Integrating ctDNA Analysis and Radiomics for Dynamic Risk Assessment in Localized Lung Cancer
(1) Check for updates
整合 ctDNA 分析和放射组学，对局部肺癌进行动态风险评估 (1) 检查更新

Everett J. Moding

^{1, 2}

, Mohammad Shahrokh Esfahani

^{2, 3}

, Cheng Jin

^{1}

, Angela B. Hui

^{1, 2}

, Barzin Y. Nabet

^{2}

, Yufei Liu

^{4}

, Jacob J. Chabon

^{2}

, Michael S. Binkley

^{1, 2}

, David M. Kurtz

^{3}

, Emily G. Hamilton

^{2}

, Aadel A. Chaudhuri

^{5}

, Chih Long Liu

^{3}

, Zhe Li

^{1}

, Rene F. Bonilla

^{1}

, Alice L. Jiang

^{1}

, Brianna C. Lau

^{1}

, Pablo Lopez

^{4}

, Jianzhong He

^{4}

, Yawei Qiao

^{4}

, Ting Xu

^{4}

, Luyang Yao

^{4}

, Saumil Gandhi

^{4}

, Zhongxing Liao

^{4}

, Millie Das

^{3.6}

, Kavitha J. Ramchandran

^{3}

, Sukhmani K. Padda

^{3}

, Joel W. Neal

^{3}

, Heather A. Wakelee

^{3}

, Michael F. Gensheimer

^{1}

, Billy W. Loo Jr

^{1}

, Ruijiang Li

^{1, 2}

, Steven H. Lin

^{4}

, Ash A. Alizadeh

^{2, 3, 7}

, and Maximilian Diehn

^{1, 2, 7}

Everett J. Moding

^{1, 2}

, Mohammad Shahrokh Esfahani

^{2, 3}

, Cheng Jin

^{1}

, Angela B. Hui

^{1, 2}

, Barzin Y. Nabet

^{2}

, Yufei Liu

^{4}

, Jacob J. Chabon

^{2}

, Michael S. Binkley

^{1, 2}

, David M. Kurtz

^{3}

, Emily G. Hamilton

^{2}

, Aadel A. Chaudhuri

^{2}

, Aadel A. Chaudhuri.Binkley

^{1, 2}

, David M. Kurtz

^{3}

, Emily G. Hamilton

^{2}

, Aadel A. Chaudhuri

^{5}

, Chih Long Liu

^{3}

, Zhe Li

^{1}

, Rene F. Bonilla

^{1}

, Alice L. Jiang

^{1}

, Brianna C. Lau

^{1}

, Pablo Lopez

^{4}

, Jianzhong He

^{4}

, Yawei Qiao

^{4}

, Ting Xu

^{4}

, Luyang Yao

^{4}

, Saumil Gandhi

^{4}

, Zhongxing Liao

^{4}

, Millie Das

^{3.6}

, Kavitha J. Ramchandran

^{3}

, Sukhmani K. Padda

^{3}

.Padda

^{3}

, Joel W. Neal

^{3}

, Heather A. Wakelee

^{3}

, Michael F. Gensheimer

^{1}

, Billy W. Loo Jr

^{1}

, Ruijiang Li

^{1, 2}

, Steven H. Lin

^{4}

, Ash A. Alizadeh

^{2, 3, 7}

, and Maximilian Diehn

^{1, 2, 7}

Abstract 摘要

The complementary and clinical utility of combining liquid biopsy with radiographic analysis has not been demonstrated。 ctDNA minimal residual disease after chemoradiotherapy (CRT) for non-small cell lung cancer (NSCLC) is highly prognostic, but on-treatment biomarkers are needed to enable response-adapted therapies. In this study, we analyzed 418 patients with NSCLC undergoing CRT to develop and validate a novel dynamic risk model that accurately predicts ultimate progression-free survival during treatment. We optimize tissue-free variant calling from plasma samples to facilitate ctDNA monitoring and demonstrate the importance of accounting for persistent clonal hematopoiesis variants. We show that mid-CRT ctDNA concentration is prognostic for disease progression and integrate additional pre-CRT risk factors, including radiomics, into a combined model that improves outcome prediction. Our results suggest that tumor features, radiomics, and mid-CRT ctDNA analysis are complementary and can identify patients at high and low risk of progression to potentially enable response-adapted therapies.
非小细胞肺癌（NSCLC）化放疗（CRT）后的 ctDNA 最小残留病变具有高度预后性，但需要治疗中的生物标志物来实现反应适应性疗法。在这项研究中，我们分析了 418 名接受 CRT 治疗的 NSCLC 患者，开发并验证了一种新型动态风险模型，该模型能准确预测治疗期间的最终无进展生存期。我们优化了血浆样本的无组织变异调用，以促进 ctDNA 监测，并证明了考虑持续性克隆造血变异的重要性。我们表明，CRT 中期的 ctDNA 浓度可预示疾病的进展，并将包括放射组学在内的其他 CRT 前风险因素整合到一个综合模型中，从而改善了结果预测。我们的研究结果表明，肿瘤特征、放射组学和 CRT 中期 ctDNA 分析是相辅相成的，可以识别进展风险高和风险低的患者，从而有可能采用适应反应的疗法。

SIGNIFICANCE: This study demonstrates that combining tumor features, radiomics, and ctDNA analysis improves outcome prediction in NSCLC treated with CRT therapy. Our integrated model could enable personalized and response-adapted therapies to reduce toxicity and improve outcomes in patients.
意义：这项研究表明，将肿瘤特征、放射组学和 ctDNA 分析结合起来，可以改善接受 CRT 治疗的 NSCLC 患者的预后预测。我们的综合模型可实现个性化和适应反应的疗法，从而降低毒性并改善患者的预后。

INTRODUCTION 引言

Lung cancer is the leading cause of cancer deaths, with approximately 230,000 new cases of lung cancer diagnosed annually in the United States leading to approximately 140,000 deaths (1). Chemoradiation therapy (CRT) plays an important role in the definitive management of unresectable locally advanced non-small cell lung cancer (NSCLC; ref. 2). Unfortunately, the majority of patients with stage III NSCLC develop progressive disease (PD) after receiving the current standard of care (3). In addition, the definitive radiation dose
肺癌是导致癌症死亡的首要原因，美国每年新确诊肺癌病例约 23 万例，导致约 14 万人死亡（1）。化学放疗（CRT）在无法切除的局部晚期非小细胞肺癌（NSCLC；参考文献 2）的最终治疗中发挥着重要作用。遗憾的是，大多数 III 期 NSCLC 患者在接受目前的标准治疗后，病情会出现进展（PD）（3）。此外，明确的放射剂量

has been unchanged for 40 years despite evidence that lower doses may be curative in a subset of patients with NSCLC (4). Although personalized medicine has revolutionized the treatment of metastatic NSCLC, all patients with unresectable locally advanced NSCLC receive the same treatment because there are currently no reliable methods for monitoring response during CRT or identifying patients who would potentially benefit from de-escalation or intensification of therapy.
尽管有证据表明，较低剂量的 CRT 可治愈部分 NSCLC 患者（4），但 40 年来 CRT 的剂量一直未变。尽管个性化医疗已经彻底改变了转移性 NSCLC 的治疗方法，但所有无法切除的局部晚期 NSCLC 患者都接受同样的治疗，因为目前还没有可靠的方法来监测 CRT 期间的反应，或识别可能从降级或强化治疗中获益的患者。

Several prior studies have attempted to identify prognostic factors in NSCLC, but the majority of these studies have focused on overall survival (OS) across diverse groups of patients, and few factors have been validated in independent data sets (5). As a result, tumor-node-metastasis (TNM) stage is the only prognostic factor currently considered in the treatment of locally advanced NSCLC, even though considerable variability in outcomes exists between uniformly treated patients (6). For locoregionally advanced NSCLC, a number of patient and tumor characteristics and more recently cancer genetic biomarkers have been proposed as prognostic factors (7). Several groups have generated models incorporating multiple pretreatment prognostic factors (8-14), but these are not currently incorporated into treatment decisions as part of the standard of care.
之前有几项研究试图确定 NSCLC 的预后因素，但这些研究大多侧重于不同患者群体的总生存期（OS），很少有因素在独立数据集中得到验证（5）。因此，肿瘤-结节-转移（TNM）分期是目前治疗局部晚期 NSCLC 唯一考虑的预后因素，尽管统一治疗的患者之间的预后存在很大差异（6）。对于局部晚期 NSCLC，许多患者和肿瘤特征以及最近提出的癌症基因生物标记物被认为是预后因素（7）。一些研究小组已经建立了包含多种治疗前预后因素的模型（8-14），但目前还没有将这些因素纳入治疗决策，作为治疗标准的一部分。

Due to their importance in cancer diagnosis, staging, and treatment planning, clinical images, including CT, are routinely acquired prior to treatment for patients with NSCLC. Advances in machine learning have enabled radiomic analysis of clinical images through the large-scale extraction and analysis of quantitative image features (15). Radiomic analysis can capture biological features in NSCLC that can serve as biomarkers to aid in diagnosis, prognostication, treatment selection, and treatment response monitoring

(16, 17)

. For example, radiomic signatures have been reported to be associated with underlying gene expression patterns in NSCLC and to be associated with patient prognosis (18). In locoregionally
由于临床图像（包括 CT）在癌症诊断、分期和治疗计划中的重要性，NSCLC 患者在治疗前都会常规采集临床图像。机器学习的进步通过大规模提取和分析定量图像特征，实现了临床图像的放射学分析（15）。放射组学分析可以捕捉 NSCLC 中的生物特征，这些特征可作为生物标记物帮助诊断、预后判断、治疗选择和治疗反应监测

(16, 17)

。例如，有报道称放射学特征与 NSCLC 潜在的基因表达模式相关，并与患者的预后有关（18）。在局部区域
advanced NSCLC, tumor shape and texture measured by radiomics have been reported to be associated with pathologic response to neoadjuvant chemoradiation (19), and combining radiomic features and clinical data improved prediction of the presence of gross residual disease (20). Furthermore, radiomic analysis may be complementary to clinical and genomic predictors, and integrating these methodologies may improve prognostication (21-25).
据报道，放射组学测量的肿瘤形状和质地与新辅助化疗的病理反应有关（19），将放射组学特征与临床数据相结合可提高对是否存在大体残留疾病的预测（20）。此外，放射组学分析可能是临床和基因组预测指标的补充，整合这些方法可能会改善预后（21-25）。

Incorporating treatment response could potentially improve risk stratification of patients with NSCLC and enable adaptive treatment strategies to improve rates of cure and decrease the risk of toxicity. Midtreatment imaging with CT and/or PET has been associated with patient outcomes (26-28). For example, a higher PET metabolic tumor volume (MTV) and total lesion glycolysis were associated with a higher risk of local recurrence after chemoradiation in one study (26), and another study reported that a higher PET maximum standardized uptake value during chemoradiation was associated with increased risk of death or tumor progression on multivariate analysis (28). However, these approaches are limited by normal tissue changes caused by CRT that can decrease the specificity of on- or post-treatment PET/CT scans (29). As a result, there are currently no reliable methods to monitor tumor response during treatment.
纳入治疗反应有可能改善 NSCLC 患者的风险分层，使适应性治疗策略能够提高治愈率并降低毒性风险。使用 CT 和/或 PET 进行治疗中期成像与患者的预后有关（26-28）。例如，在一项研究中，较高的 PET 代谢肿瘤体积（MTV）和病灶糖酵解总量与化疗后局部复发风险较高有关（26），另一项研究报告称，化疗期间较高的 PET 最大标准化摄取值与多变量分析中死亡或肿瘤进展风险增加有关（28）。然而，这些方法受到 CRT 引起的正常组织变化的限制，这些变化会降低治疗中或治疗后 PET/CT 扫描的特异性（29）。因此，目前还没有可靠的方法来监测治疗期间的肿瘤反应。

Tumors continuously shed DNA into the peripheral blood that can be noninvasively collected from liquid biopsies and quantified as ctDNA (ref. 30). On-treatment ctDNA levels and ctDNA kinetics have been shown to correlate with disease burden and treatment response in patients undergoing systemic therapy for metastatic NSCLC (31-37) and other solid tumors (38-44). For example, multiple studies have demonstrated that patients with NSCLC who benefit from immune checkpoint inhibitors or EGFR inhibitors have a decrease in their ctDNA levels shortly after starting treatment (31-35, 37, 44). Furthermore, we have previously demonstrated that ctDNA analysis in NSCLC can robustly detect molecular residual disease (MRD) within 4 months of completing CRT (45) and predict benefit from consolidation immunotherapy (46). However, ctDNA kinetics during definitive CRT and whether midtreatment ctDNA levels can predict patient outcomes have not been characterized in detail. Although the majority of prior ctDNA studies have relied on sequencing of matched tumor tissue for the identification of tumor-derived mutations, tumor tissue is frequently not available or inadequate for tumor genotyping (47), and identification of variants from plasma samples is complicated by technical artifacts and nonmalignant biological signals such as clonal hematopoiesis (CH; refs. 48-50).
肿瘤会不断向外周血中脱落 DNA，这些 DNA 可通过液体活检无创收集并量化为 ctDNA（参考文献 30）。在接受转移性 NSCLC（31-37）和其他实体瘤（38-44）全身治疗的患者中，治疗中的 ctDNA 水平和 ctDNA 动力学已被证明与疾病负担和治疗反应相关。例如，多项研究表明，从免疫检查点抑制剂或表皮生长因子受体抑制剂中获益的 NSCLC 患者在开始治疗后不久，其 ctDNA 水平就会下降（31-35、37、44）。此外，我们之前已经证实，NSCLC 中的 ctDNA 分析可以在完成 CRT 治疗后 4 个月内检测出分子残留病（MRD）（45），并预测巩固免疫疗法的获益情况（46）。然而，ctDNA 在明确 CRT 期间的动力学以及治疗中期的 ctDNA 水平是否能预测患者的预后还没有详细的描述。尽管之前的大多数 ctDNA 研究都依赖于对匹配的肿瘤组织进行测序来鉴定肿瘤衍生突变，但肿瘤组织往往无法获得或不足以进行肿瘤基因分型（47），而且血浆样本中变异的鉴定也因技术伪影和克隆造血等非恶性生物信号（CH；参考文献 48-50）而变得复杂。

We hypothesized that clinical risk factors, radiomics, and ctDNA analysis provide complementary information that can improve risk-stratification when combined. In this study, we optimize tissue-free identification of tumor-derived variants from pretreatment plasma samples to increase the clinical application of ctDNA analysis for treatment response assessment. We demonstrate ctDNA analysis during CRT for locoregionally advanced NSCLC is an early predictor of patient outcomes that could potentially enable response-adapted therapies. Furthermore, we present a novel dynamic risk model that integrates pretreatment prognostic factors, including radiomic analysis, with midtreatment
我们假设，临床风险因素、放射组学和 ctDNA 分析可提供互补信息，三者结合可改善风险分级。在本研究中，我们从治疗前血浆样本中优化了肿瘤衍生变异的无组织鉴定，以提高 ctDNA 分析在治疗反应评估中的临床应用。我们证明了在局部晚期 NSCLC CRT 治疗期间进行的 ctDNA 分析是患者预后的早期预测指标，有可能实现适应反应的疗法。此外，我们还提出了一种新型动态风险模型，该模型将包括放射学分析在内的治疗前预后因素与治疗中
ctDNA levels to improve prediction of progression-free survival (PFS) and enable real-time updating of individualized risk estimates.
ctDNA 水平，以改善无进展生存期（PFS）的预测，并实现个性化风险估计的实时更新。

RESULTS 结果

CH Dynamics during CRT
CRT 期间的 CH 动态

We set out to develop a composite biomarker based on ctDNA analysis that could enable risk stratification and adaptive treatment strategies during CRT for NSCLC (Fig. 1A). We therefore performed ctDNA analysis using cancer personalized profiling by deep sequencing (CAPP-seq) on a total of 101 patients with NSCLC with plasma samples collected 10 to 30 days into CRT (mid-CRT, Supplementary Fig. S1A and S1B; Supplementary Tables S1 and S2). Matched tumor tissue from pre-CRT biopsies was available for only 15 patients, so “tissue-free” identification of variants from pre-CRT plasma was necessary for the majority of patients. Previous studies have demonstrated that the majority of plasma cell-free DNA (cfDNA) variants in controls and patients with cancer are derived from CH (48, 50), an aging-related process in which nonmalignant hematopoietic cells acquire somatic alterations leading to clonal expansion. Per the World Health Organization’s classification of hematolymphoid tumors, CH of indeterminate significance (CHIP) is defined as the presence of mutations associated with hematologic malignances in the peripheral blood at a variant allele fraction (VAF) greater than

2 %

in patients without cytopenias or dysplastic hematopoiesis (51). However, CH variants also exist at lower VAFs in patients without CHIP (48, 50). Applying strict filters to remove variants present in matched leukocytes could reduce the sensitivity of tissue-free variant calling by removing some variants that are truly tumor-derived, and it is unclear whether CH variants persist during CRT. Therefore, we sought to determine the dynamics of CH variants in patients with NSCLC treated with CRT.
我们希望开发一种基于 ctDNA 分析的复合生物标记物，以便在 NSCLC 的 CRT 期间进行风险分层并制定适应性治疗策略（图 1A）。因此，我们利用深度测序癌症个体化图谱（CAPP-seq）对总共 101 名 NSCLC 患者进行了 ctDNA 分析，并在 CRT 开始后 10 到 30 天（CRT 中期，补充图 S1A 和 S1B；补充表 S1 和 S2）采集了血浆样本。只有 15 例患者能从 CRT 前活检中获得匹配的肿瘤组织，因此大多数患者需要从 CRT 前血浆中进行 "无组织 "变异鉴定。先前的研究表明，对照组和癌症患者血浆中的无细胞 DNA（cfDNA）变异大多来自 CH（48，50），这是一个与衰老相关的过程，在这个过程中，非恶性造血细胞发生体细胞改变，导致克隆扩增。根据世界卫生组织的血淋巴肿瘤分类，意义不确定的 CH（CHIP）被定义为外周血中存在与血液恶性肿瘤相关的变异等位基因分数（VAF）大于

2 %

的变异，且患者无细胞减少症或造血发育不良（51）。然而，在无 CHIP 的患者中，CH 变异体的 VAF 也较低（48，50）。应用严格的过滤器去除匹配白细胞中存在的变异可能会通过去除一些真正来源于肿瘤的变异而降低无组织变异调用的灵敏度，而且目前还不清楚 CH 变异是否会在 CRT 期间持续存在。因此，我们试图确定接受 CRT 治疗的 NSCLC 患者中 CH 变异的动态变化。

We began by characterizing the prevalence of CH variants in patients with locoregionally advanced NSCLC by identifying cfDNA single-nucleotide variants (SNV) also present in matched leukocytes from the same patient. Consistent with prior studies, we observed a high prevalence of CHIP and CH variants in both canonical and noncanonical CH genes in patients with locoregionally advanced NSCLC (Supplementary Fig. S2A-S2C). We next characterized the dynamics of CH variants during CRT. Mutations with allele fractions greater than

2 %

at baseline persisted in all patients mid-CRT, including a patient with a PPM1D nonsense mutation whose VAF increased from

2.3 %

pre-CRT to

6.6 %

mid-CRT (Supplementary Fig. S2D). Considering CH mutations at any allele fraction,

90 %

of variants remained detected at the mid-CRT time point (Fig. 1B and C). Because prior studies have suggested that certain mutations are more prevalent after chemotherapy and/ or radiation therapy in CH and therapy-related myeloid neoplasms (48, 52, 53), we investigated changes in VAF mid-CRT for frequently mutated canonical CH genes. Compared with DNMT3A mutations which remained stable mid-CRT, PPM1D mutations significantly increased whereas SF3B1 and TET2 mutations decreased (Fig. 1D). In patients with later samples available for analysis, CH variants persisted up to 11 months
我们首先通过鉴定同一患者匹配白细胞中也存在的 cfDNA 单核苷酸变体 (SNV)，确定了局部晚期 NSCLC 患者中 CH 变异的流行特征。与之前的研究一致，我们在局部区域晚期 NSCLC 患者中观察到了较高的 CHIP 以及规范和非规范 CH 基因中的 CH 变异（补充图 S2A-S2C）。接下来，我们研究了 CRT 期间 CH 变异的动态特征。基线等位基因分数大于

2 %

的突变在所有患者的 CRT 中期都持续存在，包括一名 PPM1D 无义突变患者，其 VAF 从 CRT 前的

2.3 %

增加到 CRT 中期的

6.6 %

（补充图 S2D）。考虑到任何等位基因分数的 CH 变异，

90 %

的变异在 CRT 中期时间点仍能被检测到（图 1B 和 C）。由于先前的研究表明，某些突变在 CH 和治疗相关的髓系肿瘤化疗和/或放疗后更为普遍（48、52、53），因此我们研究了经常突变的典型 CH 基因在 CRT 中期的 VAF 变化。与 DNMT3A 基因突变在 CRT 中期保持稳定相比，PPM1D 基因突变明显增加，而 SF3B1 和 TET2 基因突变则有所减少（图 1D）。在有较晚样本可供分析的患者中，CH 变异持续时间长达 11 个月

Figure 1. Filtering CH and applying the machine learning-based SNV score improves tissue-free variant calling.

A

, Schematic of the overall study design. Tissue-free ctDNA analysis was optimized and integrated with radiomics, biological features, and molecular features using Bayesian Cox proportional hazard modeling to build a dynamic risk prediction model for PFS in patients with NSCLC treated with CRT. B, Log

_{10}

fold change in CH VAF mid-CRT (

n = 128

total variants in 50 patients). C, Percent of CH variants detected mid-CRT (

n = 128

variants). D, Log

_{10}

fold change in CH VAF mid-CRT in canonical genes with at least three variants identified pre-CRT. P values were calculated using two-sided Mann-Whitney tests. E, Comparison of the SNV score for variants not identified in the tumor (not tumor-adjudicated,

n = 216

variants) and SNVs present in the tumor (tumor-adjudicated,

n = 203

variants) of patients with matched tumor tissue available for analysis. Medians are shown with solid lines, and quartiles are shown with dashed lines.

P

value was calculated using a two-sided Mann-Whitney test. F, Number of tumor-adjudicated variants called per patient from cfDNA and matched leukocyte sequencing using previously defined empiric filters or a SNV score threshold with equivalent positive predictive value for a variant being tumor-adjudicated. Only patients with at least one variant called by either method are plotted (

n = 49

patients).

P

value was calculated using a two-sided Wilcoxon matched-pair signed-rank test. G, Change in number of total variants or tumor-adjudicated variants called with the SNV score compared with empiric filters. Only patients with at least one variant called by either method are included (

n = 49

patients). H, Plot of sensitivity defined as the fraction of tumor-adjudicated variants called and specificity defined as the fraction of age- and risk-matched controls with no variants called vs. SNV score threshold used for filtering (

n = 203

tumor-adjudicated variants, 56 control patients). The dotted line denotes the SNV score threshold of 0.3 used in the mid-treatment ctDNA analysis. (A, Created with BioRender.com.)
图 1.过滤 CH 和应用基于机器学习的 SNV 评分可改善无组织变异的调用。

A

，总体研究设计示意图。使用贝叶斯考克斯比例危险模型对无组织 ctDNA 分析进行优化并与放射组学、生物学特征和分子特征整合，以建立 CRT 治疗的 NSCLC 患者 PFS 的动态风险预测模型。B、CRT 中期 CH VAF 的对数

_{10}

折叠变化（50 例患者中的

n = 128

总变异）。C、CRT 中期检测到的 CH 变异百分比（

n = 128

变异）。D、CRT 中期至少发现三个变异的典型基因中 CH VAF 的对数

_{10}

倍变化。P 值使用双侧 Mann-Whitney 检验计算。E、有匹配肿瘤组织可供分析的患者的肿瘤中未发现的变异（未经肿瘤鉴定，

n = 216

变异）与肿瘤中存在的 SNVs（经肿瘤鉴定，

n = 203

变异）的 SNV 评分比较。实线表示中位数，虚线表示四分位数。

P

值采用双侧曼-惠特尼检验计算。F，使用先前定义的经验过滤器或具有同等阳性预测值的 SNV 评分阈值，从 cfDNA 和匹配的白细胞测序中调用每位患者的肿瘤判定变异的数量。只有至少有一个变体被两种方法调用的患者（

n = 49

患者）才会被绘制出来。

P

值采用双侧 Wilcoxon 配对符号秩检验计算。G、与经验筛选法相比，用 SNV 评分调用的总变异数或肿瘤判断变异数的变化。只有通过任一方法至少调用了一个变体的患者才被纳入（

n = 49

患者）。H，灵敏度与用于筛选的 SNV 评分阈值的对比图（

n = 203

肿瘤判断变异，56 例对照患者），灵敏度定义为调用的肿瘤判断变异的比例，特异性定义为未调用变异的年龄和风险匹配对照的比例。虚线表示治疗中期 ctDNA 分析中使用的 SNV 评分阈值 0.3。(A，用 BioRender.com 创建）。
after starting CRT (Supplementary Fig. S2E). Because the majority of CH variants remain detected during CRT, and CRT has a differential impact on CH depending on the mutation, these results demonstrate that CH confounds ctDNA monitoring during treatment for NSCLC. We therefore performed deep sequencing of matched pretreatment leukocytes and adopted a stringent filtering strategy to remove CH variants of any VAF from our ctDNA monitoring approach.
补充图 S2E）。由于大多数 CH 变异在 CRT 期间仍能被检测到，而且 CRT 对 CH 的影响因突变而异，这些结果表明 CH 会干扰 NSCLC 治疗期间的 ctDNA 监测。因此，我们对匹配的治疗前白细胞进行了深度测序，并采取了严格的过滤策略，以从我们的 ctDNA 监测方法中剔除任何 VAF 的 CH 变异。

Tissue-Free Variant Calling with Machine Learning
利用机器学习进行无组织变异调用

We and others have previously utilized empirically defined filters for tissue-free genotyping of cfDNA to identify tumor-derived variants (49, 54, 55). Although this approach can be utilized to achieve high specificity when comparing cfDNA from patients and controls, applying it may not optimally maximize both sensitivity and specificity. We recently developed a multitiered machine learning approach to integrate
我们和其他人曾利用经验定义的过滤器对 cfDNA 进行无组织基因分型，以鉴定肿瘤衍生变异（49、54、55）。虽然这种方法在比较患者和对照组的 cfDNA 时可以达到很高的特异性，但应用这种方法可能无法最大限度地提高灵敏度和特异性。我们最近开发了一种多层次的机器学习方法，以整合
cfDNA genomic features for noninvasive lung cancer detection (50), and we hypothesized that a similar approach could improve tissue-free tumor SNV calling for ctDNA monitoring during CRT.
我们假设，类似的方法也能改善无组织肿瘤 SNV 呼唤，用于 CRT 期间的 ctDNA 监测。

We analyzed targeted deep sequencing of plasma cfDNA and matched leukocytes from 190 patients with lung cancer and 92 controls along with DNA extracted from matched tumor tissue for 119 of the patients with lung cancer (Supplementary Table S3). We extracted 25 key biological and technical features for each SNV identified in the cfDNA, such as fragment size, allele fraction, background frequencies, and mapping quality. We used machine learning using semisupervised elastic net logistic regression to train a model that integrates these features to estimate the probability of an individual SNV being tumor-derived (SNV score, Supplementary Fig. S3A-S3C).
我们对 190 名肺癌患者和 92 名对照者的血浆 cfDNA 和匹配的白细胞以及 119 名肺癌患者从匹配的肿瘤组织中提取的 DNA 进行了靶向深度测序分析（补充表 S3）。我们提取了在 cfDNA 中发现的每个 SNV 的 25 个关键生物学和技术特征，如片段大小、等位基因比例、背景频率和映射质量。我们使用半监督弹性网逻辑回归的机器学习方法训练了一个模型，该模型整合了这些特征，以估计单个 SNV 来源于肿瘤的概率（SNV 评分，补充图 S3A-S3C）。

