Integrating circulating tumor DNA analysis and radiomics for dynamic risk assessment in localized lung cancer

Everett J Moding; Mohammad Shahrokh Esfahani; Cheng Jin; Angela B Hui; Barzin Y Nabet; Yufei Liu; Jacob J Chabon; Michael S Binkley; David M Kurtz; Emily G Hamilton; Aadel A Chaudhuri; Chih Long Liu; Zhe Li; Rene F Bonilla; Alice L Jiang; Brianna C Lau; Pablo Lopez; Jianzhong He; Yawei Qiao; Ting Xu; Luyang Yao; Saumil Gandhi; Zhongxing Liao; Millie Das; Kavitha J Ramchandran; Sukhmani K Padda; Joel W Neal; Heather A Wakelee; Michael F Gensheimer; Billy W Loo, Jr; Ruijiang Li; Steven H Lin; Ash A Alizadeh; Maximilian Diehn

doi:10.1158/2159-8290.CD-24-1704

. Author manuscript; available in PMC: 2026 Feb 4.

Published in final edited form as: Cancer Discov. 2025 Aug 4;15(8):1609–1629. doi: 10.1158/2159-8290.CD-24-1704

Integrating circulating tumor DNA analysis and radiomics for dynamic risk assessment in localized lung cancer

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,^*, Maximilian Diehn ^1,^2,^7,^*

PMCID: PMC12324966 NIHMSID: NIHMS2077095 PMID: 40299851

Abstract

The complementarity and clinical utility of combining liquid biopsies and radiomic image analysis has not been demonstrated. Circulating tumor DNA (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. Here, 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.

INTRODUCTION

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) (2). Unfortunately, the majority of patients with stage III NSCLC develop progressive disease after receiving the current standard of care (3). In addition, the definitive radiation dose 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.

Several prior studies have attempted to identify prognostic factors in NSCLC, but the majority of these studies have focused on overall survival across diverse groups of patients and few factors have been validated in independent data sets (5). As a result, 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 pre-treatment prognostic factors (8–14), but these are not currently incorporated into treatment decisions as part of standard of care.

Due to their importance in cancer diagnosis, staging, and treatment planning, clinical images, including computed tomography (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 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).

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. Mid-treatment imaging with CT and/or positron emission tomography (PET) has been associated with patient outcomes (26–28). For example, a higher PET metabolic tumor volume and total lesion glycolysis were associated with a higher risk of local recurrence after chemoradiation in one study (26), and another study reported 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.

Tumors continuously shed DNA into the peripheral blood that can be non-invasively collected from liquid biopsies and quantified as circulating tumor DNA (ctDNA) (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 chemoradiation therapy and whether mid-treatment 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 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 non-malignant biological signals such as clonal hematopoiesis (CH) (48–50).

We hypothesized that clinical risk factors, radiomics, and ctDNA analysis provide complementary information that can improve risk-stratification when combined. Here, we optimize tissue-free identification of tumor-derived variants from pre-treatment 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 pre-treatment prognostic factors including radiomic analysis with mid-treatment ctDNA levels to improve prediction of progression-free survival and enable real-time updating of individualized risk estimates.

RESULTS

Clonal hematopoiesis dynamics during CRT

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 (Figure 1A). We therefore performed ctDNA analysis using CAPP-Seq on a total of 101 patients with NSCLC with plasma samples collected 10–30 days into CRT (mid-CRT, Figure S1A,B, Table 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 non-malignant hematopoietic cells acquire somatic alterations leading to clonal expansion. Per the World Health Organization’s classification of hematolymphoid tumors, clonal hematopoiesis 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 if CH variants persist during CRT. Therefore, we sought to determine the dynamics of CH variants in patients with NSCLC treated with CRT.

Figure 1: — 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 progression-free survival in patients with non-small cell lung cancer (NSCLC) treated with chemoradiation therapy (CRT). B, Log10 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, Log10 fold change in CH VAF mid-CRT in canonical genes with at least 3 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 to 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 versus 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.)

We began by characterizing the prevalence of CH variants in patients with locoregionally advanced NSCLC by identifying cfDNA single nucleotide variants (SNVs) 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 non-canonical CH genes in patients with locoregionally advanced NSCLC (Figure S2A–C). 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 variant allele fraction increased from 2.3% pre-CRT to 6.6% mid-CRT (Figure S2D). Considering CH mutations at any allele fraction, 90% of variants remained detected at the mid-CRT time point (Figure 1B,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 variant allele fraction mid-CRT for frequently mutated canonical CH genes. Compared with DNMT3A mutations which remained stable mid-CRT, PPM1D mutations significantly increased while SF3B1 and TET2 mutations decreased (Figure 1D). In patients with later samples available for analysis, CH variants persisted up to 11 months after starting CRT (Figure 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 pre-treatment leukocytes and adopted a stringent filtering strategy to remove CH variants of any VAF from our ctDNA monitoring approach.

Tissue-free variant calling with machine learning

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 multi-tiered machine learning approach to integrate cfDNA genomic features for non-invasive lung cancer detection (50), and we hypothesized that a similar approach could improve tissue-free tumor SNV calling for ctDNA monitoring during CRT.

We analyzed targeted deep sequencing of plasma cell-free DNA and matched leukocytes from 190 patients with lung cancer and 92 controls along with DNA extracted from matched tumor tissue for 119 of the lung cancer patients (Table S3). We extracted 25 key biological and technical features for each SNV identified in the cfDNA, such fragment size, allele fraction, background frequencies, and mapping quality. We employed machine learning using semi-supervised 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, Figure S3A–C).

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 tumor-derived. The SNV score was significantly higher for tumor-adjudicated variants than not tumor-adjudicated variants (median 0.91 vs. 0.09, Figure 1E) and was higher with increasing allele-fraction (Figure 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 (Figure 1F and G, Figure 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 (Figure 1H, Figure 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.

Circulating tumor DNA kinetics during CRT

Having established an improved methodology for tissue-free variant calling, we applied the SNV score for pre-treatment identification of variants in patients without matched tumor tissue available for sequencing. We then monitored for the variants identified pre-treatment 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-IIIB NSCLC with variants identified pre-treatment (Figure 2A and B, Table 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 pre-treatment 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.

