Checkpoint inhibitor–associated pneumonitis in non-small cell lung cancer: a cohort study with transcriptomic analysis of inflammatory mechanisms

Abstract

Background

Immune checkpoint inhibitors (ICIs) have significantly improved outcomes in non-small cell lung cancer (NSCLC), yet their use is associated with a notable risk of immune-related adverse events, including checkpoint inhibitor–associated pneumonitis (CIP). The real-world incidence, risk magnitude, and underlying immunopathogenesis of CIP in NSCLC remain inadequately defined.

Methods

We conducted a retrospective cohort study using electronic health records from 21,671 NSCLC patients, categorized into ICI (n = 8,744) and non-ICI (n = 12,927) groups. Incidence of pneumonitis was evaluated using propensity score–matched analysis, Kaplan-Meier curves, and Cox regression models. Subgroup analyses were performed across demographic and clinical variables. Differentially expressed genes (DEGs) from transcriptomic datasets were analyzed to explore inflammatory mechanisms, including GO/KEGG pathway enrichment, protein–protein interaction (PPI) network construction, and single-sample gene set enrichment analysis (ssGSEA).

Results

The incidence of pneumonitis was significantly higher in the ICI group (28.9%) compared to the non-ICI group (10.0%) (hazard ratio [HR] = 2.86; 95% confidence interval [CI], 2.43–3.29; P < 0.001). This elevated risk persisted across age, sex, BMI, comorbidities, and autoimmune status. Transcriptomic analysis revealed distinct upregulation of immune-related genes (e.g., TCF7L1, ATP1B4, RPL18A), with enrichment of pathways including IFN-γ signaling, Th17 differentiation, TNF and JAK-STAT signaling. ssGSEA confirmed increased immune activation scores in CIP samples. PPI network and hub gene analysis identified GHRH, ZBTB21, and PLAU as central regulators.

Conclusions

ICI use in NSCLC is associated with a markedly increased risk of pneumonitis. Transcriptomic profiling suggests that overactivation of pro-inflammatory immune pathways underlies CIP pathogenesis.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12890-025-04005-0.

Introduction

Lung cancer remains one of the most commonly diagnosed malignancies worldwide and is the leading cause of cancer-related mortality, with an estimated 2 million new cases and 1.76 million deaths annually [1]. Over the past decade, significant advancements have been achieved in the treatment of metastatic non-small cell lung cancer (mNSCLC), leading to notable improvements in patient survival [2]. In 2022, approximately 2.5 million new cases of lung cancer were reported globally, accounting for 12.4% of all new cancer diagnoses [3]. Non-small cell lung cancer (NSCLC) constitutes approximately 85% of all lung cancer cases and is frequently diagnosed at an advanced stage, when curative interventions are often no longer feasible [4]. Despite progress in chemotherapy and targeted therapies, the five-year survival rate for patients with metastatic NSCLC has historically remained below 7% [5, 6].

In recent years, the emergence of immune checkpoint inhibitors (ICIs), particularly those targeting PD-1 or PD-L1, has fundamentally reshaped the therapeutic landscape of NSCLC and significantly improved long-term outcomes for many patients [7]. Since the approval of the first ICI for NSCLC in 2015, clinical adoption has rapidly expanded [8]. A nationwide Norwegian registry study covering 2012 to 2021 showed that the proportion of patients receiving ICIs as first-line therapy rose from 19% in 2017 to 84% in 2021 [9]. Multiple landmark clinical trials have demonstrated that in biomarker-selected populations, ICIs can extend median overall survival beyond 20 months [10].

However, the success of immunotherapy has been accompanied by distinct toxicities, particularly immune-related adverse events (irAEs), which are caused by off-target immune activation. Among these, checkpoint inhibitor-associated pneumonitis (CIP) is one of the most clinically significant complications, especially in thoracic oncology. Patients with NSCLC often present with pre-existing pulmonary comorbidities such as COPD or interstitial lung disease, which further increase the risk of CIP [11, 12]. CIP typically manifests with non-specific respiratory symptoms, such as cough and dyspnea, along with variable radiographic abnormalities, making early detection challenging [13, 14]. The reported incidence of CIP varies across tumor types and treatment settings, ranging from 3% to 7% in NSCLC, with a fatality rate of up to 1% [15]. Although several clinical risk factors have been identified—such as prior thoracic radiotherapy and chronic lung diseases—the underlying immunopathogenic mechanisms remain poorly elucidated [16, 17].

