A decades-old treatment without personalization. For patients with locally advanced, unresectable non-small cell lung cancer (NSCLC), chemoradiation therapy (CRT) has been the standard treatment for over 40 years, delivered at essentially the same dose to all patients. The majority still develop progressive disease after treatment, yet there are currently no reliable methods to identify during treatment which patients will respond and which will not.
Why monitoring during treatment matters. If doctors could identify mid-treatment that a patient is responding poorly, they could potentially intensify therapy - or conversely, reduce radiation dose for patients responding exceptionally well to spare them unnecessary toxicity. This concept of 'response-adapted therapy' has succeeded in blood cancers but has never been demonstrated in solid tumors like lung cancer.
Limitations of existing monitoring tools. PET/CT scans during CRT are complicated by normal tissue reactions from radiation that can mask or mimic tumor response. No reliable imaging or blood-based test has been established for monitoring NSCLC treatment response in real time, leaving oncologists without guidance to individualize treatment during the course of therapy.
Two emerging approaches. This study focuses on combining two promising technologies: liquid biopsies measuring circulating tumor DNA (ctDNA) - fragments of tumor DNA shed into the bloodstream - and radiomics, which uses computer algorithms to extract quantitative features from CT scan images. The researchers hypothesized these two approaches provide complementary information that together could identify high- and low-risk patients during treatment.
Not all blood DNA comes from tumors. Liquid biopsies detect tumor DNA circulating in blood - but most DNA fragments in blood actually come from normal blood-producing (hematopoietic) cells. As people age, blood stem cells can accumulate mutations and form dominant clones, a process called clonal hematopoiesis (CH). These non-cancer mutations can contaminate ctDNA readings and mimic tumor signals.
CRT changes clonal hematopoiesis patterns. A key finding of this study was that chemoradiation therapy differentially affects different CH mutations within just weeks. PPM1D mutations expanded significantly during treatment, while SF3B1 and TET2 mutations decreased. This means CH cannot simply be removed once at the start and ignored - it is a dynamic, shifting source of potential interference throughout treatment.
Solution: mandatory leukocyte filtering. The researchers addressed this challenge by performing deep DNA sequencing of matched white blood cells (leukocytes) from each patient alongside their plasma samples. Any variant found in white blood cells was removed from the ctDNA analysis, regardless of its apparent allele fraction. This stringent filtering approach was essential for reliable monitoring.
Machine learning to distinguish tumor DNA from noise. Because tumor tissue biopsies were only available for 15% of patients in this study, the team developed a machine learning algorithm called the SNV Score to identify tumor-derived mutations directly from blood plasma. By analyzing 25 biological and technical features of each DNA variant, this model successfully distinguished tumor mutations from technical artifacts and CH variants, enabling 'tissue-free' ctDNA monitoring for the vast majority of patients.
ctDNA drops dramatically during CRT. Across all 61 analyzed patients, ctDNA concentration fell by a median of 13-fold between pre-treatment and mid-treatment blood draws (10-30 days into a typical 6-week course). This rapid decrease reflects the tumor cells being killed by radiation and chemotherapy, releasing less DNA into the bloodstream as the tumor shrinks.
The amount of ctDNA remaining mid-treatment is what matters. Patients whose ctDNA level fell below 3.2 haploid genome equivalents per milliliter mid-treatment had significantly better progression-free survival in both the training cohort (6.4-fold reduced risk of progression) and the independent validation cohort (4.2-fold reduction). Pre-treatment ctDNA levels were not predictive - it is the mid-treatment residual level that carries prognostic information.
Pattern of failure provides biological insight. Patients with very low mid-treatment ctDNA who still progressed tended to develop isolated local recurrences (within the radiation field) rather than distant metastases. This makes biological sense - a low systemic ctDNA level suggests low tumor burden outside the radiation zone, while local disease might persist within the treated area despite treatment.
Validated in an independent cohort. The ctDNA threshold identified in the training cohort (MD Anderson Cancer Center) was applied without modification to the independent validation cohort (Stanford University) and showed nearly identical prognostic performance. This cross-institutional validation is critical for demonstrating that findings are generalizable and not specific to a single institution's patient population or treatment practices.
Extracting quantitative features from CT images. Using pre-treatment CT scans, the researchers calculated 14 quantitative radiomic features for each tumor, including measures of tumor shape, internal texture, image density, and the relationship between the tumor and its surrounding blood vessels. These features were then tested for their ability to predict which patients would eventually have disease progression after CRT.
Two features emerged as most predictive. After rigorous feature selection, the final radiomic model incorporated just two novel measurements: the blurriness of the tumor invasive margin (how well-defined the tumor edge is on CT) and vessel scattering (how evenly blood vessels are distributed around the tumor surface). These features likely reflect the biological aggressiveness and vascular characteristics of the tumor.
Validated across independent datasets. The radiomic model was trained on a publicly available dataset of 209 NSCLC patients (RTOG 0235/ACRIN 6668 trial) and validated independently in MD Anderson patients. Patients with a high radiomic score had a hazard ratio for disease progression of 2.9 in training and 6.1 in validation - indicating very strong prognostic performance that reproduced in a completely different patient population.
Practical for routine clinical use. The radiomic score can be calculated from both contrast-enhanced and non-contrast CT scans - the type patients routinely receive for treatment planning - without requiring any additional imaging tests. Its two component features showed excellent reproducibility when calculated from independently generated tumor outlines, suggesting the measurements are robust to variations in how tumors are defined.
Combining the pieces into one model. The Continuous Individualized Risk Index for Lung Cancer Chemoradiation Therapy (CIRI-LCRT) combines three prognostic factors: tumor histology (squamous vs. non-squamous), the pre-treatment radiomic score, and mid-treatment ctDNA concentration. This model generates a personalized probability of disease progression or death that can be updated in real time as new data becomes available during treatment.
Dramatically better than any single measure. CIRI-LCRT achieved a C-statistic of 0.94 at 24 months in the independent validation cohort - far exceeding any individual biomarker (mid-CRT ctDNA alone, radiomics alone, or histology alone) and substantially outperforming all previously published prognostic models for this disease, which had C-statistics of only 0.62 to 0.76.
Real-world application illustrated. Two patients with identical Stage IIIA diagnoses showed starkly different CIRI-LCRT predictions. One had a low radiomic score that dropped to 12% predicted progression risk after low mid-treatment ctDNA, and remained disease-free 25 months later. The other had a high radiomic score and very high mid-CRT ctDNA (37.8 hGE/ml), giving 100% predicted risk - and developed local recurrence and brain metastases within 6 months.
Earlier predictions than waiting for post-treatment ctDNA. Compared to detecting residual disease after all treatment is complete, CIRI-LCRT provided predictions a median of 3 months earlier for patients who ultimately progressed. Earlier identification of high-risk patients could enable earlier escalation to additional therapies like consolidation immunotherapy or radiation boosts, potentially improving survival outcomes.
A framework for response-adapted therapy. The authors outline a practical clinical workflow: radiomic analysis from routine pre-treatment CT; blood draw at day 10 of treatment for ctDNA; results available by week 3 of a typical 6-week course. This timing would allow systemic therapy modifications or radiation boost planning for the second half of treatment in patients identified as high-risk.
Identifying candidates for additional therapies. High CIRI-LCRT risk scores could guide which patients should receive consolidation immunotherapy (like durvalumab) after CRT, which has been shown to improve survival in locally advanced NSCLC. The model's ability to identify non-squamous histology as a risk factor aligns with evidence that this patient group benefits most from consolidation immunotherapy.
Treatment de-escalation for low-risk patients. Conversely, patients identified as low-risk by CIRI-LCRT might be candidates for radiation dose reduction, sparing them toxicity. The authors note that evidence suggests some patients may be cured with lower radiation doses than currently standard, but no reliable tool existed to identify such patients until now.
Validation and next steps needed. The study's main limitations include its retrospective design, modest validation cohort size (21 patients), and the fact that most patients were treated before consolidation immunotherapy became standard care. Prospective validation in larger cohorts and ultimately randomized trials testing whether CIRI-LCRT-guided treatment adjustments actually improve patient outcomes will be required before clinical implementation.