Figure 2: — 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 MD Anderson Cancer Center (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, Log10 fold change in ctDNA concentration from the pre-CRT to mid-CRT timepoint in patients from the training cohort with (n=23) and without (n=17) progression or death during follow up. P value was calculated using a two-sided Mann-Whitney test. E, Hazard ratios with 95% confidence intervals 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, Hazard ratios with 95% confidence intervals for a multivariable Cox proportional hazards model for PFS including mid-CRT ctDNA concentration in the training cohort (n=40 patients). **G-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 patient 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.)

We first characterized ctDNA kinetics at the mid-CRT timepoint 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 (Figure 2C). This decrease was driven by a reduction in ctDNA molecules rather than a change in total cfDNA (Figure S4A and B). 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 (Figure 2D). Both ctDNA fold change and mid-CRT ctDNA concentration were statistically significantly associated with progression-free survival (PFS) in univariable and multivariable analyses, but pre-CRT concentrations were not (Figure 2E and F, Figure S4C–E).

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 (Figure 2G and H). Patients with a ctDNA concentration greater than 3.2 hGE/ml had significantly worse PFS in both the training (Hazard ratio (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 (Figure S4F) and a corresponding significantly lower mid-CRT ctDNA concentration in patients with isolated local recurrences versus other progression or death (Figure 2I). These results suggest that patients with isolated local recurrences have a lower total body disease burden than those who progress distantly. This data demonstrates that ctDNA levels during CRT are highly prognostic for PFS in patients with locoregionally advanced NSCLC.

Prediction of progression-free survival using radiomics

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 progressive disease. We first sought to determine if radiomic analysis of pre-treatment 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 RTOG 0235/ACRIN 6668 (n=209 patients, Table S6) (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 (Figure 3A). In the training cohort, 5 CT image features were associated with PFS (FDR<0.05): 1) entropy of image intensity, 2) margin blurriness, 3) number of blood vessels in contact with the tumor, 4) vessel coverage, and 5) vessel scattering. After feature selection, the final radiomic signature incorporated margin blurriness and vessel scattering (Figure 3B and Figure S5A).

In the training cohort, the radiomic score as a continuous variable was strongly associated with PFS on multivariable analysis (Figure 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 (Figure S5B–D). In addition, margin blurriness and vessel scattering displayed excellent reproducibility across independent sets of tumor contours (Figure S5E and F). 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 MD Anderson Cancer Center (Figure 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.

Biological and molecular prognostic factors in locoregionally advanced NSCLC

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 historical cohort of 108 patients with stage IIB-IIIA NSCLC from Stanford University and the Cancer Genome Atlas (TCGA) treated with radiation therapy to identify and train additional features prognostic of PFS (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 historical training cohort, pre-CRT largest lesion metabolic tumor volume (MTV), largest lesion gross tumor volume (GTV), and histology (non-squamous cell carcinoma vs. squamous cell carcinoma) were significantly associated with inferior PFS (Figure 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 (Figure S6A).

Figure 4: — A, Hazard ratios with 95% confidence intervals for univariable Cox proportional hazards models for progression-free survival (PFS) based each biological feature in a historical 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-squamous cell carcinoma vs. squamous cell carcinoma histology, 38 patients for largest lesion metabolic tumor volume (MTV) and largest lesion gross tumor volume (GTV)). **B-C,** Kaplan-Meier analysis of PFS stratified by (B) histology and (C) GTV above or below the optimal cutpoint in the historical training cohort. P values were calculated using two-sided log-rank tests. D, Hazard ratios with 95% confidence intervals for univariable Cox proportional hazards models for progression-free survival (PFS) based on mutation status of each significant molecular feature identified from Figure S6B (n=108 patients). **E-F,** Kaplan-Meier analysis of PFS stratified by (E) *KRAS* mutation status and (F) *KEAP1* mutation status in the historical training cohort. P values were calculated using two-sided log-rank tests.

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 least 5% of lung adenocarcinomas or lung squamous cell carcinomas (61). Of the 23 genes analyzed, only mutations in KRAS and KEAP1 were significantly associated with PFS (Figure 4D–F, Figure S6B). Consistent with prior reports (62), non-squamous cell carcinomas more frequently failed distantly (Figure S6C). In addition, tumors with KEAP1 mutations, but not with KRAS mutations, more frequently failed within the radiation field (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 overall survival in our cohort of patients with locoregionally advanced NSCLC treated with CRT (Figure S7A and B).

A dynamic risk index for NSCLC

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 progressive disease 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 hyper-parameters for CIRI-LCRT from the RTOG 0235/ACRIN 6668 dataset and the biological and molecular feature hyper-parameters in our historical training cohort. The MDACC training cohort was used to infer the ctDNA hyper-parameters and to train the full CIRI-LCRT model. After the full model was finalized and locked, validation was performed in the Stanford cohort.

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 (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 risk factors including mid-CRT ctDNA concentration (Figure 5A–C, Table S9 and S10). Importantly, performance was nearly identical in the independent validation cohort (Figure 5D and E). Model performance steadily improved with the addition of each feature in both the training and validation cohort (Figure S8A), demonstrating the benefit of integrating complementary data.

Figure 5: — A, Timeline of a patient with NSCLC treated with CRT. CIRI-LCRT integrates pre-CRT risk factors (non-squamous cell carcinoma vs. squamous cell carcinoma 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 standard deviation) 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 calculated empirically from 2000 bootstrap resamplings. **F-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 standard error of the mean. The slope and Y-intercept with 95% confidence intervals 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 greyed 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.

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 (Figure 5F). When we applied the same cutoffs to our validation cohort, we observed similar stratification of risk (Figure 5G). Considering only the final prediction at the mid-CRT time point, stratifying patients by ≤50% or >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 (Figure S8B,C). We observed good calibration of our model across the whole cohort when comparing predicted and observed risk of PFS at 12 months (Figure 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 (Figure 5I). LUP810 presented with a left upper lobe squamous cell carcinoma 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%. Now 25 months after starting CRT, LUP810 remains 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.

Comparing CIRI-LCRT with ctDNA MRD

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 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-statistic and performed comparably for prediction of PFS at 12 months and by Kaplan-Meier analysis (Figure 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.

