Development of a computed tomography radiomics and CD38 integrated model: predicting immunotherapy response and investigating biological implications in non-small cell lung cancer

Introduction

Lung cancer remains a leading cause of cancer-related mortality worldwide, with non-small cell lung carcinoma (NSCLC) accounting for approximately 80–85% of cases (1). Immunotherapy has advanced significantly, evolving from second-line monotherapy to first-line, dual, combination, and neoadjuvant therapies—an era termed Immunotherapy 2.0. Landmark trials such as Keynote-024, Keynote-042 (2,3), Keynote-189, Keynote-407 (4,5), Checkmate-9LA, and Checkmate-227 (6,7) have demonstrated substantial survival benefits. However, immunotherapy response rates remain limited (20–30%), with strict exclusion criteria for patients with brain metastases, epidermal growth factor receptor (EGFR)/anaplastic lymphoma kinase (ALK) mutations, prior glucocorticoid use, or advanced age. Trials targeting EGFR-mutant NSCLC, such as Keynote-789, ATLANTIC, and IMpower150, have achieved limited success (8,9). Furthermore, some programmed death-ligand 1 (PD-L1)-negative patients respond well, indicating the need for better predictive methods.

Radiomics extracts high-throughput quantitative features from medical imaging modalities [e.g., computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography/CT (PET/CT)], encompassing intensity, shape, and texture characteristics, to link imaging phenotypes with tumor biology. It has shown promise in lung cancer for diagnosis, therapeutic response assessment, and prognostication (10). Radiomics-based models demonstrated superior prognostic accuracy compared to traditional clinical predictors in NSCLC (11). Radiomics effectively stratified survival risk in stage IA pure solid NSCLC (12) and independently predicted disease-free survival (DFS) when combined with clinical factors (13). Moreover, radiomic features reflect tumor heterogeneity, correlating with molecular characteristics (14), facilitating noninvasive prediction of molecular biomarkers. This integration of imaging and genomic data—radiogenomics—has enabled subtype classification and mutation status prediction in NSCLC (15-19). Radiomics also aids immunotherapy response prediction through associations with intratumoral immune cell infiltration (20). These advances position radiomics as a critical bridge between tumor biology and clinical practice, enhancing diagnostics, prognosis, and individualized treatments. However, most radiomics models that aim to predict the efficacy of NSCLC immunotherapy are constructed as single-modality imaging signatures. Although these models can capture macroscopic tumor heterogeneity, they provide only indirect information about the dynamic tumor-immune microenvironment and generally achieve moderate and sometimes unstable predictive performance across cohorts. Furthermore, the biological mechanisms linking radiomic patterns to specific immunosuppressive pathways remain largely unclear, which limits the interpretability and clinical adoption of radiomics-only models.

CD38 is a type II transmembrane glycoprotein expressed broadly on immune cells including T cells, B cells, and NK cells (21). Elevated CD38 expression in tumors contributes significantly to immune escape through the adenosine pathway, suppressing CD8⁺ T-cell function and recruiting immunosuppressive cells [regulatory T cells (Tregs), tumor-associated macrophages (TAMs), and myeloid-derived suppressor cells (MDSCs)] (22,23). Recent studies highlight CD38’s role in lung cancer, with CD38-catalyzed cyclic ADP-ribose pathways promoting tumor progression (24). Although combining anti-CD38 antibodies with PD-L1 inhibitors showed limited efficacy in clinical trials (25), the approach remains promising. Anti-CD38 antibodies also mediate tumor control via complement-dependent cytotoxicity, antibody-dependent cellular cytotoxicity, and phagocytosis (26,27). Thus, CD38 is increasingly recognized as a critical immunotherapy target, offering new avenues for cancer treatment.

Given current challenges in identifying suitable NSCLC patients for immunotherapy, our study integrates CT radiomics and CD38 expression to create a fusion model predicting immunotherapy response. We further explore its prognostic value and biological significance to enhance patient selection and therapeutic outcomes. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1828/rc).

Methods

Development of a CT radiomics and CD38 fusion model for predicting immunotherapy response in NSCLC

Study subjects

This study retrospectively enrolled patients with pathologically confirmed NSCLC from the Affiliated Hospital of Hangzhou Normal University. Inclusion criteria were: (I) pathologically confirmed NSCLC; (II) received immunotherapy; (III) availability of CT imaging within 1 month prior to immunotherapy initiation; (IV) availability of pathological tissue samples and complete clinical data. Exclusion criteria were: (I) patients with severe comorbid cardiac or renal dysfunction, or who were critically ill with multiple malignancies; (II) patients lost to follow-up during treatment; (III) patients with CT images of insufficient quality for feature extraction; (IV) patients receiving fewer than 2 cycles of immunotherapy.

From July 2021 to December 2024, 45 out of 135 screened patients met the study criteria and were enrolled. These patients were divided into a training cohort (n=31) and a test cohort (n=14 ) in a 7:3 ratio using the “randomizr” R package. Baseline characteristics collected included: age, sex, smoking status, histological subtype, tumor-node-metastasis (TNM) stage, PD-L1 expression (≥50%/<50%), immunotherapy agent [programmed death-1 (PD-1) mAb/PD-L1 mAb/cytotoxic T-lymphocyte-associated protein 4 (CTLA-4) mAb], prior treatment regimen (surgery/non-surgical), concurrent treatment regimen (chemotherapy/targeted therapy). Treatment response was assessed using Response Evaluation Criteria in Solid Tumors (RECIST) v1.1 criteria—(I) complete response (CR): disappearance of all target lesions, no new lesions, normalization of tumor markers, maintained for ≥4 weeks; (II) partial response (PR): ≥30% decrease in the sum of diameters of target lesions, maintained for ≥4 weeks; (III) stable disease (SD): neither sufficient shrinkage for PR nor sufficient increase for PD; (IV) progressive disease (PD): ≥20% increase in the sum of diameters of target lesions, or appearance of new lesions. In this study, patients achieving CR or PR were classified as responders (R-group), while those with PD or SD were classified as non-responders (NR-group). All data were collected from the Affiliated Hospital of Hangzhou Normal University. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study protocol was approved by the Institutional Review Board of the Affiliated Hospital of Hangzhou Normal University [approval No. 2024(E2)-KS-077]. A waiver of informed consent was granted due to the retrospective nature of the study. The study flowchart for this section is presented in Figure 1.

Figure 1 Workflow of constructing a radiomics-CD38 fusion model to predict immunotherapy response in non-small cell lung cancer. CT, computed tomography; ROI, region of interest.