Despite increasing clinical recognition of CIP, large-scale real-world studies integrating clinical outcomes with molecular data are lacking. To address this gap, we conducted a matched cohort study to evaluate the incidence and risk of CIP among NSCLC patients receiving ICI therapy. Furthermore, we performed transcriptomic profiling to characterize inflammatory signatures and identify key molecular pathways involved in CIP pathogenesis.

Materials and methods

Study design and population

We conducted a retrospective cohort study using de-identified electronic health records from a large tertiary class-A hospital in China, covering the period from January 2010 to December 2023. Eligible participants were adults (≥ 18 years) with pathologically confirmed NSCLC who initiated systemic anticancer therapy and had complete follow-up information. Patients were stratified into two groups according to treatment exposure: (i) the ICI group, defined as those who received at least one FDA- or NMPA-approved immune checkpoint inhibitor (anti–PD-1, anti–PD-L1, or anti–CTLA-4 therapy); and (ii) the non-ICI group, defined as those treated exclusively with non-ICI regimens such as chemotherapy, targeted therapy, or supportive care. Patients were excluded if they had a prior diagnosis of pneumonitis, interstitial lung disease, or incomplete clinical records. This exclusion was intended to minimize misclassification of immune-related pneumonitis but may have inadvertently excluded some patients with tumor-related obstructive pneumonia, which we acknowledge as a potential limitation. An overview of the study design and transcriptomic analysis workflow is shown in Fig. 1.

Fig. 1.

Study flowchart and dataset integration. Flow diagram showing patient selection and dataset integration. A total of 21,671 NSCLC patients were included from the hospital database, comprising 8,744 patients who received immune checkpoint inhibitors (ICIs) and 12,927 who did not. Among them, 2,523 ICI users and 1,288 non-ICI patients developed pneumonitis. Transcriptomic data were obtained from the GEO database for integrative analyses, including differential gene expression, GO/KEGG enrichment, protein–protein interaction (PPI) network construction, and hub-gene identification

Exposure and outcome definitions

ICI exposure was defined as the receipt of FDA- or NMPA-approved agents targeting PD-1, PD-L1, or CTLA-4. In our clinical dataset, exposure was recorded only as a binary variable (Yes/No), and detailed subtype classification (PD-1, PD-L1, CTLA-4) was not systematically collected. Therefore, ICIs were analyzed as a composite exposure. Use of antirheumatic medications was defined based on prescription records, and these prescriptions were not restricted to rheumatoid arthritis but also included other autoimmune diseases such as systemic lupus erythematosus, psoriatic arthritis, and ankylosing spondylitis. The primary outcome of interest was CIP, identified by a combination of ICD-10 codes for drug-induced interstitial lung disease, clinical notes, and exclusion of infectious causes via microbiologic testing and imaging. To maximize sensitivity, CIP cases were identified using a broad set of criteria (ICD-10 coding, radiologic findings, and clinical documentation) with exclusion of common infectious causes when microbiological or imaging data were available. Because manual adjudication of all ~ 21,000 cases was not feasible, this strategy may have captured a wider spectrum of pneumonitis-like events, including mild or atypical cases that would not meet stricter trial definitions.

Propensity score matching

To minimize baseline confounding, propensity score matching (PSM) was conducted using nearest-neighbor matching with a caliper of 0.1 standard deviations. Before matching, the ICI group included 8,744 patients and the non-ICI group included 12,927 patients. After matching, all 8,744 ICI patients were retained and matched to an approximately equal number of non-ICI patients, while unmatched non-ICI patients were excluded from further analyses. Covariate balance was evaluated using standardized mean differences (SMD), with SMD < 0.1 considered acceptable. Baseline characteristics before and after matching are summarized in Table 1 and in descriptive comparisons available upon request.

Table 1.

