-free survival (PFS) prediction model reflecting long-term therapeutic benefit. The immune phenotyping model achieved strong predictive performance with AUCs of 0.907 in training, 0.810 in the independent test set (N=90), and 0.842 in external validation (N=76). Notably, immune phenotyping based on this model outperformed image-based prediction methods such as Lunit-SCOPE and PD-L1 tumor proportion score (TPS), achieving AUCs of 0.933 for the 1%<TPS<50% subgroup and 0.809 for TPS>50%, compared to 0.733 and 0.559, respectively. The PFS prediction model showed a high correlation between predicted and observed PFS (R = 0.94), and risk scores derived from this model demonstrated excellent predictive accuracy for ICI response (AUCs of 0.964 in training and 0.887 and 0.849 in two external validations). Biological interpretability was further supported by single-cell RNA-seq analysis, which revealed that model-derived genes were enriched in T cell activation and exhaustion compartments, reflecting immune activation linked to therapeutic response. This integrated framework demonstrates dual predictive capacity for short-term ICI response and long-term clinical benefit, offering a biologically interpretable and clinically scalable transcriptome-based platform with strong translational potential in precision immuno-oncology.
利益披露 Disclosure
K. Lee, None..
H. Park, None..
H. Lo, None..
B. Manavalan, None..
Y. Jeon, None.