Scholarlink

Background <p>Early detection of non-small cell lung cancer (NSCLC) is paramount for patient survival, but conventional diagnostics are often invasive. T-cell receptor (TCR) sequencing of peripheral blood offers a non-invasive alternative by capturing the systemic immune response, but its complex data requires advanced analytical methods.</p> Methods <p>We propose a multi-branch ensemble learning framework to diagnose NSCLC using TCR sequencing data. It synergistically integrates three analytical branches: one quantifies repertoire-level features including diversity metrics, clonality indices, and gene usage patterns; one identifies convergent TCR clusters indicative of shared antigen recognition; and one employs a Transformer-based language model to capture sequence-level patterns in CDR3 regions. All repertoire-level features were standardized using Z-score normalization, and binary classification (NSCLC vs. Healthy) was performed through a stacking ensemble classifier.</p> Results <p>The framework was validated on 150 early-stage NSCLC patients and 162 healthy controls from 7 independent sources (<i>N</i> = 312). Principal component analysis confirmed that samples cluster primarily by disease status rather than study source (Silhouette score: 0.293 for disease vs. 0.244 for study), indicating biological signal dominance over potential batch effects. To evaluate multi-center generalizability, the model was tested on two independent external cohorts: DB1 (Illumina MiSeq, <i>N</i> = 47) and DB2 (Adaptive Biotechnologies immunoSEQ, <i>N</i> = 45). The model achieved an AUC of 0.982 in DB1 and 0.941 in DB2, indicating robust performance across different clinical settings. Notably, validation on 35 independent NSCLC samples from Sun Yat-sen Memorial Hospital yielded a sensitivity of 91.4%, further supporting its potential for clinical application.</p> Conclusion <p>Our framework provides a powerful, accurate, and interpretable tool for non-invasive NSCLC detection. By capturing a holistic picture of the anti-tumor immune response through complementary analytical branches, this work offers a promising step toward liquid biopsy-based cancer screening. Further validation in larger, multi-center prospective cohorts is essential before clinical translation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A multi-branch ensemble learning framework for detection of non-small cell lung cancer via T-cell receptor sequencing

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月11日

查看全文
  • Wenjian Wang,
  • Xueting Hu,
  • Yi Luan,
  • Wenzeng Chen,
  • Qinxuan Zhu,
  • Guiping Tian,
  • Qihao Zheng,
  • Jing Meng,
  • Chuan Wang,
  • Minghui Wang

摘要

Background

Early detection of non-small cell lung cancer (NSCLC) is paramount for patient survival, but conventional diagnostics are often invasive. T-cell receptor (TCR) sequencing of peripheral blood offers a non-invasive alternative by capturing the systemic immune response, but its complex data requires advanced analytical methods.

Methods

We propose a multi-branch ensemble learning framework to diagnose NSCLC using TCR sequencing data. It synergistically integrates three analytical branches: one quantifies repertoire-level features including diversity metrics, clonality indices, and gene usage patterns; one identifies convergent TCR clusters indicative of shared antigen recognition; and one employs a Transformer-based language model to capture sequence-level patterns in CDR3 regions. All repertoire-level features were standardized using Z-score normalization, and binary classification (NSCLC vs. Healthy) was performed through a stacking ensemble classifier.

Results

The framework was validated on 150 early-stage NSCLC patients and 162 healthy controls from 7 independent sources (N = 312). Principal component analysis confirmed that samples cluster primarily by disease status rather than study source (Silhouette score: 0.293 for disease vs. 0.244 for study), indicating biological signal dominance over potential batch effects. To evaluate multi-center generalizability, the model was tested on two independent external cohorts: DB1 (Illumina MiSeq, N = 47) and DB2 (Adaptive Biotechnologies immunoSEQ, N = 45). The model achieved an AUC of 0.982 in DB1 and 0.941 in DB2, indicating robust performance across different clinical settings. Notably, validation on 35 independent NSCLC samples from Sun Yat-sen Memorial Hospital yielded a sensitivity of 91.4%, further supporting its potential for clinical application.

Conclusion

Our framework provides a powerful, accurate, and interpretable tool for non-invasive NSCLC detection. By capturing a holistic picture of the anti-tumor immune response through complementary analytical branches, this work offers a promising step toward liquid biopsy-based cancer screening. Further validation in larger, multi-center prospective cohorts is essential before clinical translation.