CT image acquisition and parameters

All CT scans followed standardized thoracoabdominal clinical protocols. Scans were performed on Siemens SOMATOM or GE Discovery series scanners (≥64-detector rows) during end-inspiration breath-hold, covering the lung apex to posterior costophrenic angles. Parameters: 1.25 mm collimation, pitch 1.1, 120 kV, 200 mA, 512×512 matrix. Images were reconstructed at 1.25 mm thickness and 1.375 mm intervals. For contrast-enhanced scans, non-ionic iodinated contrast (350 mgI/mL) was intravenously injected as a bolus at 2.5 mL/s (dose: 1.5 mL/kg). Smart bolus-tracking [aortic threshold: 150 Hounsfield units (HU)] triggered data acquisition at the portal venous phase (60 s post-injection). Automatic dose modulation (CARE Dose 4D/ASiR) was applied to optimize radiation dose while maintaining image quality.

Radiomic feature extraction

CT images in digital imaging and communications in medicine (DICOM) format were retrieved from the hospital picture archiving and communication system (PACS) and processed using a standardized radiomics workflow in 3D Slicer (v4.10.2). Preprocessing included slice thickness correction and voxel spacing normalization. Pulmonary tumor parenchyma regions of interest (ROIs) were manually segmented on axial images by an experienced respiratory radiologist (≥5 years of experience), excluding vasculature (>80 HU), air regions (<−150 HU), and adjacent inflammatory areas. In multifocal cases, the primary lesions were prioritized. Following isotropic resampling (1 mm × 1 mm × 1 mm), 3D tumor volumes were constructed for feature extraction. Extracted features comprised morphological parameters (18 geometric features), first-order statistics, and texture features derived from gray-level co-occurrence matrix (GLCM) (14 offsets), gray-level run-length matrix (GLRLM) (13 directions), gray-level size zone matrix (GLSZM), gray-level dependence matrix (GLDM) (dependency =0.1), and neighboring gray tone difference matrix (NGTDM). Multiscale analysis included Laplacian of Gaussian (LoG) filtering (σ=1.0, 1.5, 2.0, 2.5 mm) and wavelet transformation (Haar basis, eight directions). After excluding 37 DICOM header metadata features from the initial 1,260 features, a total of 1,223 image biomarker standardization initiative (IBSI)-compliant radiomic features were retained.

Development and performance evaluation of the radiomics model

The radiomics model was constructed in the training cohort using least absolute shrinkage and selection operator (LASSO) regression via the “glmnet” package in R, selecting significant features from 1,223 standardized radiomic features through a 10-fold cross-validation approach. Features with non-zero coefficients were retained and input into a binary logistic regression model. Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA) through the “ROCR”, “rms”, and “rmda” packages in R. Metrics such as sensitivity, specificity, 95% confidence intervals (CI), and area under the curve (AUC) were calculated. The model was validated in the validation cohort, assessing discriminatory ability via ROC curves, predictive accuracy through calibration curves, and clinical utility using DCA.

Development and performance evaluation of the fusion model

Formalin-fixed paraffin-embedded tumor samples from 40 randomly selected NSCLC patients prior to immunotherapy underwent standardized CD38 immunohistochemistry using a commercially available anti-CD38 antibody. Immunohistochemical staining was performed on tissue sections, and CD38 expression was quantified using the average optical density (AOD), calculated via ImageJ (v1.53t) software. The AOD value was determined by the formula: AOD = integrated density/area of positive signal, which provided a continuous quantitative measurement of CD38 expression in the tumor samples. Radiomic features and CD38 AOD values were then integrated into a logistic regression-based fusion model to predict immunotherapy response. Model performance was assessed against a radiomics-only model using ROC curves, AUC, net reclassification improvement (NRI), integrated discrimination improvement (IDI), and calibration and decision curve analyses. Significant clinical predictors of immunotherapy response were identified by univariate logistic regression (autoReg package, P≤0.05) and combined with fusion model scores to create a nomogram, converting regression coefficients into a 0–100 points scale for individualized prediction of immunotherapy response. This approach integrates imaging, molecular, and clinical parameters, facilitating optimized patient stratification. To further explore the biological relationship between imaging and the immune microenvironment, we analyzed the correlations between CD38 expression (AOD) and the radiomic texture features included in the model. Spearman’s rank correlation coefficients were calculated, and scatter plots with regression lines and 95% CIs were generated using the ggpubr package (v0.6.0) in R (v4.3.2). Statistical significance was defined as two-sided P<0.05.

Statistical analysis

All statistical analyses were performed using R (version 4.3.2). The “compareGroups” package assessed baseline comparability between training and validation cohorts. For continuous variables, two-sample t-tests were used if normally distributed, while the Mann-Whitney U test was applied for non-normal distributions. Categorical variables were compared using the chi-squared test or Fisher’s exact test. Univariate logistic regression analyzed the association between clinical variables and immunotherapy response, and multivariate logistic regression was used to build the radiomics and fusion models. Statistical significance was set at two-sided P<0.05. Key R packages included “compareGroups”, “glmnet”, “rms”, “ROCR”, “rmda”, “PredictABEL”, and “autoReg”.

Prognostic prediction using the fusion model and exploration of biological implications

Study subjects

This retrospective cohort study utilized publicly available international databases. A total of 89 pathologically confirmed NSCLC patients with multi-modal data were included through integration of two authoritative databases. CT imaging data were obtained from the NSCLC-Radiomics-Genomics cohort in The Cancer Imaging Archive (TCIA) (https://www.cancerimagingarchive.net/collection/nsclc-radiomics-genomics/), while matched gene expression profiles and clinical outcomes were retrieved from the Gene Expression Omnibus (GEO) (https://www.ncbi.nlm.nih.gov/geo/) GSE58661 dataset. Data were precisely matched using unique patient identifiers to ensure consistency across imaging, genomic, clinical, and survival information. All data collection adhered to open-access database policies.

Prognostic prediction using the fusion model and validation in clinical subgroups

Based on the previously established fusion model, each NSCLC patient (n=89) with integrated CT imaging, gene expression, clinical, and outcome data was assigned a response score. Patients were divided into high and low score groups using the mean as the cutoff. Survival analysis was performed using the “survival” and “survminer” R packages. Kaplan-Meier (K-M) curves were generated to compare overall survival (OS) between groups, and differences were assessed by log-rank test. To further assess clinical utility, subgroup analyses stratified by age (>64/≤64 years), sex, T stage (T1–2/T3–4), M stage (M0/M1), N stage (N0/N1–3), and TNM stage (stage I–II/III–IV) were conducted to evaluate the fusion model’s prognostic value across clinical subgroups. Statistical significance was set at two-sided P<0.05.