We benchmarked the SNV score using leave-one-out cross-validation in patients with matched tumor and cfDNA sequencing available in order to adjudicate SNVs as tumorderived. The SNV score was significantly higher for tumoradjudicated variants than non-tumor-adjudicated variants (median 0.91 vs. 0.09 , Fig. 1E) and was higher with increasing allele fraction (Supplementary Fig. S3D). We next compared the SNV score with empiric filters previously defined to achieve

95 %

specificity in control cfDNA samples (46). Using a SNV score threshold that achieved equivalent positive predictive value for identifying tumor-adjudicated SNVs, the SNV score significantly increased the number of total variants and tumor-adjudicated variants called per patient compared with empiric filters (Fig. 1F and G; Supplementary Fig. S3E). In addition, because the SNV score estimates the probability of a variant being tumor-derived, the filtering threshold can be tuned for the clinical application. Higher SNV scores increased specificity at the cost of worsening sensitivity and number of patients with at least one SNV available for monitoring analysis (Fig. 1H; Supplementary Fig. S3F). Balancing these factors, we established a SNV score threshold of 0.3 for our monitoring analysis. At this threshold, concordance between tissue-free variant calling and tumor sequencing is

72.1 %

. These results demonstrate that integrating SNV features with machine learning can improve tissue-free variant calling.
我们在有匹配肿瘤和 cfDNA 测序结果的患者中使用留空交叉验证法对 SNV 评分进行了基准测试，以判定 SNV 是否来自肿瘤。肿瘤判定变异的 SNV 得分明显高于非肿瘤判定变异（中位数 0.91 vs. 0.09，图 1E），并且随着等位基因比例的增加而增加（补充图 S3D）。接下来，我们将 SNV 得分与先前定义的经验过滤器进行了比较，以实现对照 cfDNA 样本中

95 %

的特异性（46）。与经验过滤器相比，SNV 评分能显著增加每位患者被调用的总变异数和肿瘤判定变异数（图 1F 和 G；补充图 S3E）。此外，由于 SNV 得分估计的是变异源于肿瘤的概率，因此可以根据临床应用调整筛选阈值。SNV 得分越高，特异性越高，但灵敏度和至少有一个 SNV 可用于监测分析的患者数量却在下降（图 1H；补充图 S3F）。在平衡这些因素后，我们将监测分析的 SNV 评分阈值定为 0.3。在此阈值下，无组织变异调用与肿瘤测序的一致性为

72.1 %

。这些结果表明，将 SNV 特征与机器学习相结合可以改善无组织变异的调用。

ctDNA Kinetics during CRT
CRT 期间的 ctDNA 动力学

Having established an improved methodology for tissuefree variant calling, we applied the SNV score for pretreatment identification of variants in patients without matched tumor tissue available for sequencing. We then monitored for the variants identified pretreatment in the mid-CRT plasma sample. We identified a training cohort of 40 patients treated at MD Anderson Cancer Center (MDACC) and a validation cohort of 21 patients treated at Stanford University for stage IIB to IIIB NSCLC with variants identified before treatment (Fig. 2A and B; Supplementary Tables S4 and S5). Across both cohorts, we tracked a median of five SNVs (range, 1-26). The genes most frequently mutated by SNVs were TP53 (44%), KRAS (10%), KEAP1 (10%), PIK3CA (8%), and EGFR (5%). The median pretreatment ctDNA allele fraction was

0.77 %

(range,

0.08 % - 15.20 %

) in the training cohort and

0.97 %

(range,

0.02 % - 16.86 %

) in the validation cohort.
在确立了无组织变异调用的改进方法后，我们将 SNV 评分用于对没有匹配肿瘤组织可供测序的患者进行治疗前变异鉴定。然后，我们在 CT 中期血浆样本中监测治疗前识别出的变异。我们确定了一个由在 MD 安德森癌症中心（MDACC）接受治疗的 40 例患者组成的训练队列和一个由在斯坦福大学接受治疗的 21 例 IIB 至 IIIB 期 NSCLC 患者组成的验证队列，这两组患者在治疗前均发现了变异（图 2A 和 B；补充表 S4 和 S5）。在这两个队列中，我们追踪到的 SNV 中位数为 5 个（范围为 1-26）。最常发生 SNV 突变的基因是 TP53（44%）、KRAS（10%）、KEAP1（10%）、PIK3CA（8%）和表皮生长因子受体（5%）。在训练队列中，治疗前 ctDNA 等位基因比例中位数为

0.77 %

（范围为

0.08 % - 15.20 %

）；在验证队列中，治疗前 ctDNA 等位基因比例中位数为

0.97 %

（范围为

0.02 % - 16.86 %

）。

We first characterized ctDNA kinetics at the mid-CRT time point in the training cohort. Across all patients in the training cohort, the ctDNA concentration decreased by a median fold of 13.3, from a median of 31.8 haploid genome equivalents per milliliter (

hGE / mL

) pre-CRT to

0.92 hGE / mL

mid-CRT (Fig. 2C). This decrease was driven by a reduction in ctDNA molecules rather than a change in total cfDNA (Supplementary Fig. S4A and S4B). Patients without disease progression or death following CRT had a larger decrease in ctDNA concentration and lower mid-CRT concentration compared with patients who experienced disease progression or death (Fig. 2D). Both ctDNA fold change and mid-CRT ctDNA concentration were statistically significantly associated with PFS in univariable and multivariable analyses, but pre-CRT concentrations were not (Fig. 2E and F; Supplementary Fig. S4C-S4E).
我们首先描述了训练队列中 CRT 中期时间点的 ctDNA 动力学特征。在训练队列的所有患者中，ctDNA 浓度的中位数下降了 13.3 倍，从 CT 前的每毫升 31.8 个单倍体基因组当量（

hGE / mL

）下降到 CT 中期的

0.92 hGE / mL

（图 2C）。这种下降是由 ctDNA 分子的减少而不是 cfDNA 总量的变化引起的（补充图 S4A 和 S4B）。与出现疾病进展或死亡的患者相比，CRT 后未出现疾病进展或死亡的患者的 ctDNA 浓度下降幅度更大，CRT 中期浓度更低（图 2D）。在单变量和多变量分析中，ctDNA 折叠变化和 CRT 中期 ctDNA 浓度均与 PFS 有显著统计学相关性，但 CRT 前浓度与 PFS 无关（图 2E 和 F；补充图 S4C-S4E）。

We next identified an optimal threshold of mid-CRT ctDNA concentration for distinguishing between patients with short and long PFS in the training cohort and applied it to patients in the validation cohort (Fig. 2G and H). Patients with a ctDNA concentration greater than

3.2 hGE / mL

had significantly worse PFS in both the training [HR

= 6.4

95 %

confidence interval (CI) 2.6-15.7] and validation cohorts (

HR = 4.2, 95 %

CI, 1.3-13.4). The majority of misclassified patients had mid-CRT ctDNA concentrations less than or equal to

3.2 hGE / mL

but ultimately developed progression. We also investigated mid-CRT ctDNA concentrations by first site of failure. We observed an increase in isolated local recurrences among patients with mid-CRT concentrations less than or equal to

3.2 hGE / mL

(Supplementary Fig. S4F) and a corresponding significantly lower mid-CRT ctDNA concentration in patients with isolated local recurrences versus other progression or death (Fig. 2I). These results suggest that patients with isolated local recurrences have a lower total body disease burden than those who progress distantly. These data demonstrate that ctDNA levels during CRT are highly prognostic for PFS in patients with locoregionally advanced NSCLC.
接下来，我们确定了用于区分训练队列中短 PFS 和长 PFS 患者的最佳阈值--CRT 中期 ctDNA 浓度，并将其应用于验证队列中的患者（图 2G 和 H）。在训练队列[HR

= 6.4

95 %

置信区间 (CI) 2.6-15.7]和验证队列（

HR = 4.2, 95 %

CI, 1.3-13.4）中，ctDNA 浓度大于

3.2 hGE / mL

的患者的 PFS 明显较差。大多数被误诊的患者在 CRT 中期的 ctDNA 浓度小于或等于

3.2 hGE / mL

，但最终病情恶化。我们还按首个失败部位调查了 CRT 中期 ctDNA 浓度。我们观察到，在 CRT 中期浓度小于或等于

3.2 hGE / mL

的患者中，孤立性局部复发有所增加（补充图 S4F），而孤立性局部复发患者的 CRT 中期 ctDNA 浓度相对于其他进展或死亡患者显著降低（图 2I）。这些结果表明，与远处进展的患者相比，孤立的局部复发患者的全身疾病负担较低。这些数据表明，CRT 期间的 ctDNA 水平对局部晚期 NSCLC 患者的 PFS 具有高度预示性。

Prediction of PFS Using Radiomics
利用放射组学预测 PFS

Although mid-CRT ctDNA concentration was strongly prognostic, we hypothesized that combining it with additional prognostic factors could reduce false negatives and improve prediction of which patients will ultimately develop PD. We first sought to determine whether radiomic analysis of pretreatment CT images could predict PFS after CRT for NSCLC. To explore multiple radiomic features and build a robust model for patient outcome prediction, we trained our radiomics model using publicly available CT images from patients treated with CRT for NSCLC on Radiation Therapy Oncology Group (RTOG) 0235/ACRIN 6668 (

n = 209

patients, Supplementary Table S6; ref. 56). After image segmentation, we calculated a total of 14 knowledge-based radiomic features, including tumor morphology, intensity, and texture, as well as quantitative characteristics of the tumor-invasive margin and tumor-associated vasculature (Fig. 3A). In the training cohort, five CT image features were associated with PFS (FDR < 0.05): (i) entropy of image intensity, (ii) margin blurriness, (iii) number of blood vessels in contact with the tumor, (iv) vessel coverage, and (v) vessel scattering.
虽然 CRT 中期的 ctDNA 浓度具有很强的预后性，但我们假设将其与其他预后因素相结合可以减少假阴性，并改善对哪些患者最终会发展为 PD 的预测。我们首先试图确定对治疗前 CT 图像的放射学分析是否能预测 NSCLC CRT 后的 PFS。为了探索多种放射组学特征并建立一个稳健的患者预后预测模型，我们使用公开的 CT 图像对放射治疗肿瘤学组（RTOG）0235/ACRIN 6668（

n = 209

患者，补充表 S6；参考文献 56）中接受 CRT 治疗的 NSCLC 患者进行了放射组学模型训练。图像分割后，我们计算出了共 14 个基于知识的放射学特征，包括肿瘤形态、强度和纹理，以及肿瘤浸润边缘和肿瘤相关血管的定量特征（图 3A）。在训练队列中，有五个 CT 图像特征与 PFS 相关（FDR < 0.05）：(i)图像强度熵，(ii)边缘模糊度，(iii)与肿瘤接触的血管数量，(iv)血管覆盖率，以及(v)血管散射。

Figure 2. ctDNA levels during CRT are prognostic of PFS. A, Schematic of genotyping and ctDNA monitoring during CRT. Tumor genotyping was performed using tumor tissue when available or using the SNV score on pre-CRT plasma in combination with peripheral blood leukocytes. Plasma samples were collected for ctDNA analysis pre-CRT and 10-30 days into CRT (mid-CRT). B, Plot of patient characteristics and tumor variants for patients in the ctDNA training cohort treated at MDACC

n = 40

) and the validation cohort treated at Stanford University (

n = 21

). C, Pre-CRT and mid-CRT ctDNA concentrations in the training cohort (

n = 40

patients). Patients with ctDNA not detected at the mid-CRT time point are plotted one log below the ctDNA limit of detection with open circles.

P

value was calculated using a two-sided Wilcoxon matched-pair signed-rank test.

D

, Log 10 fold change in ctDNA concentration from the pre-CRT to mid-CRT time point in patients from the training cohort with (

n = 23

) and without (

n = 17

) (continued on following page)
图 2：CRT 期间的 ctDNA 水平是 PFS 的预后指标。A，CRT 期间基因分型和 ctDNA 监测示意图。在有肿瘤组织的情况下使用肿瘤组织进行肿瘤基因分型，或结合外周血白细胞使用 CRT 前血浆的 SNV 评分进行肿瘤基因分型。在 CRT 前和 CRT 后 10-30 天（CRT 中期）收集血浆样本进行 ctDNA 分析。B、在 MDACC 治疗的 ctDNA 培训队列（

n = 40

）和在斯坦福大学治疗的验证队列（

n = 21

）中患者的特征和肿瘤变异图。C、训练队列中 CT 前和 CT 中的 ctDNA 浓度（

n = 40

患者）。在 CRT 中期时间点未检测到 ctDNA 的患者，其 ctDNA 浓度比 ctDNA 检测限低 1 个对数，并用开圆表示。

P

值采用双侧 Wilcoxon 配对符号秩检验计算。

D

，训练队列中有(

n = 23

)和没有(

n = 17

)的患者 ctDNA 浓度从 CRT 前到 CRT 中期时间点的对数 10 倍变化（续下页）

After feature selection, the final radiomic signature incorporated margin blurriness and vessel scattering (Fig. 3B; Supplementary Fig. S5A).
经过特征选择后，最终的放射学特征包括边缘模糊和血管散射（图 3B；补充图 S5A）。

In the training cohort, the radiomic score as a continuous variable was strongly associated with PFS on multivariable analysis (Fig. 3C). Notably, the radiomic score can be calculated for both contrast-enhanced and non-contrast-enhanced CTs, and its prognostic value was independent of whether CT contrast was administered (Supplementary Fig. S5BS5D). In addition, margin blurriness and vessel scattering displayed excellent reproducibility across independent sets of tumor contours (Supplementary Fig. S5E and S5F). To stratify patients into low-risk and high-risk groups based on the radiomic score, we determined the optimal cutoff by ROC analysis in the training cohort and applied the same cutoff to a validation cohort of patients treated with CRT for NSCLC at MDACC (Fig. 3D-G). Patients with a radiomic score greater than 1.194 had significantly worse PFS in the training (

HR = 2.9, 95 % CI, 2.1 - 4.0

) and validation cohorts (HR

= 6.1, 95 % CI, 1.8 - 20.2)

. These data demonstrate the ability of imaging features to stratify patients with locally advanced NSCLC by risk of recurrence or death after CRT.
在训练队列中，放射学评分作为连续变量与多变量分析中的 PFS 密切相关（图 3C）。值得注意的是，放射学评分既可用于造影剂增强 CT，也可用于非造影剂增强 CT，而且其预后价值与是否使用造影剂无关（补充图 S5BS5D）。此外，边缘模糊度和血管散射在独立的肿瘤轮廓中显示出极好的再现性（补充图 S5E 和 S5F）。为了根据放射学评分将患者分为低危和高危两组，我们在训练队列中通过 ROC 分析确定了最佳分界点，并将相同的分界点应用于在 MDACC 接受 CRT 治疗的 NSCLC 患者的验证队列（图 3D-G）。在训练队列（

HR = 2.9, 95 % CI, 2.1 - 4.0

）和验证队列（HR

= 6.1, 95 % CI, 1.8 - 20.2)

.这些数据表明，成像特征能够根据 CRT 后的复发或死亡风险对局部晚期 NSCLC 患者进行分层。

Biological and Molecular Prognostic Factors in Locoregionally Advanced NSCLC
局部晚期 NSCLC 的生物学和分子预后因素

We next aimed to identify additional complementary biological and molecular prognostic factors that could potentially improve risk stratification. However, few prognostic factors have been validated in patients treated with CRT for NSCLC. Therefore, we established a separate historic cohort of 108 patients with stage IIB to IIIA NSCLC from Stanford University and The Cancer Genome Atlas (TCGA) treated with radiotherapy to identify and train additional features prognostic of PFS (Supplementary Table S7). We focused on previously identified prognostic factors in NSCLC treated with CRT (7) and predictors of local recurrence (26, 57, 58). Within our historic training cohort, pre-CRT largest lesion MTV, largest lesion gross tumor volume (GTV), and histology (non-squamous cell carcinoma (non-SCC) vs. SCC) were significantly associated with inferior PFS (Fig. 4A-C). However, there was not a significant association of sex, age, or stage with PFS. As expected, largest lesion GTV and MTV were highly correlated (Supplementary Fig. S6A).
我们接下来的目标是找出更多可改善风险分层的补充性生物和分子预后因素。然而，很少有预后因素在接受 CRT 治疗的 NSCLC 患者中得到验证。因此，我们从斯坦福大学和癌症基因组图谱（The Cancer Genome Atlas，TCGA）中挑选了 108 名接受放疗的 IIB 期至 IIIA 期 NSCLC 患者建立了一个独立的历史队列，以确定和训练预后 PFS 的其他特征（补充表 S7）。我们重点研究了之前确定的接受 CRT 治疗的 NSCLC 的预后因素（7）和局部复发的预测因素（26、57、58）。在我们的历史训练队列中，CRT 前最大病灶 MTV、最大病灶肿瘤总体积 (GTV) 和组织学（非鳞状细胞癌 (non-SCC) 与 SCC）与较差的 PFS 显著相关（图 4A-C）。然而，性别、年龄或分期与 PFS 没有明显关系。不出所料，最大病灶 GTV 与 MTV 高度相关（补充图 S6A）。

Although mutations in KRAS and TP53 have been associated with poor outcomes in cohorts of NSCLC that include all stages and treatment modalities

(59, 60)

, very few studies have exclusively examined locoregionally advanced NSCLC treated with CRT. To identify recurrent mutations associated with PFS in locoregionally advanced NSCLC, we focused on previously identified lung cancer driver genes observed in at
虽然 KRAS 和 TP53 突变与包括所有分期和治疗方式

(59, 60)

的 NSCLC 队列中的不良预后有关，但很少有研究专门研究了接受 CRT 治疗的局部晚期 NSCLC。为了确定与局部晚期 NSCLC 的 PFS 相关的复发性突变，我们将重点放在了以前确定的肺癌驱动基因上。
least

5 %

of lung adenocarcinomas or lung SCCs (61). Of the 23 genes analyzed, only mutations in KRAS and KEAP1 were significantly associated with PFS (Fig. 4D-F; Supplementary Fig. S6B). Consistent with prior reports (62), non-SCCs more frequently failed distantly (Supplementary Fig. S6C). In addition, tumors with KEAP1 mutations, but not with KRAS mutations, more frequently failed within the radiation field (63).
至少

5 %

的肺腺癌或肺 SCC（61）。在分析的 23 个基因中，只有 KRAS 和 KEAP1 的突变与 PFS 显著相关（图 4D-F；补充图 S6B）。与之前的报道（62）一致，非 SCCs 更经常在远处失败（补充图 S6C）。此外，KEAP1 基因突变的肿瘤（而非 KRAS 基因突变的肿瘤）在放射野内更常发生衰竭（63）。

CHIP (i.e., mutations in CH-associated genes with VAF

> 2 %

) has previously been associated with an increased risk of cardiovascular events in the general population (64) and worse outcomes in patients with solid tumors (65). We therefore evaluated the association of CH with PFS in the ctDNA training cohort. Given the low prevalence of CHIP in our cohort, we instead focused on patients with variants in canonical CH genes at any allele fraction (

52 %

of the cohort). The presence of CH was not associated with PFS or OS in our cohort of patients with locoregionally advanced NSCLC treated with CRT (Supplementary Fig. S7A and S7B).
CHIP（即 VAF

> 2 %

的 CH 相关基因突变）曾与普通人群心血管事件风险增加（64）和实体瘤患者预后恶化（65）有关。因此，我们在 ctDNA 培训队列中评估了 CH 与 PFS 的关系。鉴于 CHIP 在我们队列中的发病率较低，我们转而关注在任何等位基因比例（队列中

52 %

）的典型 CH 基因中存在变异的患者。在我们接受 CRT 治疗的局部晚期 NSCLC 患者队列中，CH 的存在与 PFS 或 OS 无关（补充图 S7A 和 S7B）。

A Dynamic Risk Index for NSCLC
非小细胞肺癌动态风险指数

Having identified pre-CRT prognostic factors for PFS, we aimed to combine these factors with our radiomics model and mid-CRT ctDNA changes to improve prediction of PD during CRT for NSCLC. We previously described the Continuous Individualized Risk Index (CIRI), a dynamic risk model that integrates diverse biomarkers into a single patient-level risk estimate that can be updated throughout the course of treatment (66). Using this approach, we built a prognostic model for locoregionally advanced NSCLC treated with CRT that incorporated pre-CRT and mid-CRT risk factors called CIRI-LCRT. To minimize overfitting, we inferred the radiomic hyperparameters for CIRI-LCRT from the RTOG 0235/ ACRIN 6668 dataset and the biological and molecular feature hype-parameters in our historic training cohort. The MDACC training cohort was used to infer the ctDNA hyperparameters and to train the full CIRI-LCRT model. After the full model was finalized and locked, validation was performed in the Stanford cohort.
在确定了 CRT 前的 PFS 预后因素后，我们的目标是将这些因素与我们的放射组学模型和 CRT 中期的 ctDNA 变化相结合，以改善 NSCLC CRT 期间的 PD 预测。我们之前介绍过连续个体化风险指数（CIRI），这是一种动态风险模型，它将不同的生物标志物整合到单一的患者水平风险评估中，并可在整个治疗过程中进行更新（66）。利用这种方法，我们为接受 CRT 治疗的局部晚期 NSCLC 建立了一个预后模型，该模型纳入了 CRT 前和 CRT 中期的风险因素，称为 CIRI-LCRT。为了尽量减少过拟合，我们从 RTOG 0235/ ACRIN 6668 数据集中推断出了 CIRI-LCRT 的放射学超参数，并从历史训练队列中推断出了生物和分子特征超参数。MDACC 训练队列用于推断 ctDNA 超参数和训练完整的 CIRI-LCRT 模型。在最终确定并锁定完整模型后，在斯坦福队列中进行了验证。

We first evaluated all possible combinations of significant biological, molecular, radiomic, and ctDNA features. Several combinations of features performed very similarly in our training cohort, with C-statistics of the top 10 models ranging from 0.932 to 0.955 (Supplementary Table S8). To generate a model that would be as easy as possible to apply clinically, we chose the model among the top 10 in our training cohort with the fewest features. The final CIRI-LCRT model, incorporating histology, radiomics, and mid-CRT ctDNA concentration, displayed robust prediction of PFS at 12 and 24 months and significantly outperformed individual
我们首先评估了所有可能的重要生物学、分子、放射学和 ctDNA 特征组合。在我们的训练队列中，几种特征组合的表现非常相似，前 10 个模型的 C 统计量从 0.932 到 0.955 不等（补充表 S8）。为了生成一个尽可能易于临床应用的模型，我们选择了训练队列中特征最少的前 10 个模型。最终的 CIRI-LCRT 模型结合了组织学、放射组学和中期 CT CT CTDNA 浓度，对 12 个月和 24 个月的 PFS 预测效果显著，并明显优于单个模型。

Figure 2. (Continued) progression or death during follow-up. P value was calculated using a two-sided Mann-Whitney test. E, HRs with 95% CIs for univariable Cox proportional hazards models for PFS based on pre-CRT and mid-CRT ctDNA parameters as continuous variables in the training cohort (

n = 40

patients). F, HRs with 95% CIs for a multivariable Cox proportional hazards model for PFS including mid-CRT ctDNA concentration in the training cohort (

n = 40

patients).

G

and

H

, Kaplan-Meier analysis of PFS based on mid-CRT ctDNA concentration above and below the optimal cutpoint defined in the training cohort (

n = 40; G

) and applied to the validation cohort (

n = 21; H

P

values were calculated using two-sided log-rank tests. I, Mid-CRT ctDNA concentration in patients from the training cohort with progression or death by first site of recurrence (

n = 6

isolated local recurrence, 17 distant progression or death). Patients with ctDNA not detected at the mid-CRT time point are plotted 1 log below the ctDNA limit of detection with open circles. Dotted line represents the optimal ctDNA cutpoint in the training cohort.

P

value was calculated using a two-sided Mann-Whitney test. (A, Created with BioRender.com.)
图 2.（续）随访期间病情进展或死亡。P 值采用双侧 Mann-Whitney 检验计算。E、训练队列（

n = 40

患者）中基于 CRT 前和 CRT 中 ctDNA 参数作为连续变量的单变量 Cox 比例危险模型 PFS 的 HRs 及 95% CIs。F，训练队列（

n = 40

患者）中包括 CRT 中期 ctDNA 浓度的 PFS 多变量 Cox 比例危险模型的 HRs 及 95% CIs。

G

和

H

，基于训练队列（

n = 40; G

）中定义并应用于验证队列（

n = 21; H

）的高于和低于最佳切点的 CRT 中期 ctDNA 浓度的 PFS 的卡普兰-梅耶分析。

P

值采用双侧对数秩检验计算。I、训练队列中按第一复发部位（

n = 6

孤立局部复发，17 例远处复发或死亡）划分的病情进展或死亡患者的 CTDNA 中期浓度。在 CRT 中期时间点未检测到 ctDNA 的患者在 ctDNA 检测限以下 1 个对数处以开放圆圈标出。虚线代表训练队列中的最佳 ctDNA 切点。

P

值通过双侧 Mann-Whitney 检验计算得出。(A，使用 BioRender.com 创建）。

Figure 3. Radiomic analysis of pretreatment CT images predicts PFS. A, Schematic of the radiomic analysis workflow. Tumors and tumor-associated blood vessels were segmented on pre-CRT CT images, and 14 radiomic features were extracted. After feature selection, a radiomic model was constructed using Cox regression analysis prior to determining the optimal cutpoint for stratifying patients at high vs. low risk for progression or death. B, Representative images for radiomic features. CT images for RTOG23 showed a blurry tumor margin and chaotic blood vessel distribution, and this patient progressed 4 months after treatment. In contrast, CT images for RTOG34 showed more distinct tumor margins and narrowly focused blood vessel distribution, and this patient remained disease-free at 68 months. C, HRs with 95% CIs for a multivariable Cox proportional hazards model for PFS, including radiomic score, in the training cohort (

n = 209

patients). D, ROC curve for prediction of PFS at 2 years using the radiomic score. The optimal cutpoint is displayed on the graph. E, Percentage of patients with high-risk and low-risk radiomic scores who developed disease progression or died by 2 years in the training cohort. F and G, Kaplan-Meier analysis of PFS based on radiomic risk in the RTOG 0235 training cohort (

n = 209; F

) and the MDACC validation cohort

(n = 62; G) . P

values were calculated using two-sided log-rank tests.
图 3.对治疗前 CT 图像的放射学分析可预测 PFS。A，放射学分析工作流程示意图。在 CRT 前 CT 图像上分割肿瘤和肿瘤相关血管，并提取 14 个放射学特征。特征选择后，使用 Cox 回归分析构建放射组学模型，然后确定最佳切点，对进展或死亡风险高与低的患者进行分层。B，放射学特征的代表性图像。RTOG23 的 CT 图像显示肿瘤边缘模糊，血管分布混乱，该患者在治疗 4 个月后病情恶化。相比之下，RTOG34 的 CT 图像显示肿瘤边缘更清晰，血管分布更集中，该患者在 68 个月后仍未发病。C、训练队列（

n = 209

患者）中包括放射学评分在内的 PFS 多变量 Cox 比例危险模型的 HRs 及 95% CIs。D、使用放射学评分预测 2 年后 PFS 的 ROC 曲线。图中显示了最佳切点。E、在训练队列中，放射学评分为高风险和低风险的患者在 2 年后出现疾病进展或死亡的百分比。F 和 G，根据 RTOG 0235 训练队列（

n = 209; F

）和 MDACC 验证队列

(n = 62; G) . P

的放射学风险计算的 PFS 的卡普兰-秩检验值，采用双侧对数秩检验。
risk factors, including mid-CRT ctDNA concentration (Fig. 5A-C; Supplementary Tables S9 and S10). Importantly, performance was nearly identical in the independent validation cohort (Fig. 5D and E). Model performance steadily improved with the addition of each feature in both the training and validation cohorts (Supplementary Fig. S8A), demonstrating the benefit of integrating complementary data.
风险因素，包括 CT 中期 ctDNA 浓度（图 5A-C；补充表 S9 和 S10）。重要的是，独立验证队列的性能几乎相同（图 5D 和 E）。随着训练队列和验证队列中每个特征的增加，模型性能稳步提高（补充图 S8A），证明了整合互补数据的好处。

The CIRI-LCRT quantitative risk at each time point can be used to stratify patients into risk groups. Considering groups with

< 33 %, 33 %

66 %

, and

> 66 %

predicted risk of progression or death by 36 months at the pre-CRT or mid-CRT time point significantly stratified the training cohort into patients with low, medium, and high risk of progression or death (Fig. 5F). When we applied the same cutoffs to our validation
每个时间点的 CIRI-LCRT 定量风险可用于对患者进行风险分层。考虑到

< 33 %, 33 %

至

66 %

组和

> 66 %

组在 CRT 前或 CRT 中期时间点预测的 36 个月进展或死亡风险，可将训练队列显著分为低、中和高进展或死亡风险患者组（图 5F）。当我们将相同的截断值应用于验证

Figure 4. Biological and molecular prognostic factors in patients with NSCLC treated with CRT. A, HRs with 95% CIs for univariable Cox proportional hazards models for PFS based each biological feature in a historic training cohort of patients from TCGA and Stanford University who did not undergo ctDNA analysis (

n = 108

patients for male vs. female sex, age as continuous variable, stage III vs. II, and non-SCC vs. SCC histology; 38 patients for largest lesion MTV and largest lesion GTV). B and C, Kaplan-Meier analysis of PFS stratified by (B) histology and © GTV above or below the optimal cutpoint in the historic training cohort.