Baseline characteristics of patients with NSCLC after propensity score matching, stratified by pneumonitis severity (No pneumonitis, grade 1–2, and grade ≥ 3) 

Variable	No pneumonitis (n = 6559)	Grade ≥ 3 (n = 915)	Grade 1–2 (n = 14 197)	p value	SMD
Age, years (mean ± SD)	55.80 (9.70)	65.93 (9.36)	60.89 (9.73)	< 0.001	0.705
Length of stay, days (mean ± SD)	4.06 (2.87)	4.10 (2.90)	4.02 (2.83)	0.574	0.018
Follow-up, months (mean ± SD)	20.88 (8.66)	20.73 (8.65)	21.02 (8.66)	0.389	0.023
Sex
Female	3314 (50.5)	442 (48.3)	7201 (50.7)	0.207	0.022
Male	3245 (49.5)	473 (51.7)	6996 (49.3)
Immune checkpoint inhibitor				< 0.001	0.179
No	4113 (62.7)	453 (49.5)	8361 (58.9)
Yes	2446 (37.3)	462 (50.5)	5836 (41.1)
Smoking status				< 0.001	0.497
No	6477 (98.7)	698 (76.3)	13 194 (92.9)
Yes	82 (1.3)	217 (23.7)	1003 (7.1)
Nicotine dependence				0.311	0.029
No	6208 (94.6)	860 (94.0)	13 483 (95.0)
Yes	351 (5.4)	55 (6.0)	714 (5.0)
Alcohol-related disorder				0.906	0.007
No	6388 (97.4)	892 (97.5)	13 815 (97.3)
Yes	171 (2.6)	23 (2.5)	382 (2.7)
Hypertension				< 0.001	0.845
No	5767 (87.9)	312 (34.1)	9185 (64.7)
Yes	792 (12.1)	603 (65.9)	5012 (35.3)
Diabetes mellitus				< 0.001	0.743
No	6076 (92.6)	432 (47.2)	10 794 (76.0)
Yes	483 (7.4)	483 (52.8)	3403 (24.0)
Ischemic heart disease				< 0.001	0.356
No	6457 (98.4)	772 (84.4)	13 497 (95.1)
Yes	102 (1.6)	143 (15.6)	700 (4.9)
Liver disease				< 0.001	0.289
No	6484 (98.9)	815 (89.1)	13 628 (96.0)
Yes	75 (1.1)	100 (10.9)	569 (4.0)
Cerebrovascular disease				< 0.001	0.315
No	6485 (98.9)	803 (87.8)	13 696 (96.5)
Yes	74 (1.1)	112 (12.2)	501 (3.5)
Rheumatoid arthritis				< 0.001	0.276
No	6543 (99.8)	844 (92.2)	13 995 (98.6)
Yes	16 (0.2)	71 (7.8)	202 (1.4)
Systemic lupus erythematosus				< 0.001	0.169
No	6554 (99.9)	888 (97.0)	14 106 (99.4)
Yes	5 (0.1)	27 (3.0)	91 (0.6)
Systemic sclerosis				< 0.001	0.172
No	6554 (99.9)	887 (96.9)	14 102 (99.3)
Yes	5 (0.1)	28 (3.1)	95 (0.7)
Antirheumatic drug use				0.836	0.01
No	5865 (89.4)	814 (89.0)	12 660 (89.2)
Yes	694 (10.6)	101 (11.0)	1537 (10.8)
Depression				0.444	0.012
No	6118 (93.3)	855 (93.4)	13 308 (93.7)
Yes	441 (6.7)	60 (6.6)	889 (6.3)
Tumor stage				0.18	0.064
I	923 (14.1)	108 (11.8)	1968 (13.9)
II	1206 (18.4)	171 (18.7)	2508 (17.7)
III	2149 (32.8)	286 (31.3)	4604 (32.4)
IV	2281 (34.8)	350 (38.3)	5117 (36.0)

Continuous variables are expressed as mean ± standard deviation (SD), and categorical variables as counts with percentages (%)

P values were obtained from χ² tests or ANOVA, as appropriate

Standardized mean differences (SMD) < 0.10 indicate acceptable balance across groups

Tumor stage (I–IV) was summarized to provide clinical context; stage distribution did not differ significantly between pneumonitis subgroups (P = 0.18, SMD = 0.064)

The number of patients receiving antirheumatic medications exceeded those diagnosed with rheumatoid arthritis because these drugs are also used to treat other autoimmune conditions

ICI immune checkpoint inhibitor, RA rheumatoid arthritis, SLE systemic lupus erythematosus