We selected two patients vignettes to help illustrate the ability of CIRI-LCRT to provide an earlier prediction of PFS than ctDNA MRD (Figure 6E). LUP238 underwent CRT for a stage IIIA right middle lobe squamous cell carcinoma 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 squamous cell carcinoma of the left lower lobe and remained alive and progression free 24 months later. His mid-CRT CIRI-LCRT predicted 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 de-intensification.

DISCUSSION

By accounting for clonal hematopoiesis 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 pre-treatment 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).

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 (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 et al. 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 (73). In contrast to our study where only 25% of patients had ctDNA undetected mid-CRT, 79% of patients were undetected using droplet digital PCR for 1–3 mutations per patient, suggesting assay sensitivity could be important for mid-CRT ctDNA analysis. Pan et al. recently reported superior outcomes for patients with undetectable ctDNA in plasma samples collected near the end of induction chemotherapy followed by CRT in NSCLC (74). The plasma samples on our study were collected ~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.

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 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 overall survival with reported validation C-statistics/areas under the curve ranging from 0.62–0.76 (8,10,11).

Our final CIRI-LCRT model incorporated tumor histology (non-squamous cell carcinoma vs. squamous cell carcinoma), 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 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 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.

Similar to most prior studies of stage III NSCLC treated with CRT, patients with squamous cell carcinomas 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 overall survival in patients with non-squamous histologies (85). Notably, durvalumab significantly improved overall survival in non-squamous histologies but not squamous cell carcinoma, 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 since 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. tumor tissue not available to assess histology).

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 overall survival in unselected patients with NSCLC (60), and one prior study associated EGFR 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 historical 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 overall survival 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.

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 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.

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 clonal hematopoiesis (48,50). Cytotoxic therapy increases the prevalence of clonal hematopoiesis (48,65), which is associated with an increased risk of therapy-related myeloid neoplasms (91). Recently, repeat sampling from patients prior to and over a year following chemotherapy or radiation therapy demonstrated expansion of clonal hematopoiesis 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). Here, we demonstrate for the first time that CRT can differentially select for or against clonal hematopoiesis 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 clonal hematopoiesis 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.

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 pre-treatment 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). 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.

Strengths of our study include accounting for clonal hematopoiesis 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 progression-free survival 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.

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.

MATERIALS AND METHODS

Study design and patients

All samples analyzed in this study were prospectively collected with written informed consent from subjects enrolled on protocols approved by Institutional Review Boards at MD Anderson Cancer Center (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 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.

A total of 101 patients with American Joint Committee on Cancer (AJCC) 7^th edition stage IIA-IIIB NSCLC from MDACC and Stanford University were retrospectively identified who had a mid-CRT plasma sample collected 10–30 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 stage IIB-IIIB and had variants identified pre-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 timepoint (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.

Finally, we identified a separate historical 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-IIIB NSCLC described previously (58,95) who had undergone tumor genotyping using a clinical hybrid-capture based sequencing assay covering 130–198 genes (96) and did not undergo mid-CRT ctDNA analysis. In addition, we identified 70 patients with IIB-IIIB NSCLC from TCGA with high quality progression-free survival data available (97) by filtering for patients who had undergone whole exome DNA sequencing, received radiation therapy, and did not undergo surgical resection.

Power considerations

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 hazard ratio 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 alpha of 0.05 (98).

Chemoradiation therapy and blood collection

Patients treated with chemoradiation therapy for locoregionally advanced NSCLC at MD Anderson Cancer Center and Stanford University completed pre-treatment 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 10–30 days into chemoradiation therapy. For a subset of patients, a third peripheral blood sample was collected within 4 months of completing radiation and chemotherapy for MRD analysis.

Library preparation and sequencing

DNA samples from plasma, leukocytes, and tumors were sequenced using cancer personalized profiling by deep sequencing (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 storage at −80 °C prior to DNA isolation. Tumor DNA was isolated from formalin-fixed, paraffin embedded sections when available. Cell-free DNA was extracted from plasma using the QIAmp Circulating Nucleic Acid Kit (QIAGEN) according to the manufacturer’s instructions. A median of 32 ng cfDNA was prepared for sequencing using the 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 with FLEX sequencing adapters both described previously (49,50). Both adapters incorporate dual-index sample barcodes for demultiplexing and unique identifiers (UIDs) 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) (100). Samples were sequenced on an Illumina HiSeq 4000 System using 2 × 150 base pair paired-end reads with eight base indexing. Sequencing data were processed using a custom bioinformatics pipeline to identify putative SNVs as described previously (50).

Clonal hematopoiesis analysis

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 below.

SNV score tissue-free variant calling

We adapted the framework from our previously described Lung Cancer Likelihood in Plasma (Lung-CLiP) (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 pre-processed to remove 1) SNPs from any individual in the study, 2) mutations in oncogenes without any cases in COSMIC v85, 3) variants present in matched leukocyte DNA by Monte Carlo monitoring at a detection index of less than 0.1, 4) single-nucleotide polymorphisms from any patient in the cohort, 5) variants lying in repeat, intronic, intergenic, or pseudogene regions, 6) variants less than or equal to 50% of the median sample depth, 7) variants with a population allele frequency greater than or equal to 0.1% in the gnomAD database (102), 8) non-coding variants, 9) mutations in canonical CH genes, and 10) recurrent background artefacts specific to each targeted sequencing space.

For each SNV, the following features were annotated: 1) Bayesian background of the variant in germline samples, 2) variant allele frequency, 3) germline depth, 4) mean barcode family size, 5) short fragment score 1, 6) short fragment score 2, 7) transition/transversion, 8) duplex support, 9) pass outlier cutoff, 10) mapping quality, 11) lung cancer hotspot, 12) UID errors corrected, 13) mean Phred quality, 14) mean variant position in read, 15) power to detect the variant in the matched germline, 16) variant frequency in the cohort, 17) selector tile mutation frequency in the cohort, 18) Monte Carlo detection index for the variant in the germline, 19) distribution of reads mapped to the plus and minus strand, 20) mean number of non-reference bases in reads with the variant, 21) normalized depth, 22) normalized barcode family size, 23) gnomAD population allele frequency, 24) lung cancer driver gene, and 25) non-synonymous 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.