Performance evaluation of the fusion model for prognostic prediction

We first standardized the data and performed principal component analysis (PCA) using the prcomp function. The clustering of high- and low-score groups was visualized with the “scatterplot3d” R package to assess the fusion model’s classification ability. Time-dependent ROC curve analysis for response score and clinical variables (age, sex, T, M, N, and TNM stage) was conducted using the “timeROC” package, with 3-year survival as the endpoint. We compared the predictive performance of the fusion model with clinical indicators and further evaluated its prognostic accuracy at 1, 3, and 5 years. Finally, we visualized survival status and duration relative to response score, providing more precise prognostic insights for clinical practice.

Biological exploration in high- and low-response score groups

After stratifying patients into high- and low-score groups using the fusion model, we systematically explored biological heterogeneity between the groups. Differentially expressed genes (DEGs) were identified with the “limma” package, using thresholds of P<0.05 and |logFC| >0.5; genes with logFC >0.5 and P<0.05 were upregulated in the high-score group, while logFC <−0.5 and P<0.05 indicated upregulation in the low-score group. Volcano plots highlighting significant genes were generated using “ggpubr” and “ggthemes”. Functional enrichment analysis of DEGs was conducted with “clusterProfiler” to assess Gene Ontology (GO) terms, including molecular function, biological process, and cellular component, with visualization by “GOplot”. Gene set enrichment analysis (GSEA) was performed with “GSEABase”, evaluating all genes in each group to identify pathways associated with treatment response, visualized with “enrichplot”. Finally, the tumor immune microenvironment was characterized using the CIBERSORT algorithm and single-sample GSEA (ssGSEA) from the “IOBR” package, comprehensively assessing immune cell infiltration in both groups.

Results

Baseline consistency assessment of training and validation cohorts

A total of 45 eligible patients were included and randomly assigned to a training cohort (n=31) and a validation cohort (n=14) at a 7:3 ratio. To ensure balanced baseline characteristics, systematic comparability testing was performed. Certain clinical variables (e.g., PD-L1 status, treatment regimen) exhibited extreme skewness or absolute homogeneity and were excluded from analysis. Five key variables—age, sex, smoking status, histological subtype, and TNM stage—were compared between groups. No significant differences were observed in age, sex, smoking status, histological subtype, or TNM stage (all P>0.05), confirming comparability of baseline characteristics between cohorts and supporting the validity of model development and evaluation (Table 1).

Radiomics model construction and evaluation

In the training cohort (n=31), 1,223 radiomic features were standardized and dimensionality reduction was performed using LASSO regression (Figure 2A,2B). Six features with zero variance were removed. After parameter optimization (λ=0.1, 500 iterations), three nonzero features (wavelet-LHH ngtdm Contrast, wavelet-HHH glcm ClusterProminence, original glcm InverseVariance) were selected and incorporated into a multivariable logistic regression to construct a CT-based radiomics model for predicting immunotherapy response. The model achieved an AUC of 0.734 (95% CI: 0.626–0.842) in the training cohort (Figure 2C). The optimal cutoff (Youden index) was 0.368, with sensitivity of 0.775 and specificity of 0.673. Calibration curves showed good agreement between predicted and observed probabilities, with a mean absolute error (MAE) of 0.076 after bootstrap correction (Figure 2D). DCA demonstrated that the radiomics model provided greater standardized net benefit than both “treat all” and “treat none” strategies across a wide range of threshold probabilities, indicating stable clinical utility (Figure 2E). External validation in an independent cohort (n=14) confirmed robust model performance, with an AUC of 0.878 (Figure 2F) and excellent calibration (MAE =0.034) (Figure 2G). DCA (Figure 2H) demonstrated that across most threshold probabilities (0–0.8), the radiomics model (red curve) offered higher standardized net benefit than “treat all” and “treat none” strategies. This indicates that the model can improve clinical decision-making and reduce unnecessary interventions.

Figure 2 Performance evaluation of the radiomics model based on LASSO and logistic regression. (A) LASSO coefficient profiles of the 1,223 radiomic features. (B) 10-fold cross-validation to select the optimal lambda value based on minimum binomial deviance. (C) ROC curve showing the discriminatory ability of the radiomics model in the training cohort. (D) Calibration curve assessing the agreement between predicted and actual responses in the training cohort. (E) DCA evaluating the clinical utility of the radiomics model in the training cohort. (F) ROC curve of the model in the validation cohort. (G) Calibration curve of the validation cohort showing good agreement. (H) DCA curve for the validation cohort indicating favorable net clinical benefit. Model 1: radiomics model. AUC, area under the curve; DCA, decision curve analysis; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic.

Fusion model construction and evaluation

A fusion model was developed by combining radiomic features and the immune biomarker CD38 from 40 randomly selected patients. Immunohistochemistry showed significantly higher CD38 expression in non-responders, supporting its role in immune suppression (Figure 3A,3B). The model integrated three radiomic features (wavelet-LHH ngtdm Contrast, original glcm InverseVariance, wavelet-HHH glcm ClusterProminence) and CD38 quantitative values (AOD), using multivariable logistic regression (Table 2). In the training set, the fusion model achieved an AUC of 0.801 (95% CI: 0.664–0.937), outperforming the radiomics-only model (AUC 0.771) (Figure 3C). It showed high specificity (1.000) and good sensitivity (0.679) at the optimal cutoff. Continuous NRI (0.9524, P<0.001) and IDI (0.1482, P=0.001) confirmed improved discrimination and risk stratification (Table 3). Calibration analysis showed good agreement (MAE =0.033) (Figure 3D), and DCA demonstrated superior net benefit across a wide range of thresholds, highlighting the model’s clinical utility (Figure 3E). Overall, the fusion model provides an interpretable and practical tool for individualized immunotherapy decision-making in NSCLC. In the subgroup analysis based on tumor subtype, the fusion model demonstrated an AUC of 0.875 for lung adenocarcinoma (LUAD) patients, whereas the lung squamous cell carcinoma (LUSC) group showed a slightly lower AUC of 0.75 (Figure 3F). This indicates that the fusion model performed better in LUAD patients, but still provided moderate predictive accuracy for LUSC patients. Correlation analysis further demonstrated that CD38 expression was positively correlated with wavelet-LHH ngtdm Contrast (R=0.53, P<0.001) and wavelet-HHH glcm ClusterProminence (R=0.55, P<0.001), but negatively correlated with original glcm InverseVariance (R=–0.35, P=0.03) (Figure 3G). These findings suggest that higher CD38 expression is associated with increased radiomic heterogeneity and reduced gray-level uniformity, indicating that CD38-mediated immunosuppression may manifest as distinct imaging texture patterns. Additionally, Significant clinical predictors of immunotherapy response—including sex, smoking history, age, TNM stage, and histological subtype—were identified by univariate logistic regression (Table 4). These variables, together with the fusion model, were integrated into a multivariable nomogram, which allows individualized risk prediction (Figure 3H) and supports precise clinical stratification and decision-making in NSCLC.