Statistical analysis

Descriptive statistics summarized baseline characteristics, with continuous variables expressed as mean ± SD and categorical variables as counts and percentages. Between-group differences were assessed using Student’s t-test or chi-square tests as appropriate. Kaplan–Meier curves were generated to estimate pneumonitis-free survival, with group comparisons by log-rank test. Cox proportional hazards models were used to calculate hazard ratios (HRs) and 95% confidence intervals (CIs). Three sequential multivariable models were constructed: Model 1 adjusted for age and sex; Model 2 additionally for BMI and major comorbidities (hypertension, diabetes); and Model 3 further for immunosuppressant or antirheumatic medication use. Subgroup analyses explored heterogeneity across sex, age, smoking history (ever vs. never), autoimmune disease, and medication use. Stratification by current/former/never smoking was not possible, representing a limitation. Pneumonitis cases were also categorized by severity (Grade 1–2 vs. ≥3) to provide clinical context and reduce detection bias. The proportional hazards assumption was evaluated using Schoenfeld residuals for covariates and the global model, with no significant violations detected. Analyses were performed in R (version 4.2.1), with P < 0.05 considered statistically significant. Model performance was evaluated using Harrell’s C-index and time-dependent receiver operating characteristic (ROC) curves at 1 and 3 years. Calibration was assessed by comparing predicted and observed survival probabilities.

Transcriptomic data and differential expression analysis

For mechanistic insights into CIP, transcriptomic data were obtained from the publicly available Gene Expression Omnibus (GEO) under accession number GSE216329. This dataset comprised peripheral blood mononuclear cell (PBMC)-derived T cells from 24 NSCLC patients (18 with lung adenocarcinoma), of whom 22 received anti–PD-1 therapy. Seven patients developed checkpoint inhibitor pneumonitis (CIP), while the remainder did not or experienced other immune-related adverse events such as inflammatory arthritis and thyroiditis. Single-cell RNA sequencing was performed to characterize systemic immune transcriptional changes. Raw data were normalized and analyzed using the limma package in R, and differentially expressed genes (DEGs) were identified using |log2 fold change| > 1 and adjusted P < 0.05 (Benjamini–Hochberg method).

Baseline characteristics

Baseline demographic and clinical variables included age, sex, body mass index (BMI), smoking history (ever vs. never smokers), comorbidities (hypertension, diabetes, ischemic heart disease, liver disease, cerebrovascular disease, autoimmune disorders, and depression), and medication use (antirheumatics, immunosuppressants). Tumor stage (I–IV) was also extracted to provide oncologic context. Prior chemotherapy, radiotherapy, and surgery were variably documented but largely unavailable in this dataset and therefore not included in the primary analysis. These variables are summarized in Table 1.

Pathway enrichment and network analysis

Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were performed using the “clusterProfiler” package to identify biological processes associated with the DEGs. Protein–protein interaction (PPI) networks were constructed using the STRING database and visualized in Cytoscape. The PPI networks helped identify key interacting proteins related to CIP. Hub genes within the network were identified using CytoHubba based on degree centrality. Single-sample gene set enrichment analysis (ssGSEA) was performed to quantify immune pathway activity, focusing on pathways such as IFN-γ, TNF, IL-17, JAK-STAT, and NF-κB signaling.

Results

Patient characteristics

Before matching, there were 8,744 patients in the immune checkpoint inhibitor (ICI) group and 12,927 patients in the non-ICI group. After propensity score matching, 8,744 ICI-treated patients were matched to an equal number of non-ICI controls, yielding two balanced cohorts for subsequent analyses.

To provide additional clinical context and mitigate potential detection bias, patients were further stratified according to pneumonitis severity (No pneumonitis, Grade 1–2, and Grade ≥ 3) (Table 1). After matching, most baseline characteristics were generally comparable across groups. However, smoking history remained notably inconsistent across pneumonitis groups, a pattern likely reflecting underreporting in electronic health records. The tumor stage distribution (I–IV) was similar among pneumonitis strata, with most patients presenting with stage III–IV disease, and no significant difference was observed (P = 0.18, SMD = 0.064).

Patients who developed Grade ≥ 3 pneumonitis tended to be older (65.9 ± 9.4 years) and exhibited a higher prevalence of hypertension (65.9% vs. 12.1%), diabetes mellitus (52.8% vs. 7.4%), and autoimmune diseases such as rheumatoid arthritis and systemic lupus erythematosus. In contrast, patients without pneumonitis had the lowest comorbidity burden. The number of individuals prescribed antirheumatic medications exceeded those diagnosed with rheumatoid arthritis, consistent with their broader use in other autoimmune conditions.