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 1) an allele fraction less than or equal to 2%, 2) Monte Carlo detection index less than or equal to 0.05 when monitoring in matched germline, 3) population allele frequency greater than or equal to 0.1% in the gnomAD database, and 4) 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 non-adjudicated variants in patients with matched tumors, and this model was used to label variants from NSCLC patients 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 to 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.

Circulating tumor DNA monitoring

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 500x 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 measured by Qubit (Thermo Fisher Scientific) to calculate the ctDNA concentration with each haploid genomic equivalent assumed to have a mass of 3.3 picograms. Tumor genotype-informed CAPP-Seq was utilized for ctDNA MRD analysis with variants called from tumor tissue or pre-treatment plasma as described previously (46).

Radiomic analysis

The radiomics model was constructed in accordance with the guidelines recommended by the Image Biomarker Standardization Initiative (IBSI) (103) using the RTOG 0235/ACRIN 6668 dataset (n=209) that is publicly available at the cancer imaging archive (TCIA) (104). The model was then independently validated in the MDACC cohort (n=62). The treatment planning CT scans were downloaded from TCIA for the RTOG cohort or retrieved from the local institutional PACS. The gross tumor volume was manually delineated for radiation treatment planning by the treating physician. A total of 14 knowledge-based radiomic features were calculated: 4 morphology features (volume, surface area, surface to volume ratio, and sphericity), 2 image intensity features (energy and entropy), 4 texture features (contrast, correlation, homogeneity, and joint entropy of the gray level co-occurrence matrix), tumor invasive margin blurriness, and 3 tumor-associated blood vessel features (number of blood vessels in direct contact with the tumor, vessel coverage, and vessel scattering).

Morphology features, image intensity features, and texture features were extracted based on the gross tumor volumes using an open-source python package PyRadiomics (V3.0.1) (105). To capture information on the invasive tumor margin, we computed a quantitative feature to evaluate the blurriness of a 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 morphological dilation and erosion operations. We then computed the 3D image gradient summed along three orthogonal axes to measure the blurriness of 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 two-dimensional entropy is calculated as follows:

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

where, $i$ represents the gray value of the pixel, $j$ represents the mean gray value of its neighborhood. The probability $P_{i, j}$ is defined as follows:

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

where $f (i, j)$ is the number of occurrences of two tuples $(i, j)$ , and $W$ and $H$ are the image sizes. The neighborhood consists of the 8 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.

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 false-discovery rate using the Benjamini-Hochberg procedure. For feature selection, we applied an elastic net model with the mixing parameter set to 0.5 and selected the optimal features at 1 standard error 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 × margin blurriness + 1.67 × 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.

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 with those based on the original gross tumor volume from the RTOG 0235 dataset. We evaluated reproducibility of image features using intra-class correlation coefficient (ICC), which varies between 0 and 1, with 1 indicating perfect agreement in measurements. According to the Cicchetti critera(110), ICC values above 0.85 are considered excellent.

Identification of pre-treatment prognostic factors

Sex, age as a continuous variable, stage III vs. stage II, largest lesion MTV as a continuous variable, largest lesion GTV as a continuous variable, and non-squamous cell carcinoma vs. squamous cell carcinoma histology were evaluated as pre-CRT biological prognostic factors for PFS in locoregionally advanced NSCLC using univariable Cox proportional hazards models in the historical training cohort of patients from TCGA (n=70) and a previously reported cohort of patients from Stanford (58,95) who did not undergo mid-CRT ctDNA testing as described above (n=38). GTV and MTV data was 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 squamous cell carcinoma or lung adenocarcinoma (61). Because three different sequencing panels were applied in the historical training cohort, we further limited our analysis to genes with at least 3 patients having pathogenic mutations (n=23 genes). Mutations in the oncogenes KRAS, EGFR, BRAF, and PIK3CA were considered pathogenic if at least 1 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 (CADD) PHRED scores ≥20 (111). All other mutations were considered pathogenic if their CADD PHRED score was ≥20. The association of pathogenic mutations in each gene were associated with PFS using univariate Cox proportional hazards models and P values were corrected for multiple hypothesis testing.

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 Greenwood’s formula. We defined a baseline survival function by combining all of the patients with locoregionally advanced NSCLC from the RTOG 0235, historical, and ctDNA training cohorts (n=357 patients). Because of the strong correlation between GTV and MTV (Figure S6A) and ease of measuring GTV in the clinic (since unlike MTV it is already being defined for every patient receiving CRT), we chose to evaluate GTV in the model. Survival curves and confidence intervals from the historical training cohort (TCGA and Stanford patients without ctDNA analysis) were used as prior knowledge to infer the hyper-parameters (mean $μ$ and variance $σ^{2}$ of each coefficient $β$ in the model) for histology, largest lesion GTV, KEAP1 status, and KRAS status. The hyper-parameters for the radiomic model were inferred from the RTOG 0235 dataset, and the hyper-parameters for mid-CRT ctDNA concentration and log fold change in ctDNA concentration were inferred from the ctDNA training cohort (MDACC patients). Having defined the hyper-parameters for each covariate, Cox partial likelihood was used as the likelihood function and Markov Chain Monte Carlo sampling was employed 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 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% confidence intervals. 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 hyper-parameter, choosing the model with fewest features, and validating the model in an independent cohort.

Statistics

PFS was defined as the time from the start of chemoradiation therapy 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 hazard ratios were calculated using Cox regression. Circulating tumor DNA concentrations were log-transformed for regression analyses to produce normally distributed data. All variables were standardized to enable comparison of hazard ratios and 95% confidence intervals from Cox models. The optimal cutoff for stratifying PFS by mid-CRT 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’s 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 confidence intervals and empiric P values performed from 2000 bootstrap resamplings. P values were corrected for multiple hypothesis testing using the Benjamini-Hochberg procedure with the ‘stats’ R package. 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).

Data and code availability

Anonymized clinical and demographic data for the patients in this study as well as cell-free DNA metrics, somatic mutation data, radiomic, biological, and molecular metrics, and CIRI-LCRT predicted progression-free survival are provided in the Supplementary Tables. DNA sequencing data generated in this study from patients enrolled at MD Anderson Cancer Center 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-CLIP 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.

Supplementary Material

NIHMS2077095-supplement-1.xlsx^{(547.1KB, xlsx)}

NIHMS2077095-supplement-2.pdf^{(2MB, pdf)}

STATEMENT OF SIGNIFICANCE.