P

values were calculated using two-sided log-rank tests. D, HRs with 95% CIs for univariable Cox proportional hazards models for PFS based on mutation status of each significant molecular feature identified from Supplementary Fig. S6B (

n = 108

patients).

E

and

F

, Kaplan-Meier analysis of PFS stratified by (E) KRAS mutation status and (F) KEAP1 mutation status in the historic training cohort.

P

values were calculated using two-sided log-rank tests.
图 4：接受 CRT 治疗的 NSCLC 患者的生物学和分子预后因素接受 CRT 治疗的 NSCLC 患者的生物学和分子预后因素。A、基于 TCGA 和斯坦福大学未进行 ctDNA 分析的患者历史训练队列（

n = 108

患者为男性 vs. 女性、年龄为连续变量、III 期 vs. II 期、非 SCC vs. SCC 组织学；38 名患者为最大病灶 MTV 和最大病灶 GTV）中每个生物学特征的单变量 Cox 比例危险模型 PFS 的 HRs 及 95% CIs。B 和 C，按(B)组织学和© GTV 高于或低于历史训练队列中的最佳切点对 PFS 进行分层的卡普兰-梅厄分析。

P

值采用双侧对数秩检验计算。D，基于补充图 S6B 中确定的每个重要分子特征的突变状态（

n = 108

患者）的 PFS 单变量 Cox 比例危险模型的 HRs 及 95% CI。

E

和

F

，根据历史训练队列中(E)KRAS 突变状态和(F)KEAP1 突变状态对 PFS 进行分层的卡普兰-梅耶分析。

P

值采用双侧对数秩检验计算。
cohort, we observed similar stratification of risk (Fig. 5G). Considering only the final prediction at the mid-CRT time point, stratifying patients by

\leq 50 %

> 50 %

risk of progression or death by 36 months significantly separated the training cohort into low- and high-risk groups, and applying this cutoff to the validation cohort achieved similar results (Supplementary Fig. S8B and S8C). We observed good calibration of our model across the whole cohort when comparing predicted and observed risk of PFS at 12 months (Fig. 5H).
我们观察到类似的风险分层（图 5G）。仅考虑 CRT 中期时间点的最终预测，按

\leq 50 %

或

> 50 %

到 36 个月时的进展或死亡风险对患者进行分层，可显著地将训练队列分为低风险组和高风险组，将这一分界线应用于验证队列也取得了相似的结果（补充图 S8B 和 S8C）。在比较 12 个月时预测的 PFS 风险和观察到的 PFS 风险时，我们观察到整个队列的模型校准良好（图 5H）。

CIRI-LCRT enabled individualized real-time updating of the probability of PFS as model features became available over the course of CRT. For example, two patients in the validation cohort, LUP810 and LUP235, were both treated with CRT for stage IIIA NSCLC (Fig. 5I). LUP810 presented with a left upper lobe SCC with a low risk radiomic score, corresponding to a

41 %

CIRI-LCRT pre-CRT risk of progression or death at 24 months. Mid-CRT, the patient’s ctDNA concentration was

1.7 hGE / mL

, lowering his CIRI-LCRT risk to

12 %

. Twenty-five months after starting CRT, LUP810 remained
在 CRT 治疗过程中，随着模型特征的出现，CIRI-LCRT 能够对 PFS 概率进行个性化的实时更新。例如，验证队列中的两名患者 LUP810 和 LUP235 都接受了 IIIA 期 NSCLC 的 CRT 治疗（图 5I）。LUP810 患有左上叶 SCC，放射学评分为低风险，对应于 CIRI-LCRT 前的

41 %

24 个月时进展或死亡风险。CRT 中期，患者的 ctDNA 浓度为

1.7 hGE / mL

，使其 CIRI-LCRT 风险降至

12 %

。在开始 CRT 25 个月后，LUP810 仍然
disease-free. In contrast, LUP235 presented with a central adenocarcinoma with a high risk radiomic score, leading to a 95% CIRI-LCRT pre-CRT risk. At his mid-CRT blood draw, his ctDNA concentration was

37.8 hGE / mL

, corresponding to a 100% CIRI-LCRT risk of progression. He ultimately developed a local recurrence and distant brain metastases 6 months after starting CRT. Taken together, these results demonstrated that CIRI-LCRT improves prediction of PFS over individual biomarkers alone, enabling accurate risk stratification during CRT for NSCLC.
无疾病。相比之下，LUP235 患有中心腺癌，放射学评分风险较高，CIRI-LCRT 前风险为 95%。在 CT 中期抽血时，他的 ctDNA 浓度为

37.8 hGE / mL

，对应的 CIRI-LCRT 进展风险为 100%。他最终在开始 CRT 6 个月后出现局部复发和远处脑转移。综上所述，这些结果表明，CIRI-LCRT 比单独使用单个生物标志物更能预测 PFS，从而在 NSCLC 的 CRT 治疗过程中实现准确的风险分层。

Comparing CIRI-LCRT with ctDNA MRD
CIRI-LCRT 与 ctDNA MRD 的比较

Given the excellent performance of CIRI-LCRT for predicting PFS during CRT for NSCLC, we performed an exploratory analysis to compare CIRI-LCRT with detection of ctDNA MRD after completion of all treatment. We identified 37 patients across the training and validation cohorts with plasma samples available for analysis from the first follow-up visit
鉴于 CIRI-LCRT 在预测 NSCLC CRT 期间的 PFS 方面表现出色，我们进行了一项探索性分析，将 CIRI-LCRT 与完成所有治疗后的 ctDNA MRD 检测进行比较。我们在训练队列和验证队列中确定了 37 名患者，这些患者的血浆样本可在首次随访时进行分析

Figure 5. Training and validation of CIRI-LCRT for prediction of PFS during CRT for NSCLC. A, Timeline of a patient with NSCLC treated with CRT. CIRI-LCRT integrates pre-CRT risk factors (non-SCC vs. SCC histology and radiomics) and the mid-CRT risk factor (mid-CRT ctDNA concentration) as information is obtained to generate individualized PFS curve predictions. The performance of CIRI-LCRT is evaluated 12 months (PFS12) and 24 months (PFS24) after starting CRT. B-E, Bar plots of the C-statistic (mean and SD) for PFS at each time interval in the training and validation cohorts for each individual risk factor and the full CIRI-LCRT model after integration of pre-CRT and mid-CRT risk factors.

P

values were (continued on following page)
图 5.用于预测 NSCLC CRT 期间 PFS 的 CIRI-LCRT 的训练和验证。A，接受 CRT 治疗的 NSCLC 患者的时间表。CIRI-LCRT 将 CRT 前的风险因素（非 SCC 与 SCC 组织学和放射组学）和 CRT 中期的风险因素（CRT 中期 ctDNA 浓度）作为信息整合在一起，生成个性化的 PFS 曲线预测。开始 CRT 后 12 个月（PFS12）和 24 个月（PFS24），对 CIRI-LCRT 的性能进行评估。B-E，训练队列和验证队列中每个单个风险因素以及整合 CRT 前和 CRT 中期风险因素后的完整 CIRI-LCRT 模型在每个时间间隔的 PFS C 统计量（平均值和标度）柱状图。

P

值为（续下页）
after completion of all chemotherapy and radiation. Despite the mid-CRT plasma sample being collected a median of 2.1 months prior to the MRD plasma sample, CIRI-LCRT outperformed ctDNA MRD for prediction of PFS at 24 months by C-statistics and performed comparably for prediction of PFS at 12 months and by Kaplan-Meier analysis (Fig. 6A-D). In patients who ultimately progressed or died who were correctly predicted by both approaches, CIRI-LCRT provided a 3.0-month median improvement in lead time over ctDNA MRD.
在完成所有化疗和放疗后。尽管 CIRI-LCRT 中期的血浆样本采集时间比 MRD 血浆样本采集时间中位早 2.1 个月，但 CIRI-LCRT 在预测 24 个月的 PFS 方面的 C 统计量优于 ctDNA MRD，在预测 12 个月的 PFS 方面和 Kaplan-Meier 分析中的表现相当（图 6A-D）。在两种方法都能正确预测的最终进展或死亡患者中，CIRI-LCRT 比 ctDNA MRD 的中位预测时间延长了 3.0 个月。

We selected two patients’ vignettes to help illustrate the ability of CIRI-LCRT to provide an earlier prediction of PFS than ctDNA MRD (Fig. 6E). LUP238 underwent CRT for a stage IIIA right middle lobe SCC and ultimately developed local and distant disease progression 10 months after starting treatment. Four months prior to having ctDNA MRD detected, he had a 99% CIRI-LCRT risk of progression or death by 24 months based on pre-CRT risk factors and mid-CRT ctDNA analysis. In contrast, LUP141 completed CRT for a stage IIB SCC of the left lower lobe and remained alive and progression free 24 months later. His mid-CRT CIRI-LCRTpredicted risk of progression or death by 24 months was

31 %

, and ctDNA MRD was not detected 4 months later. Overall, these findings illustrate the potential for CIRI-LCRT to provide a substantially earlier prediction of disease progression or death over ctDNA MRD, potentially enabling improved patient outcomes through earlier treatment escalation or deintensification.
我们选择了两名患者的病例来帮助说明 CIRI-LCRT 比 ctDNA MRD 更早预测 PFS 的能力（图 6E）。LUP238 因右肺中叶 SCC IIIA 期而接受 CRT 治疗，最终在开始治疗 10 个月后出现局部和远处疾病进展。在检测到 ctDNA MRD 的 4 个月前，根据 CRT 前的风险因素和 CRT 中期的 ctDNA 分析，他在 24 个月前病情进展或死亡的 CIRI-LCRT 风险为 99%。与此形成鲜明对比的是，LUP141 因左下叶 IIB 期 SCC 而完成了 CRT，24 个月后仍然存活且无进展。他的 CRT 中期 CIRI-LCRT 预测的 24 个月进展或死亡风险为

31 %

，4 个月后未检测到 ctDNA MRD。总之，这些研究结果表明，CIRI-LCRT 有可能比 ctDNA MRD 更早预测疾病进展或死亡，从而通过更早升级治疗或减量治疗改善患者预后。

DISCUSSION 讨论

By accounting for CH and optimizing tissue-free tumor variant calling from plasma samples, we demonstrated that mid-CRT ctDNA levels measured before the midway point of CRT are prognostic of disease progression in locoregionally advanced NSCLC. Furthermore, we integrated pretreatment risk factors, including radiomic analysis with mid-CRT ctDNA analysis, to build a dynamic risk model that can be updated real-time as model features become available over the course of treatment. In an independent validation cohort, CIRI-LCRT improved prediction of PFS over ctDNA analysis alone with substantially better performance than prior prognostic models for locoregionally advanced NSCLC (8-12).
通过考虑 CH 并优化血浆样本中的无组织肿瘤变异调用，我们证明了在 CRT 中点之前测量的 CRT 中期 ctDNA 水平是局部区域晚期 NSCLC 疾病进展的预后指标。此外，我们还整合了治疗前的风险因素，包括放射学分析和 CRT 中期 ctDNA 分析，从而建立了一个动态风险模型，该模型可在治疗过程中根据模型特征进行实时更新。在一个独立的验证队列中，CIRI-LCRT 比单独的 ctDNA 分析提高了对 PFS 的预测，其性能大大优于之前的局部晚期 NSCLC 预后模型（8-12）。

In contrast to hematopoietic malignancies, the ability to adapt treatment based on mid-treatment PET/CT has not been demonstrated in NSCLC or other solid cancers (67). As a result, new techniques to monitor response to therapy will be critical for adaptive treatment approaches in solid tumors. Early ctDNA kinetics have been associated with patient outcomes in metastatic NSCLC treated with targeted therapies
与造血恶性肿瘤相比，根据治疗中期 PET/CT 调整治疗的能力尚未在 NSCLC 或其他实体瘤中得到证实（67）。因此，监测治疗反应的新技术对于实体瘤的适应性治疗方法至关重要。在接受靶向治疗的转移性 NSCLC 患者中，早期 ctDNA 动力学与患者预后有关

(31, 32)

and immunotherapy (33-35), and we have previously demonstrated that ctDNA kinetics during definitive chemotherapy for diffuse large B-cell lymphoma are prognostic of patient outcomes (68). In patients receiving definitive therapy for solid cancers, several prior studies have demonstrated that ctDNA MRD after completion of therapy is highly prognostic (45, 69-72), but few studies have investigated the association of ctDNA levels during definitive treatment with patient outcomes. Khakoo and colleagues (73) previously examined mid-treatment ctDNA levels in patients undergoing CRT for localized rectal cancer but found no significant association of ctDNA detection with patient outcomes. In contrast to our study in which only

25 %

of patients had ctDNA undetected mid-CRT,

79 %

of patients were undetected using droplet digital PCR for 1 to 3 mutations per patient, suggesting assay sensitivity could be important for mid-CRT ctDNA analysis. Pan and colleagues (74) recently reported superior outcomes for patients with undetectable ctDNA in plasma samples collected near the end of induction chemotherapy followed by CRT in NSCLC. The plasma samples on our study were collected

\sim 9

weeks closer to the start of treatment, leaving more time for possible treatment personalization. The optimal threshold for stratifying PFS was

3.2 hGE / mL

, which corresponds to a ctDNA allele fraction of

0.1 %

when considering the median mid-CRT cfDNA concentration of

10.3 ng / mL

for the patients on our study. Although prior studies have reported concordant ctDNA levels across assays

(75, 76)

, this allele fraction is near the limit of detection for commercial tumor genotype-naïve ctDNA assays (30). As a result, tumor genotype-informed assays will likely be necessary, and further work is needed to ensure the generalizability of this threshold across different assays.

(31, 32)

和免疫疗法（33-35），我们以前曾证实，弥漫大 B 细胞淋巴瘤明确化疗期间的 ctDNA 动力学对患者的预后有影响（68）。对于接受确定性治疗的实体瘤患者，之前的几项研究表明，治疗结束后的 ctDNA MRD 对预后有很大影响（45, 69-72），但很少有研究调查确定性治疗期间的 ctDNA 水平与患者预后的关系。Khakoo 及其同事（73）曾对接受 CRT 治疗局部直肠癌患者的治疗中期 ctDNA 水平进行了研究，但发现 ctDNA 检测与患者预后无明显关联。在我们的研究中，只有

25 %

的患者在 CRT 中期未检测到 ctDNA，与此不同的是，使用液滴数字 PCR 检测时，

79 %

的患者未检测到 1 到 3 个突变，这表明检测灵敏度对 CRT 中期 ctDNA 分析很重要。Pan 及其同事（74）最近报告称，在 NSCLC 诱导化疗接近尾声时采集的血浆样本中检测不到 ctDNA，但随后进行 CRT 治疗的患者疗效更佳。我们研究中的血浆样本是在治疗开始前

\sim 9

周采集的，为可能的个性化治疗留出了更多时间。对 PFS 进行分层的最佳阈值为

3.2 hGE / mL

，考虑到我们研究中患者的 CRT 中期 cfDNA 浓度中位数为

10.3 ng / mL

，这相当于 ctDNA 等位基因分数为

0.1 %

。虽然之前的研究报告了不同检测方法的 ctDNA 水平一致

(75, 76)

，但这一等位基因分数已接近商业化肿瘤基因型鉴定 ctDNA 检测方法的检测极限（30）。因此，肿瘤基因型检测可能是必要的，还需要进一步的工作来确保这一阈值在不同检测中的通用性。

Despite several prior studies attempting to identify prognostic factors in locoregionally advanced NSCLC, patients currently all receive the same dose of radiation therapy with concurrent platinum doublet chemotherapy. CIRI-LCRT was highly prognostic of PFS with a C-statistic of 0.94 at 24 months in an independent validation cohort, suggesting that CIRI-LCRT-predicted risk could be used to guide adaptive therapy approaches such as changing concurrent systemic therapy during radiation. Remarkably, CIRI-LCRT during CRT performed similarly to detection of ctDNA after completion of all therapy despite being performed more than 2 months earlier than MRD analysis. Earlier initiation of salvage therapies has bene shown to improve outcomes in other cancers (77-79), and a post hoc analysis of the phase III PACIFIC trial that compared consolidation durvalumab versus placebo after CRT for stage III NSCLC found a greater benefit in patients randomized within 14 days of radiotherapy (3). As a result, the improved lead time with CIRI-LCRT
尽管之前有几项研究试图确定局部晚期 NSCLC 的预后因素，但目前所有患者都在接受相同剂量的放疗的同时接受铂双联化疗。在一个独立的验证队列中，CIRI-LCRT 对 PFS 有高度的预后作用，24 个月时的 C 统计量为 0.94，这表明 CIRI-LCRT 预测的风险可用于指导适应性治疗方法，如在放疗期间改变同时进行的全身治疗。值得注意的是，尽管 CIRI-LCRT 比 MRD 分析早 2 个多月进行，但在 CRT 期间进行 CIRI-LCRT 的效果与完成所有治疗后检测 ctDNA 的效果相似。在其他癌症中，更早地启动挽救疗法已被证明能改善预后（77-79），III 期 PACIFIC 试验比较了 III 期 NSCLC CRT 后 Durvalumab 与安慰剂的巩固治疗，结果发现在放疗后 14 天内随机接受治疗的患者获益更大（3）。因此，CIRI-LCRT 延长了准备时间。

Figure 5. (Continued) calculated empirically from 2,000 bootstrap resamplings. F and G, Kaplan-Meier analysis of PFS stratified by CIRI-LCRT-predicted risk of progression or death by 36 months at either the pre-CRT or mid-CRT time point in (F) the training cohort and (G) the validation cohort.

P

values were calculated using two-sided log-rank tests for trend. H, Calibration plot for CIRI-LCRT including pre-CRT and mid-CRT predictions demonstrating predicted and observed PFS at 12 months. Predictions were grouped by predicted risk, and the observed risk of each group was calculated by Kaplan-Meier analysis. Error bars represent the SEM. The slope and Y-intercept with 95% CIs are shown for the line of best fit by linear regression. Perfect calibration would be represented by a slope of 1 and a Y-intercept of 0. I, Vignettes demonstrating CIRI-LCRT-predicted survival as information becomes available over the course of CRT for a patient with no progression at 25 months (LUP810) and a patient with local and distant progression 6 months after starting CRT (LUP235). Updated survival curves with prior predictions grayed out and favorable and unfavorable risk factors are shown on the plots at the time of integration into the model. Tumors are indicated with orange arrows on PET/CT images. Conc., concentration.
图 5.（续）由 2,000 次引导重采样经验计算得出。F 和 G，在(F)训练队列和(G)验证队列中，按 CIRI-LCRT 预测的 36 个月前 CRT 前或 CRT 中期的进展或死亡风险对 PFS 进行分层的 Kaplan-Meier 分析。

P

值通过双侧对数秩检验趋势计算得出。H，CIRI-LCRT 的校准图，包括 CRT 前和 CRT 中的预测，显示 12 个月时的预测和观察 PFS。预测结果按预测风险分组，每组的观察风险通过 Kaplan-Meier 分析计算。误差条代表 SEM。斜率和 Y-截距以及 95% CIs 显示的是线性回归最佳拟合线。斜率为 1 和 Y-截距为 0 代表完美的校准。 I, 随着 CRT 过程中信息的增加，CIRI-LCRT 预测存活率的小插图显示了一名在 25 个月时没有进展的患者（LUP810）和一名在开始 CRT 6 个月后出现局部和远处进展的患者（LUP235）的存活率。更新后的生存曲线中，先前的预测结果变灰，图中显示的是纳入模型时的有利和不利风险因素。PET/CT 图像上的肿瘤用橙色箭头表示。Conc.，浓度。

Figure 6. CIRI-LCRT during CRT performs comparably with ctDNA MRD after completion of treatment. A and B, Bar plots of the C-statistic (mean and SD) for PFS based on CIRI-LCRT-predicted risk and ctDNA MRD detection at (A) 12 months and (B) 24 months in patients from the whole cohort with ctDNA MRD samples available for analysis. P values were calculated empirically from 2,000 bootstrap resamplings. C and D, Kaplan-Meier analysis of PFS stratified by © CIRI-LCRT risk prediction after integration of pre-CRT and mid-CRT risk factors and (D) ctDNA MRD detection after completion of all therapy.

P

values were calculated using two-sided log-rank tests. E, Vignettes showing the timing of CIRI-LCRT and ctDNA MRD analysis and PET/CT or CT imaging prior to treatment and at last follow-up in two patients correctly predicted by both methods. LUP238 developed local and distant progression 10 months after starting CRT. LUP141 had no evidence of progression 24 months after starting CRT. In both cases, CIRI-LCRT correctly predicted the patient outcome more than 4 months prior to ctDNA MRD analysis. Tumors are indicated with orange arrows. Conc., concentration. (E, Created with BioRender.com.)
图 6.CRT 期间的 CIRI-LCRT 与完成治疗后的 ctDNA MRD 的表现相当。A 和 B，在(A) 12 个月和(B) 24 个月时，基于 CIRI-LCRT 预测风险和 ctDNA MRD 检测的 PFS 的 C 统计量（平均值和标度）柱状图，这些患者来自有 ctDNA MRD 样本可供分析的整个队列。P 值根据 2000 次引导重采样经验计算得出。C 和 D，根据© CIRI-LCRT 风险预测整合 CRT 前和 CRT 中期风险因素后的 PFS 分层的 Kaplan-Meier 分析，以及（D）完成所有治疗后的 ctDNA MRD 检测。

P

值采用双侧对数秩检验计算。E、小插图显示了两种方法均正确预测的两名患者在治疗前和最后一次随访时进行 CIRI-LCRT 和 ctDNA MRD 分析以及 PET/CT 或 CT 成像的时间。LUP238 在开始 CRT 治疗 10 个月后出现局部和远处进展。LUP141 在开始 CRT 治疗 24 个月后无进展迹象。在这两个病例中，CIRI-LCRT 都在进行 ctDNA MRD 分析前 4 个多月正确预测了患者的预后。肿瘤用橙色箭头表示。Conc.，浓度。(E，用 BioRender.com 绘制）。
could also help to improve outcomes by enabling earlier initiation of consolidation or salvage therapy. Ultimately, validation in larger cohorts and prospective trials will be necessary to demonstrate improved outcomes with individualized treatment based on CIRI-LCRT. CIRI-LCRT performs substantially better than prior NSCLC prognostic models using only pre-CRT factors to predict OS with reported validation C-statistics/areas under the curve ranging from 0.62 to 0.76 (8, 10, 11).
此外，CIRI-LCRT 还能使巩固或挽救治疗更早开始，从而有助于改善疗效。最终，需要在更大的队列和前瞻性试验中进行验证，以证明基于 CIRI-LCRT 的个体化治疗能改善预后。CIRI-LCRT 在预测 OS 方面的表现大大优于之前仅使用 CRT 前因素的 NSCLC 预后模型，已报道的验证 C 统计量/曲线下面积从 0.62 到 0.76 不等（8、10、11）。

Our final CIRI-LCRT model incorporated tumor histology (non-SCC vs. SCC), radiomic analysis, and mid-CRT ctDNA concentration. Although tumor size as measured by GTV or MTV and KEAP1 and KRAS mutation status were significantly associated with PFS, they did not substantially improve
我们的最终 CIRI-LCRT 模型纳入了肿瘤组织学（非 SCC 与 SCC）、放射学分析和 CRT 中期 ctDNA 浓度。虽然以 GTV 或 MTV 测量的肿瘤大小以及 KEAP1 和 KRAS 突变状态与 PFS 显著相关，但它们并没有显著改善 PFS。
the CIRI-LCRT model, suggesting other features such as the radiomic score may provide similar information. Numerous prior studies have demonstrated the potential for radiomic analysis to predict patient outcomes after CRT for NSCLC (17). These studies have primarily focused on tumor morphology and textural features (19, 20, 23, 80). Our radiomic model performed the best when incorporating two novel features quantifying the blurriness of the tumor-invasive margin and tumor-associated blood vessel scattering. Previous studies have correlated radiomic features with ctDNA levels in patients with cancer

(81, 82)

. To our knowledge, this is the first study to integrate ctDNA and radiomic analyses to predict outcomes of patients with cancer. Although prior studies have suggested that changes in radiomics features may help
CIRI-LCRT 模型，这表明放射学评分等其他特征也能提供类似的信息。之前的许多研究已经证明了放射学分析在预测 NSCLC CRT 患者预后方面的潜力（17）。这些研究主要关注肿瘤形态和纹理特征（19、20、23、80）。我们的放射学模型在结合了量化肿瘤浸润边缘模糊性和肿瘤相关血管散射的两个新特征后表现最佳。以前的研究已将癌症

(81, 82)

患者的放射学特征与 ctDNA 水平相关联。据我们所知，这是第一项整合 ctDNA 和放射学分析来预测癌症患者预后的研究。尽管之前的研究表明，放射组学特征的变化可能有助于
to predict treatment response (83), mid-CRT radiographic imaging is not currently a standard of care and was not routinely acquired for patients in our cohort.
83），CRT 中期放射成像目前还不是治疗标准，我们队列中的患者也没有常规获得该成像。