Within the dataset, ICI use was recorded as a binary variable (Yes/No), and specific subtypes (PD-1, PD-L1, or CTLA-4 inhibitors) were not available, precluding subtype-specific analysis. Detailed baseline characteristics before propensity score matching are provided in Supplementary Table S1

Incidence and timing of pneumonitis

Over a median follow-up of 4.8 years, pneumonitis occurred in 28.9% of patients receiving ICIs, compared to 10.0% among non-ICI users. This incidence likely reflects an upper-bound estimate, influenced by the broad case definition and high sensitivity of our electronic health record screening approach. Pneumonitis events accumulated early, with divergence in pneumonitis-free survival emerging within the first few months of therapy, as shown in the time-to-event curves (Fig. 2). In multivariable Cox regression, ICI therapy was associated with a significantly elevated risk of pneumonitis (hazard ratio 2.86, 95% CI: 2.43–3.29). The risk estimates remained consistent across various adjustment models (Table 2). No significant violations of the proportional hazards assumption were observed for any covariates or for the global model (all P > 0.05; Supplementary Figure S1).

Fig. 2.

Pneumonitis-free survival according to ICI exposure. Kaplan–Meier curves depicting pneumonitis-free survival in ICI-treated and non-ICI patients. The ICI group exhibited a significantly higher cumulative risk of pneumonitis during follow-up (log-rank P < 0.001)

Table 2.

Incidence and risk of checkpoint inhibitor–associated pneumonitis (CIP) in NSCLC patients, overall and in clinical subgroups

Group	N	Follow-up time (person-years)	No. of pneumonitis	Cumulative incidence (%)	Incidence rate (cases/1000 person-years)	HR (95% CI)
Model 1
Non-ICI	12,927	22,468	1288	9.96	57.33	Reference 2.86 (2.43–3.29)
ICI	8744	15400.3	2523	28.85	163.83
Model 2
Non-ICI	12,927	22,468	1288	9.96	57.33	Reference 2.86 (2.43–3.29)
ICI	8744	15400.3	2523	28.85	163.83
Model 3
Non-ICI	12,927	22,468	1288	9.96	57.33	Reference 2.86 (2.43–3.29)
ICI	8744	15400.3	2523	28.85	163.83
Group
Autoimmune disease (RA) - No	12,805	22240.9	1281	10	57.6	Reference 0.54 (0.46–0.62)
Autoimmune disease (RA) - Yes	1257	227.1	7	5.6	30.83
Group
Medication (Antirheumatics) - No	11,875	20597.1	1199	10.1	58.21	Reference 0.82 (0.7–0.94)
Medication (Antirheumatics) - Yes	1055	1870.8	89	8.44	47.57
Group
Psychiatric comorbidity (Depressive episode) - No	12,169	21122.9	1220	10.03	57.76	Reference 0.88 (0.75–1.01)
Psychiatric comorbidity (Depressive episode) - Yes	761	1345.1	68	8.94	50.55

Model 1: unadjusted; Model 2: adjusted for age and sex; Model 3: fully adjusted for age, sex, hypertension, diabetes, smoking, and comorbidities

Cumulative incidence (%) was calculated as (number of CIP cases/total patients) × 100

Incidence rates are expressed as cases per 1000 person-years

HR (95% CI) = hazard ratio and 95% confidence interval from Cox proportional-hazards models

ICI immune checkpoint inhibitor, RA rheumatoid arthritis

Reference groups: Non-ICI therapy (for overall models), RA-negative patients (for autoimmune subgroup), and non-antirheumatic users (for medication subgroup)

Risk stratification by subgroup

Subgroup analyses demonstrated that the elevated risk of pneumonitis among ICI-treated patients was generally consistent across demographic and clinical categories (Fig. 3). Older age (≥ 65 years) was associated with higher risk compared with younger patients, whereas no significant difference was observed between male and female patients. Among lifestyle factors, smoking history was strongly associated with increased pneumonitis risk, while alcohol use showed no significant impact.

Fig. 3.

Subgroup-specific risk factors associated with checkpoint inhibitor–associated pneumonitis (CIP). Forest plot showing hazard ratios (HRs) with 95% confidence intervals for CIP across demographic, lifestyle, comorbidity, autoimmune, and tumor-stage subgroups. Older age (≥ 65 years), hypertension, cerebrovascular and ischemic heart disease, autoimmune diseases (RA, SLE, systemic sclerosis), and advanced tumor stage (III–IV) were independently associated with increased CIP risk

With respect to comorbidities, patients with hypertension, diabetes, ischemic heart disease, and cerebrovascular disease exhibited significantly elevated risks, whereas those with liver disease did not show a clear association. Notably, autoimmune conditions—including rheumatoid arthritis, systemic lupus erythematosus (SLE), and systemic sclerosis—were linked to particularly high risks of pneumonitis. A modest, non-significant trend toward increased risk was also observed among patients with depressive disorders. Finally, advanced tumor stage (III–IV) conferred higher risk compared with early-stage disease.