This study demonstrates that combining tumor features, radiomics, and ctDNA analysis improves outcome prediction in non-small cell lung cancer treated with chemoradiation therapy. Our integrated model could enable personalized and response-adapted therapies to reduce toxicity and improve outcomes in patients.

ACKNOWLEDGEMENTS

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. Moding), the Radiological Society of North America (E. Moding), Conquer Cancer supported by GO2 Foundation for Lung Cancer (E. Moding), the National Cancer Institute (M. Diehn and A. Alizadeh: R01CA188298, R01CA244526, R01CA254179, and R. Li, M. Diehn and B. Loo: R01CA233578), the Virginia and D.K. Ludwig Fund for Cancer Research (M. Diehn and A. Alizadeh), the Bakewell Foundation (M. Diehn and A. Alizadeh), the SDW/DT and Shanahan Family Foundations (A. Alizadeh), and the CRK Faculty Scholar Fund (M. Diehn). 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. Schematics were created with BioRender.com.

CONFLICTS OF INTEREST

E.J.M. has served as a paid consultant for Guidepoint and GLG. M.S.E. has served as a paid consultant for Foresight Diagnostics. B.Y.N. is currently an employee and stockholder at Roche/Genentech. A.A.C. has served as an advisor/consultant for Roche, Tempus Labs, Geneoscopy, and Oscar Health, and has received speaker honoraria from Roche, Varian Medical Systems, and Foundation Medicine. M.Das has received research funding from Novartis, Abbvie, United Therapeutics, CellSight, Varian, and Verily and has served as a consultant for Astra Zeneca, Beigene, Sanofi, and Jazz Pharmaceuticals. S.K.P. has received research funding from BioAlta and has served as a consultant for AstraZeneca, Nanobiotix, Janssen Pharmaceuticals, Jazz Pharmaceuticals, Mirati Therapeutics, and Genentech. J.W.N. has served as an advisor/consultant for ARIAD/Takeda, AstraZeneca, Genentech/Roche, Lilly, Exelixis, Loxo, and Jounce. J.W.N. has received research funding from Genentech/Roche, Merck, Novartis, Boehringer Ingelheim, Exelixis, ARIAD/Takeda, and Nektar. H.A.W. has served on the advisory board for AstraZeneca, Xcovery, Janssen, Mirati, Merck, Takeda, and Genentech/Roche and has received compensation from AstraZeneca, Xcovery, Janssen, and Mirati. H.A.W. has received research funding from ACEA Biosciences, Arrys Therapeutics, AstraZeneca/Medimmune, BMS, Celgene, Clovis Oncology, Exelixis, Genentech/Roche, Gilead, Lilly, Merck, Novartis, Pfizer, Pharmacyclics, and Xcovery. S.H.L. receives grant funding from STCube Pharmaceuticals, Nektar Therapeutics, and Beyond Spring Pharmaceuticals, has served as a consultant for XRAD Therapeutics, has served on the advisory board for AstraZeneca and Creatv Microtech, and is a co-founder, scientific advisor and has stock options for Scenexo, Inc. A.A.A. and M.Diehn are co-inventors on patent applications related to CAPP-Seq. M.S.E., D.M.K., A.A.A., and M. Diehn are co-inventors on a patent application related to CIRI. M.S.E., J.J.J., D.M.K., A.A.A., and M. Diehn are co-inventors on a patent application related to the SNV score. A.A.A. has equity in CiberMed and Foresight Diagnostics and has served as a consultant for Roche, Genentech, Chugai, and Pharmacyclics. M.Diehn has equity in CiberMed and Foresight Diagnostics, has received research funding from AstraZeneca, Genentech, Illumina, and Varian Medical Systems, and has served as a paid consultant for Roche, AstraZeneca, Novartis, Genentech, Boehringer Ingelheim, BioNTech, Gritstone Oncology, Illumina, and RefleXion. The other authors declare no competing interests.

