Consolidation immunotherapy improves survival but benefits are not universal. Limited stage small-cell lung cancer, representing inoperable stage III disease, has a median overall survival of 12 to 18 months -- substantially worse than stage I/II disease. The ADRIATIC trial demonstrated that post-definitive chemoradiotherapy consolidation with durvalumab improved median overall survival from 33.4 to 55.9 months (hazard ratio 0.73) and progression-free survival from 9.2 to 16.6 months. Despite this population-level benefit, not all patients benefit equally, and biomarkers to identify who will benefit from consolidation immunotherapy are urgently needed.
Tissue-based biomarker research in LS-SCLC is limited by a prevalent practical constraint: small amounts of available baseline tumor tissue. While recurrent genetic alterations are well characterized in SCLC -- with TP53 and RB1 mutations detected in approximately 93% and 82% of tumors respectively -- most other mutations have low prevalence and limited biomarker utility. The scarcity of tissue restricts the depth of exploratory biomarker analysis that can be conducted at baseline.
Plasma circulating tumor DNA provides a non-invasive liquid biopsy alternative with high accuracy and concordance compared to tissue biopsies. Prior studies have demonstrated prognostic value of ctDNA detection in LS-SCLC patients undergoing definitive chemoradiotherapy, but large cohort studies examining ctDNA dynamics at multiple treatment time points and linking them to both prognosis and immunotherapy benefit had not been conducted before this work.
PTEN alterations associate with immune-active tumors and improved survival. Analysis of baseline tumor tissue from 203 LS-SCLC patients identified PTEN mutations as the only gene mutated in at least 5% of patients that was significantly associated with both better progression-free survival (p = 0.047) and overall survival (p = 0.040) in patients receiving definitive chemoradiotherapy. A consistent trend toward improved OS was observed in an independent external cohort of 218 LS-SCLC patients from cBioPortal (p = 0.064).
Transcriptomic analysis of the external cohort revealed that PTEN-mutant tumors were enriched in multiple immune-related signaling pathways compared to PTEN wild-type tumors. Enriched pathways included antigen processing and presentation (normalized enrichment score 1.840), T cell-mediated immunity (NES 1.185), response to interferon-gamma (NES 2.014), positive regulation of T-cell activation (NES 1.590), and positive regulation of immune response (NES 1.446). This immune enrichment profile suggests that PTEN mutation marks an immunologically active tumor phenotype with enhanced T-cell surveillance and better responsiveness to chemoradiotherapy.
Despite its biological significance, PTEN mutation's low prevalence limits its utility as a broadly applicable prognostic biomarker. This limitation drove the development of a more universally applicable ctDNA-based algorithm that could provide prognostic stratification independent of which specific mutations were present in each patient's tumor.
Serial ctDNA from two treatment time points and a four-variable prediction model. The ctDNA analysis enrolled 81 inoperable LS-SCLC patients who received definitive chemoradiotherapy alone and had plasma ctDNA samples at both the post-induction chemotherapy (post-ICT) and thoracic radiotherapy (TRT) time points. The cohort was split into a training cohort (n = 49) and a test cohort (n = 32) by enrollment date. Plasma cfDNA underwent high-depth targeted NGS sequencing at 30,000x depth using the customized Pulmocan panel, with ctDNA positivity defined by detection of at least one somatic alteration after filtering clonal hematopoiesis variants.
ctDNA detection at post-ICT (p less than 0.01) and during TRT (p = 0.012) was each independently associated with significantly worse progression-free survival. Elastic-net penalized Cox regression repeated 100 times with variable selection for inclusion at 40% or higher frequency selected four predictors: tumor shrinkage after ICT, prophylactic cranial irradiation treatment, post-ICT ctDNA detection, and TRT ctDNA detection. Greater tumor shrinkage and PCI receipt were associated with improved PFS, while ctDNA detection at either time point associated with unfavorable PFS.
The initial Elastic-net Cox model achieved time-dependent AUC of 0.794 for predicting 3-year progression in the training cohort and 0.651 in leave-one-out cross-validation. In the test cohort, performance was modestly reduced with AUC of 0.671 at 36 months. A key finding was that using a 12-month rather than 36-month cutoff improved discrimination in the test cohort, because a higher risk score cutoff better identifies patients with aggressive disease prone to early relapse -- the dominant failure pattern in LS-SCLC where most relapses occur within the first two years of treatment.
A learning model designed for small LS-SCLC cohorts. Given the universal challenge of small available cohorts in rare and aggressive diseases like LS-SCLC, the study developed a Bayesian survival prediction algorithm based on lognormal family modeling with 10,000 iterations and non-informative priors, trained on the same four-variable predictor set. The Bayesian approach supports sequential learning: posterior distributions from training serve as priors for updating when new evidence (the test cohort) is incorporated, enabling continuous knowledge refinement as data accumulates.
The Bayesian algorithm achieved AUC of 0.796 for predicting progression by end of follow-up in the training cohort and 0.788 in LOOCV. Time-dependent AUC at 36 months was 0.763 training and 0.796 LOOCV -- slightly superior to the Elastic-net model. Risk score cutoffs were determined using 1,000 bootstrap iterations: a 36-month cutoff of 0.235 and a 12-month cutoff of 0.463 were established, with lower scores indicating lower progression risk and significantly superior PFS.
The posterior Bayesian algorithm incorporating test cohort data as new evidence produced updated cutoffs of 0.319 (36-month) and 0.457 (12-month). Applied to all patients with available variables, the updated algorithm significantly separated PD from non-PD patients (p less than 0.001) and risk classification remained an independent predictor of both PFS and OS in multivariable Cox regression controlling for age, sex, smoking status, disease stage, and PCI treatment. The 2-year OS rate of patients in the low-risk subgroup was approximately 75%, comparable to stage I SCLC patients eligible for surgical treatment.
Consolidation immunotherapy benefit is exclusive to high-risk patients. The central clinical application of the Bayesian algorithm was tested in 86 LS-SCLC patients who received post-definitive chemoradiotherapy consolidation immunotherapy with anti-PD-L1 agents (predominantly durvalumab, serplulimab, or sugemalimab). The updated Bayesian algorithm was applied to predict counterfactual progression risk in these patients, and their outcomes were compared to those of dCRT-only patients within each risk stratum.
Significantly improved progression-free survival from consolidation immunotherapy was exclusively observed in patients with Bayesian prediction scores above 0.457 -- the highest-risk group (p less than 0.001, confirmed in multivariable analysis). No significant improvement in PFS was observed in patients with scores between 0.319 and 0.457 (p = 0.629), nor in patients with scores below 0.319 (p = 0.16). This pattern was consistent with a dose-response relationship: adjusted hazard ratios for immunotherapy benefit decreased progressively as the high-risk cutoff was raised, meaning that patients more likely to develop progression gained the most benefit from consolidation immunotherapy.
This finding is biologically coherent: ctDNA detection during definitive chemoradiotherapy reflects residual tumor burden and minimal residual disease status after treatment. Patients with detectable ctDNA -- who are more likely to be classified as high-risk -- have residual viable cancer cells that immune checkpoint blockade may eliminate. The parallel with the ADRIATIC trial, which showed more obvious durvalumab efficacy in patients with partial response or stable disease to concurrent chemoradiation, supports the hypothesis that high-risk ctDNA-positive patients represent the population with most to gain from adding immunotherapy.
Small cohorts, retrospective design, and the need for prospective validation. The primary limitation of this study is the relatively small sample size of the ctDNA cohorts, which is common in LS-SCLC research due to the rarity of the disease subtype and its aggressive course. To mitigate overfitting, the study used leave-one-out cross-validation, repeated variable selection across 100 elastic-net runs with a 40% inclusion threshold, and the Bayesian framework's inherent handling of uncertainty in small datasets. These methodological safeguards reduce but do not eliminate the risk of overfitting.
Absence of pre-treatment baseline ctDNA samples for the dCRT and immunotherapy cohorts prevents comparison of on-treatment ctDNA dynamics against baseline tumor burden. Detailed clinical data such as baseline tumor volume were not uniformly available in the retrospective records, precluding their inclusion in the prognostic algorithm. The time points of ctDNA testing during thoracic radiotherapy were not uniform across patients, limiting the ability to investigate whether ctDNA dynamics during TRT affect model performance differently depending on when in the treatment course sampling occurred.
Prospective cohort studies or clinical trials are required to confirm that high-risk patients identified by the ctDNA-based prediction algorithm benefit meaningfully from post-definitive chemoradiotherapy consolidation immunotherapy. External cohorts recruited at other centers with larger sample sizes are needed to mitigate overfitting concerns. The Bayesian framework's sequential updating capability positions it well for multi-institutional data harmonization, where patient data from different centers can be incorporated progressively to refine posterior estimates.
A dual-purpose biomarker for prognosis and immunotherapy selection. This study demonstrates that serial ctDNA monitoring during definitive chemoradiotherapy, combined with prophylactic cranial irradiation receipt and tumor shrinkage assessment, can effectively predict disease progression in inoperable LS-SCLC. The four-variable algorithm provides individualized risk scores that stratify patients into clinically meaningful subgroups with substantially different survival trajectories.
The clinical application extends beyond prognosis prediction. High-risk patients identified by the Bayesian algorithm -- those most likely to progress under definitive chemoradiotherapy alone -- are also the specific subgroup who derive statistically significant survival benefit from adding consolidation immunotherapy. Low-risk patients, conversely, show outcomes comparable to surgically eligible stage I SCLC patients and may not need the additional toxicity burden and cost of consolidation immunotherapy.
Together, these findings support a biomarker-informed treatment selection strategy for LS-SCLC: ctDNA monitoring during definitive chemoradiotherapy, combined with clinical variables, identifies patients who should be prioritized for consolidation immunotherapy versus those who may be managed adequately with definitive chemoradiotherapy alone. The identification of PTEN mutations as a favorable tissue-based prognostic signal linked to immune enrichment adds a complementary molecular layer for future integration into comprehensive biomarker panels.