Model performance evaluation

The prognostic model demonstrated good discrimination in both TCGA and CGGA cohorts. In the TCGA training cohort, the C-index was 0.703, with 1- and 3-year AUCs of 0.706 and 0.817, respectively. In the CGGA validation cohort, the C-index was 0.727, with 1- and 3-year AUCs of 0.749 and 0.803. These results indicate consistent predictive performance and support the robustness of the model (Fig. 4).

Fig. 4.

Model performance in predicting CIP risk. A Time-dependent ROC curves in the TCGA training cohort, with C-index = 0.703 and 1-year and 3-year AUCs of 0.706 and 0.817. B Time-dependent ROC curves in the CGGA validation cohort, with C-index = 0.727 and 1-year and 3-year AUCs of 0.749 and 0.803

Transcriptional profiles of CIP

Differential gene expression analysis was performed using the GEO dataset GSE216329 (Bukhari et al., Cell Reports Medicine, 2023), which comprised PBMC-derived T cells from 24 NSCLC patients (18 with lung adenocarcinoma) treated with anti–PD-1 therapy. Seven patients developed checkpoint inhibitor pneumonitis (CIP), while the remaining patients either did not develop CIP or experienced other immune-related adverse events. A total of 47 genes were significantly upregulated and 36 downregulated in CIP cases. Notable upregulated genes included TCF7L1, ATP1B4, and RPL18A (Fig. 5A). GO and KEGG enrichment analyses highlighted multiple immune-related pathways, including interferon-γ signaling, Th17 cell differentiation, TNF signaling, and JAK-STAT signaling (Figs. 5B–C). Among the genes enriched in these pathways, glutaminase (GLS), a key enzyme in glutamine metabolism, has been increasingly recognized for its role in supporting immune cell activation and cytokine production. Therefore, the observed upregulation of GLS in our dataset may reflect enhanced metabolic–immune coupling during checkpoint inhibitor–associated pneumonitis, rather than a direct causal mechanism. Given that only seven CIP cases were included in the transcriptomic dataset, these findings should be interpreted with caution and regarded as exploratory and hypothesis-generating rather than confirmatory.

Fig. 5.

Transcriptomic profiling and functional enrichment in CIP (GSE216329). A Bar plot showing GO/KEGG enrichment results of differentially expressed genes (DEGs) between CIP and non-CIP samples. Immune-related pathways, including cytokine-mediated signaling, T-cell activation, Th17 differentiation, TNF signaling, and interferon-γ response, were prominently enriched. B Dot plot ranking enriched pathways by GeneRatio, where circle size represents gene counts and color indicates adjusted P-values. Pathways involving T-cell receptor signaling, chemokine signaling, NF-κB signaling, and IFN-γ response showed strong enrichment. C Volcano plot displaying DEGs between CIP and non-CIP cases. Significantly upregulated genes included TCF7L1, ATP1B4, and RPL18A, while several ribosomal genes were downregulated

Immune activation and network analysis

Using the GSE216329 dataset, protein–protein interaction (PPI) network analysis was performed to explore functional relationships among the differentially expressed genes. Hub genes such as GHRH, ZBTB21, and PLAU emerged as central regulators within the PBMC-derived T cell transcriptomic profiles (Figs. 6A–B). Immune enrichment analysis with ssGSEA demonstrated elevated activity in pathways including IFN-γ, TNF, and NF-κB signaling in CIP cases (Fig. 6C). Hierarchical clustering further distinguished CIP from non-CIP samples based on systemic immune transcriptional signatures (Fig. 6D).

Fig. 6.