Figure 3 Development and validation of the CD38-radiomics fusion model and nomogram. (A) Representative immunohistochemical images of CD38 expression in immunotherapy responders (magnification ×40). (B) Representative CD38 staining in non-responders to immunotherapy, showing high CD38 expression (magnification ×40). (C) ROC curves comparing the radiomics model and the CD38-radiomics fusion model in predicting immunotherapy response. (D) Calibration curve of the fusion model demonstrating good agreement between predicted and observed response probabilities. (E) DCA indicating greater net clinical benefit of the fusion model over the radiomics-only model. (F) Performance of the fusion model for immunotherapy prediction across tumor subtypes (LUAD and LUSC). (G) Spearman correlation plots showing associations between CD38 expression (AOD) and radiomic features (wavelet-LHH ngtdm Contrast, wavelet-HHH glcm ClusterProminence, original glcm InverseVariance). (H) Nomogram incorporating clinical predictors (gender, smoking status, age, TNM stage, histological subtype) and the fusion model score to estimate individualized probability of immunotherapy response in NSCLC patients. Model 1 and Model 2 represent radiomics model and fusion model, respectively. AOD, average optical density; DCA, decision curve analysis; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; NSCLC, non-small cell lung cancer; ROC, receiver operating characteristic; TNM, tumor-node-metastasis.

Prognostic analysis and performance evaluation of the fusion model

The study workflow for prognosis prediction and biological interpretation using the CD38-radiomics fusion model is illustrated in Figure 4. Among 89 NSCLC patients, individuals were stratified into high-score (High Score, n=55) and low-score (Low Score, n=34) groups based on the mean response score from the fusion model. K-M survival analysis revealed that the High Score group had significantly longer OS compared to the Low Score group (P=0.001; Figure 5A), demonstrating strong prognostic discrimination by the fusion model. Subgroup analyses indicated that the High Score group consistently showed improved OS across most clinical subgroups—including age >64 years, ≤64 years, males, T1–2, M0, N0, and stage I–II (all P<0.05)—whereas no significant difference was observed in females, T3–4, M1, N1–3, or stage III–IV, likely due to small sample sizes (Figure 5B). PCA showed clear separation between the two groups in three-dimensional space (Figure 5C). For 3-year survival prediction, the fusion model achieved an AUC of 0.71, outperforming individual clinical factors (AUC=0.207–0.605; Figure 5D). The AUCs at 1, 3, and 5 years were 0.70, 0.71, and 0.72, respectively, indicating stable and reliable performance over time (Figure 5E). Visualization of survival outcomes further demonstrated that higher response scores were associated with longer survival and lower mortality (Figure 5F). Collectively, the fusion model enables effective prognostic stratification and may inform individualized clinical decision-making in NSCLC.

Figure 4 Workflow for prognosis prediction and biological interpretation using the CD38-radiomics fusion model. AUC, area under the curve.

Figure 5 Prognostic value and predictive performance of the CD38-radiomics fusion model in NSCLC. (A) Kaplan-Meier curve showing longer overall survival in the high-score group. (B) Subgroup survival analyses confirm the model’s prognostic value across clinical strata. (C) PCA shows clear separation between high- and low-score patients. (D) Time-dependent ROC at 3 years demonstrates better performance of the fusion model versus clinical variables. (E) ROC curves at 1, 3, and 5 years show stable predictive ability. (F) Risk and survival plots indicate that high-score patients have better prognosis. Model 2: fusion model. AUC, area under the curve; M, metastasis; N, node; NSCLC, non-small cell lung cancer; PCA, principal component analysis; ROC, receiver operating characteristic; T, tumor.

Biological insights of the fusion model

Integrating genomic and immune microenvironment analyses, this study systematically elucidates the biological underpinnings of the fusion model for stratifying NSCLC patients’ response to immunotherapy. Differential gene expression analysis (DESeq2) identified 113 significantly altered genes between groups defined by the fusion model, with 63 upregulated in the High Score group and 50 in the Low Score group (Figure 6A). Genes upregulated in the High Score group are associated with epithelial polarity and angiogenesis, while those elevated in the Low Score group (e.g., ZNF521, FAP, NPR2) are linked to fibroblast activation and extracellular matrix (ECM) remodeling. GO enrichment revealed these genes are significantly involved in ECM organization and collagen fibril assembly (Figure 6B), suggesting that ECM remodeling in the Low Score group may hinder immune cell infiltration, whereas epithelial gene activation in the High Score group may enhance antigen presentation. GSEA further showed that the High Score group is enriched in epithelial differentiation, epidermis development, and hormone metabolic processes (Figure 6C,6D), while the Low Score group is dominated by pathways involved in collagen metabolism and ECM remodeling, consistent with an immunosuppressive milieu. Cibersort and ssGSEA analyses revealed the High Score group exhibited increased infiltration of plasma cells, CD8⁺ T cells, activated CD4⁺ memory T cells, and M1 macrophages (Figure 6E,6F), representing a “hot tumor” immune phenotype. In contrast, the Low Score group was characterized by resting CD4⁺ memory T cells and M2 macrophages, reflecting an immunosuppressive environment. The enrichment of MDSCs and Tregs in the High Score group suggests a potential immune feedback mechanism, whereas the accumulation of resting T cells and M2 macrophages in the Low Score group may contribute to immune evasion. These findings provide a mechanistic basis for the prognostic stratification achieved by the fusion model.

Figure 6 Biological differences between high and low score groups in NSCLC. (A) Volcano plot showing DEGs between high and low-score groups. (B) GO enrichment analysis of DEGs across biological process, cellular component, and molecular function categories. (C,D) GSEA plots showing significantly enriched biological pathways in high (C) and low (D) score groups. (E) Immune infiltration analysis using CIBERSORT algorithm, indicating differences in immune cell composition between groups. (F) ssGSEA analysis confirms distinct immune landscape, with variation in immune activation and suppression signatures. ns, not significant; *, P<0.05; **, P<0.01; ***, P<0.001. CIBERSORT, Cell-type Identification By Estimating Relative Subsets Of RNA Transcripts; DEGs, differentially expressed genes; GO, Gene Ontology; NSCLC, non-small cell lung cancer; ssGSEA, single-sample gene set enrichment analysis.