REFERENCES

1.Siegel RL, Miller KD, Jemal A. Cancer statistics, 2019. CA Cancer J Clin. 2019;69:7–34. [DOI] [PubMed] [Google Scholar]
2.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. [DOI] [PubMed] [Google Scholar]
3.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. [DOI] [PubMed] [Google Scholar]
4.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. [DOI] [PubMed] [Google Scholar]
5.Brundage MD, Davies D, Mackillop WJ. Prognostic factors in non-small cell lung cancer: a decade of progress. Chest. 2002;122:1037–57. [DOI] [PubMed] [Google Scholar]
6.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
7.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
8.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. [DOI] [PubMed] [Google Scholar]
9.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. [DOI] [PubMed] [Google Scholar]
10.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. Br J Cancer. 2017;117:744–51. [DOI] [PMC free article] [PubMed] [Google Scholar]
11.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
12.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. [DOI] [PubMed] [Google Scholar]
13.van Laar M, van Amsterdam WAC, van Lindert ASR, de Jong PA, Verhoeff JJC. Prognostic factors for overall survival of stage III non-small cell lung cancer patients on computed tomography: A systematic review and meta-analysis. Radiother Oncol. 2020;151:152–75. [DOI] [PubMed] [Google Scholar]
14.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
15.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. [DOI] [PubMed] [Google Scholar]
16.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
17.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
18.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
19.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
20.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
21.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. [DOI] [PubMed] [Google Scholar]
22.Gatta R, Depeursinge A, Ratib O, Michielin O, Leimgruber A. Integrating radiomics into holomics for personalised oncology: from algorithms to bedside. European Radiology Experimental. 2020;4:11. [DOI] [PMC free article] [PubMed] [Google Scholar]
23.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
24.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. [DOI] [PubMed] [Google Scholar]
25.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
26.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. [DOI] [PubMed] [Google Scholar]
27.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. [DOI] [PubMed] [Google Scholar]
28.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. [DOI] [PubMed] [Google Scholar]
29.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. [DOI] [PubMed] [Google Scholar]
30.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
31.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. [DOI] [PubMed] [Google Scholar]
32.Sorensen BS, Wu L, Wei W, Tsai J, Weber B, Nexo E, et al. Monitoring of epidermal growth factor receptor tyrosine kinase inhibitor-sensitizing 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. [DOI] [PMC free article] [PubMed] [Google Scholar]
33.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. Clinical Cancer Research. 2018;24:6212–22. [DOI] [PubMed] [Google Scholar]
34.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 Discovery. 2017;7:264–76. [DOI] [PMC free article] [PubMed] [Google Scholar]
35.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
36.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
37.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;183:363–376.e13. [DOI] [PMC free article] [PubMed] [Google Scholar]
38.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
39.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
40.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
41.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
42.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. [DOI] [PubMed] [Google Scholar]
43.Parikh AR, Mojtahed A, Schneider JL, Kanter K, Seventer EEV, Fetter IJ, et al. Serial ctDNA Monitoring to Predict Response to Systemic Therapy in Metastatic Gastrointestinal Cancers. Clin Cancer Res. 2020;26:1877–85. [DOI] [PMC free article] [PubMed] [Google Scholar]
44.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. Nature Cancer. 2020;1:873–81. [DOI] [PubMed] [Google Scholar]
45.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
46.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
47.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. [DOI] [PubMed] [Google Scholar]
48.Razavi P, Li BT, Brown DN, Jung B, Hubbell E, Shen R, et al. High-intensity sequencing reveals the sources of plasma circulating cell-free DNA variants. Nat Med. 2019;25:1928–37. [DOI] [PMC free article] [PubMed] [Google Scholar]
49.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
50.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
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. [DOI] [PMC free article] [PubMed] [Google Scholar]
52.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;23:700–713.e6. [DOI] [PMC free article] [PubMed] [Google Scholar]
53.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
54.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. [DOI] [PubMed] [Google Scholar]
55.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
56.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
57.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. [DOI] [PubMed] [Google Scholar]
58.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
59.Huncharek M 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. [DOI] [PubMed] [Google Scholar]
60.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. European Respiratory Journal. 2001;18:705–19. [DOI] [PubMed] [Google Scholar]
61.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;173:371–385.e18. [DOI] [PMC free article] [PubMed] [Google Scholar]
62.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. [DOI] [PubMed] [Google Scholar]
63.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
64.Jaiswal S, Fontanillas P, Flannick J, Manning A, Grauman PV, Mar BG, et al. Age-Related Clonal Hematopoiesis Associated with Adverse Outcomes. New England Journal of Medicine. 2014;371:2488–98. [DOI] [PMC free article] [PubMed] [Google Scholar]
65.Coombs CC, Zehir A, Devlin SM, Kishtagari A, Syed A, Jonsson P, et al. Therapy-Related Clonal Hematopoiesis in Patients with Non-hematologic Cancers Is Common and Associated with Adverse Clinical Outcomes. Cell Stem Cell. 2017;21:374–382.e4. [DOI] [PMC free article] [PubMed] [Google Scholar]
66.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
67.Cliffe H, Patel C, Prestwich R, Scarsbrook A. Radiotherapy response evaluation using FDG PET-CT—established and emerging applications. The British Journal of Radiology. 2017;90:20160764. [DOI] [PMC free article] [PubMed] [Google Scholar]
68.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
69.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
70.Garcia-Murillas I, Chopra N, Comino-Méndez I, Beaney M, Tovey H, Cutts RJ, et al. Assessment of Molecular Relapse Detection in Early-Stage Breast Cancer. JAMA Oncol. 2019;5:1473–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
71.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. [DOI] [PubMed] [Google Scholar]
72.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
73.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. [DOI] [PubMed] [Google Scholar]
74.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;41:1763–1773.e4. [DOI] [PubMed] [Google Scholar]
75.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
76.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
77.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. [DOI] [PubMed] [Google Scholar]
78.Gökbuget N, Kneba M, Raff T, Trautmann H, Bartram C-R, Arnold R, et al. Adult patients with acute lymphoblastic leukemia and molecular failure display a poor prognosis and are candidates for stem cell transplantation and targeted therapies. Blood. 2012;120:1868–76. [DOI] [PubMed] [Google Scholar]
79.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. [DOI] [PubMed] [Google Scholar]
80.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
81.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
82.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
83.Fave X, Zhang L, Yang J, Mackin D, Balter P, Gomez D, et al. Delta-radiomics features for the prediction of patient outcomes in non-small cell lung cancer. Sci Rep. 2017;7:588. [DOI] [PMC free article] [PubMed] [Google Scholar]
84.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 Patients. Cancer. 2011;117:2995–3003. [DOI] [PubMed] [Google Scholar]
85.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
86.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. Journal of Thoracic Oncology. 2015;10:1720–5. [DOI] [PubMed] [Google Scholar]
87.Alexander BM, Othus M, Caglar HB, Allen AM. Tumor Volume Is a Prognostic Factor in Non–Small-Cell Lung Cancer Treated With Chemoradiotherapy. International Journal of Radiation Oncology*Biology*Physics. 2011;79:1381–7. [DOI] [PubMed] [Google Scholar]
88.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
89.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. [DOI] [PubMed] [Google Scholar]
90.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. Nature Genetics. 2018;50:1735–43. [DOI] [PMC free article] [PubMed] [Google Scholar]
91.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
92.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. [DOI] [PubMed] [Google Scholar]
93.Daly ME, Singh N, Ismaila N, Management of Stage III NSCLC Guideline Expert Panel. Management of Stage III Non-Small Cell Lung Cancer: ASCO Guideline Rapid Recommendation Update. J Clin Oncol. 2024;42:3058–60. [DOI] [PubMed] [Google Scholar]
94.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. [DOI] [PubMed] [Google Scholar]
95.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
96.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. Journal of Thoracic Oncology. 2017;12:833–42. [DOI] [PMC free article] [PubMed] [Google Scholar]
97.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;173:400–416.e11. [DOI] [PMC free article] [PubMed] [Google Scholar]
98.Dupont WD, Plummer WD. Power and sample size calculations. A review and computer program. Control Clin Trials. 1990;11:116–28. [DOI] [PubMed] [Google Scholar]
99.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
100.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
101.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
102.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
103.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
104.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
105.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
106.Kahala G, Sklair M, Spitzer H. Multi-scale Blood Vessel Detection and Segmentation in Breast MRIs. J Med Biol Eng. 2019;39:424–30. [Google Scholar]
107.Liu G, Zheng X. Fabric defect detection based on information entropy and frequency domain saliency. Vis Comput. 2021;37:515–28. [Google Scholar]
108.Heagerty PJ, Lumley T, Pepe MS. Time-dependent ROC curves for censored survival data and a diagnostic marker. Biometrics. 2000;56:337–44. [DOI] [PubMed] [Google Scholar]
109.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. [DOI] [PMC free article] [PubMed] [Google Scholar]
110.Cicchetti DV. Guidelines, criteria, and rules of thumb for evaluating normed and standardized assessment instruments in psychology. Psychological Assessment. 1994;6:284–90. [Google Scholar]
111.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