Similar to most prior studies of stage III NSCLC treated with CRT, patients with SCCs had better PFS and fewer distant metastases than patients with non-squamous histologies (62, 84). In contrast, a recent exploratory analysis of the phase III PACIFIC trial that compared consolidation durvalumab versus placebo after CRT for stage III NSCLC reported better PFS and OS in patients with non-squamous histologies (85). Notably, durvalumab significantly improved OS in non-squamous histologies but not SCC, suggesting this result was likely driven by response to consolidation durvalumab. The majority of the patients on our study did not receive consolidation immunotherapy, and patients with a ctDNA response to consolidation immunotherapy were excluded from analysis. Based on the results of the PACIFIC trial, patients with a high CIRI-LCRT risk of progression due to non-squamous histology would likely be good candidates for consolidation durvalumab. A key feature of CIRI-LCRT is that because it was constructed using Bayesian Cox proportional hazard modeling, patient-level risk can be calculated even if all prognostic factors are not available (e.g., if tumor tissue is not available to assess histology).
与之前大多数接受 CRT 治疗的 III 期 NSCLC 研究相似，与非鳞状组织学患者相比，SCC 患者的 PFS 更佳，远处转移更少（62、84）。与此相反，最近一项对 III 期 PACIFIC 试验的探索性分析比较了 CRT 治疗 III 期 NSCLC 后巩固使用 durvalumab 与安慰剂的效果，结果显示非鳞癌组织学患者的 PFS 和 OS 更好（85）。值得注意的是，durvalumab 能显著改善非鳞癌组织学患者的 OS，但不能改善 SCC 患者的 OS，这表明这一结果很可能是由对巩固性 durvalumab 的反应驱动的。我们研究中的大多数患者没有接受巩固性免疫疗法，对巩固性免疫疗法有 ctDNA 反应的患者未纳入分析。根据 PACIFIC 试验的结果，因非鳞癌组织学而导致 CIRI-LCRT 进展风险较高的患者很可能是巩固性杜瓦鲁单抗的理想候选者。CIRI-LCRT 的一个主要特点是，由于它是采用贝叶斯考克斯比例危险模型构建的，因此即使无法获得所有预后因素（如无法获得肿瘤组织以评估组织学），也能计算出患者水平的风险。

Consistent with our prior study (63), patients with KEAP1 mutations had higher rates of local failure that translated into worse PFS after CRT. TP53 status has previously been associated with OS in unselected patients with NSCLC (60), and one prior study associated

E G F R

mutations with worse PFS in patients treated with CRT for stage III NSCLC (86). However, neither TP53 or EGFR mutations were prognostic of PFS in our historic training cohort. GTV and MTV were highly correlated, and both were significantly associated with PFS. GTV has previously been associated with local relapse, PFS, and OS in NSCLC treated with CRT (87, 88). In our cohort, there was no difference in site of first failure by GTV. Interestingly, absolute mid-CRT ctDNA concentration was a stronger predictor of PFS than log fold change in ctDNA concentration, and the ctDNA parameters were redundant in the CIRI-LCRT model. It is possible that a decrease in disease burden within the radiation field accounts for the majority of the decrease in ctDNA concentration during CRT, and log fold change in ctDNA concentration is less informative about the risk of distant metastasis which is the predominant mode of failure in NSCLC. Notably, although TNM stage is a strong prognostic factor in stage III lung cancer because it can identify patients eligible for surgical resection (89), TNM stage was not a significant prognostic factor in the uniformly treated patient population in this study.
与我们之前的研究（63）一致的是，KEAP1 突变患者的局部失败率较高，从而导致 CRT 治疗后的 PFS 较差。TP53 状态曾与未入选的 NSCLC 患者的 OS 相关（60），一项先前的研究显示，在接受 CRT 治疗的 III 期 NSCLC 患者中，

E G F R

突变与较差的 PFS 相关（86）。然而，在我们的历史训练队列中，TP53 或 EGFR 突变都不是 PFS 的预后因素。GTV 和 MTV 高度相关，两者都与 PFS 显著相关。在接受 CRT 治疗的 NSCLC 患者中，GTV 曾与局部复发、PFS 和 OS 相关（87, 88）。在我们的队列中，GTV 与首次失败的部位没有差异。有趣的是，CRT 中期 ctDNA 绝对浓度比 ctDNA 浓度对折变化对 PFS 的预测作用更强，而且 ctDNA 参数在 CIRI-LCRT 模型中是多余的。可能是在 CRT 期间，放射野内疾病负担的减少占了 ctDNA 浓度减少的大部分，而 ctDNA 浓度对折变化对远处转移风险的信息量较小，而远处转移是 NSCLC 的主要失败模式。值得注意的是，尽管 TNM 分期是 III 期肺癌的一个重要预后因素，因为它可以确定符合手术切除条件的患者（89），但在本研究中，TNM 分期在接受统一治疗的患者群体中并不是一个重要的预后因素。

As is often the case for research studies, tumor tissue was only available for

15 %

of the patients in this study. To overcome this limitation, we developed the SNV score by integrating variant features with machine learning to determine the probability of an individual variant identified in cfDNA being derived from a patient’s tumor. The SNV score enabled identification of more tumor-adjudicated variants than our best empirically defined filters. Because sensitivity and specificity can be tuned by changing the SNV score threshold, this approach could be useful for diverse applications. For example, high specificity is critical when identifying variants
与研究中常见的情况一样，本研究中只有

15 %

名患者可获得肿瘤组织。为了克服这一限制，我们开发了 SNV 评分，将变异特征与机器学习相结合，以确定在 cfDNA 中识别出的单个变异来自患者肿瘤的概率。与我们根据经验定义的最佳过滤器相比，SNV 评分能识别更多的肿瘤判断变异。由于灵敏度和特异性可通过改变 SNV 评分阈值进行调整，因此这种方法可用于多种应用。例如，在识别变异时，高特异性是至关重要的。
predictive of response to systemic therapy, so a higher SNV score threshold (greater probability of being tumor-derived) would be desirable for this application. Other groups have previously used machine learning to identify variants from next-generation sequencing of tumor tissue (90) and to suppress sequencing noise to improve identification of low-allele fraction variants (48). However, to our knowledge, the SNV score is the first machine learning approach to leverage a combination of technical features and biological differences in tumor-derived cfDNA to improve tissue-free genotyping from plasma samples. We anticipate that this approach could increase the clinical utilization of ctDNA analysis when tumor tissue is not available or insufficient for sequencing.
因此，对于这种应用来说，较高的 SNV 评分阈值（肿瘤来源的可能性更大）是可取的。其他研究小组以前曾利用机器学习从肿瘤组织的下一代测序中识别变异（90），并抑制测序噪音以提高低等位基因变异的识别率（48）。然而，据我们所知，SNV 评分是第一种利用肿瘤来源的 cfDNA 的技术特征和生物学差异来改进血浆样本无组织基因分型的机器学习方法。我们预计，当肿瘤组织无法获得或不足以进行测序时，这种方法能提高 ctDNA 分析的临床利用率。

The majority of cfDNA in plasma samples is derived from hematopoietic cells in circulation, and recent studies have demonstrated that most cfDNA variants are derived from CH (48, 50). Cytotoxic therapy increases the prevalence of CH

(48, 65)

, which is associated with an increased risk of therapyrelated myeloid neoplasms (91). Recently, repeat sampling from patients prior to and over a year following chemotherapy or radiotherapy demonstrated expansion of CH variants in the DNA damage response genes TP53, PPM1D, and CHEK2 (53). Furthermore, Ppm1d mutations have previously been shown to expand during cytotoxic therapy in mouse models (52). In this study, we demonstrate for the first time that CRT can differentially select for or against CH mutations within weeks of starting therapy. We observed a decrease in the allele fraction of SF3B1 and TET2 mutations and an expansion in PPM1D mutations, consistent with the changes observed during chemotherapy in mice. During CRT, we did not observe a significant change in the allele fraction of TP53 mutations. However, it is possible that some variants could be selected for over longer time periods. These findings suggest that efforts to suppress therapy-related myeloid neoplasms likely need to occur during therapy, and subsequent risk likely depends on which CH variants are present prior to treatment. In contrast to prior studies

(64, 65)

, we did not observe an association between CH and outcomes for patients with locoregionally advanced NSCLC treated with CRT. As a result, CH was not included in our CIRI-LCRT model. Limitations of this analysis included that due to the low prevalence of patients with CHIP (i.e., CH mutations with VAF > 2%), we focused on mutations in canonical CH genes with any VAF and the relatively small sample size. Therefore, additional studies in larger cohorts are warranted to further explore associations between CHIP and patient outcomes after CRT for NSCLC.
血浆样本中的 cfDNA 大部分来自循环中的造血细胞，最近的研究表明，大部分 cfDNA 变体来自 CH（48、50）。细胞毒治疗会增加 CH

(48, 65)

的发病率，这与治疗相关的髓系肿瘤风险增加有关（91）。最近，对化疗或放疗前及化疗或放疗后一年以上的患者进行的重复采样显示，DNA 损伤反应基因 TP53、PPM1D 和 CHEK2 中的 CH 变异体增加了（53）。此外，Ppm1d 突变也曾在小鼠模型的细胞毒治疗过程中被证明会扩大（52）。在本研究中，我们首次证明了 CRT 可在开始治疗的数周内对 CH 突变进行不同程度的选择或抑制。我们观察到 SF3B1 和 TET2 突变的等位基因比例下降，PPM1D 突变扩大，这与小鼠化疗期间观察到的变化一致。在 CRT 期间，我们没有观察到 TP53 突变等位基因比例的显著变化。不过，某些变异可能会在更长的时间内被选择。这些研究结果表明，抑制与治疗相关的骨髓肿瘤的努力可能需要在治疗期间进行，而随后的风险可能取决于治疗前存在哪些 CH 变异。与之前的研究

(64, 65)

不同，我们没有观察到 CH 与接受 CRT 治疗的局部晚期 NSCLC 患者的预后之间存在关联。因此，我们的 CIRI-LCRT 模型中未包括 CH。这项分析的局限性包括：由于 CHIP 患者的发病率较低（即 VAF>2%的 CH 基因突变），我们重点研究了具有任何 VAF 的典型 CH 基因突变，而且样本量相对较小。因此，有必要在更大的队列中进行更多的研究，以进一步探讨 CHIP 与 NSCLC CRT 后患者预后之间的关系。

Prior to routine clinical implementation, CIRI-LCRT should be validated in prospective cohorts, and the clinical utility of personalizing treatment based on CIRI-LCRT should be tested in a randomized interventional study. In terms of feasibility, tumor histology is already routinely assessed by pathologists on pretreatment biopsies, and the radiomic score can be obtained from routine diagnostic or radiation treatment planning CT scans. A key factor for adaptive therapy based on CIRI-LCRT will be assay turnaround time for ctDNA analysis. In our cohort, we included patients analyzed as early as 10 days into CRT. Turnaround times continue to improve for ctDNA analysis with current turnaround times reported to be approximately 9 days for commercial assays (92).
在常规临床应用之前，CIRI-LCRT 应在前瞻性队列中得到验证，基于 CIRI-LCRT 的个性化治疗的临床实用性应在随机干预研究中得到检验。在可行性方面，病理学家已经对治疗前活检的肿瘤组织学进行了常规评估，放射学评分也可从常规诊断或放射治疗计划 CT 扫描中获得。基于 CIRI-LCRT 的适应性治疗的一个关键因素是 ctDNA 分析的检测周转时间。在我们的队列中，早在 CRT 开始的 10 天，我们就对患者进行了分析。ctDNA 分析的周转时间在不断缩短，据报道，目前商业测定的周转时间约为 9 天（92）。

As a result, mid-CRT blood samples could be drawn at day 10, and mid-CRT CIRI-LCRT risk could be determined by the end of the third week of CRT. This would enable systemic therapy to be changed for the second half of the typical 6-week radiation therapy course, and/or a cone down scan could be obtained to plan a radiation boost to residual disease.
因此，可以在 CRT 第 10 天抽取 CRT 中期的血液样本，并在 CRT 第三周结束时确定 CRT 中期的 CIRI-LCRT 风险。这样就可以在典型的 6 周放疗疗程的后半段改变全身疗法，和/或进行锥形向下扫描，以计划对残留疾病进行放射增量。

Strengths of our study include accounting for CH when identifying tumor variants pre-CRT, thorough evaluation of multiple previously described prognostic factors, the ability to incorporate group-level prior data from multiple training sets to integrate diverse outcome predictors, and validation of our model in an independent cohort across academic institutions. Limitations of our study include the modest sample size of our validation cohort, the retrospective nature of our analysis, and incomplete information available for all datasets. Potential biases from retrospective analyses include information bias due to inaccurate data collection in the medical record, survivorship bias from including only patients with mid-CRT blood samples available for analysis, and confounding due to unmeasured variables. Most of our patients were treated prior to consolidation therapy with durvalumab or more recently osimertinib becoming standard of care (93), and both of these treatments have been shown to improve PFS in patients with stage III lung NSCLC (85, 94). Furthermore, our cohorts were selected from patients receiving treatment from specialists at comprehensive cancer centers, which may not reflect the broader population of patients with locoregionally advanced NSCLC. Although treatment regimens were not standardized across institutions and providers, our model performed strongly and comparably in both the training and validation cohorts. Future research in larger prospectively collected cohorts will be required for validation before clinical implementation of CIRI-LCRT.
我们研究的优点包括：在确定 CRT 前的肿瘤变异时考虑了 CH 因素；对以前描述过的多种预后因素进行了全面评估；能够纳入来自多个训练集的组级先验数据，以整合多种结局预测因素；在跨学术机构的独立队列中验证了我们的模型。我们研究的局限性包括验证队列的样本量不大、我们的分析具有回顾性以及所有数据集的信息不完整。回顾性分析可能存在的偏差包括：病历数据收集不准确导致的信息偏差、仅纳入有 CRT 中期血液样本可供分析的患者导致的存活率偏差以及未测量变量导致的混杂。我们的大多数患者都是在使用达伐单抗或最近成为标准治疗方法的奥希替尼进行巩固治疗之前接受治疗的（93），而这两种治疗方法都被证明可以改善 III 期肺 NSCLC 患者的 PFS（85，94）。此外，我们的队列选自接受综合癌症中心专科医生治疗的患者，这可能无法反映更广泛的局部晚期 NSCLC 患者群体。虽然不同机构和医疗机构的治疗方案不尽相同，但我们的模型在训练队列和验证队列中的表现都很好，具有可比性。在临床实施 CIRI-LCRT 之前，还需要在更大规模的前瞻性队列中进行研究验证。

In summary, we have demonstrated that mid-treatment ctDNA analysis can be used to monitor the response of locoregionally advanced NSCLC to CRT. Furthermore, we have integrated mid-CRT ctDNA analysis with pre-CRT risk factors, including radiomic analysis, to build a highly prognostic model for PFS. We propose that CIRI-LCRT could enable personalized and response-adapted therapies to reduce toxicity and improve outcomes in patients with unresectable NSCLC treated with CRT. However, validation in larger cohorts and prospective randomized trials will be necessary to test the benefit of modifying therapy based on CIRI-LCRT.
总之，我们已经证明，治疗中期的 ctDNA 分析可用于监测局部晚期 NSCLC 对 CRT 的反应。此外，我们还将 CRT 中期的 ctDNA 分析与 CRT 前的风险因素（包括放射学分析）相结合，建立了一个 PFS 的高度预后模型。我们认为，CIRI-LCRT 可以实现个性化和适应反应的疗法，从而降低毒性，改善接受 CRT 治疗的不可切除 NSCLC 患者的预后。然而，要检验基于 CIRI-LCRT 调整疗法的益处，还需要在更大的队列和前瞻性随机试验中进行验证。

METHODS 方法

Study Design and Patients
研究设计和患者

All samples analyzed in this study were prospectively collected with written informed consent from subjects enrolled on protocols approved by institutional review boards at MDACC and Stanford University. All studies were conducted in accordance with the Declaration of Helsinki. A subset of the patients were analyzed in prior studies (37, 45, 46, 49, 50). Two SNV score models were trained in cohorts of patients with lung cancer and control donors without a diagnosis of cancer prepared with two different types of sequencing adapters described in detail below. A total of 93 patients and 36 controls were used to train the tandem sequencing adapter SNV
本研究中分析的所有样本均为前瞻性采集，受试者在知情同意的情况下按照 MDACC 和斯坦福大学机构审查委员会批准的方案注册。所有研究均按照《赫尔辛基宣言》进行。先前的研究（37、45、46、49、50）对部分患者进行了分析。在使用两种不同类型的测序适配器制备的肺癌患者和未确诊癌症的对照供体队列中训练了两种 SNV 评分模型，详情如下。共有 93 名患者和 36 名对照者被用于训练串联测序适配器 SNV
score, and 97 patients and 56 controls were used to train the FLEX sequencing adapter SNV score. These patients were also used for mid-CRT ctDNA analysis if matched tumor tissue was available for tumor genotyping.
得分，97 例患者和 56 例对照用于训练 FLEX 测序适配器 SNV 评分。如果有匹配的肿瘤组织可用于肿瘤基因分型，这些患者还被用于 CRT 中期 ctDNA 分析。

A total of 101 patients with American Joint Committee on Cancer seventh edition stage IIA-IIIB NSCLC from MDACC and Stanford University were retrospectively identified who had a mid-CRT plasma sample collected 10 to 30 days into CRT for cfDNA analysis. All of these patients were used to characterize the prevalence of CH and monitor CH allele fraction changes during CRT. A subset of these patients (

n = 61

) were divided into ctDNA training (MDACC,

n = 40

) and validation cohorts (Stanford,

n = 21

) based on the institutions where they received CRT if they were of stages IIB to IIIB and had variants identified before treatment by tumor genotyping or SNV score analysis on pre-CRT plasma. In addition, patients who received consolidation anti-PD-L1 immunotherapy were not included in the mid-CRT ctDNA analysis if they had a ctDNA response during consolidation immunotherapy because we have previously shown these patients have improved outcomes that would not be captured at the mid-CRT time point (46). The ctDNA training cohort was used to identify the mid-CRT ctDNA concentration cutpoint and train the mid-CRT portion of the CIRI-LCRT model.
我们回顾性地确定了 101 名来自 MDACC 和斯坦福大学的美国癌症联合委员会第七版 IIA-IIIB 期 NSCLC 患者，这些患者在 CRT 进行 10 到 30 天后采集了 CRT 中期血浆样本，用于 cfDNA 分析。所有这些患者都被用来描述 CH 的患病率，并监测 CRT 期间 CH 等位基因分数的变化。这些患者中的一部分（

n = 61

）如果属于 IIB 至 IIIB 期，并且在治疗前通过肿瘤基因分型或 CRT 前血浆 SNV 评分分析发现了变异，则根据他们接受 CRT 的机构将其分为 ctDNA 培训队列（MDACC，

n = 40

）和验证队列（斯坦福大学，

n = 21

）。此外，接受巩固性抗 PD-L1 免疫疗法的患者如果在巩固性免疫疗法期间出现了 ctDNA 反应，则不纳入 CRT 中期 ctDNA 分析，因为我们之前已经证明，这些患者的预后有所改善，但在 CRT 中期时间点却无法捕捉到（46）。ctDNA 训练队列用于确定 CRT 中期 ctDNA 浓度切点和训练 CIRI-LCRT 模型的 CRT 中期部分。

Finally, we identified a separate historic training cohort (

n = 108

NSCLC patients) to identify and train prognostic pre-CRT factors for the CIRI-LCRT model. This cohort included 38 patients treated with CRT at Stanford University for stage IIB to IIIB NSCLC described previously

(58, 95)

who had undergone tumor genotyping using a clinical hybrid capture-based sequencing assay covering 130 to 198 genes (96) and did not undergo mid-CRT ctDNA analysis. In addition, we identified 70 patients with stage IIB to IIIB NSCLC from TCGA with high-quality PFS data available (97) by filtering for patients who had undergone whole-exome DNA sequencing, received radiotherapy, and did not undergo surgical resection.
最后，我们确定了一个单独的历史训练队列（

n = 108

NSCLC 患者），用于识别和训练 CIRI-LCRT 模型的 CRT 前预后因素。该队列包括 38 例在斯坦福大学接受 CRT 治疗的 IIB 至 IIIB 期 NSCLC 患者，这些患者之前曾接受过

(58, 95)

所述的临床混合捕获测序法进行的肿瘤基因分型，涵盖 130 至 198 个基因 (96)，但未接受中期 CRT ctDNA 分析。此外，我们还从 TCGA 中筛选出了 70 例 IIB 期至 IIIB 期 NSCLC 患者，这些患者均接受过全外显子组 DNA 测序，接受过放疗，且未进行手术切除（97）。

Power Considerations 电源考虑因素

We performed a power analysis to determine the appropriate size for the ctDNA validation cohort to detect a significant difference in PFS between patients above and below a CIRI-LCRT-predicted risk of progression or death at 36 months of

50 %

. Assuming a HR of 9 for progression or death in patients with a CIRI-LCRT-predicted risk greater than

50 %, 20

total patients would achieve

95 %

power to detect a significant different in PFS between the two groups with a median follow-up of 24 months at an

α

of 0.05 (98).
我们进行了一项功率分析，以确定 ctDNA 验证队列的适当规模，从而检测出高于和低于 CIRI-LCRT 预测的 36 个月进展或死亡风险

50 %

的患者在 PFS 方面的显著差异。假定 CIRI-LCRT 预测风险高于

50 %, 20

的患者进展或死亡的 HR 为 9，那么在

α

为 0.05 的条件下，在中位随访 24 个月的情况下，总患者数将达到

95 %

的功率，以检测两组患者 PFS 的显著差异 (98)。

CRT and Blood Collection
CRT 和采血

Patients treated with CRT for locoregionally advanced NSCLC at MDACC and Stanford University completed pretreatment staging with chest CT, whole-body PET/CT, and brain MRI, and the diagnosis of NSCLC was confirmed by pathology review at the treating institution. Patients were treated with a median of 66 Gy in 30 fractions with concurrent platinum doublet chemotherapy. None of the patients received tyrosine kinase inhibitors or other targeted agents as part of initial definitive treatment. Peripheral blood samples were collected for plasma and leukocyte isolation prior to starting treatment and 10 to 30 days into CRT. For a subset of patients, a third peripheral blood sample was collected within 4 months of completing radiotherapy and chemotherapy for MRD analysis.
在 MDACC 和斯坦福大学接受 CRT 治疗的局部晚期 NSCLC 患者均通过胸部 CT、全身 PET/CT 和脑部 MRI 进行了预处理分期，NSCLC 的诊断由治疗机构的病理复查确认。患者接受了中位 66 Gy、30 次分割的治疗，并同时接受了铂双联化疗。在最初的确定性治疗中，患者均未接受酪氨酸激酶抑制剂或其他靶向药物治疗。在开始治疗前和开始 CRT 的 10 到 30 天内，采集外周血样本进行血浆和白细胞分离。对于一部分患者，在完成放疗和化疗后的 4 个月内采集第三次外周血样本进行 MRD 分析。

Library Preparation and Sequencing
文库制备和测序

DNA samples from plasma, leukocytes, and tumors were sequenced using cancer personalized profiling by CAPP-seq as described previously (49, 50). Briefly, peripheral venous blood was collected in K2EDTA (Becton Dickinson) or CellSave (Menarini Silicon Biosystems) tubes and centrifuged to separate plasma and leukocytes for
血浆、白细胞和肿瘤的 DNA 样本通过 CAPP-seq 进行癌症个性化图谱测序，如前所述（49, 50）。简而言之，外周静脉血收集于 K2EDTA（Becton Dickinson 公司）或 CellSave（Menarini Silicon Biosystems 公司）试管中，离心分离血浆和白细胞后，用 CAPP-seq 进行测序。
storage at

- 80^{\circ} C

prior to DNA isolation. Tumor DNA was isolated from formalin-fixed, paraffin embedded sections when available. cfDNA was extracted from plasma using QIAmp Circulating Nucleic Acid Kit (QIAGEN) according to the manufacturer’s instructions. A median of 32 ng cfDNA was prepared for sequencing using KAPA LTP Library Prep Kit (Kapa Biosystems) with minor modifications to the manufacturer’s instructions. Samples sequenced prior to November 2017 were prepared using tandem sequencing adapters, and samples sequenced subsequently were prepared using FLEX sequencing adapters, both described previously

(49, 50)

. Both adapters incorporate dual-index sample barcodes for demultiplexing and unique identifiers (UID) to enable the tracking of individual molecules. However, FLEX adapters utilize error-correcting UIDs and separate the UID portion of the adapter from the portion containing the sample barcode. Hybrid capture-based target enrichment was performed using one of three previously described custom selector pools of biotinylated oligonucleotides (Roche NimbleGen) targeting 170, 302, or 355 kilobase pairs frequently mutated in NSCLC

(50, 99)

. The 355 kilobase selector also included 11 genes canonically associated with CH (ASXL1, CBL, DNMT3A, GNB1, GNAS, MYD88, PPM1D, SF3B1, STAT3, TET2, and TP53; ref. 100). Samples were sequenced on an Illumina HiSeq 4000 System using

2 \times 150

base pair pairedend reads with eight base indexing. Sequencing data were processed using a custom bioinformatics pipeline to identify putative SNVs as described previously (50).
在分离 DNA 之前，将其储存在

- 80^{\circ} C

中。根据制造商的说明，使用 QIAmp Circulating Nucleic Acid Kit (QIAGEN)从血浆中提取 cfDNA。使用 KAPA LTP 文库制备试剂盒（Kapa Biosystems）制备中位数为 32 ng 的 cfDNA 用于测序，并对生产商的说明稍作修改。2017 年 11 月之前测序的样本使用串联测序适配器制备，之后测序的样本使用 FLEX 测序适配器制备，这两种适配器均在之前

(49, 50)

进行过描述。这两种适配器都包含用于解复用的双索引样本条形码和唯一标识符（UID），以便追踪单个分子。不过，FLEX 适配器使用的是纠错 UID，并将适配器的 UID 部分与包含样品条形码的部分分开。基于混合捕获的靶标富集是使用先前描述的三种定制选择池之一进行的，这三种选择池是生物素化寡核苷酸（罗氏 NimbleGen），靶标是在 NSCLC

(50, 99)

中经常突变的 170、302 或 355 千碱基对。355 千碱基选择器还包括 11 个与 CH 典型相关的基因（ASXL1、CBL、DNMT3A、GNB1、GNAS、MYD88、PPM1D、SF3B1、STAT3、TET2 和 TP53；参考文献 100）。100).样本在 Illumina HiSeq 4000 系统上进行测序，使用

2 \times 150

碱基对配对末端读数和 8 个碱基索引。测序数据使用定制的生物信息学管道进行处理，以确定推定的 SNV，如前所述（50）。

CH Analysis CH 分析

Coding variants associated with CH were first identified in the cfDNA. SNPs were removed by filtering out variants with an allele fraction between

40 %

and

60 %

or greater than

90 %

, rescuing potential CHIP by keeping variants in canonical CH genes observed in at least one case of hematologic malignancy in the Catalogue of Somatic Mutations in Cancer (COSMIC) v85 (RRID: SCR_002260). Variants were also required to have a population allele frequency of less than

0.1 %

in the Genome Aggregation Database (Broad Institute) and to have a sequencing depth greater than

