Biomarker exploration for immunotherapy plus chemotherapy following resistance to third-generation EGFR-TKIs in lung adenocarcinoma Authors: Lige Wu, Tongji Xie, Yan Li, Xuezhi Hao, Xin Zhang, Jianming Ying, Junling Li, Puyuan Xing Published: 2026 ABSTRACT BackgroundThe efficacy of immune checkpoint inhibitor plus chemotherapy (ICI-chemo) following third-generation EGFR-TKI resistance in EGFR-mutant lung adenocarcinoma (LUAD) is variable and lacks reliable predictive biomarkers. This study investigated associations between treatment-naïve primary tumor transcriptomic profiles, tumor immune microenvironment (TIME) characteristics, and subsequent ICI-chemo efficacy.MethodsThis retrospective study enrolled 12 advanced LUAD patients who received ICI-chemo after progressing on third-generation EGFR-TKIs. Patients were stratified into Better (N = 7) and Worse (N = 5) efficacy groups using 6-month progression-free survival (PFS) as cutoff. Treatment-naïve tumor specimens were analyzed using Digital Spatial Profiling (DSP) whole transcriptome sequencing and multiplex immunofluorescence (mIF) to assess compartment-specific gene expression, pathway enrichment, and infiltration of 19 immune cell types. LASSO-Cox regression was performed to identify key prognostic genes.ResultsSFTPC expression was significantly elevated across all compartments in the Worse efficacy group. Enrichment analysis revealed that the efficacy of ICI-chemo in EGFR-mutant patients arises from a combination of proliferative tumor cells and an activated TIME. A total of eight key prognostic genes were identified: DNM1L_Immune, ADPGK_Immune, SLC12A4_Stroma, PLA2G6_Immune, FKBP1A_Immune, CHST15_Tumor, TRMT5_Immune, TAX1BP1_Immune. Complex correlations were observed between these key genes and the expression of immune checkpoints. mIF revealed significantly higher infiltration of cancer-associated fibroblast subtype I (CAFI) in the Worse group, while the Better efficacy group exhibited a more activated antitumor TIME.ConclusionsBaseline upregulation of SFTPC, enrichment of CAFI, and specific immune checkpoint expression patterns are associated with poor response to ICI-chemo following EGFR-TKI resistance in EGFR-mutant LUAD. The eight spatially resolved genes identified may serve as potential predictive biomarkers for patient stratification. These findings underscore the functional interdependence between tumor-intrinsic properties and TIME characteristics in determining immunotherapy outcomes, and provide biologically plausible candidates warranting independent validation in larger cohorts.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12885-026-16027-w. Background Lung cancer is the leading cause of cancer-related mortality worldwide [1]. Adenocarcinoma is the most common histopathological subtype of lung cancer, accounting for approximately 40% of all cases [2]. The discovery of EGFR activating mutations and immune checkpoints has revolutionized the field of lung cancer therapy, particularly in the case of lung adenocarcinoma (LUAD) [3, 4]. In recent years, EGFR tyrosine kinase inhibitors (EGFR-TKIs) have undergone rapid evolution, and third-generation agents such as osimertinib have become the standard first-line therapy for advanced non-small cell lung cancer (NSCLC) harboring EGFR sensitizing mutations [5]. Known resistance mechanisms to third-generation EGFR-TKIs include on-target EGFR-dependent alterations (e.g., EGFR C797S, EGFR exon 20 insertions, EGFR amplification) and off-target EGFR-independent pathways (e.g., MET amplification, FGFR amplification, HER2 amplification, and small cell lung cancer transformation) [6]. Nevertheless, 30% − 50% of resistance cases remain mechanistically uncharacterized. The combination of immune checkpoint inhibitor (ICI) represented by pembrolizumab and chemotherapy (chemo) is currently the first-line standard treatment for advanced NSCLC without driver gene mutations [5]. Resistance to PD-(L)1 blockade is associated with multiple factors, including a lack of or low PD-L1 expression, absence of tumor-infiltrating lymphocytes, inhibition of antigen presentation, a suppressive tumor immune microenvironment (TIME), epigenetic modifications, and dysregulation of interferon signaling [7]. Treatment-naïve patients with EGFR-mutant NSCLC typically exhibit poor responses to immunotherapy, which may be attributed to the suppressive TIME in EGFR-mutated NSCLC [8, 9]. Targeted therapy and immunotherapy are two cornerstone treatment modalities for patients with advanced NSCLC. Whether immunotherapy can be effectively sequenced following resistance to targeted agents remains a key focus of ongoing research worldwide. A network meta-analysis of immunotherapy after EGFR-TKI resistance [10] demonstrated that, compared with chemo, ICI monotherapy is associated with shorter progression-free survival (PFS), whereas combinations—particularly