Discussion

This study retrospectively enrolled 45 NSCLC patients receiving immunotherapy and systematically developed and validated a multimodal predictive model. Following rigorous radiomics modeling guidelines, stratified sampling was used to ensure balanced core baseline characteristics between the training set (n=31) and validation set (n=14) (P>0.05), thereby minimizing cohort heterogeneity (28). Due to sample size and clinical data distribution, certain key variables (such as PD-L1 expression and treatment regimen) were not included in the model. Notably, recent studies have highlighted the significant spatial and temporal heterogeneity of PD-L1 expression, presenting a major challenge for immunotherapy prediction models (29). Future efforts should integrate multi-region or liquid biopsies to incorporate multidimensional immune microenvironment data, enhancing the model’s biological interpretability. During feature selection, default LASSO regression parameters led to all coefficients being shrunk to zero, indicating excessive penalization. By narrowing the λ range (0.001–0.1) and increasing the number of iterations, three radiomic features were ultimately retained, striking a balance between model complexity and predictive accuracy (30-32). These three features demonstrated clear biological relevance: wavelet-LHH ngtdm Contrast [quantifying local gray-level variations at tumor margins, reflecting metabolic heterogeneity (33)], Original glcm InverseVariance [measuring gray-level uniformity, associated with immune microenvironment heterogeneity (34)], wavelet-HHH glcm ClusterProminence [a three-dimensional texture feature capturing spatial heterogeneity (35)]. These findings are consistent with prior studies, supporting their role as noninvasive imaging surrogates for tumor molecular phenotypes (36,37). The radiomics model constructed from these three features achieved moderate predictive performance in the training cohort (AUC =0.734), with sensitivity (0.775) notably exceeding specificity (0.673), indicating a tendency to reduce false negatives—an important attribute for clinical immunotherapy decision-making. The model showed minimal calibration bias in low- to intermediate-risk intervals, and DCA demonstrated that the radiomics model provided greater standardized net benefit than both “treat all” and “treat none” strategies across a wide range of threshold probabilities. In the independent validation cohort, the AUC increased to 0.878, likely attributable to higher clinical homogeneity and dimensionality reduction, with DCA further confirming broad applicability. Compared to traditional biomarkers, this multiparameter model enables individualized, probability-based risk stratification and clinical decision-making. Future multicenter studies with larger cohorts and the integration of spatial transcriptomics are needed to further elucidate the relationship between imaging features and the immune microenvironment.

In this study, we developed a highly specific and clinically practical fusion prediction model by integrating CT radiomic features with the immune microenvironment marker CD38. Radiomics quantified tumor heterogeneity, while CD38 reflected the immunosuppressive state. Immunohistochemistry revealed significantly higher CD38 expression in non-responders compared to responders, consistent with the mechanism whereby CD38 mediates T cell inhibition through the adenosine pathway (38). Using AOD as a quantitative, continuous metric for CD38 expression further improved the precision of immunotherapy response prediction. Multivariate regression demonstrated that low CD38 expression independently predicted benefit from immunotherapy, in line with its immunosuppressive function (39), and enhanced the interpretability of the model. The fusion model improved the AUC from 0.771 to 0.801, with clear gains in both sensitivity and calibration accuracy. Mechanistically, the synergy between radiomic features and CD38 captured both tumor structural and immune functional dimensions. From a biological perspective, the texture features retained in the model primarily capture gray-level non-uniformity and edge heterogeneity, reflecting stromal remodeling, hypoxia, and heterogeneous immune-cell infiltration in the tumor microenvironment (TME). CD38, as an immunosuppressive molecule broadly expressed on immune cell subsets, promotes an adenosine-rich milieu favoring regulatory and myeloid suppressor cells. In line with this, correlation analysis showed that higher CD38 expression was associated with greater texture heterogeneity and reduced gray-level uniformity. These findings suggest that the radiomic features may serve as noninvasive surrogates of a CD38-driven “cold” immune phenotype, supporting the biological rationale of our image–immunology dual-modal model. In the logistic regression, the negative coefficient of Original glcm InverseVariance (–5.81784) combined with the positive effect of CD38 (+0.15235) amplified risk prediction for non-response, indicating that reduced texture uniformity together with elevated CD38 marks an immunosuppressive microenvironment (40,41). Substantial increases in continuous NRI (0.9524, P<0.001) and IDI (0.1482, P=0.001) further demonstrated the superiority of the fusion model in risk stratification and dynamic monitoring. After 1,000 bootstrap calibrations, the MAE of the fusion model was reduced to 0.033, showing enhanced calibration, especially in high-risk probability intervals. DCA showed that, compared to treat-all or treat-none strategies, the fusion model provided a greater net clinical benefit across most probability thresholds. This suggests the model can more accurately identify patients likely to benefit from immunotherapy, supporting its value in guiding clinical decisions for NSCLC. By incorporating clinical variables such as sex, age, smoking history, TNM stage, and histological subtype, we constructed a comprehensive nomogram that enables individualized, visualized risk assessment across molecular, imaging, and clinical dimensions, providing more precise clinical decision support for immunotherapy.

The fusion model integrating radiomic features, CD38, and clinical variables was further validated for prognostic stratification in NSCLC patients receiving immunotherapy. K-M analysis showed significantly longer OS in the high-score group compared to the low-score group (P=0.001), with this benefit consistent across age, male sex, and early-stage disease subgroups (all P<0.05), indicating good generalizability. In contrast, performance declined in advanced-stage (Stage III–IV, P=0.24) and metastatic patients (M1, n=4), consistent with reports that increased heterogeneity in metastatic NSCLC reduces model accuracy (42). PCA demonstrated clear separation between groups along PC1–PC3, suggesting that the fusion features capture TME heterogeneity and spatial associations between imaging texture and CD38 (43,44). Time-dependent ROC analysis yielded AUCs of 0.70, 0.71, and 0.72 at 1, 3, and 5 years, outperforming individual clinical variables and aligning with studies showing improved performance through multimodal integration (45). Limitations include reduced performance in advanced disease, possibly due to greater genomic instability and heterogeneity (46), and exclusion of key biomarkers such as PD-L1 and tumor mutational burden (TMB)—previous work indicates that combining TMB with radiomics can enhance predictive accuracy (47,48). For high-risk patients in the low-score group, integrating dynamic imaging with liquid biopsy may enable early intervention, as circulating tumor DNA (ctDNA) clearance has been associated with improved OS and progression-free survival (PFS) (49). Overall, the model offers multidimensional, mechanism-based prognostic stratification and supports individualized immunotherapy planning.

Through differential gene expression, functional enrichment, and immune infiltration analyses, this study not only validated the predictive performance of the fusion model but also elucidated the molecular basis underlying TME heterogeneity between high- and low-score groups and its impact on immunotherapy outcomes. High-score tumors showed upregulation of epithelial polarity-related genes (PM20D2, THSD4), potentially enhancing epithelial differentiation and barrier function to stabilize antigen presentation and activate adaptive immunity, whereas low-score tumors exhibited elevated fibroblast activation marker FAP and ECM remodeling genes (NPR2, ZNF521), which may promote ECM deposition and fibrosis, creating a physical immune barrier (50-52). GO enrichment confirmed low-score group enrichment in ECM organization and collagen assembly processes, consistent with resistance mechanisms involving restricted T-cell migration (53-55), while high-score tumors were enriched in epithelial differentiation and hormone receptor-related pathways, potentially enhancing tumor immunogenicity and immune cell recruitment (56,57). GSEA further indicated activation of epithelial differentiation/hormone metabolism pathways in high responders and collagen metabolic pathways in low-response patients. Immune profiling revealed that high responders were enriched in CD8⁺ T cells, activated CD4⁺ memory T cells, and M1 macrophages, reflecting a “hot tumor” phenotype (58-60), with additional increases in plasma cells and follicular helper T cells, suggesting coordinated humoral and cellular immunity (61,62). In contrast, low-response patients were dominated by resting CD4⁺ memory T cells and M2 macrophages, the latter potentially suppressing T-cell function and promoting Treg expansion via interleukin-10 (IL-10) and transforming growth factor-β (TGF-β) (63-65). ssGSEA showed enhanced effector T-cell, NK-cell, and dendritic-cell functions in high responders, while low-response patients had elevated MDSCs and neutrophils, which may inhibit T-cell proliferation through arginase-1 or reactive oxygen species (66,67). Notably, the concurrent enrichment of MDSCs and Tregs in high responders may reflect negative feedback regulation following immune activation. Collectively, the fusion model predicts immunotherapy response by capturing the dynamic balance between ECM remodeling and epithelial differentiation, along with immune infiltration patterns, offering a mechanistic basis for developing dual-targeted strategies against the ECM-immune axis.

Conclusions

The radiomics model based on tumor heterogeneity demonstrated high predictive performance and stability in forecasting immunotherapy response in NSCLC, with ROC, calibration, and decision curve analyses confirming strong discrimination and consistency across training and validation cohorts. Incorporation of the immune microenvironment marker CD38 yielded a fusion model with significantly improved specificity and enhanced ability to identify low-response patients, as evidenced by IDI and NRI analyses. Integration with key clinical variables into a visualized nomogram provided an intuitive tool for individualized risk stratification and flexible threshold adjustment, offering practical support for precision treatment planning. The fusion model also proved robust in prognostic stratification, with high-response patients achieving significantly longer OS than low-response patients, and consistent results across age and stage subgroups. PCA, ROC, and survival score correlations confirmed predictive stability. Mechanistically, high responders showed activation of epithelial differentiation genes enhancing antigen presentation, accompanied by enrichment of CD8⁺ T cells, M1 macrophages, and effector immune modules, representing a “hot tumor” phenotype. In contrast, low-response patients exhibited ECM remodeling-driven collagen deposition forming an immune-suppressive barrier, with infiltration of resting T cells, M2 macrophages, and MDSCs, consistent with a “cold tumor” state. Limitations include a relatively small sample size, particularly in the metastatic subgroup, which may affect generalizability; the absence of key immunotherapy biomarkers such as PD-L1 and TMB, potentially limiting comprehensiveness; and the retrospective, single-center design. Future work should involve multicenter, prospective studies and multi-omics integration to further validate and optimize the model.

Acknowledgments

We thank the TCIA and GEO databases for generously sharing a large amount of data.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1828/rc

Data Sharing Statement: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1828/dss