50 %

of the median sample depth. Variants were considered associated with CH if they were also present in pre-CRT leukocytes using the Monte Carlo monitoring approach described previously at a detection index of less than 0.05 (101). CHIP was further defined as a mutation in a canonical CH gene previously observed in at least one case of hematologic malignancy in COSMIC v85 and present at an allele fraction of greater than

2 %

in pre-CRT leukocytes without other diagnostic criteria for a hematologic malignancy (100). CH and CHIP variants identified pre-CRT were queried in mid-CRT cfDNA samples by Monte Carlo monitoring with variants considered detected at a detection index of less than 0.1 , consistent with the ctDNA analysis described in “ctDNA Monitoring.”
首先在 cfDNA 中确定与 CH 相关的编码变异。通过过滤等位基因频率在

40 %

和

60 %

之间或大于

90 %

的变异，保留在《癌症中的体细胞突变目录》（COSMIC）v85（RRID：SCR_002260）中至少一例血液恶性肿瘤中观察到的典型 CH 基因中的变异，从而挽救潜在的 CHIP。此外，还要求变异在基因组聚合数据库（Broad Institute）中的群体等位基因频率小于

0.1 %

，且测序深度大于中位样本深度的

50 %

。使用蒙特卡洛监测方法（Monte Carlo monitoring approach），如果 CRT 前白细胞中也存在检测指数小于 0.05 的变异，则认为这些变异与 CH 相关（101）。CHIP 的进一步定义是：在 COSMIC v85 中至少在一例血液恶性肿瘤中观察到的典型 CH 基因突变，且在无其他血液恶性肿瘤诊断标准的 CT 前白细胞中等位基因比例大于

2 %

（100）。通过 Monte Carlo 监测在 CT 中期的 cfDNA 样本中查询 CT 前发现的 CH 和 CHIP 变异，变异检测指数小于 0.1 时视为检测到，这与 "ctDNA 监测 "中描述的 ctDNA 分析一致。

SNV Score Tissue-Free Variant Calling
SNV 评分无组织变异调用

We adapted the framework from our previously described Lung Cancer Likelihood in Plasma framework (ref. 50) to develop the SNV score, which estimates the probability that a SNV is tumor-derived. To maximize the number of patients and variants available for training, the model was constructed and evaluated using leave-one-out cross-validation. Putative SNVs were initially called using the custom adaptive variant caller described previously (49). Variant calls were preprocessed to remove (i) SNPs from any individual in the study, (ii) mutations in oncogenes without any cases in COSMIC v85, (iii) variants present in matched leukocyte DNA by Monte Carlo monitoring at a detection index of less than 0.1, (iv) SNPs from any patient in the cohort, (v) variants lying in repeat, intronic, intergenic, or pseudogene regions, (vi) variants less than or equal to

50 %

of the median sample depth, (vii) variants with a population allele frequency greater than or equal to

0.1 %

in the gnom AD database (102),
我们从之前描述的血浆中肺癌可能性框架（参考文献 50）中改编了这一框架，以开发 SNV 评分，该评分可估算 SNV 来源于肿瘤的概率。为了最大限度地增加可用于训练的患者和变异体的数量，模型的构建和评估采用了留空交叉验证的方法。推测的 SNV 最初使用之前描述过的定制自适应变异调用器（49）进行调用。对变异调用进行了预处理，以剔除：(i) 来自研究中任何个体的 SNPs；(ii) 在 COSMIC v85 中没有任何病例的肿瘤基因突变；(iii) 通过蒙特卡洛监测在匹配的白细胞 DNA 中发现的检测指数小于 0.1 的变异；(iv) 在 COSMIC v85 中没有任何病例的肿瘤基因突变。1，(iv) 来自队列中任何患者的 SNP，(v) 位于重复、内含子、基因间或假基因区域的变异，(vi) 小于或等于中位样本深度

50 %

的变异，(vii) 在 gnom AD 数据库（102）中人群等位基因频率大于或等于

0.1 %

的变异、
(viii) noncoding variants, (ix) mutations in canonical CH genes, and (x) recurrent background artefacts specific to each targeted sequencing space.
(viii) 非编码变异；(ix) 典型 CH 基因突变；(x) 每个目标测序空间特有的经常性背景伪差。

For each SNV, the following features were annotated: (i) Bayesian background of the variant in germline samples, (ii) variant allele frequency, (iii) germline depth, (iv) mean barcode family size, (v) short fragment score 1, (vi) short fragment score 2, (vii) transition/transversion, (viii) duplex support, (ix) pass outlier cutoff, (x) mapping quality, (xi) lung cancer hotspot, (xii) UID errors corrected, (xiii) mean Phred quality, (xiv) mean variant position in read, (xv) power to detect the variant in the matched germline, (xvi) variant frequency in the cohort, (xvii) selector tile mutation frequency in the cohort, (xviii) Monte Carlo detection index for the variant in the germline, (xix) distribution of reads mapped to the plus and minus strand, (xx) mean number of nonreference bases in reads with the variant, (xxi) normalized depth, (xxii) normalized barcode family size, (xxiii) gnomAD population allele frequency, (xxiv) lung cancer driver gene, and (xxv) nonsynonymous mutation in lung cancer driver gene. In contrast to FLEX adapters, tandem adapters do not enable correction of errors in UIDs. Due to this difference and possible technical differences between samples prepared between these workflows, separate SNV score models were trained for the two adapter schemata, excluding the UID error corrected feature from the tandem adapter model.
对于每个 SNV，都注释了以下特征：(i) 种系样本中变异的贝叶斯背景，(ii) 变异等位基因频率，(iii) 种系深度，(iv) 平均条形码家族大小，(v) 短片段得分 1，(vi) 短片段得分 2，(vii) 过渡/转换、(xii)肺癌热点，(xii)UID 错误校正，(xiii)平均 Phred 质量，(xiv)读数中的平均变异位置，(xv)在匹配种系中检测变异的能力、(xvi) 队列中的变异频率， (xvii) 队列中的选择瓦突变频率， (xviii) 种系中变异的蒙特卡罗检测指数， (xix) 映射到正负链的读数分布、(xx)带有变异的读数中的非参考碱基平均数；(xxi)归一化深度；(xxii)归一化条形码族大小；(xxiii)gnomAD 群体等位基因频率；(xxiv)肺癌驱动基因；(xxv)肺癌驱动基因中的非同义突变。与 FLEX 适配体不同，串联适配体无法纠正 UID 中的错误。由于这种差异以及在这些工作流程中制备的样本之间可能存在的技术差异，因此针对这两种适配器模式训练了不同的 SNV 评分模型，串联适配器模型不包括 UID 错误校正特征。

In the subset of patients with matched tumors sequenced using the same targeted selector, tumor-adjudicated variants were identified by filtering putative SNVs from tumor genomic DNA to remove variants with (i) an allele fraction less than or equal to

2 %

, (ii) a Monte Carlo detection index less than or equal to 0.05 when monitoring in matched germline, (iii) population allele frequency greater than or equal to

0.1 %

in the gnomAD database, and (iv) mutations in canonical CH genes. Using a semi-supervised learning framework, an elastic net logistic regression model was trained to distinguish tumor-adjudicated variants from nonadjudicated variants in patients with matched tumors, and this model was used to label variants from patients with NSCLC without matched tumor samples and to assign a weight based on the probability of the variant being tumor-adjudicated. SNVs from patients and controls were combined with their labels and weights to make a final feature matrix that was used within an elastic net for logistic regression with a cross-validation regularization parameter corresponding to the minimum cross-validation for the final model. The trained model was then used to assign a SNV score to held out variants for the leave-one-out cross-validation or patients in the ctDNA training and validation cohorts without matched tumor. For benchmarking, the SNV score was compared with previously defined empiric filters (46) in patients with matched tumor using a SNV score threshold that achieved equivalent positive predictive value for a SNV being tumor-adjudicated.
在使用相同靶向选择器测序的匹配肿瘤患者子集中，通过过滤肿瘤基因组 DNA 中的推测 SNV 来确定肿瘤判断变异，以剔除具有以下特征的变异：(i) 等位基因分数小于或等于

2 %

；(ii) Monte Carlo 检测指数小于或等于 0.05，(iii) gnomAD 数据库中的群体等位基因频率大于或等于

0.1 %

，(iv) 典型 CH 基因中的突变。利用半监督学习框架训练了一个弹性网逻辑回归模型，以区分匹配肿瘤患者中的肿瘤判定变异和非判定变异，并用该模型标记来自无匹配肿瘤样本的 NSCLC 患者的变异，并根据变异为肿瘤判定变异的概率分配权重。来自患者和对照组的 SNV 与其标签和权重相结合，形成最终的特征矩阵，该矩阵被用于逻辑回归的弹性网中，其交叉验证正则化参数与最终模型的最小交叉验证相对应。然后，训练好的模型将被用于为剔除交叉验证的变异体或无匹配肿瘤的 ctDNA 训练队列和验证队列中的患者分配 SNV 得分。为了进行基准比较，在有匹配肿瘤的患者中将 SNV 得分与之前定义的经验过滤器（46）进行了比较，使用的 SNV 得分阈值与肿瘤判断 SNV 的阳性预测值相当。

ctDNA Monitoring ctDNA 监测

SNVs were identified from tumor tissue and filtered as described previously (46) or from pre-CRT plasma using a SNV score threshold of 0.3 . The presence of ctDNA was queried during treatment using Monte Carlo-based monitoring, and ctDNA was considered to be detected at a detection index of less than 0.1. Mid-CRT samples with a median deduped depth less than

500 \times

were excluded from analysis. The allele fraction of all monitored variants was averaged to determine the ctDNA allele fraction for each sample. Because the limit of detection is lower, and therefore confidence that ctDNA is not present is higher when more variants are monitored or sequencing depth is increased, samples with ctDNA not detected were assigned an allele fraction one log below the limit of detection calculated using a Poisson distribution as described previously (101). The mean allele fraction was multiplied by the plasma cfDNA concentration
从肿瘤组织中鉴定 SNV，并按照之前的描述（46）进行筛选，或使用 SNV 评分阈值 0.3 从 CRT 前的血浆中进行筛选。在治疗过程中，使用基于蒙特卡洛的监测对 ctDNA 的存在进行查询，检测指数小于 0.1 时视为检测到 ctDNA。中位深度小于

500 \times

的 CRT 中期样本被排除在分析之外。对所有监测变异的等位基因分数进行平均，以确定每个样本的 ctDNA 等位基因分数。由于检测限较低，因此当监测到更多变异或测序深度增加时，对不存在 ctDNA 的置信度就会更高，因此未检测到 ctDNA 的样本的等位基因分数会比检测限低一个对数，计算方法采用泊松分布，如前所述（101）。等位基因平均分数乘以血浆 cfDNA 浓度
measured by Qubit (Thermo Fisher Scientific) to calculate the ctDNA concentration with each haploid genomic equivalent assumed to have a mass of 3.3 pg . Tumor genotype-informed CAPP-seq was utilized for ctDNA MRD analysis with variants called from tumor tissue or pretreatment plasma as described previously (46).
通过 Qubit（赛默飞世尔科技公司）测量来计算 ctDNA 浓度，假设每个单倍体基因组当量的质量为 3.3 pg。如前所述（46），ctDNA MRD 分析使用了肿瘤基因型信息 CAPP-seq，从肿瘤组织或治疗前血浆中调用变体。

Radiomic Analysis 放射原子分析

The radiomics model was constructed in accordance with the guidelines recommended by the Image Biomarker Standardization Initiative (ref. 103) using the RTOG 0235/ACRIN 6668 dataset (

n = 209

) that is publicly available at The Cancer Imaging Archive (ref. 104). The model was then independently validated in the MDACC cohort (

n = 62

). The treatment planning CT scans were downloaded from The Cancer Imaging Archive for the RTOG cohort or retrieved from the local institutional Picture Archiving and Communication System (PACS). The GTV was manually delineated for radiation treatment planning by the treating physician. A total of 14 knowledgebased radiomic features were calculated: four morphology features (volume, surface area, surface to volume ratio, and sphericity), two image intensity features (energy and entropy), four texture features (contrast, correlation, homogeneity, and joint entropy of the gray level co-occurrence matrix), tumor-invasive margin blurriness, and three tumor-associated blood vessel features (number of blood vessels in direct contact with the tumor, vessel coverage, and vessel scattering).
放射组学模型是根据图像生物标记标准化倡议（Image Biomarker Standardization Initiative）（参考文献 103）推荐的指南，使用 RTOG 0235/ACRIN 6668 数据集（

n = 209

）构建的，该数据集可在癌症成像档案（The Cancer Imaging Archive）（参考文献 104）中公开获取。该模型随后在 MDACC 队列中进行了独立验证（

n = 62

）。RTOG 队列的治疗计划 CT 扫描是从癌症影像档案馆下载的，或者是从当地机构的图像存档和通信系统（PACS）中获取的。GTV 由主治医生手动划定，用于制定放射治疗计划。共计算了 14 个基于知识的放射学特征：四个形态特征（体积、表面积、表面体积比和球形度）、两个图像强度特征（能量和熵）、四个纹理特征（灰度共现矩阵的对比度、相关性、同质性和联合熵）、肿瘤浸润边缘模糊度和三个肿瘤相关血管特征（与肿瘤直接接触的血管数量、血管覆盖率和血管散射）。

Morphology features, image intensity features, and texture features were extracted based on the GTVs using an open-source Python package PyRadiomics (v3.0.1; ref. 105). To capture information on the invasive tumor margin, we computed a quantitative feature to evaluate the blurriness of a three-dimensional (3D) ring region defined around the tumor boundary. The 3D ring region was formed by extending a radial distance of 20 mm inward and 20 mm outward from the tumor boundary via morphologic dilation and erosion operations. We then computed the 3D image gradient summed along three orthogonal axes to measure the blurriness of the tumor-invasive margin. To quantify the relationship between each tumor and associated vasculature, we first used a previously validated Hessian-based approach to detect and segment the blood vessels around the tumor in 3D (106). Vessel coverage was defined as the ratio between the tumor-vasculature interface (i.e., the overlap area of blood vessels on the tumor surface) and total surface area of the tumor. Vascular scattering was computed as the entropy of a binary image formed by the tumor-vasculature interface (107). In detail, we first transformed the 3D tumor surface image into a 2D binary surface image, which was divided into two categories: areas overlapping with blood vessels and tumor surface without overlapping blood vessels. Given the binary surface image represented by a matrix P , the 2D entropy is calculated as follows:
我们使用开源 Python 软件包 PyRadiomics（v3.0.1；参考文献 105）根据 GTV 提取了形态特征、图像强度特征和纹理特征。为了获取浸润性肿瘤边缘的信息，我们计算了一个定量特征来评估肿瘤边界周围三维（3D）环形区域的模糊程度。三维环形区域是通过形态学扩张和侵蚀操作从肿瘤边界向内和向外延伸 20 毫米的径向距离形成的。然后，我们计算了沿三个正交轴相加的三维图像梯度，以测量肿瘤浸润边缘的模糊程度。为了量化每个肿瘤与相关血管之间的关系，我们首先使用了之前验证过的基于黑森的方法来检测和分割肿瘤周围的三维血管（106）。血管覆盖率被定义为肿瘤-血管界面（即肿瘤表面血管的重叠面积）与肿瘤总表面积之间的比率。血管散射是根据肿瘤-血管界面形成的二值图像的熵计算出来的（107）。具体来说，我们首先将三维肿瘤表面图像转化为二维二进制表面图像，分为两类：与血管重叠的区域和没有重叠血管的肿瘤表面。二维表面图像由矩阵 P 表示，二维熵的计算方法如下：

E = - \sum_{i} \sum_{j} P (i, j) \log (P (i, j))

In which i represents the gray value of the pixel and j represents the mean gray value of its neighborhood. The probability

P_{i, j}

is defined as follows:
其中 i 代表像素的灰度值，j 代表其附近像素的平均灰度值。概率

P_{i, j}

的定义如下：

P_{i, j} = (f (i, j)) / (W \cdot H)

In which

f (i, j)

is the number of occurrences of two tuples

(i, j)

and

W

and

H

are the image sizes. The neighborhood consists of the eight pixels that are immediately adjacent to the target pixel. When blood vessels are distributed evenly throughout the tumor surface, vascular scattering is high. By contrast, when the distribution of blood vessels is limited to a narrow range/area around the tumor, vascular scattering is low.
其中，

f (i, j)

是两个图元

(i, j)

的出现次数，

W

和

H

是图像大小。邻域由紧邻目标像素的八个像素组成。当血管均匀分布在整个肿瘤表面时，血管散射较高。相比之下，当血管的分布局限于肿瘤周围的一个狭窄范围/区域时，血管散射较低。

The association of the 14 radiomic features with PFS was assessed using univariable Cox regression models, and

P

values were adjusted for multiple testing correction by the FDR using the BenjaminiHochberg procedure. For feature selection, we applied an elastic net
我们使用单变量 Cox 回归模型评估了 14 个放射学特征与 PFS 的关系，并使用 BenjaminiHochberg 程序根据 FDR 调整了

P

值，以进行多重检验校正。对于特征选择，我们采用了弹性网
model with the mixing parameter set to 0.5 and selected the optimal features at 1 SE beyond the minimum partial likelihood deviance to alleviate overfitting. Finally, we constructed a radiomic model by performing Cox regression analysis with ridge regularization using the RTOG cohort. The regularization strength parameter was estimated using five-fold cross-validation in which the number of PFS events was balanced in each fold. In the final radiomic score, two features (margin blurriness and vessel scattering) were included. The final radiomic score was defined as

1.75 \times

margin blurriness

+ 1.67 \times

vessel scattering. ROC analysis was performed using the “survivalROC” R package (108) at 2 years to define the optimal cutoff value for stratifying patients as high versus low risk for progression or death. The radiomic model and cutoff value were locked prior to independent validation in the MDACC cohort.
在混合参数设置为 0.5 的模型中，我们选择了超出最小偏似然偏差 1 SE 的最佳特征，以减轻过度拟合。最后，我们利用 RTOG 队列通过脊正则化进行 Cox 回归分析，构建了放射组学模型。正则化强度参数是通过五倍交叉验证估算的，其中每倍的 PFS 事件数量是平衡的。最终放射学评分包括两个特征（边缘模糊和血管散乱）。最终放射学评分定义为

1.75 \times

边缘模糊

+ 1.67 \times

血管散乱。使用 "survivalROC "R 软件包（108）进行了 2 年的 ROC 分析，以确定将患者分层为进展或死亡高风险和低风险的最佳临界值。在 MDACC 队列中进行独立验证之前，已锁定放射学模型和临界值。

We assessed the reproducibility of the margin blurriness and vessel scattering radiomics features using an independent set of tumor contours generated using a deep learning model for automated lung tumor segmentation (109). We recomputed the radiomic features using the new segmentation and compared them with those based on the original GTV from the RTOG 0235 dataset. We evaluated reproducibility of image features using intraclass correlation coefficient, which varies between 0 and 1 , with 1 indicating perfect agreement in measurements. According to the Cicchetti criteria (110), intraclass correlation coefficient values above 0.85 are considered excellent.
我们使用一组独立的肿瘤轮廓图评估了边缘模糊度和血管散射放射组学特征的可重复性，这组轮廓图是使用深度学习模型生成的，用于自动肺部肿瘤分割（109）。我们使用新的分割方法重新计算了放射组学特征，并与基于 RTOG 0235 数据集原始 GTV 的特征进行了比较。我们使用类内相关系数（intraclass correlation coefficient）评估了图像特征的再现性，类内相关系数介于 0 和 1 之间，1 表示测量结果完全一致。根据 Cicchetti 标准（110），类内相关系数值超过 0.85 即为优秀。

Identification of Pretreatment Prognostic Factors
确定治疗前的预后因素

Sex, age as a continuous variable, stage III versus stage II, largest lesion MTV as a continuous variable, largest lesion GTV as a continuous variable, and non-SCC versus SCC histology were evaluated as pre-CRT biological prognostic factors for PFS in locoregionally advanced NSCLC using univariable Cox proportional hazards models in the historic training cohort of patients from TCGA (

n = 70

) and a previously reported cohort of patients from Stanford University

(58, 95)

who did not undergo mid-CRT ctDNA testing as described above

(n = 38)

. GTV and MTV data were not available for patients from TCGA. To identify molecular prognostic factors in locoregionally advanced NSCLC, we considered lung cancer driver genes mutated in at least

5 %

of patients with either lung SCC or lung adenocarcinoma (61). Because three different sequencing panels were applied in the historic training cohort, we further limited our analysis to genes with at least three patients having pathogenic mutations (

n = 23

genes). Mutations in the oncogenes KRAS, EGFR, BRAF, and PIK3CA were considered pathogenic if at least one other patient with cancer had a mutation in the same codon in COSMIC v85. Mutations in the Neh2 domain (amino acids 16-86) of NFE2L2 were considered pathogenic if they had Combined Annotation-Dependent Depletion PHRED scores

\geq 20

(111). All other mutations were considered pathogenic if their Combined Annotation-Dependent Depletion PHRED score was

\geq 20

. The association of pathogenic mutations in each gene was associated with PFS using univariate Cox proportional hazards models, and

P

values were corrected for multiple hypothesis testing.
性别、年龄为连续变量、III 期与 II 期、最大病灶 MTV 为连续变量、最大病灶 GTV 为连续变量、在 TCGA（

n = 70

）患者的历史训练队列和斯坦福大学

(58, 95)

先前报道的未进行 CRT 中期 ctDNA 检测的患者队列中，使用单变量 Cox 比例危险模型评估了非 SCC 与 SCC 组织学作为局部区域晚期 NSCLC PFS 的 CRT 前生物学预后因素

(n = 38)

。TCGA 未提供患者的 GTV 和 MTV 数据。为了确定局部晚期 NSCLC 的分子预后因素，我们考虑了至少在

5 %

肺 SCC 或肺腺癌患者中发生突变的肺癌驱动基因（61）。由于在历史训练队列中应用了三种不同的测序面板，我们进一步将分析范围限制在至少有三名患者发生致病基因突变的基因（

n = 23

基因）。如果至少有一名癌症患者在 COSMIC v85 中的相同密码子中发生突变，那么致癌基因 KRAS、EGFR、BRAF 和 PIK3CA 中的突变即被视为致病基因。NFE2L2 的 Neh2 结构域（氨基酸 16-86）中的突变如果具有联合注释依赖性删除 PHRED 评分

\geq 20

（111），则被视为致病性突变。如果其他所有突变的联合注释依赖性删除 PHRED 评分为

\geq 20

，则被认为是致病性的。使用单变量 Cox 比例危险模型对每个基因的致病突变与 PFS 的相关性进行分析，并对

P

值进行多重假设检验校正。

CIRI-LCRT

We constructed CIRI-LCRT to generate personalized survival functions pre-CRT and mid-CRT using Bayesian Cox proportional hazard modeling as described previously (66). In this approach, the baseline survival function is defined across the full training cohort before considering individual risk factors, and Cox coefficients for each covariate are defined in the subset of the training cohort where prior knowledge is available. The mean and variance for each individual risk factor are estimated from the uncertainty around the prior survival curves based on the number of patients at risk at each time point using the Greenwood formula. We defined a baseline survival function by combining all of the patients with locoregionally
我们构建了 CIRI-LCRT，利用贝叶斯 Cox 比例危险模型生成个性化的 CT 前和 CT 中期生存功能，如前所述（66）。在这种方法中，在考虑个体风险因素之前，先在整个训练队列中定义基线生存功能，然后在训练队列的子集中定义每个协变量的 Cox 系数，这些子集中的先验知识是可用的。每个单个风险因素的均值和方差是根据每个时间点的风险患者人数，利用格林伍德公式从先验生存曲线周围的不确定性中估算出来的。我们定义了一个基线生存率函数，将所有局部区域性癌症患者的基线生存率结合在一起。
advanced NSCLC from the RTOG 0235, historic, and ctDNA training cohorts (

n = 357

patients). Because of the strong correlation between GTV and MTV (Supplementary Fig. S6A) and ease of measuring GTV in the clinic (because unlike MTV, it is already being defined for every patient receiving CRT), we chose to evaluate GTV in the model. Survival curves and CIs from the historic training cohort (TCGA and Stanford University patients without ctDNA analysis) were used as prior knowledge to infer the hyperparameters (mean

μ

and variance

σ^{2}

of each coefficient

β

in the model) for histology, largest lesion GTV, KEAP1 status, and KRAS status. The hyperarameters for the radiomic model were inferred from the RTOG 0235 dataset, and the hyperparameters for mid-CRT ctDNA concentration and log fold change in ctDNA concentration were inferred from the ctDNA training cohort (MDACC patients). Having defined the hyperparameters for each covariate, Cox partial likelihood was used as the likelihood function, and Markov chain Monte Carlo sampling was used to calculate the individual survival curves based on the data available pre-CRT or mid-CRT for each patient. We evaluated CIRI models including all possible combinations of histology, GTV, KEAP1 status, KRAS status, radiomic model, mid-CRT ctDNA concentration, and log fold change in ctDNA concentration in the ctDNA training cohort. The model among the top 10 in the training cohort with the fewest features (histology, radiomics model, and mid-CRT ctDNA) was used for the final CIRI-LCRT model, which was applied to the validation cohort (Stanford University ctDNA patients). To assess model calibration, predictions were divided into groups based on CIRI-LCRT estimated risk, and the observed risk of progression or death calculated by the Kaplan-Meier method was plotted versus the CIRI-LCRT-predicted risk. Linear regression was performed for the observed versus predicted risk, and the slope and intercept were calculated with

95 %

CIs. Perfect calibration would be represented by a slope of 1 and an intercept of 0 . Overfitting was minimized by considering the uncertainty around prior survival curves, maximizing the cohort size when inferring each hyperparameter, choosing the model with fewest features, and validating the model in an independent cohort.
这些患者来自 RTOG 0235、历史性和 ctDNA 培训队列中的晚期 NSCLC（

n = 357

患者）。由于 GTV 与 MTV 之间存在很强的相关性（补充图 S6A），而且在临床上测量 GTV 很容易（因为与 MTV 不同，每一位接受 CRT 治疗的患者都已经定义了 GTV），因此我们选择在模型中评估 GTV。我们将历史训练队列（TCGA 和斯坦福大学未进行 ctDNA 分析的患者）的生存曲线和 CI 作为先验知识，推断组织学、最大病灶 GTV、KEAP1 状态和 KRAS 状态的超参数（模型中各系数

β

的均值

μ

和方差

σ^{2}

）。放射学模型的超参数由 RTOG 0235 数据集推断，CRT 中期 ctDNA 浓度和 ctDNA 浓度对折变化的超参数由 ctDNA 培训队列（MDACC 患者）推断。在定义了每个协变量的超参数后，我们使用 Cox 部分似然法作为似然函数，并使用马尔科夫链蒙特卡洛抽样法，根据每位患者 CT 前或 CT 中期的可用数据计算个体生存曲线。我们评估了 CIRI 模型，包括组织学、GTV、KEAP1 状态、KRAS 状态、放射学模型、CRT 中期 ctDNA 浓度以及 ctDNA 培训队列中 ctDNA 浓度的对数折叠变化的所有可能组合。最终的 CIRI-LCRT 模型采用了训练队列中特征（组织学、放射学模型和 CT 中期 ctDNA）最少的前 10 个模型，并将其应用于验证队列（斯坦福大学 ctDNA 患者）。为评估模型校准情况，根据 CIRI-LCRT 估计风险将预测结果分为若干组，并将 Kaplan-Meier 法计算出的观察到的进展或死亡风险与 CIRI-LCRT 预测风险进行对比。对观察到的风险与预测的风险进行线性回归，计算出斜率和截距以及

95 %

CIs。斜率为 1、截距为 0 则表示完全校准。考虑到先前生存曲线的不确定性，在推断每个超参数时最大化队列规模，选择特征最少的模型，并在独立队列中验证模型，从而将过拟合降到最低。

Statistics 统计资料

PFS was defined as the time from the start of CRT to the date of any progression or death. PFS was calculated using the Kaplan-Meier method, censoring patients without progression or death at the time of last imaging follow-up. Statistical significance for Kaplan-Meier analyses was determined using two-sided log-rank tests when comparing two groups or two-sided log-rank tests for trend when comparing multiple CIRI-LCRT groups with increasing predicted risk of progression or death. Univariable and multivariable Cox proportional hazards models were fit with the “coxph” function from the “survival” R package, and the significance of individual variables was assessed using two-sided Wald tests. All HRs were calculated using Cox regression. ctDNA concentrations were log-transformed for regression analyses to produce normally distributed data. All variables were standardized to enable comparison of HRs and 95% CIs from Cox models. The optimal cutoff for stratifying PFS by midCRT ctDNA concentration was defined in the training cohort using the “surv_cutpoint” function from the “survminor” R package and applied to the validation cohort. Two-sided Mann-Whitney U-tests were used to compare distributions, and two-sided Fisher exact tests were used to compare proportions. Paired distributions were compared using two-sided Wilcoxon matched-pair signed-rank tests. Correlation between variables was assessed using the Pearson correlation coefficient. C-statistics for individual risk factors and CIRI models were calculated using the “survivalROC” R package (108) with CIs and empiric

P

values performed from 2,000 bootstrap resamplings.

P

values were corrected for multiple hypothesis testing using the Benjamini-Hochberg procedure with the “stats” R package.
PFS 定义为从开始 CRT 到出现任何进展或死亡的时间。PFS 采用 Kaplan-Meier 法计算，剔除最后一次影像学随访时无进展或死亡的患者。在比较两组患者时，采用双侧对数秩检验确定 Kaplan-Meier 分析的统计学意义；在比较多个 CIRI-LCRT 组患者时，采用双侧对数秩检验确定趋势，预测进展或死亡风险不断增加。使用 "survival "R 软件包中的 "coxph "函数拟合单变量和多变量 Cox 比例危险模型，并使用双侧 Wald 检验评估单个变量的显著性。ctDNA 浓度在回归分析中进行了对数变换，以产生正态分布数据。所有变量均进行了标准化处理，以便比较来自 Cox 模型的 HRs 和 95% CIs。使用 "survminor "R 软件包中的 "surv_cutpoint "函数在训练队列中定义了根据 CRT 中期 ctDNA 浓度对 PFS 进行分层的最佳切点，并将其应用于验证队列。双侧曼-惠特尼 U 检验用于比较分布，双侧费雪精确检验用于比较比例。使用双侧 Wilcoxon 配对符号秩检验比较配对分布。变量之间的相关性采用皮尔逊相关系数进行评估。使用 "survivalROC "R 软件包（108）计算单个风险因素和 CIRI 模型的 C 统计量，CI 和经验

P

值由 2000 次引导重采样得出。

P

值使用 "stats "R 软件包中的 Benjamini-Hochberg 程序进行多重假设检验校正。

Statistical significance was assumed at

P < 0.05

. Statistical analyses were performed with Prism 8 (GraphPad Software, RRID: SCR_002798) or R version 3.6.2 through the RStudio environment (RRID: SCR_000432).
统计显著性假定为

P < 0.05

。统计分析使用 Prism 8（GraphPad Software，RRID：SCR_002798）或通过 RStudio 环境（RRID：SCR_000432）使用 R 3.6.2 版进行。

Data Availability 数据可用性

Anonymized clinical and demographic data for the patients in this study as well as cfDNA metrics, somatic mutation data, radiomic, biological, and molecular metrics, and CIRI-LCRT-predicted PFS are provided in the Supplementary Tables. DNA sequencing data generated in this study from patients enrolled at MDACC are publicly available in the Database of Genotypes and Phenotypes at phs003947.v1.p1. Due to restrictions related to dissemination of germline sequence information included in the informed consent forms used to enroll study subjects at Stanford University, we are unable to provide access to raw sequencing data. Reasonable requests for additional data will be reviewed by the senior authors to determine whether they can be fulfilled in accordance with these privacy restrictions. Sequencing data were processed using a custom bioinformatics pipeline available at http://cappseq.stanford.edu. The SNV score was calculated using custom code from the SNV model from the Lung Cancer Likelihood in Plasma framework available at http://clip. stanford.edu. The code used to calculate the radiomic score is available at https://github.com/lilab-stanford/lung-radiomics.
本研究中患者的匿名临床和人口统计学数据、cfDNA 指标、体细胞突变数据、放射学、生物学和分子指标以及 CIRI-LCRT 预测的 PFS 见补充表格。本研究中 MDACC 入组患者的 DNA 测序数据可在 phs003947.v1.p1 的基因型和表型数据库中公开获取。由于斯坦福大学用于研究对象入组的知情同意书中包含种系序列信息的传播限制，我们无法提供原始测序数据。对额外数据的合理要求将由资深作者进行审查，以确定是否能根据这些隐私限制予以满足。测序数据使用 http://cappseq.stanford.edu 上的定制生物信息学管道进行处理。SNV 得分是使用血浆中肺癌可能性框架中 SNV 模型的定制代码计算得出的，该框架可在 http://clip. stanford.edu 获取。用于计算辐射组学得分的代码可从 https://github.com/lilab-stanford/lung-radiomics 网站获取。

Authors' Disclosures 作者披露

E.J. Moding reports grants from the American Society for Radiation Oncology, the Radiological Society of North America, and Conquer Cancer supported by the GO2 Foundation for Lung Cancer during the conduct of the study, as well as personal fees from Guidepoint and GLG outside the submitted work. B.Y. Nabet reports other support from Genentech/Roche outside the submitted work. J.J. Chabon reports personal fees and other support from Foresight Diagnostics outside the submitted work; in addition, J.J. Chabon has patent filings pending, issued, licensed, and with royalties paid from Yes and reports employment and equity interest in Foresight Diagnostics. D.M. Kurtz reports personal fees and other support from Foresight Diagnostics outside the submitted work; in addition, D.M. Kurtz has patents pertaining to detection of cfDNA pending, issued, licensed, and with royalties paid from Foresight Diagnostics; Dr. D.M. Kurtz is a founder and holds equity in Foresight Diagnostics. E.G. Hamilton reports personal fees from Foresight Diagnostics outside the submitted work; in addition, E.G. Hamilton has a patent for “Systems and methods for cell-free nucleic acid methylation assessment” pending. A.A. Chaudhuri reports personal fees and nonfinancial support from Roche, grants, personal fees, and nonfinancial support from Tempus, nonfinancial support from Guardant Health and Caris, other support from Geneoscopy and LiquidCell Dx, and personal fees and nonfinancial support from Illumina, personal fees from Myriad Genetics, Invitae, Daiichi Sankyo, AstraZeneca, AlphaSights, DeciBio, and Guidepoint outside the submitted work; in addition, A.A. Chaudhuri has a patent for filings related to cancer biomarkers pending, issued, licensed, and with royalties paid. M. Das reports grants from Merck, Genentech, CellSight, and Varian, personal fees from Regeneron, Sanofi/Genzyme, Bristol Myer Squibb, Janssen, Jazz Pharmaceuticals, Catalyst Pharmaceuticals, Guardant, Natera, OncoHost, Merus, Summit Pharmaceuticals, Novocure, Gilead, AstraZeneca, EMD Sereno, Abbvie, and Daiichi Sankyo, and grants and personal fees from Novartis outside the submitted work. K.J. Ramchandran reports grants from Varian outside the submitted work. S.K. Padda reports personal fees from Regeneron, Bayer, AstraZeneca, Eli Lily, Bristol Myers Squibb, Takeda Pharma, Summit Therapeutics, Sanofi Genzyme, Amgen, Mirati, Janssen (J&J), Rayze Biotech, Jazz Pharma, Nanobiotix, and Genentech outside the
E.J. Moding 报告在研究期间获得了美国放射肿瘤学会（American Society for Radiation Oncology）、北美放射学会（Radiological Society of North America）和 GO2 肺癌基金会（GO2 Foundation for Lung Cancer）支持的 Conquer Cancer 的资助，以及 Guidepoint 和美国格理集团（GLG）在所提交工作之外的个人酬金。B.Y. Nabet 报告在提交的工作之外还获得了基因泰克/罗氏公司的其他资助。此外，J.J. Chabon 的专利申请正在等待批准、已获批准、已获许可，并已从中获得版税，他还报告了在 Foresight Diagnostics 的工作和股权。此外，D.M. Kurtz 博士还拥有与 cfDNA 检测相关的专利申请，这些专利申请正在等待批准、已获授权，并由 Foresight Diagnostics 公司支付专利使用费；D.M. Kurtz 博士是 Foresight Diagnostics 公司的创始人，并持有该公司的股份。此外，E.G. Hamilton 正在申请 "无细胞核酸甲基化评估系统和方法 "的专利。A.A. Chaudhuri 报告了罗氏公司的个人酬金和非财务支持，Tempus 公司的赠款、个人酬金和非财务支持，Guardant Health 公司和 Caris 公司的非财务支持，Geneoscopy 公司和 LiquidCell Dx 公司的其他支持，Illumina 公司的个人酬金和非财务支持，Myriad Genetics 公司、Invitae 公司、第一三共公司、阿斯利康公司、AlphaSights 公司、DeciBio 公司和 Guidepoint 公司的个人酬金。A. Chaudhuri 有一项与癌症生物标志物有关的专利申请，该专利已申请、颁发、许可，并已支付版税。M. Das 报告获得了默克公司、基因泰克公司、CellSight 公司和瓦里安公司的资助，并从 Regeneron 公司、赛诺菲/Genzyme 公司、Bristol Myer Squibb 公司、杨森公司、Jazz Pharmaceuticals 公司、Catalyst Pharmaceuticals 公司、Guardant 公司、Natera 公司、OncoHost 公司、Merus 公司、Summit Pharmaceuticals 公司、Novocure 公司、吉利德公司、阿斯利康公司、EMD Sereno 公司、艾伯维公司和第一三共公司领取了个人酬金，此外还从诺华公司领取了资助和个人酬金。K.J. Ramchandran 报告了所提交工作之外的瓦里安公司的资助。S.K.Padda 报告的个人酬金来自 Regeneron、Bayer、AstraZeneca、Eli Lily、Bristol Myers Squibb、Takeda Pharma、Summit Therapeutics、Sanofi Genzyme、Amgen、Mirati、Janssen (J&J)、Rayze Biotech、Jazz Pharma、Nanobiotix 和 Genentech。
submitted work. J.W. Neal reports grants and personal fees from Genentech/Roche, Novartis, and Exelixis, personal fees from AstraZeneca, Takeda Pharmaceuticals, Eli Lilly and Company, Amgen, Iovance Biotherapeutics, Blueprint Pharmaceuticals, Regeneron Pharmaceuticals, Natera, Sanofi/Regeneron, D2G Oncology, Surface Oncology, Turning Point Therapeutics, Mirati Therapeutics, Gilead Sciences, Abbvie, Summit Therapeutics, Novocure, Janssen Oncology, Anheart Therapeutics, Bristol Myers Squibb, Daiichi Sankyo/ AstraZeneca, and Nuvation Bio, and grants from Revolution Medicines, Nuvalent, Inc., Merck, Boehringer Ingelheim, Nektar Therapeutics, Takeda Pharmaceuticals, Adaptimmune, GSK, and Janssen outside the submitted work. H.A. Wakelee reports grants from the NIH during the conduct of the study, as well as grants from AstraZeneca, Bayer, Bristol Myers Squibb, Genentech/Roche, Helsinn, Merck, SeaGen, Xcovery, and Gilead (via an IIT mechanism), personal fees from Chugai, OncoC4, IOBiotech, Mirati, Beigene, and GSK, and other support from Bristol Myers Squibb, Genentech/Roche, Merck, and AstraZeneca outside the submitted work. M.F. Gensheimer reports grants from Varian Medical Systems and XRad Therapeutics outside the submitted work and stock ownership in Amgen. R. Li reports grants from the NIH during the conduct of the study. A.A. Alizadeh reports other support from Foresight, Capstan, and CiberMed during the conduct of the study; in addition, A.A. Alizadeh has a patent for CIRI US20220392605A1 pending and licensed. M. Diehn reports grants and personal fees from AstraZeneca, other support from Gritstone Bio, personal fees and nonfinancial support from Regeneron, personal fees from Bristol Myers Squibb, nonfinancial support and other support from Foresight Diagnostics, and other support from CiberMed and Perception Medicine outside the submitted work; in addition, M. Diehn has a patent for ctDNA methods issued, licensed, and with royalties paid from Roche and a patent for ctDNA methods issued, licensed, and with royalties paid from Foresight Diagnostics and Invited Faculty Member at Hokkaido University. No disclosures were reported by the other authors.
已提交作品。J.W.Neal 报告了来自 Genentech/Roche、Novartis 和 Exelixis 的资助和个人酬金，以及来自 AstraZeneca、Takeda Pharmaceuticals、Eli Lilly and Company、Amgen、Iovance Biotherapeutics、Blueprint Pharmaceuticals、Regeneron Pharmaceuticals、Natera、Sanofi/Regeneron、D2G Oncology、Surface Oncology、Turning Point Therapeutics、Mirati Therapeutics、Gilead Sciences、Abbvie、Summit Therapeutics、Novocure 的个人酬金、Surface Oncology、Turning Point Therapeutics、Mirati Therapeutics、Gilead Sciences、Abbvie、Summit Therapeutics、Novocure、Janssen Oncology、Anheart Therapeutics、Bristol Myers Squibb、Daiichi Sankyo/ AstraZeneca 和 Nuvation Bio，以及 Revolution Medicines、Nuvalent, Inc.,默克（Merck）、勃林格殷格翰（Boehringer Ingelheim）、Nektar Therapeutics、武田制药（Takeda Pharmaceuticals）、Adaptimmune、葛兰素史克（GSK）和杨森（Janssen）的资助。H.A.Wakelee 报告在研究期间获得了美国国立卫生研究院的资助，以及阿斯利康、拜耳、百时美施贵宝、基因泰克/罗氏、Helsinn、默克、SeaGen、Xcovery 和吉利德（通过 IIT 机制）的资助，中外制药、OncoC4、IOBiotech、Mirati、Beigene 和葛兰素史克的个人酬金，以及百时美施贵宝、基因泰克/罗氏、默克和阿斯利康在所提交工作之外的其他支持。M.F. Gensheimer 在所提交的工作之外还报告了来自瓦里安医疗系统公司和 XRad Therapeutics 公司的资助，以及在安进公司的股票所有权。R. Li 报告在研究期间获得了美国国立卫生研究院（NIH）的资助。此外，A.A. Alizadeh 的 CIRI US20220392605A1 专利正在申请和授权中。M. Diehn 报告了阿斯利康的资助和个人酬金、Gritstone Bio 的其他资助、Regeneron 的个人酬金和非财务资助、Bristol Myers Squibb 的个人酬金、Foresight Diagnostics 的非财务资助和其他资助，以及 CiberMed 和 Perception Medicine 的其他资助；此外，M.Diehn 拥有罗氏公司颁发、许可和支付版税的 ctDNA 方法专利，以及 Foresight Diagnostics 公司颁发、许可和支付版税的 ctDNA 方法专利，并且是北海道大学的特邀教员。其他作者未披露任何信息。

Authors' Contributions 作者的贡献

E.J. Moding: Conceptualization, data curation, software, formal analysis, funding acquisition, validation, investigation, visualization, methodology, writing-original draft, writing-review and editing. M. Shahrokh Esfahani: Conceptualization, data curation, software, formal analysis, validation, investigation, methodology, writingoriginal draft, writing-review and editing. C. Jin: Data curation, software, formal analysis, validation, investigation, visualization, writing-review and editing. A.B. Hui: Data curation, validation, investigation, writing-review and editing. B.Y. Nabet: Data curation, investigation, visualization, writing-review and editing. Y. Liu: Data curation, investigation, writing-review and editing. J.J. Chabon: Data curation, software, formal analysis, methodology, writingreview and editing. M.S. Binkley: Data curation, formal analysis, investigation, methodology, writing-review and editing. D.M. Kurtz: Software, methodology, writing-review and editing. E.G. Hamilton: Data curation, software, methodology, writing-review and editing. A.A. Chaudhuri: Conceptualization, data curation, investigation, writing-review and editing. C.L. Liu: Resources, software, project administration, writing-review and editing. Z. Li: Formal analysis, validation, investigation, writing-review and editing. R.F. Bonilla: Resources, data curation, writing-review and editing. A.L. Jiang: Resources, data curation, writing-review and editing. B.C. Lau: Resources, data curation, writing-review and editing. P. Lopez: Resources, data curation, writing-review and editing. J. He: Resources, data curation, writing-review and editing. Y. Qiao: Resources, data curation, writing-review and editing. T. Xu: Resources, data curation, writing-review and editing. L. Yao: Resources, data curation, writing-review and editing. S. Gandhi: Resources, writingreview and editing. Z. Liao: Resources, writing-review and editing.
E.J. Moding：概念化、数据整理、软件、形式分析、资金获取、验证、调查、可视化、方法论、写作-原稿、写作-审阅和编辑。M. Shahrokh Esfahani：概念化、数据整理、软件、形式分析、验证、调查、方法论、撰写原稿、撰写-审阅和编辑。C. Jin：数据整理、软件、形式分析、验证、调查、可视化、撰写-审阅和编辑。A.B. Hui：数据整理、验证、调查、撰写-审阅和编辑。B.Y. Nabet：数据整理、调查、可视化、撰写-审阅和编辑。Y. Liu：数据整理、调查、撰写-审阅和编辑。J.J. Chabon：数据整理、软件、形式分析、方法论、撰写-审阅和编辑。M.S. Binkley：数据整理、形式分析、调查、方法论、写作-审查和编辑。D.M. Kurtz：软件、方法论、写作-审查和编辑。E.G. Hamilton：数据整理、软件、方法论、撰写-审阅和编辑。A.A. Chaudhuri：概念化、数据整理、调查、撰写-审阅和编辑。C.L. Liu：资源、软件、项目管理、撰写-审阅和编辑。Z. Li：形式分析、验证、调查、撰写-审阅和编辑。R.F. Bonilla：资源、数据整理、撰写-审阅和编辑。A.L. Jiang：资源、数据整理、撰写-审阅和编辑。B.C. Lau：资源、数据整理、撰写-审阅和编辑。P. Lopez：资源、数据整理、撰写-审阅和编辑。J. He：资源、数据整理、撰写-审阅和编辑。Y.Qiao：资源、数据整理、撰写-审阅和编辑。T. Xu：资源、数据整理、撰写-审阅和编辑。L. 姚：资源、数据整理、写作-审阅和编辑。S. Gandhi：资源、撰写-审阅和编辑。Z. Liao：资源、撰写-审阅和编辑。
M. Das: Resources, writing-review and editing. K.J. Ramchandran: Resources, writing-review and editing. S.K. Padda: Resources, writing-review and editing. J.W. Neal: Resources, writing-review and editing. H.A. Wakelee: Resources, writing-review and editing. M.F. Gensheimer: Resources, writing-review and editing. B.W. Loo: Resources, funding acquisition, writing-review and editing. R. Li: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing. S.H. Lin: Conceptualization, resources, data curation, supervision, validation, investigation, methodology, writing-original draft, project administration, writing-review and editing. A.A. Alizadeh: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing. M. Diehn: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing-original draft, project administration, writing-review and editing.
M. Das：资源、写作-审阅和编辑。K.J. Ramchandran：资源、撰写-审阅和编辑。S.K. Padda：资源、撰写-审阅和编辑。J.W. Neal：资源、撰写-审阅和编辑。H.A. Wakelee：资源、写作审阅和编辑。M.F. Gensheimer：资源、撰写-审阅和编辑。B.W. Loo：资源、资金获取、撰写-审阅和编辑。R. Li：构思、资源、数据整理、软件、形式分析、监督、资金获取、验证、调查、可视化、方法论、撰写-原稿、项目管理、撰写-审阅和编辑。S.H. Lin：概念化、资源、数据整理、监督、验证、调查、方法论、撰写-原稿、项目管理、撰写-审阅和编辑。A.A. Alizadeh：概念化、资源、数据整理、软件、形式分析、监督、资金获取、验证、调查、可视化、方法论、撰写原稿、项目管理、撰写-审阅和编辑。M. Diehn：概念化、资源、数据整理、软件、形式分析、监督、资金获取、验证、调查、可视化、方法论、撰写原稿、项目管理、撰写-审阅和编辑。

Acknowledgments 致谢

We thank the patients and families who participated in this study. This work was supported by grants from the American Society for Radiation Oncology (E.J. Moding), the Radiological Society of North America (E.J. Moding), Conquer Cancer supported by GO2 Foundation for Lung Cancer (E.J. Moding), the NCI (M. Diehn and A.A. Alizadeh: R01CA188298, R01CA244526, and R01CA254179; R. Li, M. Diehn, and B.W. Loo: R01CA233578), the Virginia and D.K. Ludwig Fund for Cancer Research (M. Diehn and A.A. Alizadeh), the Bakewell Foundation (M. Diehn and A.A. Alizadeh), the SDW/DT and Shanahan Family Foundations (A.A. Alizadeh), and the CRK Faculty Scholar Fund (M. Diehn). A.A. Alizadeh is a Scholar of The Leukemia & Lymphoma Society. Any opinions, findings, and conclusions expressed in this material are those of the author(s) and do not necessarily reflect those of the American Society of Clinical Oncology, Conquer Cancer, or GO2 Foundation for Lung Cancer.
我们感谢参与这项研究的患者和家属。这项工作得到了美国放射肿瘤学会（E.J. Moding）、北美放射学会（E.J. Moding）、GO2 肺癌基金会支持的征服癌症（E.J. Moding）、美国国家癌症研究所（NCI）（M. Diehn 和 A.A. Alizadeh：R01CA188298, R01CA244526, and R01CA254179; R. Li, M. Diehn, and B.W. Loo：R01CA233578), Virginia and D.K. Ludwig Fund for Cancer Research (M. Diehn and A.A. Alizadeh), the Bakewell Foundation (M. Diehn and A.A. Alizadeh), the SDW/DT and Shanahan Family Foundation (A.A. Alizadeh), and the CRK Faculty Scholar Fund (M. Diehn).A.A. Alizadeh 是白血病和淋巴瘤协会的学者。本资料中表达的任何观点、发现和结论均为作者个人观点、发现和结论，不代表美国临床肿瘤学会、征服癌症组织或 GO2 肺癌基金会的观点、发现和结论。

Note 备注

Supplementary data for this article are available at Cancer Discovery Online (http://cancerdiscovery.aacrjournals.org/).
本文的补充数据可在 Cancer Discovery Online ( http://cancerdiscovery.aacrjournals.org/) 上查阅。

Received December 3, 2024; revised February 25, 2025; accepted April 16, 2025; posted first April 29, 2025.
2024 年 12 月 3 日收到；2025 年 2 月 25 日修订；2025 年 4 月 16 日接受；2025 年 4 月 29 日首次发布。