Immune activation and network characteristics in CIP. A Core protein–protein interaction (PPI) network constructed from differentially expressed genes (DEGs), highlighting key immune-related nodes such as GHRH, ZBTB21, and PLAU. B Extended PPI network visualization showing broader interaction patterns and the connectivity landscape of CIP-associated genes within the immune-activation module. C Single-sample GSEA (ssGSEA) analysis indicating significantly higher pathway activities in CIP, including IFN-γ, TNF, IL-17, T-cell receptor, and NF-κB signaling pathways. D Cluster-based transcriptomic profiling of immune-related genes, revealing distinct expression patterns that differentiate CIP from non-CIP patients

Discussion

In this real-world cohort, ICI therapy was linked to a significantly higher pneumonitis risk in NSCLC, with an incidence of ~ 30% versus 10% in non-ICI controls. This exceeds the 4–7% reported in clinical trials such as KEYNOTE-024 and CheckMate-057 [18, 19]. The pronounced discrepancy suggests trial settings may underestimate ICI-related toxicity, while our findings—consistent across subgroups and multivariate models—underscore the complementary value of real-world evidence. It should also be noted that our dataset recorded immune checkpoint inhibitor (ICI) exposure as a binary variable without distinguishing between PD-1, PD-L1, and CTLA-4 inhibitors. Because these subtypes differ in their immune-related toxicity profiles, the lack of subclassification may limit the clinical interpretability and granularity of our findings. A temporal trend analysis from 2010 to 2023 revealed a modest increase in CIP incidence after 2020, which is further illustrated in Supplementary Figure S1.

To explore the underlying mechanisms, we performed transcriptomic integration and identified several immune-related signaling pathways and key upstream regulators potentially involved in CIP pathogenesis. The incidence of pneumonitis in our ICI-treated NSCLC cohort reached nearly 30%, markedly higher than the rates reported in pivotal trials such as KEYNOTE-024 (overall pneumonitis rate of 5.8%, with 2.6% at grade 3–4 severity) and CheckMate-057 (pneumonitis rate approximately 4.6%) [20, 21]. In addition, multiple meta-analyses have consistently shown that the incidence of pneumonitis among NSCLC patients receiving PD-1/PD-L1 monotherapy typically ranges between 3% and 5%. In contrast, the much higher incidence observed in our cohort suggests that the real-world burden of ICI-related pneumonitis may be underestimated in trial-based assessments. In addition, comparisons with other Chinese real-world studies suggest that our findings represent the higher end but are not entirely inconsistent. For example, Zhai et al. reported CIP incidence of 11–14% in NSCLC patients receiving ICIs [16], while Cho et al. and Suresh et al. observed higher pneumonitis rates among Asian cohorts compared with Western populations [17, 22]. A nationwide multi-center analysis from China further demonstrated inter-institutional heterogeneity, with rates ranging from 8% to 19% [23]. Furthermore, we identified upregulated expression of upstream regulatory genes such as GHRH, ZBTB21, and PLAU, not only in our cohort but also across other lung inflammation datasets from the GEO repository.

Compared with previous studies, the incidence of CIP observed in our study was significantly higher [24]. For instance, an analysis by Lee et al. based on the U.S. FDA Adverse Event Reporting System (FAERS) reported a 4.1% incidence of ICI-related pneumonitis among NSCLC patients [23], a rate consistent with many clinical trials. Similarly, Suresh et al. [22] reported a pneumonitis incidence of approximately 19% among ICI-treated NSCLC patients, while Tiu et al. [25] found a 1-year attributable risk of 2.49% (95% CI: 1.50–3.47%) in a multi-institutional claims and EHR-based analysis. Existing systematic reviews have also indicated that the incidence of CIP in NSCLC patients receiving PD-1/PD-L1 monotherapy remains generally low in clinical trial settings, commonly ranging from 3% to 5% [26, 27], substantially lower than what is seen in some real-world studies. In our study, the nearly 30% incidence may reflect more complex patient conditions, higher baseline comorbidity burden, increased management challenges, or more sensitive recognition of atypical symptoms in clinical practice.

GLS plays a central role in glioma beyond basic metabolism [28]. Pharmacological inhibition of GLS eliminates glioblastoma stem-like cells (GSCs), which are closely linked to resistance and recurrence [29, 30]. Glutaminolysis also sustains redox balance and biosynthesis, enabling rapid proliferation and metabolic adaptability [31]. GLS activity differs across GBM subgroups, reflecting tumor heterogeneity. Moreover, compensatory upregulation of GLS after mTOR inhibition supports therapy resistance by maintaining metabolic flux [32]. Together, these findings highlight GLS as both a prognostic marker and a promising therapeutic target, with potential benefit from GLS inhibition in combination strategies.