NIHMS2077095-supplement-1.xlsx^{(547.1KB, xlsx)}

NIHMS2077095-supplement-2.pdf^{(2MB, pdf)}

Data Availability Statement

[R1] 1.Siegel RL, Miller KD, Jemal A. Cancer statistics, 2019. CA Cancer J Clin. 2019;69:7–34. [DOI] [PubMed] [Google Scholar]

[R2] 2.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. [DOI] [PubMed] [Google Scholar]

[R3] 3.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. [DOI] [PubMed] [Google Scholar]

[R4] 4.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. [DOI] [PubMed] [Google Scholar]

[R5] 5.Brundage MD, Davies D, Mackillop WJ. Prognostic factors in non-small cell lung cancer: a decade of progress. Chest. 2002;122:1037–57. [DOI] [PubMed] [Google Scholar]

[R6] 6.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R7] 7.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R8] 8.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. [DOI] [PubMed] [Google Scholar]

[R9] 9.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. [DOI] [PubMed] [Google Scholar]

[R10] 10.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. Br J Cancer. 2017;117:744–51. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R11] 11.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R12] 12.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. [DOI] [PubMed] [Google Scholar]

[R13] 13.van Laar M, van Amsterdam WAC, van Lindert ASR, de Jong PA, Verhoeff JJC. Prognostic factors for overall survival of stage III non-small cell lung cancer patients on computed tomography: A systematic review and meta-analysis. Radiother Oncol. 2020;151:152–75. [DOI] [PubMed] [Google Scholar]

[R14] 14.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R15] 15.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. [DOI] [PubMed] [Google Scholar]

[R16] 16.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R17] 17.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R18] 18.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R19] 19.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R20] 20.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R21] 21.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. [DOI] [PubMed] [Google Scholar]

[R22] 22.Gatta R, Depeursinge A, Ratib O, Michielin O, Leimgruber A. Integrating radiomics into holomics for personalised oncology: from algorithms to bedside. European Radiology Experimental. 2020;4:11. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R23] 23.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R24] 24.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. [DOI] [PubMed] [Google Scholar]

[R25] 25.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R26] 26.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. [DOI] [PubMed] [Google Scholar]

[R27] 27.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. [DOI] [PubMed] [Google Scholar]

[R28] 28.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. [DOI] [PubMed] [Google Scholar]

[R29] 29.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. [DOI] [PubMed] [Google Scholar]

[R30] 30.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R31] 31.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. [DOI] [PubMed] [Google Scholar]

[R32] 32.Sorensen BS, Wu L, Wei W, Tsai J, Weber B, Nexo E, et al. Monitoring of epidermal growth factor receptor tyrosine kinase inhibitor-sensitizing 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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R33] 33.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. Clinical Cancer Research. 2018;24:6212–22. [DOI] [PubMed] [Google Scholar]

[R34] 34.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 Discovery. 2017;7:264–76. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R35] 35.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R36] 36.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R37] 37.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;183:363–376.e13. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R38] 38.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R39] 39.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R40] 40.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R41] 41.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R42] 42.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. [DOI] [PubMed] [Google Scholar]

[R43] 43.Parikh AR, Mojtahed A, Schneider JL, Kanter K, Seventer EEV, Fetter IJ, et al. Serial ctDNA Monitoring to Predict Response to Systemic Therapy in Metastatic Gastrointestinal Cancers. Clin Cancer Res. 2020;26:1877–85. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R44] 44.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. Nature Cancer. 2020;1:873–81. [DOI] [PubMed] [Google Scholar]

[R45] 45.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R46] 46.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R47] 47.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. [DOI] [PubMed] [Google Scholar]

[R48] 48.Razavi P, Li BT, Brown DN, Jung B, Hubbell E, Shen R, et al. High-intensity sequencing reveals the sources of plasma circulating cell-free DNA variants. Nat Med. 2019;25:1928–37. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R49] 49.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R50] 50.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R51] 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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R52] 52.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;23:700–713.e6. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R53] 53.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R54] 54.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. [DOI] [PubMed] [Google Scholar]

[R55] 55.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R56] 56.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R57] 57.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. [DOI] [PubMed] [Google Scholar]

[R58] 58.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R59] 59.Huncharek M 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. [DOI] [PubMed] [Google Scholar]

[R60] 60.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. European Respiratory Journal. 2001;18:705–19. [DOI] [PubMed] [Google Scholar]

[R61] 61.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;173:371–385.e18. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R62] 62.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. [DOI] [PubMed] [Google Scholar]

[R63] 63.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R64] 64.Jaiswal S, Fontanillas P, Flannick J, Manning A, Grauman PV, Mar BG, et al. Age-Related Clonal Hematopoiesis Associated with Adverse Outcomes. New England Journal of Medicine. 2014;371:2488–98. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R65] 65.Coombs CC, Zehir A, Devlin SM, Kishtagari A, Syed A, Jonsson P, et al. Therapy-Related Clonal Hematopoiesis in Patients with Non-hematologic Cancers Is Common and Associated with Adverse Clinical Outcomes. Cell Stem Cell. 2017;21:374–382.e4. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R66] 66.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R67] 67.Cliffe H, Patel C, Prestwich R, Scarsbrook A. Radiotherapy response evaluation using FDG PET-CT—established and emerging applications. The British Journal of Radiology. 2017;90:20160764. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R68] 68.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R69] 69.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R70] 70.Garcia-Murillas I, Chopra N, Comino-Méndez I, Beaney M, Tovey H, Cutts RJ, et al. Assessment of Molecular Relapse Detection in Early-Stage Breast Cancer. JAMA Oncol. 2019;5:1473–8. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R71] 71.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. [DOI] [PubMed] [Google Scholar]

[R72] 72.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R73] 73.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. [DOI] [PubMed] [Google Scholar]