REFERENCES 参考文献

Siegel RL, Miller KD, Jemal A. Cancer statistics, 2019. CA Cancer J Clin 2019;69:7-34.
Siegel RL, Miller KD, Jemal A. 2019 年癌症统计数据。CA Cancer J Clin 2019;69:7-34.
Ettinger DS, Wood DE, Aisner DL, Akerley W, Bauman J, Chirieac LR, et al. Non-small cell lung cancer, version 5.2017, NCCN clinical practice guidelines in oncology. J Natl Compr Canc Netw 2017;15: 504-35.
Ettinger DS, Wood DE, Aisner DL, Akerley W, Bauman J, Chirieac LR, et al. 非小细胞肺癌，5.2017 版，NCCN 肿瘤学临床实践指南。J Natl Compr Canc Netw 2017;15: 504-35.
Antonia SJ, Villegas A, Daniel D, Vicente D, Murakami S, Hui R, et al. Overall survival with durvalumab after chemoradiotherapy in stage III NSCLC. N Engl J Med 2018;379:2342-50.
Antonia SJ, Villegas A, Daniel D, Vicente D, Murakami S, Hui R, et al. III 期 NSCLC 化疗后使用 durvalumab 的总生存率。N Engl J Med 2018;379:2342-50.
Perez CA, Stanley K, Rubin P, Kramer S, Brady L, Perez-Tamayo R, et al. A prospective randomized study of various irradiation doses and fractionation schedules in the treatment of inoperable non-oat-cell carcinoma of the lung. Preliminary report by the Radiation Therapy Oncology Group. Cancer 1980;45:2744-53.
Perez CA、Stanley K、Rubin P、Kramer S、Brady L、Perez-Tamayo R 等. 在治疗无法手术的肺部非扁平细胞癌时采用不同照射剂量和分次照射计划的前瞻性随机研究。放射治疗肿瘤学小组的初步报告。癌症 1980；45：2744-53。
Brundage MD, Davies D, Mackillop WJ. Prognostic factors in non-small cell lung cancer: a decade of progress. Chest 2002;122: 1037-57.
Brundage MD, Davies D, Mackillop WJ.非小细胞肺癌的预后因素：十年进展。Chest 2002; 122: 1037-57.
Huber RM, De Ruysscher D, Hoffmann H, Reu S, Tufman A. Interdisciplinary multimodality management of stage III nonsmall cell lung cancer. Eur Respir Rev 2019;28:190024.
Huber RM、De Ruysscher D、Hoffmann H、Reu S、Tufman A. III 期非小细胞肺癌的多学科多模式管理。Eur Respir Rev 2019;28:190024.
Berghmans T, Paesmans M, Sculier J-P. Prognostic factors in stage III non-small cell lung cancer: a review of conventional, metabolic and new biological variables. Ther Adv Med Oncol 2011;3:127-38.
Berghmans T, Paesmans M, Sculier J-P.III 期非小细胞肺癌的预后因素：传统、代谢和新生物变量综述》。Ther Adv Med Oncol 2011; 3:127-38.
Dehing-Oberije C, Aerts H, Yu S, De Ruysscher D, Menheere P, Hilvo M, et al. Development and validation of a prognostic model using blood biomarker information for prediction of survival of non-small-cell lung cancer patients treated with combined chemotherapy and radiation or radiotherapy alone (NCT00181519, NCT00573040, and NCT00572325). Int J Radiat Oncol Biol Phys 2011;81:360-8.
Dehing-Oberije C、Aerts H、Yu S、De Ruysscher D、Menheere P、Hilvo M 等：利用血液生物标志物信息开发和验证预后模型，用于预测接受联合化疗和放疗或单独放疗的非小细胞肺癌患者的生存期（NCT00181519、NCT00573040 和 NCT00572325）。Int J Radiat Oncol Biol Phys 2011;81:360-8.
Dehing-Oberije C, De Ruysscher D, van der Weide H, Hochstenbag M, Bootsma G, Geraedts W, et al. Tumor volume combined with number of positive lymph node stations is a more important prognostic factor than TNM stage for survival of non-small-cell lung cancer patients treated with (chemo)radiotherapy. Int J Radiat Oncol Biol Phys 2008;70:1039-44.
Dehing-Oberije C, De Ruysscher D, van der Weide H, Hochstenbag M, Bootsma G, Geraedts W, et al.与 TNM 分期相比，肿瘤体积和淋巴结阳性数目是非小细胞肺癌放（化）疗患者生存率的更重要预后因素。Int J Radiat Oncol Biol Phys 2008;70:1039-44.
Alexander M, Wolfe R, Ball D, Conron M, Stirling RG, Solomon B, et al. Lung cancer prognostic index: a risk score to predict overall survival after the diagnosis of non-small-cell lung cancer. BrJ Cancer 2017;117:744-51.
Alexander M, Wolfe R, Ball D, Conron M, Stirling RG, Solomon B, et al. 肺癌预后指数：预测非小细胞肺癌诊断后总生存期的风险评分。BrJ Cancer 2017;117:744-51.
Jochems A, Deist TM, El Naqa I, Kessler M, Mayo C, Reeves J, et al. Developing and validating a survival prediction model for NSCLC patients through distributed learning across 3 countries. Int J Radiat Oncol Biol Phys 2017;99:344-52.
Jochems A, Deist TM, El Naqa I, Kessler M, Mayo C, Reeves J, et al.Int J Radiat Oncol Biol Phys 2017;99:344-52.
Jayasurya K, Fung G, Yu S, Dehing-Oberije C, De Ruysscher D, Hope A, et al. Comparison of Bayesian network and support vector machine models for two-year survival prediction in lung cancer patients treated with radiotherapy. Med Phys 2010;37:1401-7.
Jayasurya K、Fung G、Yu S、Dehing-Oberije C、De Ruysscher D、Hope A 等：贝叶斯网络和支持向量机模型在肺癌放疗患者两年生存预测中的比较。医学物理 2010；37：1401-7。
van Laar M, van Amsterdam WAC, van Lindert ASR, de Jong PA, Verhoeff JJC. Prognostic factors for overall survival of stage III nonsmall cell lung cancer patients on computed tomography: a systematic review and meta-analysis. Radiother Oncol 2020;151:152-75.
van Laar M, van Amsterdam WAC, van Lindert ASR, de Jong PA, Verhoeff JJC.计算机断层扫描显示 III 期非小细胞肺癌患者总生存期的预后因素：系统综述与荟萃分析》（Radiother Oncol 2020;151:152-75）。Radiother Oncol 2020;151:152-75.
Wang C, Fang J, Jiang T, Hu S, Wang P, Liu X, et al. Development and validation of a prognostic nomogram model in locally advanced NSCLC based on metabolic features of PET/CT and hematological inflammatory indicators. EJNMMI Phys 2024;11:24.
Wang C, Fang J, Jiang T, Hu S, Wang P, Liu X, et al.基于 PET/CT 代谢特征和血液学炎症指标的局部晚期 NSCLC 预后提名图模型的建立与验证.EJNMMI Phys 2024;11:24.
Lambin P, Leijenaar RTH, Deist TM, Peerlings J, de Jong EEC, van Timmeren J, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017;14:749-62.
Lambin P, Leijenaar RTH, Deist TM, Peerlings J, de Jong EEC, van Timmeren J, et al. Radiomics: the bridge between medical imaging and personalized medicine.Nat Rev Clin Oncol 2017;14:749-62.
Pan F, Feng L, Liu B, Hu Y, Wang Q. Application of radiomics in diagnosis and treatment of lung cancer. Front Pharmacol 2023;14:1295511.
Pan F, Feng L, Liu B, Hu Y, Wang Q. 放射组学在肺癌诊治中的应用。Front Pharmacol 2023;14:1295511.
Shi L, He Y, Yuan Z, Benedict S, Valicenti R, Qiu J, et al. Radiomics for response and outcome assessment for non-small cell lung cancer. Technol Cancer Res Treat 2018;17:1533033818782788.
Shi L, He Y, Yuan Z, Benedict S, Valicenti R, Qiu J, et al. Radiomics for response and outcome assessment for non-small cell lung cancer.Technol Cancer Res Treat 2018;17:1533033818782788.
Aerts HJWL, Velazquez ER, Leijenaar RTH, Parmar C, Grossmann P, Carvalho S, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun 2014;5:4006.
Aerts HJWL, Velazquez ER, Leijenaar RTH, Parmar C, Grossmann P, Carvalho S, et al. 使用定量放射组学方法通过无创成像解码肿瘤表型。Nat Commun 2014; 5:4006.
Coroller TP, Agrawal V, Narayan V, Hou Y, Grossmann P, Lee SW, et al. Radiomic phenotype features predict pathological response in non-small cell lung cancer. Radiother Oncol 2016;119:480-6.
Coroller TP, Agrawal V, Narayan V, Hou Y, Grossmann P, Lee SW, et al. 放射表型特征可预测非小细胞肺癌的病理反应。Radiother Oncol 2016;119:480-6.
Coroller TP, Agrawal V, Huynh E, Narayan V, Lee SW, Mak RH, et al. Radiomic-based pathological response prediction from primary tumors and lymph nodes in NSCLC. J Thorac Oncol 2017;12: 467-76.
Coroller TP, Agrawal V, Huynh E, Narayan V, Lee SW, Mak RH, et al. 基于放射组学的 NSCLC 原发肿瘤和淋巴结病理反应预测。J Thorac Oncol 2017;12: 467-76.
Emaminejad N, Qian W, Guan Y, Tan M, Qiu Y, Liu H, et al. Fusion of quantitative image and genomic biomarkers to improve prognosis assessment of early stage lung cancer patients. IEEE Trans Biomed Eng 2016;63:1034-43.
Emaminejad N, Qian W, Guan Y, Tan M, Qiu Y, Liu H, et al. Fusion of quantitative image and genomic biomarkers to improve prognosis assessment of early stage lung cancer patients.IEEE Trans Biomed Eng 2016;63:1034-43.
Gatta R, Depeursinge A, Ratib O, Michielin O, Leimgruber A. Integrating radiomics into holomics for personalised oncology: from algorithms to bedside. Eur Radiol Exp 2020;4:11.
Gatta R、Depeursinge A、Ratib O、Michielin O、Leimgruber A.将放射组学融入整体组学以实现个性化肿瘤学：从算法到床边。Eur Radiol Exp 2020;4:11.
Fried DV, Tucker SL, Zhou S, Liao Z, Mawlawi O, Ibbott G, et al. Prognostic value and reproducibility of pretreatment CT texture features in stage III non-small cell lung cancer. Int J Radiat Oncol Biol Phys 2014;90:834-42.
Fried DV, Tucker SL, Zhou S, Liao Z, Mawlawi O, Ibbott G, et al. III 期非小细胞肺癌治疗前 CT 纹理特征的预后价值和可重复性。Int J Radiat Oncol Biol Phys 2014;90:834-42.
Lee J, Li B, Cui Y, Sun X, Wu J, Zhu H, et al. A quantitative CT imaging signature predicts survival and complements established prognosticators in stage I non-small cell lung cancer. Int J Radiat Oncol Biol Phys 2018;102:1098-106.
Lee J, Li B, Cui Y, Sun X, Wu J, Zhu H, et al. 定量 CT 成像特征可预测 I 期非小细胞肺癌患者的生存并补充已有的预后指标 Int J Radiat Oncol Biol Phys 2018;102:1098-106.
Wu J, Aguilera T, Shultz D, Gudur M, Rubin DL, Loo BW, et al. Early-stage non-small cell lung cancer: quantitative imaging characteristics of (18)F fluorodeoxyglucose PET/CT allow prediction of distant metastasis. Radiology 2016;281:270-8.
Wu J, Aguilera T, Shultz D, Gudur M, Rubin DL, Loo BW, et al. 早期非小细胞肺癌：(18)F 氟脱氧葡萄糖 PET/CT 的定量成像特征可预测远处转移。放射学》2016;281:270-8。
Gensheimer MF, Hong JC, Chang-Halpenny C, Zhu H, Eclov NCW, To J, et al. Mid-radiotherapy PET/CT for prognostication and detection of early progression in patients with stage III non-small cell lung cancer. Radiother Oncol 2017;125:338-43.
Gensheimer MF, Hong JC, Chang-Halpenny C, Zhu H, Eclov NCW, To J, et al. III 期非小细胞肺癌患者放疗中期 PET/CT 用于预后判断和早期进展检测。Radiother Oncol 2017;125:338-43.
Kong F-MS, Frey KA, Quint LE, Ten Haken RK, Hayman JA, Kessler M, et al. A pilot study of [18F]fluorodeoxyglucose positron emission tomography scans during and after radiation-based therapy in patients with non small-cell lung cancer. J Clin Oncol 2007;25:3116-23.
Kong F-MS、Frey KA、Quint LE、Ten Haken RK、Hayman JA、Kessler M 等：非小细胞肺癌患者放疗期间和放疗后[18F]氟脱氧葡萄糖正电子发射断层扫描试验研究。J Clin Oncol 2007;25:3116-23.
Vera P, Mezzani-Saillard S, Edet-Sanson A, Ménard J-F, Modzelewski R, Thureau S, et al. FDG PET during radiochemotherapy is predictive of outcome at 1 year in non-small-cell lung cancer patients: a prospective multicentre study (RTEP2). Eur J Nucl Med Mol Imaging 2014;41:1057-65.
Vera P、Mezzani-Saillard S、Edet-Sanson A、Ménard J-F、Modzelewski R、Thureau S 等：《放化疗期间的 FDG PET 可预测非小细胞肺癌患者 1 年后的预后：一项前瞻性多中心研究 (RTEP2)》。Eur J Nucl Med Mol Imaging 2014;41:1057-65.
Kocak Z, Evans ES, Zhou S-M, Miller KL, Folz RJ, Shafman TD, et al. Challenges in defining radiation pneumonitis in patients with lung cancer. Int J Radiat Oncol Biol Phys 2005;62:635-8.
Kocak Z、Evans ES、Zhou S-M、Miller KL、Folz RJ、Shafman TD 等：《界定肺癌患者放射性肺炎的挑战》。Int J Radiat Oncol Biol Phys 2005;62:635-8.
Moding EJ, Nabet BY, Alizadeh AA, Diehn M. Detecting liquid remnants of solid tumors: circulating tumor DNA minimal residual disease. Cancer Discov 2021;11:2968-86.
Moding EJ, Nabet BY, Alizadeh AA, Diehn M. 《检测实体肿瘤的液体残留：循环肿瘤 DNA 最小残留病》。Cancer Discov 2021;11:2968-86.
Mok T, Wu Y-L, Lee JS, Yu C-J, Sriuranpong V, Sandoval-Tan J, et al. Detection and dynamic changes of EGFR mutations from circulating tumor DNA as a predictor of survival outcomes in NSCLC patients treated with first-line intercalated erlotinib and chemotherapy. Clin Cancer Res 2015;21:3196-203.
Mok T, Wu Y-L, Lee JS, Yu C-J, Sriuranpong V, Sandoval-Tan J, et al. 循环肿瘤 DNA 中表皮生长因子受体突变的检测和动态变化可预测接受厄洛替尼和化疗一线间期治疗的 NSCLC 患者的生存结果。Clin Cancer Res 2015;21:3196-203.
Sorensen BS, Wu L, Wei W, Tsai J, Weber B, Nexo E, et al. Monitoring of epidermal growth factor receptor tyrosine kinase inhibitorsensitizing and resistance mutations in the plasma DNA of patients with advanced non-small cell lung cancer during treatment with erlotinib. Cancer 2014;120:3896-901.
Sorensen BS、Wu L、Wei W、Tsai J、Weber B、Nexo E 等：在厄洛替尼治疗期间监测晚期非小细胞肺癌患者血浆 DNA 中表皮生长因子受体酪氨酸激酶抑制剂致敏突变和耐药突变。Cancer 2014;120:3896-901.
Raja R, Kuziora M, Brohawn PZ, Higgs BW, Gupta A, Dennis PA, et al. Early reduction in ctDNA predicts survival in patients with lung and bladder cancer treated with durvalumab. Clin Cancer Res 2018;24:6212-22.
Raja R, Kuziora M, Brohawn PZ, Higgs BW, Gupta A, Dennis PA, et al. ctDNA 的早期减少可预测接受度伐单抗治疗的肺癌和膀胱癌患者的生存率。Clin Cancer Res 2018;24:6212-22.
Anagnostou V, Smith KN, Forde PM, Niknafs N, Bhattacharya R, White J, et al. Evolution of neoantigen landscape during immune checkpoint blockade in non-small cell lung cancer. Cancer Discov 2017;7:264-76.
Anagnostou V, Smith KN, Forde PM, Niknafs N, Bhattacharya R, White J, et al. 非小细胞肺癌免疫检查点阻断过程中新抗原景观的演变。Cancer Discov 2017;7:264-76.
Goldberg SB, Narayan A, Kole AJ, Decker RH, Teysir J, Carriero NJ, et al. Early assessment of lung cancer immunotherapy response via circulating tumor DNA. Clin Cancer Res 2018;24:1872-80.
Goldberg SB、Narayan A、Kole AJ、Decker RH、Teysir J、Carriero NJ 等：通过循环肿瘤 DNA 早期评估肺癌免疫治疗反应。Clin Cancer Res 2018;24:1872-80.
Assaf ZJF, Zou W, Fine AD, Socinski MA, Young A, Lipson D, et al. A longitudinal circulating tumor DNA-based model associated with survival in metastatic non-small-cell lung cancer. Nat Med 2023;29:859-68.
Assaf ZJF, Zou W, Fine AD, Socinski MA, Young A, Lipson D, et al.Nat Med 2023;29:859-68.
Nabet BY, Esfahani MS, Moding EJ, Hamilton EG, Chabon JJ, Rizvi H, et al. Noninvasive early identification of therapeutic benefit from immune checkpoint inhibition. Cell 2020;183: 363-76.e13.
Nabet BY、Esfahani MS、Moding EJ、Hamilton EG、Chabon JJ、Rizvi H 等：无创早期识别免疫检查点抑制的治疗效果。Cell 2020;183：363-76.e13.
Wyatt AW, Azad AA, Volik SV, Annala M, Beja K, McConeghy B, et al. Genomic alterations in cell-free DNA and enzalutamide resistance in castration-resistant prostate cancer. JAMA Oncol 2016;2: 1598-606.
Wyatt AW, Azad AA, Volik SV, Annala M, Beja K, McConeghy B, et al. 无细胞 DNA 中的基因组改变与去势抵抗性前列腺癌的恩杂鲁胺耐药性 JAMA Oncol 2016;2: 1598-606.
Tie J, Kinde I, Wang Y, Wong HL, Roebert J, Christie M, et al. Circulating tumor DNA as an early marker of therapeutic response in patients with metastatic colorectal cancer. Ann Oncol 2015;26: 1715-22.
Tie J, Kinde I, Wang Y, Wong HL, Roebert J, Christie M, et al. 循环肿瘤 DNA 作为转移性结直肠癌患者治疗反应的早期标志物。Ann Oncol 2015; 26: 1715-22.
Del Re M, Vivaldi C, Rofi E, Vasile E, Miccoli M, Caparello C, et al. Early changes in plasma DNA levels of mutant KRAS as a sensitive marker of response to chemotherapy in pancreatic cancer. Sci Rep 2017;7:7931.
Del Re M, Vivaldi C, Rofi E, Vasile E, Miccoli M, Caparello C, et al.突变型 KRAS 血浆 DNA 水平的早期变化作为胰腺癌化疗反应的敏感标志物 Sci Rep 2017;7:7931.
O’Leary B, Hrebien S, Morden JP, Beaney M, Fribbens C, Huang X, et al. Early circulating tumor DNA dynamics and clonal selection with palbociclib and fulvestrant for breast cancer. Nat Commun 2018;9:896.
O'Leary B, Hrebien S, Morden JP, Beaney M, Fribbens C, Huang X, et al.使用帕博西尼（palbociclib）和氟维司群治疗乳腺癌的早期循环肿瘤 DNA 动态和克隆选择。Nat Commun 2018; 9:896.
Wang X-Y, Zhang R, Han J-H, Chen S-Q, Zhao F-L, Chen H, et al. Early circulating tumor DNA dynamics predict neoadjuvant therapy response and recurrence in colorectal liver metastases: a prospective study. Ann Surg Oncol 2023;30:5252-63.
Wang X-Y，Zhang R，Han J-H，Chen S-Q，Zhao F-L，Chen H，等. 早期循环肿瘤 DNA 动态预测结直肠肝转移的新辅助治疗反应和复发：一项前瞻性研究。Ann Surg Oncol 2023;30:5252-63.
Parikh AR, Mojtahed A, Schneider JL, Kanter K, Van Seventer EE, Fetter IJ, et al. Serial ctDNA monitoring to predict response to systemic therapy in metastatic gastrointestinal cancers. Clin Cancer Res 2020;26:1877-85.
Parikh AR、Mojtahed A、Schneider JL、Kanter K、Van Seventer EE、Fetter IJ 等：预测转移性胃肠癌对全身治疗反应的连续 ctDNA 监测。Clin Cancer Res 2020;26:1877-85.
Bratman SV, Yang SYC, Iafolla MAJ, Liu Z, Hansen AR, Bedard PL, et al. Personalized circulating tumor DNA analysis as a predictive biomarker in solid tumor patients treated with pembrolizumab. Nat Cancer 2020;1:873-81.
Bratman SV、Yang SYC、Iafolla MAJ、Liu Z、Hansen AR、Bedard PL 等. 个性化循环肿瘤 DNA 分析作为接受 pembrolizumab 治疗的实体瘤患者的预测性生物标记物。Nat Cancer 2020; 1:873-81.
Chaudhuri AA, Chabon JJ, Lovejoy AF, Newman AM, Stehr H, Azad TD, et al. Early detection of molecular residual disease in localized lung cancer by circulating tumor DNA profiling. Cancer Discov 2017;7:1394-403.
Chaudhuri AA, Chabon JJ, Lovejoy AF, Newman AM, Stehr H, Azad TD, et al. 循环肿瘤 DNA 图谱对局部肺癌分子残留病的早期检测。Cancer Discov 2017;7:1394-403.
Moding EJ, Liu Y, Nabet BY, Chabon JJ, Chaudhuri AA, Hui AB, et al. Circulating tumor DNA dynamics predict benefit from consolidation immunotherapy in locally advanced non-small cell lung cancer. Nat Cancer 2020;1:176-83.
Moding EJ, Liu Y, Nabet BY, Chabon JJ, Chaudhuri AA, Hui AB, et al. 循环肿瘤 DNA 动态预测局部晚期非小细胞肺癌巩固免疫疗法的获益。Nat Cancer 2020;1:176-83.
Lim C, Tsao MS, Le LW, Shepherd FA, Feld R, Burkes RL, et al. Biomarker testing and time to treatment decision in patients with advanced nonsmall-cell lung cancer. Ann Oncol 2015;26:1415-21.
Lim C, Tsao MS, Le LW, Shepherd FA, Feld R, Burkes RL, et al. 晚期非小细胞肺癌患者的生物标志物检测与治疗决策时间。Ann Oncol 2015;26:1415-21.
Razavi P, Li BT, Brown DN, Jung B, Hubbell E, Shen R, et al. Highintensity sequencing reveals the sources of plasma circulating cellfree DNA variants. Nat Med 2019;25:1928-37.
Razavi P, Li BT, Brown DN, Jung B, Hubbell E, Shen R, et al. 高强度测序揭示血浆循环无细胞 DNA 变异的来源。Nat Med 2019;25:1928-37.
Newman AM, Lovejoy AF, Klass DM, Kurtz DM, Chabon JJ, Scherer F, et al. Integrated digital error suppression for improved detection of circulating tumor DNA. Nat Biotechnol 2016;34:547-55.
Newman AM, Lovejoy AF, Klass DM, Kurtz DM, Chabon JJ, Scherer F, et al. Integrated digital error suppression for improved detection of circulating tumor DNA.Nat Biotechnol 2016;34:547-55.
Chabon JJ, Hamilton EG, Kurtz DM, Esfahani MS, Moding EJ, Stehr H, et al. Integrating genomic features for non-invasive early lung cancer detection. Nature 2020;580:245-51.
Chabon JJ, Hamilton EG, Kurtz DM, Esfahani MS, Moding EJ, Stehr H, et al.自然 2020；580：245-51。
Khoury JD, Solary E, Abla O, Akkari Y, Alaggio R, Apperley JF, et al. The 5th edition of the World Health Organization classification of haematolymphoid tumours: myeloid and histiocytic/dendritic neoplasms. Leukemia 2022;36:1703-19.
Khoury JD、Solary E、Abla O、Akkari Y、Alaggio R、Apperley JF 等：世界卫生组织血液淋巴肿瘤分类第 5 版：髓细胞和组织细胞/树突状肿瘤。白血病 2022；36：1703-19。
Hsu JI, Dayaram T, Tovy A, De Braekeleer E, Jeong M, Wang F, et al. PPM1D mutations drive clonal hematopoiesis in response to cytotoxic chemotherapy. Cell Stem Cell 2018;23:700-13.e6.
Hsu JI, Dayaram T, Tovy A, De Braekeleer E, Jeong M, Wang F, et al. PPM1D 突变驱动细胞毒性化疗反应中的克隆造血。细胞干细胞》2018;23:700-13.e6。
Bolton KL, Ptashkin RN, Gao T, Braunstein L, Devlin SM, Kelly D, et al. Cancer therapy shapes the fitness landscape of clonal hematopoiesis. Nat Genet 2020;52:1219-26.
Bolton KL, Ptashkin RN, Gao T, Braunstein L, Devlin SM, Kelly D, et al. 癌症治疗塑造了克隆性造血的健康状况。Nat Genet 2020;52:1219-26.
Murtaza M, Dawson S-J, Tsui DWY, Gale D, Forshew T, Piskorz AM, et al. Non-invasive analysis of acquired resistance to cancer therapy by sequencing of plasma DNA. Nature 2013;497:108-12.
Murtaza M、Dawson S-J、Tsui DWY、Gale D、Forshew T、Piskorz AM 等：通过血浆 DNA 测序对癌症治疗获得性抗药性的非侵入性分析。自然》2013；497：108-12。
Sausen M, Phallen J, Adleff V, Jones S, Leary RJ, Barrett MT, et al. Clinical implications of genomic alterations in the tumour and circulation of pancreatic cancer patients. Nat Commun 2015;6:7686.
Sausen M, Phallen J, Adleff V, Jones S, Leary RJ, Barrett MT, et al. 胰腺癌患者肿瘤和血液循环中基因组改变的临床意义。Nat Commun 2015; 6:7686.
Machtay M, Duan F, Siegel BA, Snyder BS, Gorelick JJ, Reddin JS, et al. Prediction of survival by [18F]fluorodeoxyglucose positron emission tomography in patients with locally advanced non-small-cell lung cancer undergoing definitive chemoradiation therapy: results of the ACRIN 6668/RTOG 0235 trial. J Clin Oncol 2013;31:3823-30.
Machtay M, Duan F, Siegel BA, Snyder BS, Gorelick JJ, Reddin JS, et al. 接受确定性化放疗的局部晚期非小细胞肺癌患者的[18F]氟脱氧葡萄糖正电子发射断层成像预测生存率：ACRIN 6668/RTOG 0235 试验结果。J Clin Oncol 2013;31:3823-30.
van Diessen JNA, Chen C, van den Heuvel MM, Belderbos JSA, Sonke J-J. Differential analysis of local and regional failure in locally advanced non-small cell lung cancer patients treated with concurrent chemoradiotherapy. Radiother Oncol 2016;118:447-52.
van Diessen JNA, Chen C, van den Heuvel MM, Belderbos JSA, Sonke J-J.同期化放疗治疗局部晚期非小细胞肺癌患者局部和区域失败的差异分析。Radiother Oncol 2016;118:447-52.
Binkley MS, Koenig JL, Kashyap M, Xiang M, Liu Y, Sodji Q, et al. Predicting per-lesion local recurrence in locally advanced non-small cell lung cancer following definitive radiation therapy using pre- and mid-treatment metabolic tumor volume. Radiat Oncol 2020;15:114.
Binkley MS, Koenig JL, Kashyap M, Xiang M, Liu Y, Sodji Q, et al.使用治疗前和治疗中期代谢肿瘤体积预测局部晚期非小细胞肺癌确定性放疗后每个病灶的局部复发。Radiat Oncol 2020;15:114.
Huncharek M., Muscat J, Geschwind JF. K-ras oncogene mutation as a prognostic marker in non-small cell lung cancer: a combined analysis of 881 cases. Carcinogenesis 1999;20:1507-10.
Huncharek M.、Muscat J、Geschwind JF。K-ras 癌基因突变作为非小细胞肺癌的预后标志：对 881 个病例的综合分析。Carcinogenesis 1999;20:1507-10.
Steels E, Paesmans M, Berghmans T, Branle F, Lemaitre F, Mascaux C, et al. Role of p53 as a prognostic factor for survival in lung cancer: a systematic review of the literature with a meta-analysis. Eur Respir J 2001;18:705-19.
Steels E, Paesmans M, Berghmans T, Branle F, Lemaitre F, Mascaux C, et al. p53 作为肺癌生存预后因素的作用：文献系统回顾与荟萃分析。Eur Respir J 2001; 18:705-19.
Bailey MH, Tokheim C, Porta-Pardo E, Sengupta S, Bertrand D, Weerasinghe A, et al. Comprehensive characterization of cancer driver genes and mutations. Cell 2018;173:371-85.e18.
Bailey MH, Tokheim C, Porta-Pardo E, Sengupta S, Bertrand D, Weerasinghe A, et al. 癌症驱动基因和突变的综合表征。Cell 2018;173:371-85.e18.
Cox JD, Scott CB, Byhardt RW, Emami B, Russell AH, Fu KK, et al. Addition of chemotherapy to radiation therapy alters failure patterns by cell type within non-small cell carcinoma of lung (NSCCL): analysis of radiation therapy oncology group (RTOG) trials. Int J Radiat Oncol Biol Phys 1999;43:505-9.
Cox JD、Scott CB、Byhardt RW、Emami B、Russell AH、Fu KK 等：在放疗中加入化疗改变非小细胞肺癌（NSCCL）细胞类型的失败模式：肿瘤放疗组（RTOG）试验分析。Int J Radiat Oncol Biol Phys 1999;43:505-9.
Jeong Y, Hoang NT, Lovejoy A, Stehr H, Newman AM, Gentles AJ, et al. Role of KEAP1/NRF2 and TP53 mutations in lung squamous cell carcinoma development and radiation resistance. Cancer Discov 2017;7:86-101.
Jeong Y, Hoang NT, Lovejoy A, Stehr H, Newman AM, Gentles AJ, et al. Role of KEAP1/NRF2 and TP53 mutations in lung squamous cell carcinoma development and radiation resistance.Cancer Discov 2017;7:86-101.
Jaiswal S, Fontanillas P, Flannick J, Manning A, Grauman PV, Mar BG, et al. Age-related clonal hematopoiesis associated with adverse outcomes. N Engl J Med 2014;371:2488-98.
Jaiswal S, Fontanillas P, Flannick J, Manning A, Grauman PV, Mar BG, et al. 与年龄相关的克隆性造血与不良后果相关。N Engl J Med 2014;371:2488-98.
Coombs CC, Zehir A, Devlin SM, Kishtagari A, Syed A, Jonsson P, et al. Therapy-related clonal hematopoiesis in patients with nonhematologic cancers is common and associated with adverse clinical outcomes. Cell Stem Cell 2017;21:374-82.e4.
Coombs CC, Zehir A, Devlin SM, Kishtagari A, Syed A, Jonsson P, et al. 非血癌患者中与治疗相关的克隆性造血很常见，且与不良临床结果相关。Cell Stem Cell 2017;21:374-82.e4.
Kurtz DM, Esfahani MS, Scherer F, Soo J, Jin MC, Liu CL, et al. Dynamic risk profiling using serial tumor biomarkers for personalized outcome prediction. Cell 2019;178:699-713.e19.
Kurtz DM, Esfahani MS, Scherer F, Soo J, Jin MC, Liu CL, et al.Cell 2019;178:699-713.e19.
Cliffe H, Patel C, Prestwich R, Scarsbrook A. Radiotherapy response evaluation using FDG PET-CT-established and emerging applications. Br J Radiol 2017;90:20160764.
Cliffe H、Patel C、Prestwich R、Scarsbrook A.使用 FDG PET-CT 进行放疗反应评估--成熟和新兴应用。Br J Radiol 2017;90:20160764.
Kurtz DM, Scherer F, Jin MC, Soo J, Craig AFM, Esfahani MS, et al. Circulating tumor DNA measurements as early outcome predictors in diffuse large B-cell lymphoma. J Clin Oncol 2018;36: 2845-53.
Kurtz DM, Scherer F, Jin MC, Soo J, Craig AFM, Esfahani MS, et al. 循环肿瘤 DNA 测量作为弥漫大 B 细胞淋巴瘤的早期预后指标。J Clin Oncol 2018;36: 2845-53.
Abbosh C, Birkbak NJ, Wilson GA, Jamal-Hanjani M, Constantin T, Salari R, et al. Phylogenetic ctDNA analysis depicts early-stage lung cancer evolution. Nature 2017;545:446-51.
Abbosh C, Birkbak NJ, Wilson GA, Jamal-Hanjani M, Constantin T, Salari R, et al. 系统发育 ctDNA 分析描绘了早期肺癌的演变过程。自然》2017;545:446-51。
Garcia-Murillas I, Chopra N, Comino-Méndez I, Beaney M, Tovey H, Cutts RJ, et al. Assessment of molecular relapse detection in earlystage breast cancer. JAMA Oncol 2019;5:1473-8.
Garcia-Murillas I, Chopra N, Comino-Méndez I, Beaney M, Tovey H, Cutts RJ, et al. 早期乳腺癌分子复发检测评估。JAMA Oncol 2019;5:1473-8.
Christensen E, Birkenkamp-Demtröder K, Sethi H, Shchegrova S, Salari R, Nordentoft I, et al. Early detection of metastatic relapse and monitoring of therapeutic efficacy by ultra-deep sequencing of plasma cell-free DNA in patients with urothelial bladder carcinoma. J Clin Oncol 2019;37:1547-57.
Christensen E, Birkenkamp-Demtröder K, Sethi H, Shchegrova S, Salari R, Nordentoft I, et al. 尿路上皮膀胱癌患者血浆无细胞 DNA 超深度测序的转移性复发早期检测和疗效监测。J Clin Oncol 2019;37:1547-57.
Tie J, Wang Y, Tomasetti C, Li L, Springer S, Kinde I, et al. Circulating tumor DNA analysis detects minimal residual disease and predicts recurrence in patients with stage II colon cancer. Sci Transl Med 2016;8:346ra92.
Tie J, Wang Y, Tomasetti C, Li L, Springer S, Kinde I, et al. 循环肿瘤 DNA 分析检测最小残留病灶并预测 II 期结肠癌患者的复发。Sci Transl Med 2016;8:346ra92.
Khakoo S, Carter PD, Brown G, Valeri N, Picchia S, Bali MA, et al. MRI tumor regression grade and circulating tumor DNA as complementary tools to assess response and guide therapy adaptation in rectal cancer. Clin Cancer Res 2020;26:183-92.
Khakoo S, Carter PD, Brown G, Valeri N, Picchia S, Bali MA, et al. MRI 肿瘤回归分级和循环肿瘤 DNA 作为评估直肠癌反应和指导治疗适应性的互补工具。Clin Cancer Res 2020;26:183-92.
Pan Y, Zhang J-T, Gao X, Chen Z-Y, Yan B, Tan P-X, et al. Dynamic circulating tumor DNA during chemoradiotherapy predicts clinical outcomes for locally advanced non-small cell lung cancer patients. Cancer Cell 2023;41:1763-73.e4.
Pan Y, Zhang J-T, Gao X, Chen Z-Y, Yan B, Tan P-X, et al. 化放疗期间动态循环肿瘤 DNA 预测局部晚期非小细胞肺癌患者的临床结局。Cancer Cell 2023;41:1763-73.e4.
Deveson IW, Gong B, Lai K, LoCoco JS, Richmond TA, Schageman J, et al. Evaluating the analytical validity of circulating tumor DNA sequencing assays for precision oncology. Nat Biotechnol 2021;39: 1115-28.
Deveson IW, Gong B, Lai K, LoCoco JS, Richmond TA, Schageman J, et al.Nat Biotechnol 2021; 39: 1115-28.
Santonja A, Cooper WN, Eldridge MD, Edwards PAW, Morris JA, Edwards AR, et al. Comparison of tumor-informed and tumor-naïve sequencing assays for ctDNA detection in breast cancer. EMBO Mol Med 2023;15:e16505.
Santonja A, Cooper WN, Eldridge MD, Edwards PAW, Morris JA, Edwards AR, et al.EMBO Mol Med 2023;15:e16505.
Zhang C, Gu R, Wang H, Zhou C, Li Y, Liu Y, et al. Risk stratification in the clinical application of minimal residual disease assessment in acute myeloid leukemia. Cancer 2025;131:e35641.
Zhang C, Gu R, Wang H, Zhou C, Li Y, Liu Y, et al. 急性髓性白血病最小残留病评估临床应用中的风险分层。Cancer 2025;131:e35641.
Gökbuget N, Kneba M, Raff T, Trautmann H, Bartram C-R, Arnold $R$ , et al. Adult patients with acute lymphoblastic leukemia and
Gökbuget N, Kneba M, Raff T, Trautmann H, Bartram C-R, Arnold $R$ , et al.
molecular failure display a poor prognosis and are candidates for stem cell transplantation and targeted therapies. Blood 2012;120: 1868-76.
分子衰竭患者预后较差，是干细胞移植和靶向治疗的候选者。血液 2012; 120: 1868-76。
Stish BJ, Pisansky TM, Harmsen WS, Davis BJ, Tzou KS, Choo R, et al. Improved metastasis-free and survival outcomes with early salvage radiotherapy in men with detectable prostate-specific antigen after prostatectomy for prostate cancer. J Clin Oncol 2016;34: 3864-71.
Stish BJ, Pisansky TM, Harmsen WS, Davis BJ, Tzou KS, Choo R, et al. 前列腺癌前列腺切除术后可检测到前列腺特异性抗原的男性早期挽救性放疗改善了无转移和生存预后。J Clin Oncol 2016; 34: 3864-71.
Coroller TP, Grossmann P, Hou Y, Rios Velazquez E, Leijenaar RTH, Hermann G, et al. CT-based radiomic signature predicts distant metastasis in lung adenocarcinoma. Radiother Oncol 2015;114: 345-50.
Coroller TP, Grossmann P, Hou Y, Rios Velazquez E, Leijenaar RTH, Hermann G, et al. 基于 CT 的放射学特征预测肺腺癌的远处转移。Radiother Oncol 2015;114: 345-50.
Gill AB, Rundo L, Wan JCM, Lau D, Zawaideh JP, Woitek R, et al. Correlating radiomic features of heterogeneity on CT with circulating tumor DNA in metastatic melanoma. Cancers (Basel) 2020;12:E3493.
Gill AB, Rundo L, Wan JCM, Lau D, Zawaideh JP, Woitek R, et al.Cancers (Basel) 2020;12:E3493.
Lafata KJ, Corradetti MN, Gao J, Jacobs CD, Weng J, Chang Y, et al. Radiogenomic analysis of locally advanced lung cancer based on CT imaging and intratreatment changes in cell-free DNA. Radiol Imaging Cancer 2021;3:e200157.
Lafata KJ, Corradetti MN, Gao J, Jacobs CD, Weng J, Chang Y, et al. 基于 CT 成像和治疗内无细胞 DNA 变化的局部晚期肺癌放射基因组学分析。Radiol Imaging Cancer 2021;3:e200157.
Fave X, Zhang L, Yang J, Mackin D, Balter P, Gomez D, et al. Deltaradiomics features for the prediction of patient outcomes in nonsmall cell lung cancer. Sci Rep 2017;7:588.
Fave X, Zhang L, Yang J, Mackin D, Balter P, Gomez D, et al. Deltaradiomics features for the prediction of patient outcomes in nonsmall cell lung cancer.Sci Rep 2017;7:588.
Jeremić B, Miličić B, Milisavljevic S. Clinical prognostic factors in patients with locally advanced (stage III) nonsmall cell lung cancer treated with hyperfractionated radiation therapy with and without concurrent chemotherapy: single-institution experience in 600 Pa tients. Cancer 2011;117:2995-3003.
Jeremić B、Miličić B、Milisavljevic S.局部晚期（III 期）非小细胞肺癌患者接受超分割放疗联合或不联合化疗的临床预后因素：600 名患者的单机构经验。癌症 2011；117：2995-3003.
Spigel DR, Faivre-Finn C, Gray JE, Vicente D, Planchard D, Paz-Ares L, et al. Five-year survival outcomes from the PACIFIC trial: durvalumab after chemoradiotherapy in stage III non-small-cell lung cancer. J Clin Oncol 2022;40:1301-11.
Spigel DR、Faivre-Finn C、Gray JE、Vicente D、Planchard D、Paz-Ares L 等：PACIFIC 试验：III 期非小细胞肺癌化疗后的杜伐单抗五年生存结果。J Clin Oncol 2022;40:1301-11.
Tanaka K, Hida T, Oya Y, Oguri T, Yoshida T, Shimizu J, et al. EGFR mutation impact on definitive concurrent chemoradiation therapy for inoperable stage III adenocarcinoma. J Thorac Oncol 2015;10:1720-5.
Tanaka K, Hida T, Oya Y, Oguri T, Yoshida T, Shimizu J, et al. EGFR 突变对无法手术的 III 期腺癌确定性同期化疗的影响。J Thorac Oncol 2015;10:1720-5.
Alexander BM, Othus M, Caglar HB, Allen AM. Tumor volume is a prognostic factor in non-small-cell lung cancer treated with chemoradiotherapy. Int J Radiat Oncol Biol Phys 2011;79:1381-7.
Alexander BM, Othus M, Caglar HB, Allen AM.肿瘤体积是非小细胞肺癌化疗的预后因素之一 Int J Radiat Oncol Biol Phys 2011;79:1381-7.
Koo TR, Moon SH, Lim YJ, Kim JY, Kim Y, Kim TH, et al. The effect of tumor volume and its change on survival in stage III non-small cell lung cancer treated with definitive concurrent chemoradiotherapy. Radiat Oncol 2014;9:283.
Koo TR, Moon SH, Lim YJ, Kim JY, Kim Y, Kim TH, et al. 确定性同期化放疗治疗 III 期非小细胞肺癌的肿瘤体积及其变化对生存期的影响。Radiat Oncol 2014;9:283.
Goldstraw P, Chansky K, Crowley J, Rami-Porta R, Asamura H, Eberhardt WEE, et al. The IASLC lung cancer staging project: proposals for revision of the TNM stage groupings in the forthcoming (eighth) edition of the TNM classification for lung cancer. J Thorac Oncol 2016;11:39-51.
Goldstraw P, Chansky K, Crowley J, Rami-Porta R, Asamura H, Eberhardt WEE, et al. IASLC 肺癌分期项目：即将出版的（第八版）肺癌 TNM 分类中 TNM 分期分组的修订建议。J Thorac Oncol 2016;11:39-51.
Ainscough BJ, Barnell EK, Ronning P, Campbell KM, Wagner AH, Fehniger TA, et al. A deep learning approach to automate refinement of somatic variant calling from cancer sequencing data. Nat Genet 2018;50:1735-43.
Ainscough BJ, Barnell EK, Ronning P, Campbell KM, Wagner AH, Fehniger TA, et al.Nat Genet 2018; 50:1735-43.
Takahashi K, Wang F, Kantarjian H, Doss D, Khanna K, Thompson E, et al. Preleukaemic clonal haemopoiesis and risk of therapy-related myeloid neoplasms: a case-control study. Lancet Oncol 2017;18: 100-11.
Takahashi K, Wang F, Kantarjian H, Doss D, Khanna K, Thompson E, et al. 白血病前克隆造血与治疗相关髓系肿瘤风险：一项病例对照研究。Lancet Oncol 2017;18: 100-11.
Leighl NB, Page RD, Raymond VM, Daniel DB, Divers SG, Reckamp KL, et al. Clinical utility of comprehensive cell-free DNA analysis to identify genomic biomarkers in patients with newly diagnosed metastatic non-small cell lung cancer. Clin Cancer Res 2019;25:4691-700.
Leighl NB、Page RD、Raymond VM、Daniel DB、Divers SG、Reckamp KL 等：综合无细胞 DNA 分析在新诊断转移性非小细胞肺癌患者中识别基因组生物标志物的临床实用性。Clin Cancer Res 2019;25:4691-700.
Daly ME, Singh N, Ismaila N; Management of Stage III NSCLC Guideline Expert Panel. Management of stage III non-small cell
Daly ME、Singh N、Ismaila N；III 期 NSCLC 管理指南专家小组。III 期非小细胞
lung cancer: ASCO guideline rapid recommendation update. J Clin Oncol 2024;42:3058-60.
肺癌：ASCO 指南快速建议更新。J Clin Oncol 2024；42：3058-60。
Lu S, Ahn M-J, Reungwetwattana T, Özgüroğlu M, Kato T, Yang JC-H, et al. Osimertinib after definitive chemoradiotherapy in unresectable stage III epidermal growth factor receptor-mutated non-small-cell lung cancer: analyses of central nervous system efficacy and distant progression from the phase III LAURA study. Ann Oncol 2024;35:1116-25.
Lu S, Ahn M-J, Reungwetwattana T, Özgüroğlu M, Kato T, Yang JC-H, et al. Osimertinib after definitive chemoradiotherapy in unresectable stage III epidermal growth factor receptor-mutated non-small-cell lung cancer: analyses of central nervous system efficacy and distant progression from the phase III LAURA study.Ann Oncol 2024;35:1116-25.
Binkley MS, Jeon Y-J, Nesselbush M, Moding EJ, Nabet BY, Almanza D, et al. KEAP1/NFE2L2 mutations predict lung cancer radiation resistance that can be targeted by glutaminase inhibition. Cancer Discov 2020;10:1826-41.
Binkley MS, Jeon Y-J, Nesselbush M, Moding EJ, Nabet BY, Almanza D, et al. KEAP1/NFE2L2 突变可预测肺癌放射耐药性，谷氨酰胺酶抑制剂可作为靶点。Cancer Discov 2020;10:1826-41.
Chuang JC, Stehr H, Liang Y, Das M, Huang J, Diehn M, et al. ERBB2 -mutated metastatic non-small cell lung cancer: response and resistance to targeted therapies. J Thorac Oncol 2017;12:833-42.
Chuang JC, Stehr H, Liang Y, Das M, Huang J, Diehn M, et al. ERBB2 突变的转移性非小细胞肺癌：对靶向治疗的反应和耐药性。J Thorac Oncol 2017;12:833-42.
Liu J, Lichtenberg T, Hoadley KA, Poisson LM, Lazar AJ, Cherniack AD, et al. An integrated TCGA pan-cancer clinical data resource to drive high-quality survival outcome analytics. Cell 2018;173: 400-16.e11.
Liu J, Lichtenberg T, Hoadley KA, Poisson LM, Lazar AJ, Cherniack AD, et al. An integrated TCGA pan-cancer clinical data resource to drive high-quality survival outcome analytics.Cell 2018;173: 400-16.e11.
Dupont WD, Plummer WD. Power and sample size calculations. A review and computer program. Control Clin Trials 1990;11: 116-28.
Dupont WD, Plummer WD.功率和样本量计算。回顾与计算机程序。Control Clin Trials 1990; 11: 116-28.
Chabon JJ, Simmons AD, Lovejoy AF, Esfahani MS, Newman AM, Haringsma HJ, et al. Circulating tumour DNA profiling reveals heterogeneity of EGFR inhibitor resistance mechanisms in lung cancer patients. Nat Commun 2016;7:11815.
Circulating tumour DNA profiling reveals heterogeneity of EGFR inhibitor resistance mechanisms in lung cancer patients.Nat Commun 2016;7:11815.
Steensma DP, Bejar R, Jaiswal S, Lindsley RC, Sekeres MA, Hasserjian RP, et al. Clonal hematopoiesis of indeterminate potential and its distinction from myelodysplastic syndromes. Blood 2015;126: 9-16.
Steensma DP, Bejar R, Jaiswal S, Lindsley RC, Sekeres MA, Hasserjian RP, et al. 不确定潜能的克隆性造血及其与骨髓增生异常综合征的区别。Blood 2015; 126: 9-16.
Newman AM, Bratman SV, To J, Wynne JF, Eclov NCW, Modlin LA, et al. An ultrasensitive method for quantitating circulating tumor DNA with broad patient coverage. Nat Med 2014;20:548-54.
Newman AM, Bratman SV, To J, Wynne JF, Eclov NCW, Modlin LA, et al.Nat Med 2014; 20:548-54.
Karczewski KJ, Francioli LC, Tiao G, Cummings BB, Alföldi J, Wang Q, et al. The mutational constraint spectrum quantified from variation in 141,456 humans. Nature 2020;581:434-43.
Karczewski KJ, Francioli LC, Tiao G, Cummings BB, Alföldi J, Wang Q, et al.自然 2020; 581:434-43。
Zwanenburg A, Vallières M, Abdalah MA, Aerts HJWL, Andrearczyk V, Apte A, et al. The image biomarker standardization initiative: standardized quantitative radiomics for high-throughput image-based phenotyping. Radiology 2020;295:328-38.
Zwanenburg A, Vallières M, Abdalah MA, Aerts HJWL, Andrearczyk V, Apte A, et al. 图像生物标记标准化倡议：基于图像的高通量表型标准化定量放射组学。放射学 2020；295：328-38。
Clark K, Vendt B, Smith K, Freymann J, Kirby J, Koppel P, et al. The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository. J Digit Imaging 2013;26:1045-57.
Clark K, Vendt B, Smith K, Freymann J, Kirby J, Koppel P, et al. The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository.J Digit Imaging 2013;26:1045-57.
van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res 2017;77:e104-7.
van Griethuysen JJM、Fedorov A、Parmar C、Hosny A、Aucoin N、Narayan V 等：解码放射表型的计算放射组学系统。Cancer Res 2017;77:e104-7.
Kahala G, Sklair M, Spitzer H. Multi-scale blood vessel detection and segmentation in breast MRIs. J Med Biol Eng 2019;39:424-30.
Kahala G, Sklair M, Spitzer H. 乳房磁共振成像中的多尺度血管检测与分割。J Med Biol Eng 2019;39:424-30.
Liu G, Zheng X. Fabric defect detection based on information entropy and frequency domain saliency. Vis Comput 2021;37: 515-28.
Liu G, Zheng X. 基于信息熵和频域显著性的织物缺陷检测。视觉计算 2021; 37: 515-28.
Heagerty PJ, Lumley T, Pepe MS. Time-dependent ROC curves for censored survival data and a diagnostic marker. Biometrics 2000;56: 337-44.
Heagerty PJ, Lumley T, Pepe MS.删减生存数据和诊断标记的时间依赖性 ROC 曲线。生物计量学》2000；56：337-44。
Kashyap M, Wang X, Panjwani N, Hasan M, Zhang Q, Huang C, et al. Automated deep learning-based detection and segmentation of lung tumors at CT. Radiology 2025;314:e233029.
Kashyap M, Wang X, Panjwani N, Hasan M, Zhang Q, Huang C, et al. 基于深度学习的 CT 肺部肿瘤自动检测与分割。放射学》2025;314:e233029。
Cicchetti DV. Guidelines, criteria, and rules of thumb for evaluating normed and standardized assessment instruments in psychology. Psychol Assess 1994;6:284-90.
Cicchetti DV.评估心理学规范化和标准化评估工具的指南、标准和经验法则。Psychol Assess 1994;6:284-90.
Rentzsch P, Witten D, Cooper GM, Shendure J, Kircher M. CADD: predicting the deleteriousness of variants throughout the human genome. Nucleic Acids Res 2019;47:D886-94.
Rentzsch P, Witten D, Cooper GM, Shendure J, Kircher M. CADD：预测整个人类基因组中变异的有害性。Nucleic Acids Res 2019;47:D886-94.

$^{1}$ Department of Radiation Oncology, Stanford University, Stanford, California. $^{2}$ Stanford Cancer Institute, Stanford University, Stanford, California. $^{3}$ Division of Oncology, Department of Medicine, Stanford University, Stanford, California. $^{4}$ Department of Radiation Oncology, University of Texas MD Anderson Cancer Center, Houston, Texas. $^{5}$ Department of Radiation Oncology, Mayo Clinic, Rochester, Minnesota. $^{6}$ Department of Medicine, VA Palo Alto Health Care System, Palo Alto, California. $^{7}$ Institute for Stem Cell Biology and Regenerative Medicine, Stanford University, Stanford, California.
$^{1}$ 加利福尼亚州斯坦福大学放射肿瘤学系。 $^{2}$ 加利福尼亚州斯坦福市斯坦福大学斯坦福癌症研究所。 $^{3}$ 加利福尼亚州斯坦福斯坦福大学医学系肿瘤科。 $^{4}$ 得克萨斯大学 MD 安德森癌症中心放射肿瘤学系，得克萨斯州休斯顿。 $^{5}$ 明尼苏达州罗切斯特梅奥诊所放射肿瘤学系。 $^{6}$ 加利福尼亚州帕洛阿尔托市退伍军人帕洛阿尔托医疗保健系统医学部。 $^{7}$ 加利福尼亚州斯坦福大学干细胞生物学和再生医学研究所。
Corresponding Authors: Maximilian Diehn, Department of Radiation Oncology, Stanford Cancer Institute, and Institute for Stem Cell Biology & Regenerative Medicine, 875 Blake Wilbur Drive, Stanford, CA 943055847. E-mail: diehn@stanford.edu; Ash A. Alizadeh, Division of Oncology, Department of Medicine, Stanford Cancer Institute, and Institute for Stem Cell Biology & Regenerative Medicine, 875 Blake Wilbur Drive, Stanford, CA 94305-5847. E-mail: arasha@stanford.edu; Steven H. Lin, Department of Radiation Oncology, Unit 1422, The University of Texas MD Anderson Cancer Center, 1515 Holcombe Boulevard, Houston, TX 77030. E-mail: SHLin@mdanderson.org; Ruijiang Li, Department of Radiation Oncology, Stanford University School of Medicine, 1070 Arastradero Road, Palo Alto, CA 94304. E-mail: rli2@stanford.edu; and Everett J. Moding, Department of Radiation Oncology, Stanford Cancer Institute, 875 Blake Wilbur Drive, Stanford, CA 94305-5847. E-mail: emoding@stanford.edu
通讯作者：Maximilian Diehn，斯坦福大学癌症研究所放射肿瘤部，干细胞生物学与再生医学研究所，875 Blake Wilbur Drive, Stanford, CA 943055847。电子邮件：diehn@stanford.edu；Ash A. Alizadeh，斯坦福大学癌症研究所医学系肿瘤科以及干细胞生物学与再生医学研究所，地址：875 Blake Wilbur Drive, Stanford, CA 94305-5847。电子邮件：arasha@stanford.edu；Steven H. Lin，德克萨斯大学 MD 安德森癌症中心放射肿瘤系 1422 室，1515 Holcombe Boulevard, Houston, TX 77030。电子邮件：SHLin@mdanderson.org; Ruijiang Li, Department of Radiation Oncology, Stanford University School of Medicine, 1070 Arastradero Road, Palo Alto, CA 94304.电子邮件：rli2@stanford.edu；斯坦福大学癌症研究所放射肿瘤学系 Everett J. Moding，地址：875 Blake Wilbur Drive, Stanford, CA 94305-5847。电子邮件：emoding@stanford.edu
Cancer Discov 2025;XX:1-21
doi: 10.1158/2159-8290.CD-24-1704 ©2025 American Association for Cancer Research
doi: 10.1158/2159-8290.CD-24-1704 ©2025 美国癌症研究协会

Integrating ctDNA Analysis and Radiomics for Dynamic Risk Assessment in Localized Lung Cancer (1) Check for updates 整合 ctDNA 分析和放射组学，对局部肺癌进行动态风险评估 (1) 检查更新

Abstract 摘要

INTRODUCTION 引言

RESULTS 结果

CH Dynamics during CRTCRT 期间的 CH 动态

Tissue-Free Variant Calling with Machine Learning利用机器学习进行无组织变异调用

ctDNA Kinetics during CRTCRT 期间的 ctDNA 动力学

Prediction of PFS Using Radiomics利用放射组学预测 PFS

Biological and Molecular Prognostic Factors in Locoregionally Advanced NSCLC局部晚期 NSCLC 的生物学和分子预后因素

A Dynamic Risk Index for NSCLC非小细胞肺癌动态风险指数

Comparing CIRI-LCRT with ctDNA MRDCIRI-LCRT 与 ctDNA MRD 的比较

DISCUSSION 讨论

METHODS 方法

Study Design and Patients研究设计和患者

Power Considerations 电源考虑因素

CRT and Blood CollectionCRT 和采血

Library Preparation and Sequencing文库制备和测序

CH Analysis CH 分析

SNV Score Tissue-Free Variant CallingSNV 评分 无组织变异调用

ctDNA Monitoring ctDNA 监测

Radiomic Analysis 放射原子分析

Identification of Pretreatment Prognostic Factors确定治疗前的预后因素

CIRI-LCRT

Statistics 统计资料

Data Availability 数据可用性

Authors' Disclosures 作者披露

Authors' Contributions 作者的贡献

Acknowledgments 致谢

Note 备注

REFERENCES 参考文献

Integrating ctDNA Analysis and Radiomics for Dynamic Risk Assessment in Localized Lung Cancer
(1) Check for updates
整合 ctDNA 分析和放射组学，对局部肺癌进行动态风险评估 (1) 检查更新

CH Dynamics during CRT
CRT 期间的 CH 动态

Tissue-Free Variant Calling with Machine Learning
利用机器学习进行无组织变异调用

ctDNA Kinetics during CRT
CRT 期间的 ctDNA 动力学

Prediction of PFS Using Radiomics
利用放射组学预测 PFS

Biological and Molecular Prognostic Factors in Locoregionally Advanced NSCLC
局部晚期 NSCLC 的生物学和分子预后因素

A Dynamic Risk Index for NSCLC
非小细胞肺癌动态风险指数

Comparing CIRI-LCRT with ctDNA MRD
CIRI-LCRT 与 ctDNA MRD 的比较

Study Design and Patients
研究设计和患者

CRT and Blood Collection
CRT 和采血

Library Preparation and Sequencing
文库制备和测序

SNV Score Tissue-Free Variant Calling
SNV 评分无组织变异调用

Identification of Pretreatment Prognostic Factors
确定治疗前的预后因素