Subgroup analyses showed that patients with autoimmune diseases, psychiatric disorders, or on immunomodulatory therapy were more susceptible to CIP, a trend rarely seen in clinical trials due to exclusion of these high-risk groups. The consistent model performance in both training and validation cohorts, with C-index values above 0.70 and AUCs ranging from 0.70 to 0.82, underscores the robustness of the prognostic signature.

Despite the strengths of integrating clinical and transcriptomic data, several limitations should be noted. Residual bias remains possible despite propensity score matching. The transcriptomic dataset (GSE216329) involved a small sample and PBMC-derived T cells rather than lung tissue, limiting generalizability and requiring validation in larger cohorts and tissue samples. CIP diagnosis was based on ICD coding and chart review, which may be subject to misclassification. Because manual adjudication of all cases was not feasible, some infectious pneumonia events may have been misclassified as CIP. Therefore, our reported incidence (≈ 29%) may represent an overestimate compared with adjudicated clinical trials. Importantly, however, the relative risk elevation observed in ICI versus non-ICI groups is less likely to be affected by this limitation. Excluding patients with prior pneumonitis may also have removed cases of tumor-related obstructive pneumonia, leading to underestimation. Importantly, the Cox model met proportional hazards assumptions, supporting the robustness of our survival analyses. Calendar-year data were not directly available in the primary dataset, and detailed temporal analysis was therefore provided as Supplementary Table S2. Previous studies have demonstrated that GLS-driven glutamine metabolism promotes T-cell activation and inflammatory cytokine release, which may partially explain the immune-metabolic link observed in our transcriptomic findings.

Limitations

This study has several limitations that should be acknowledged.

First, the electronic health record dataset recorded immune checkpoint inhibitor (ICI) exposure as a binary variable (Yes/No) without subclassification by PD-1, PD-L1, or CTLA-4 inhibitors, which may limit direct comparisons across ICI subtypes.

Second, smoking history might have been incompletely documented in the clinical records, as over 89% of patients were categorized as never-smokers, inconsistent with general NSCLC epidemiology.

Third, pneumonitis events were identified based on ICD-10 coding and clinical documentation rather than standardized radiologic adjudication, which could have led to inclusion of mild or atypical cases.

Fourth, the transcriptomic analysis involved a relatively small number of CIP cases (n = 7) derived from peripheral blood mononuclear cells rather than lung tissue, and the findings should therefore be interpreted as exploratory and hypothesis-generating.

Finally, because the dataset spans 2010–2023 and lacks explicit SARS-CoV-2 infection identifiers, a potential overlap between COVID-19-related pulmonary events and ICI-related pneumonitis cannot be completely ruled out.

Conclusion

In this real-world cohort, ICI therapy was associated with a markedly increased risk of pneumonitis in NSCLC, with an incidence of 29% compared to 10% in non-ICI patients. This elevated risk remained consistent across major subgroups, underscoring the need for proactive monitoring, particularly in patients with autoimmune or psychiatric comorbidities. Transcriptomic analysis revealed upregulation of immune-related pathways such as IFN-γ, Th17, and TNF signaling, and identified candidate regulatory genes potentially linked to pneumonitis susceptibility. Nevertheless, the transcriptomic analysis was based on a limited number of PBMC-derived samples (n = 7), and further validation in larger, lung-tissue-based cohorts is warranted.

Authors’ contributions

Wenbo Zhao and Shiju Shen contributed equally to this work. Wenbo Zhao was responsible for data analysis, manuscript drafting, and visualization. Shiju Shen participated in study design, methodology optimization, and interpretation of results. Huafang Zhao supervised the project, provided clinical resources, and critically revised the manuscript for important intellectual content. All authors read and approved the final manuscript.

Funding

This study is supported by grants No. 82204136 from the National Natural Science Foundation of China.

Data availability

The transcriptomic dataset (GSE216329) used in this study is publicly available in the GEO repository. Other anonymized data generated or analyzed during this study are available from the corresponding authors upon reasonable request.This study was approved by the Ethics Committee of China Medical University (Approval No. 2025410). All procedures involving human participants were conducted in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments. Informed consent was waived due to the retrospective design and use of de-identified clinical data.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of China Medical University (Approval No. 2025410). All procedures involving human participants were conducted in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments. Informed consent was waived due to the retrospective design and use of de-identified clinical data.

Consent for publication

Not applicable. This study does not contain any individual person’s data in any form (including images or videos).

Competing interests

The authors declare no competing interests.

Competing interests

The authors declare that they have no competing interests.