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1828/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1828/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments, and approved by the Institutional Review Board of Affiliated Hospital of HangZhou Normal University [2024(E2)-KS-077]. A waiver of informed consent was granted due to the retrospective nature of the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021;71:209-49. [Crossref] [PubMed]
2. Reck M, Rodríguez-Abreu D, Robinson AG, et al. Pembrolizumab versus Chemotherapy for PD-L1-Positive Non-Small-Cell Lung Cancer. N Engl J Med 2016;375:1823-33. [Crossref] [PubMed]
3. Mok TSK, Wu YL, Kudaba I, et al. Pembrolizumab versus chemotherapy for previously untreated, PD-L1-expressing, locally advanced or metastatic non-small-cell lung cancer (KEYNOTE-042): a randomised, open-label, controlled, phase 3 trial. Lancet 2019;393:1819-30. [Crossref] [PubMed]
4. Gandhi L, Rodríguez-Abreu D, Gadgeel S, et al. Pembrolizumab plus Chemotherapy in Metastatic Non-Small-Cell Lung Cancer. N Engl J Med 2018;378:2078-92. [Crossref] [PubMed]
5. Paz-Ares L, Luft A, Vicente D, et al. Pembrolizumab plus Chemotherapy for Squamous Non-Small-Cell Lung Cancer. N Engl J Med 2018;379:2040-51. [Crossref] [PubMed]
6. Paz-Ares L, Ciuleanu TE, Cobo M, et al. First-line nivolumab plus ipilimumab combined with two cycles of chemotherapy in patients with non-small-cell lung cancer (CheckMate 9LA): an international, randomised, open-label, phase 3 trial. Lancet Oncol 2021;22:198-211. [Crossref] [PubMed]
7. Hellmann MD, Ciuleanu TE, Pluzanski A, et al. Nivolumab plus Ipilimumab in Lung Cancer with a High Tumor Mutational Burden. N Engl J Med 2018;378:2093-104. [Crossref] [PubMed]
8. Garassino MC, Cho BC, Kim JH, et al. Durvalumab as third-line or later treatment for advanced non-small-cell lung cancer (ATLANTIC): an open-label, single-arm, phase 2 study. Lancet Oncol 2018;19:521-36. [Crossref] [PubMed]
9. Socinski MA, Jotte RM, Cappuzzo F, et al. Atezolizumab for First-Line Treatment of Metastatic Nonsquamous NSCLC. N Engl J Med 2018;378:2288-301. [Crossref] [PubMed]
10. Chetan MR, Gleeson FV. Radiomics in predicting treatment response in non-small-cell lung cancer: current status, challenges and future perspectives. Eur Radiol 2021;31:1049-58. [Crossref] [PubMed]
11. Taheri P, Golden A. A prognostic PET radiomic model for risk stratification in non-small cell lung cancer: integrating radiogenomics and clinical features to predict survival and uncover tumor biology insights. J Cancer Res Clin Oncol 2025;151:180. [Crossref] [PubMed]
12. Wang T, She Y, Yang Y, et al. Radiomics for Survival Risk Stratification of Clinical and Pathologic Stage IA Pure-Solid Non-Small Cell Lung Cancer. Radiology 2022;302:425-34. [Crossref] [PubMed]
13. Yu X, Zhang Z, Zou Y, et al. CT-based radiomics model for noninvasive prediction of progression-free survival in high-grade serous ovarian carcinoma: a multicenter study incorporating preoperative and postoperative clinical factors. BMC Med Imaging 2025;25:320. [Crossref] [PubMed]
14. Park D, Jang R, Chung MJ, et al. Development and validation of a hybrid deep learning-machine learning approach for severity assessment of COVID-19 and other pneumonias. Sci Rep 2023;13:13420. [Crossref] [PubMed]
15. Zhou M, Leung A, Echegaray S, et al. Non-Small Cell Lung Cancer Radiogenomics Map Identifies Relationships between Molecular and Imaging Phenotypes with Prognostic Implications. Radiology 2018;286:307-15. [Crossref] [PubMed]
16. Chen W, Qiao X, Yin S, et al. Integrating Radiomics with Genomics for Non-Small Cell Lung Cancer Survival Analysis. J Oncol 2022;2022:5131170. [Crossref] [PubMed]
17. Yang F, Chen W, Wei H, et al. Machine Learning for Histologic Subtype Classification of Non-Small Cell Lung Cancer: A Retrospective Multicenter Radiomics Study. Front Oncol 2020;10:608598. [Crossref] [PubMed]
18. Gevaert O, Echegaray S, Khuong A, et al. Predictive radiogenomics modeling of EGFR mutation status in lung cancer. Sci Rep 2017;7:41674. [Crossref] [PubMed]
19. Shiri I, Maleki H, Hajianfar G, et al. Next-Generation Radiogenomics Sequencing for Prediction of EGFR and KRAS Mutation Status in NSCLC Patients Using Multimodal Imaging and Machine Learning Algorithms. Mol Imaging Biol 2020;22:1132-48. [Crossref] [PubMed]
20. Min J, Dong F, Chen Y, et al. The NSCLC immunotherapy response predicted by tumor-infiltrating T cells via a non-invasive radiomic approach. Front Immunol 2024;15:1379812. [Crossref] [PubMed]
21. Dwivedi S, Rendón-Huerta EP, Ortiz-Navarrete V, et al. CD38 and Regulation of the Immune Response Cells in Cancer. J Oncol 2021;2021:6630295. [Crossref] [PubMed]
22. Sarkar I, Basak D, Ghosh P, et al. CD38-mediated metabolic reprogramming promotes the stability and suppressive function of regulatory T cells in tumor. Sci Adv 2025;11:eadt2117. [Crossref] [PubMed]
23. Li W, Li Y, Jin X, et al. CD38: A Significant Regulator of Macrophage Function. Front Oncol 2022;12:775649. [Crossref] [PubMed]
24. Gao L, Liu Y, Du X, et al. The intrinsic role and mechanism of tumor expressed-CD38 on lung adenocarcinoma progression. Cell Death Dis 2021;12:680. [Crossref] [PubMed]
25. Pillai RN, Ramalingam SS, Thayu M, et al. Daratumumab Plus Atezolizumab in Previously Treated Advanced or Metastatic NSCLC: Brief Report on a Randomized, Open-Label, Phase 1b/2 Study (LUC2001 JNJ-54767414). JTO Clin Res Rep 2021;2:100104. [Crossref] [PubMed]
26. Hiemstra IH, Santegoets KCM, Janmaat ML, et al. Preclinical anti-tumour activity of HexaBody-CD38, a next-generation CD38 antibody with superior complement-dependent cytotoxic activity. EBioMedicine 2023;93:104663. [Crossref] [PubMed]
27. Liu W, Yu J, Sun K, et al. Preclinical characterization of a novel investigational monoclonal antibody CM313 with potent CD38-positive cell killing activity. Front Immunol 2024;15:1410457. [Crossref] [PubMed]
28. Hoye J, Smith T, Abadi E, et al. Correction for Systematic Bias in Radiomics Measurements Due to Variation in Imaging Protocols. Acad Radiol 2022;29:e61-72. [Crossref] [PubMed]
29. Wang Y, Zhou Y, Yang L, et al. Challenges Coexist with Opportunities: Spatial Heterogeneity Expression of PD-L1 in Cancer Therapy. Adv Sci (Weinh) 2024;11:e2303175. [Crossref] [PubMed]
30. Vabalas A, Gowen E, Poliakoff E, et al. Machine learning algorithm validation with a limited sample size. PLoS One 2019;14:e0224365. [Crossref] [PubMed]
31. Huang H. Controlling the false discoveries in LASSO. Biometrics 2017;73:1102-10. [Crossref] [PubMed]
32. Benner P. Computing Leapfrog Regularization Paths with Applications to Large-Scale K-mer Logistic Regression. J Comput Biol 2021;28:560-9. [Crossref] [PubMed]
33. Suga M, Nishii R, Miwa K, et al. Differentiation between non-small cell lung cancer and radiation pneumonitis after carbon-ion radiotherapy by (18)F-FDG PET/CT texture analysis. Sci Rep 2021;11:11509. [Crossref] [PubMed]