[R74] 74.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;41:1763–1773.e4. [DOI] [PubMed] [Google Scholar]

[R75] 75.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R76] 76.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R77] 77.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. [DOI] [PubMed] [Google Scholar]

[R78] 78.Gökbuget N, Kneba M, Raff T, Trautmann H, Bartram C-R, Arnold R, et al. Adult patients with acute lymphoblastic leukemia and molecular failure display a poor prognosis and are candidates for stem cell transplantation and targeted therapies. Blood. 2012;120:1868–76. [DOI] [PubMed] [Google Scholar]

[R79] 79.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. [DOI] [PubMed] [Google Scholar]

[R80] 80.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R81] 81.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R82] 82.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R83] 83.Fave X, Zhang L, Yang J, Mackin D, Balter P, Gomez D, et al. Delta-radiomics features for the prediction of patient outcomes in non-small cell lung cancer. Sci Rep. 2017;7:588. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R84] 84.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 Patients. Cancer. 2011;117:2995–3003. [DOI] [PubMed] [Google Scholar]

[R85] 85.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R86] 86.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. Journal of Thoracic Oncology. 2015;10:1720–5. [DOI] [PubMed] [Google Scholar]

[R87] 87.Alexander BM, Othus M, Caglar HB, Allen AM. Tumor Volume Is a Prognostic Factor in Non–Small-Cell Lung Cancer Treated With Chemoradiotherapy. International Journal of Radiation Oncology*Biology*Physics. 2011;79:1381–7. [DOI] [PubMed] [Google Scholar]

[R88] 88.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R89] 89.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. [DOI] [PubMed] [Google Scholar]

[R90] 90.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. Nature Genetics. 2018;50:1735–43. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R91] 91.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R92] 92.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. [DOI] [PubMed] [Google Scholar]

[R93] 93.Daly ME, Singh N, Ismaila N, Management of Stage III NSCLC Guideline Expert Panel. Management of Stage III Non-Small Cell Lung Cancer: ASCO Guideline Rapid Recommendation Update. J Clin Oncol. 2024;42:3058–60. [DOI] [PubMed] [Google Scholar]

[R94] 94.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. [DOI] [PubMed] [Google Scholar]

[R95] 95.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R96] 96.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. Journal of Thoracic Oncology. 2017;12:833–42. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R97] 97.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;173:400–416.e11. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R98] 98.Dupont WD, Plummer WD. Power and sample size calculations. A review and computer program. Control Clin Trials. 1990;11:116–28. [DOI] [PubMed] [Google Scholar]

[R99] 99.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R100] 100.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R101] 101.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R102] 102.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R103] 103.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R104] 104.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R105] 105.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R106] 106.Kahala G, Sklair M, Spitzer H. Multi-scale Blood Vessel Detection and Segmentation in Breast MRIs. J Med Biol Eng. 2019;39:424–30. [Google Scholar]

[R107] 107.Liu G, Zheng X. Fabric defect detection based on information entropy and frequency domain saliency. Vis Comput. 2021;37:515–28. [Google Scholar]

[R108] 108.Heagerty PJ, Lumley T, Pepe MS. Time-dependent ROC curves for censored survival data and a diagnostic marker. Biometrics. 2000;56:337–44. [DOI] [PubMed] [Google Scholar]

[R109] 109.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R110] 110.Cicchetti DV. Guidelines, criteria, and rules of thumb for evaluating normed and standardized assessment instruments in psychology. Psychological Assessment. 1994;6:284–90. [Google Scholar]

[R111] 111.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. [DOI] [PMC free article] [PubMed] [Google Scholar]

PERMALINK

Integrating circulating tumor DNA analysis and radiomics for dynamic risk assessment in localized lung cancer

Everett J Moding

Mohammad Shahrokh Esfahani

Cheng Jin

Angela B Hui

Barzin Y Nabet

Yufei Liu

Jacob J Chabon

Michael S Binkley

David M Kurtz

Emily G Hamilton

Aadel A Chaudhuri

Chih Long Liu

Zhe Li

Rene F Bonilla

Alice L Jiang

Brianna C Lau

Pablo Lopez

Jianzhong He

Yawei Qiao

Ting Xu

Luyang Yao

Saumil Gandhi

Zhongxing Liao

Millie Das

Kavitha J Ramchandran

Sukhmani K Padda

Joel W Neal

Heather A Wakelee

Michael F Gensheimer

Billy W Loo Jr

Ruijiang Li

Steven H Lin

Ash A Alizadeh

Maximilian Diehn

Abstract

INTRODUCTION

RESULTS

Clonal hematopoiesis dynamics during CRT

Figure 1: Filtering clonal hematopoiesis (CH) and applying the machine learning-based SNV score improves tissue-free variant calling.

Tissue-free variant calling with machine learning

Circulating tumor DNA kinetics during CRT

Figure 2: Circulating tumor DNA levels during chemoradiation (CRT) are prognostic of progression-free survival (PFS).

Prediction of progression-free survival using radiomics

Figure 3: Radiomic analysis of pre-treatment CT images predicts progression-free survival (PFS).

Biological and molecular prognostic factors in locoregionally advanced NSCLC

Figure 4: Biological and molecular prognostic factors in patients with non-small cell lung cancer (NSCLC) treated with chemoradiation therapy (CRT).

A dynamic risk index for NSCLC

Figure 5: Training and validation of CIRI-LCRT for prediction of progression-free survival (PFS) during chemoradiation (CRT) for non-small cell lung cancer (NSCLC).

Comparing CIRI-LCRT with ctDNA MRD

Figure 6: CIRI-LCRT during chemoradiation therapy (CRT) performs comparably to ctDNA MRD after completion of treatment.

DISCUSSION

MATERIALS AND METHODS

Study design and patients

Power considerations

Chemoradiation therapy and blood collection

Library preparation and sequencing

Clonal hematopoiesis analysis

SNV score tissue-free variant calling

Circulating tumor DNA monitoring

Radiomic analysis

Identification of pre-treatment prognostic factors

CIRI-LCRT

Statistics

Data and code availability

Supplementary Material

STATEMENT OF SIGNIFICANCE.

ACKNOWLEDGEMENTS

CONFLICTS OF INTEREST

REFERENCES

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases