Biomarkers matter but are hard to obtain. PD-L1 expression and tumor mutational burden (TMB) are the two most important biomarkers for selecting lung cancer patients for immunotherapy. However, obtaining these measurements requires invasive tissue biopsy, which is not always feasible due to contraindications, inaccessible tumor locations, or insufficient sample quality.
PET/CT as a non-invasive window. 18F-FDG PET/CT - a nuclear imaging technique that measures how actively tumors consume glucose - has shown promise as a non-invasive surrogate for these biomarkers. Prior work demonstrated that PET metabolic parameters could predict major pathological responses to neoadjuvant immunotherapy with 100% sensitivity and specificity.
The gap this study fills. While previous models used PET/CT features to predict PD-L1 expression, no models had yet been developed to predict TMB, and none had attempted to simultaneously predict both PD-L1 and TMB in a single patient. This combined prediction is clinically valuable because using both biomarkers together offers superior predictive value for immunotherapy selection compared to either alone.
Adenocarcinoma versus squamous cell carcinoma. Lung adenocarcinoma (ADC) and squamous cell carcinoma (SCC) have distinct biology, immune landscapes, and treatment response patterns. ADC shows greater treatment variability due to higher intratumoral heterogeneity and driver gene mutations, making biomarker-guided patient selection especially critical for this subtype.
Three-cohort structure. The primary cohort included 305 NSCLC patients (183 ADC, 122 SCC) from a single center between January 2017 and April 2024, all of whom had both PD-L1 expression and TMB tested simultaneously. ADC patients were randomly split 70:30 into training (n=128) and validation (n=55) cohorts for model development. An independent test cohort of 29 ADC patients receiving neoadjuvant immunotherapy was used to assess real-world clinical utility.
Biomarker definitions. PD-L1 positivity was defined as tumor proportion score (TPS) of 1% or greater, with high expression at 50% or greater. TMB was measured by next-generation sequencing, with high vs. low classification based on the median TMB value within each histological subtype (5 mutations/Mb for ADC, 10 mutations/Mb for SCC).
PET/CT acquisition and analysis. Whole-body PET/CT was performed after at least 6 hours of fasting. Metabolic parameters derived from automated tumor segmentation included SULmax (maximum standardized uptake value corrected for lean body mass), SULmean, SULpeak, metabolic tumor volume (MTV), and total lesion glycolysis (TLG). These were analyzed alongside semantic CT features assessed by two blinded experienced radiologists.
Model construction strategy. Three model types were built for each prediction target: a clinical-only model, a SULmax-only model, and a combined clinical-SULmax model. Nomograms were used to visualize the models for practical clinical use, with performance assessed by AUC, calibration curves, and decision curve analysis.
TMB differs substantially by subtype. While PD-L1 expression levels were similar between ADC and SCC (p=0.905), TMB was significantly higher in squamous cell carcinoma (median 10 vs. 5 mutations/Mb, p less than 0.001). This confirms that the two subtypes carry fundamentally different genomic profiles that affect how well they respond to different immunotherapy strategies.
PD-L1 and TMB are independent of each other. In both ADC and SCC, PD-L1 expression and TMB were not correlated with each other (Spearman's rho of -0.042 for ADC, 0.047 for SCC). This independence means measuring both biomarkers together provides complementary information, capturing different aspects of the tumor's immune response landscape.
Metabolic parameters correlate with biomarkers in ADC only. In ADC, PET metabolic parameters showed significant positive correlations with both PD-L1 expression (r=0.15-0.31) and TMB (r=0.18-0.30). Crucially, no such correlations were found in SCC. This subtype-specific relationship explains why PET/CT-based biomarker prediction models could be developed for ADC but not for SCC.
Why ADC differs from SCC metabolically. ADC and SCC exhibit distinct driver gene profiles that directly modulate tumor metabolism. In SCC, glucose uptake genes are predominantly expressed by malignant cells themselves, while in ADC they are expressed by surrounding stromal cells. Additionally, ADC has a more active tumor immune microenvironment with greater diversity of tumor-infiltrating immune cells compared to SCC.
SULmax predicts PD-L1 expression. Multivariate logistic regression identified SULmax as an independent predictor of PD-L1-Positive status in ADC (odds ratio 1.157, p less than 0.001). Patients with higher maximum glucose uptake on PET/CT were significantly more likely to express PD-L1 on their tumor cells, reflecting the biological link between metabolic activity and immune checkpoint expression.
SULmax predicts TMB. SULmax was also an independent predictor of TMB-High status in ADC (odds ratio 1.165, p less than 0.001), alongside older age (odds ratio 1.039) and EGFR-negative mutation status (odds ratio 0.330). Patients with EGFR mutations were significantly less likely to have high TMB, consistent with prior evidence that EGFR-driven tumors tend to have fewer random mutations.
Simultaneous prediction of both biomarkers. For the combined PD-L1-Positive and TMB-High prediction, age and SULmax were independent predictors (both p less than 0.05). For PD-L1-Negative and TMB-Low prediction, non-smoking status and lower SULmax were the key predictors, effectively identifying patients least likely to benefit from immunotherapy.
Why SULmax was preferred over other PET parameters. Among all PET metabolic parameters, SULmax was selected through backward stepwise regression because it contributed most to model discrimination. Other parameters including SULmean, SULpeak, MTV, and TLG were excluded as collinear variables that added no discriminatory value. SULmax also has practical advantages: it is the most widely used PET parameter clinically and requires only identification of the single highest uptake voxel, without specialized post-processing.
Combined models outperform single-biomarker models. The clinical-SULmax combined models for predicting both PD-L1 and TMB simultaneously outperformed models predicting either biomarker alone. For the PD-L1-Positive and TMB-High model, the combined approach achieved an AUC of 0.805 in training and 0.724 in validation. For the PD-L1-Negative and TMB-Low model, AUCs were 0.798 and 0.744 respectively.
Single-biomarker model performance. When predicting each biomarker separately, the combined clinical-SULmax models achieved more limited but still useful performance: AUC of 0.655 (training) and 0.689 (validation) for PD-L1-Positive alone, and 0.765 (training) and 0.705 (validation) for TMB-High alone. The superior joint prediction performance confirms that clinical and SULmax features together capture the multi-molecular characteristics of tumor biology more comprehensively.
Nomograms for clinical use. The predictive models were visualized as nomograms - graphical tools that allow clinicians to estimate a patient's probability of being in each biomarker category by summing weighted scores for each variable. This format makes the models practical for bedside use without requiring computational tools.
Calibration and clinical utility confirmed. Calibration curves and decision curve analysis confirmed that the combined models had good agreement between predicted and observed biomarker status and provided net clinical benefit over both treating all patients as biomarker-positive or all as biomarker-negative across a range of clinical decision thresholds.
Testing models against actual treatment outcomes. In 29 ADC patients receiving neoadjuvant immunotherapy, the predictive models were applied to test whether they could stratify actual pathological responses after treatment. The percentage of viable residual tumor tissue after surgery provided an objective measure of how well immunotherapy had worked.
Both models confirmed clinically useful. Patients predicted to be PD-L1-Positive and TMB-High showed significantly higher pathological regression (greater tumor destruction) compared to those predicted otherwise (p=0.035). More strikingly, patients predicted to be PD-L1-Negative and TMB-Low showed significantly lower pathological regression (p=0.001), confirming the model's ability to identify patients unlikely to respond.
Bridging prediction and outcomes. This real-world validation is particularly important because it demonstrates that the model does not merely predict biomarker test results but actually stratifies meaningful clinical outcomes. In practice, biomarker testing is often incomplete in clinical settings - the PET/CT-based model can serve as a practical substitute when tissue testing is unavailable or incomplete.
An assistive tool for patient selection. The authors position the combined clinical-SULmax model as an assistive tool for selecting ADC patients for immunotherapy, complementing or substituting for tissue biomarker testing. High-SULmax patients with compatible clinical features can be prioritized for immunotherapy, while low-SULmax patients may warrant evaluation for alternative treatment strategies.
A practical non-invasive biomarker tool. This study establishes SULmax from 18F-FDG PET/CT as a key independent predictor of both PD-L1 expression and TMB in lung adenocarcinoma, and demonstrates that the combined clinical-SULmax nomogram model can effectively predict biomarker status and stratify immunotherapy responses in real-world clinical settings.
Advantages over complex radiomics approaches. While deep radiomics models using PET/CT can achieve higher AUCs (up to 0.97), their clinical adoption is limited by complex image segmentation requirements, poor reliability with small datasets, and lack of interpretability due to the black-box effect. The clinical-SULmax nomogram offers a transparent, accessible alternative that any clinician can apply using standard PET/CT reports.
Subtype-specific findings. The study found that PET/CT metabolic parameters are useful for biomarker prediction in ADC but not in SCC, underscoring the importance of analyzing histological subtypes separately rather than treating all NSCLC as a single entity. This subtype specificity reflects genuine biological differences in tumor metabolism and immune microenvironment between the two cancer types.
Limitations and future needs. The study is limited by its retrospective single-center design, the small independent test cohort, and the use of three different PD-L1 antibody clones. Future prospective, large-scale, multicenter studies are needed to validate these models before broad clinical deployment. The authors also note that artificial intelligence and machine learning approaches incorporating radiomics features represent a promising future direction for enhancing model performance.