34. Zhu ZC, Chen MJ, Song L, et al. CT-Based Weighted Radiomic Score Predicts Tumor Response to Immunotherapy in Non-Small Cell Lung Cancer. Zhongguo Yi Xue Ke Xue Yuan Xue Bao 2023;45:794-802. [PubMed]
35. Lin Q, Wu HJ, Song QS, et al. CT-based radiomics in predicting pathological response in non-small cell lung cancer patients receiving neoadjuvant immunotherapy. Front Oncol 2022;12:937277. [Crossref] [PubMed]
36. Zhang H, Lei B, Wang L, et al. Development of an 18F-fluorodeoxyglucose positron emission tomography/computed tomography-based visceral pleural invasion prediction index in peripheral subpleural non-small cell lung cancer ≤3 cm. Quant Imaging Med Surg 2025;15:9667-84. [Crossref] [PubMed]
37. Sethy PK, Geetha Devi A, Padhan B, et al. Lung cancer histopathological image classification using wavelets and AlexNet. J Xray Sci Technol 2023;31:211-21. [Crossref] [PubMed]
38. Lin L, Roccuzzo G, Yakymiv Y, et al. The CD39/CD73/Adenosine and NAD/CD38/CD203a/CD73 Axis in Cutaneous T-Cell Lymphomas. Cells 2025;14:309. [Crossref] [PubMed]
39. Kar A, Mehrotra S, Chatterjee S. CD38: T Cell Immuno-Metabolic Modulator. Cells 2020;9:1716. [Crossref] [PubMed]
40. Huang Q, Mo J, Deng X, et al. NHOC/NHOP as novel biomarkers for predicting lymph node metastasis in NSCLC using PET/CT radiomics and machine learning: a two-center retrospective study. Eur J Nucl Med Mol Imaging 2026;53:2387-402. [Crossref] [PubMed]
41. Ng HHM, Lee RY, Goh S, et al. Immunohistochemical scoring of CD38 in the tumor microenvironment predicts responsiveness to anti-PD-1/PD-L1 immunotherapy in hepatocellular carcinoma. J Immunother Cancer 2020;8:e000987. [Crossref] [PubMed]
42. Ye G, Wu G, Zhang C, et al. CT-based quantification of intratumoral heterogeneity for predicting pathologic complete response to neoadjuvant immunochemotherapy in non-small cell lung cancer. Front Immunol 2024;15:1414954. [Crossref] [PubMed]
43. Ramlee S, Hulse D, Bernatowicz K, et al. Radiomic Signatures Associated with CD8(+) Tumour-Infiltrating Lymphocytes: A Systematic Review and Quality Assessment Study. Cancers (Basel) 2022;14:3656. [Crossref] [PubMed]
44. Chen Y, Xu T, Jiang C, et al. CT-based radiomics signature to predict CD8+ tumor infiltrating lymphocytes in non-small-cell lung cancer. Acta Radiol 2023;64:1390-9. [Crossref] [PubMed]
45. Vanguri RS, Luo J, Aukerman AT, et al. Multimodal integration of radiology, pathology and genomics for prediction of response to PD-(L)1 blockade in patients with non-small cell lung cancer. Nat Cancer 2022;3:1151-64. [Crossref] [PubMed]
46. Tayoun T, Faugeroux V, Oulhen M, et al. Targeting genome integrity dysfunctions impedes metastatic potency in non-small cell lung cancer circulating tumor cell-derived explants. JCI Insight 2022;7:e155804. [Crossref] [PubMed]
47. He B, Dong D, She Y, et al. Predicting response to immunotherapy in advanced non-small-cell lung cancer using tumor mutational burden radiomic biomarker. J Immunother Cancer 2020;8:e000550. [Crossref] [PubMed]
48. Tang X, Qian WL, Yan WF, et al. Radiomic assessment as a method for predicting tumor mutation burden (TMB) of bladder cancer patients: a feasibility study. BMC Cancer 2021;21:823. [Crossref] [PubMed]
49. Wei J, Feng J, Weng Y, et al. The Prognostic Value of ctDNA and bTMB on Immune Checkpoint Inhibitors in Human Cancer. Front Oncol 2021;11:706910. [Crossref] [PubMed]
50. Nicolas-Boluda A, Vaquero J, Vimeux L, et al. Tumor stiffening reversion through collagen crosslinking inhibition improves T cell migration and anti-PD-1 treatment. Elife 2021;10:e58688. [Crossref] [PubMed]
51. Horn LA, Chariou PL, Gameiro SR, et al. Remodeling the tumor microenvironment via blockade of LAIR-1 and TGF-β signaling enables PD-L1-mediated tumor eradication. J Clin Invest 2022;132:e155148. [Crossref] [PubMed]
52. Jensen C, Nissen NI, Von Arenstorff CS, et al. Serological assessment of collagen fragments and tumor fibrosis may guide immune checkpoint inhibitor therapy. J Exp Clin Cancer Res 2021;40:326. [Crossref] [PubMed]
53. Mai Z, Lin Y, Lin P, et al. Modulating extracellular matrix stiffness: a strategic approach to boost cancer immunotherapy. Cell Death Dis 2024;15:307. [Crossref] [PubMed]
54. Hu Q, Zhu Y, Mei J, et al. Extracellular matrix dynamics in tumor immunoregulation: from tumor microenvironment to immunotherapy. J Hematol Oncol 2025;18:65. [Crossref] [PubMed]
55. Balduit A, Agostinis C, Bulla R. Beyond the Norm: The emerging interplay of complement system and extracellular matrix in the tumor microenvironment. Semin Immunol 2025;77:101929. [Crossref] [PubMed]
56. Albini A, Bruno A, Noonan DM, et al. Contribution to Tumor Angiogenesis From Innate Immune Cells Within the Tumor Microenvironment: Implications for Immunotherapy. Front Immunol 2018;9:527. [Crossref] [PubMed]
57. Masaoutis C, Palamaris K, Kokkali S, et al. Unraveling the Immune Microenvironment of Thymic Epithelial Tumors: Implications for Autoimmunity and Treatment. Int J Mol Sci 2022;23:7864. [Crossref] [PubMed]
58. Reschke R, Olson DJ. Leveraging STING, Batf3 Dendritic Cells, CXCR3 Ligands, and Other Components Related to Innate Immunity to Induce A "Hot" Tumor Microenvironment That Is Responsive to Immunotherapy. Cancers (Basel) 2022;14:2458. [Crossref] [PubMed]
59. Peng M. Immune landscape of distinct subtypes in urothelial carcinoma based on immune gene profile. Front Immunol 2022;13:970885. [Crossref] [PubMed]
60. Kienzl M, Maitz K, Sarsembayeva A, et al. Comparative Study of the Immune Microenvironment in Heterotopic Tumor Models. Cancers (Basel) 2024;16:295. [Crossref] [PubMed]
61. Niogret J, Berger H, Rebe C, et al. Follicular helper-T cells restore CD8(+)-dependent antitumor immunity and anti-PD-L1/PD-1 efficacy. J Immunother Cancer 2021;9:e002157. [Crossref] [PubMed]
62. Qi H. T follicular helper cells in space-time. Nat Rev Immunol 2016;16:612-25. [Crossref] [PubMed]
63. Wen ZF, Liu H, Gao R, et al. Tumor cell-released autophagosomes (TRAPs) promote immunosuppression through induction of M2-like macrophages with increased expression of PD-L1. J Immunother Cancer 2018;6:151. [Crossref] [PubMed]
64. Eisel D, Das K, Dickes E, et al. Cognate Interaction With CD4(+) T Cells Instructs Tumor-Associated Macrophages to Acquire M1-Like Phenotype. Front Immunol 2019;10:219. [Crossref] [PubMed]
65. Shao Y, Lan Y, Chai X, et al. CXCL8 induces M2 macrophage polarization and inhibits CD8(+) T cell infiltration to generate an immunosuppressive microenvironment in colorectal cancer. FASEB J 2023;37:e23173. [Crossref] [PubMed]
66. Gabrilovich DI. Myeloid-Derived Suppressor Cells. Cancer Immunol Res 2017;5:3-8. [Crossref] [PubMed]
67. Zhang J, Xu X, Shi M, et al. CD13(hi) Neutrophil-like myeloid-derived suppressor cells exert immune suppression through Arginase 1 expression in pancreatic ductal adenocarcinoma. Oncoimmunology 2017;6:e1258504. [Crossref] [PubMed]