ICI-chemo and ICI plus antiangiogenesis plus chemo (ICI-antiangio-chemo)—are linked to prolonged PFS. Currently, monoimmunotherapy after EGFR-TKI resistance is not considered standard practice, with the central debate centering on the efficacy of combination strategies. However, pivotal trials such as CheckMate-722 [11] and Keynote-789 [12] failed to show a significant survival benefit with ICI-chemo in this population. In contrast, the ORIENT-31 [13] and ATTLAS [14] trials demonstrated that ICI-antiangio-chemo can significantly improve survival outcomes. Despite these advances, robust head-to-head comparisons among ICI-antiangio, chemo-antiangio, and ICI-chemo remain lacking. Furthermore, validated biomarkers to identify patients most likely to benefit from combination immunotherapy after EGFR-TKI resistance are still absent, highlighting an unmet clinical need. Recent studies have found that EGFR-TKI treatment can remodel the TIME. Comparative analysis of immune cell infiltration before and after EGFR-TKI treatment shows changes in lymphocyte subtypes and macrophage proportions [15]. EGFR-TKIs induce RGS5 + MYL9+ cancer-associated fibroblasts (CAFs) to form tunneling nanotubes that accept damaged mitochondria from tumor cells, thereby contributing to therapeutic resistance [16]. Additionally, EGFR-TKI may promote immune escape by upregulating PD-L1 expression [17, 18]. Although using post-resistance tumor tissue to identify biomarkers for subsequent immunotherapy is theoretically ideal, its real-world application remains challenging due to practical limitations. First, disease progression following EGFR-TKI resistance can manifest as either primary tumor enlargement or the emergence of new metastatic lesions, with each distant metastatic site potentially harboring a distinct TIME. Second, EGFR-TKI treatment may exacerbate intratumoral heterogeneity, rendering biopsy samples subject to sampling bias. Third, resistant tumor tissue is not always accessible or obtainable from patients after disease progression. Collectively, these limitations impede the widespread adoption and clinical translation of resistant tumor tissue analysis for identifying predictive biomarkers of immunotherapy response following EGFR-TKI failure. Given these considerations, this study aims to investigate the associations between transcriptomic profiles, TIME characteristics in treatment-naïve primary tumor specimens of patients with EGFR-mutant LUAD and the clinical efficacy of subsequent immunotherapy combinations. Materials and methods Subject investigated This retrospective study enrolled patients with advanced LUAD who developed resistance to third-generation EGFR-TKIs and subsequently received ICI-chemo treatment at our hospital between January 2019 and January 2025. Clinical and pathological data were obtained from the Electronic Medical Record system in our hospital. This study was conducted using residual tumor tissues and anonymized clinical data from patients with LUAD. This study collected and analyzed tumor tissue samples from patients with EGFR-mutant NSCLC before they received any systemic therapy. This study is a retrospective, non-interventional investigation that neither interferes with standard clinical diagnosis and treatment nor compromises patients’ medical rights or safety. No additional risks were imposed on participants. Patient identifiers were replaced with anonymized codes, and all reported findings are derived exclusively from aggregated, de-identified, and statistically analyzed data—ensuring no disclosure of personally identifiable information. Given that the majority of eligible patients had either died or were lost to follow-up prior to study initiation, direct informed consent was objectively unattainable. Accordingly, the Ethics Committee of the National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College granted a formal waiver of informed consent and approved the study protocol (Approval Number: 25/599–5545), in full compliance with national regulations and the Declaration of Helsinki. Digital Spatial Profiling (DSP) Whole Transcriptome Atlas (WTA) First, formalin-fixed paraffin-embedded (FFPE) tissue sections were hybridized with a panel of mRNA-targeting oligonucleotide probes equipped with photocleavable linkers. After hybridization, the sections were subjected to multiplex fluorescent morphological staining (including SYTO13, panCK, and CD45) to enable precise histological annotation (e.g., tumor, immune, and stroma types). Based on high-resolution fluorescence images combined with H&E staining, regions of interest (ROIs) were selected by the researchers. Using the GeoMx® DSP platform [19], each predefined ROI was irradiated with ultraviolet (UV) light to selectively cleave and release the index oligonucleotides from the probes bound specifically within the illuminated area. The released oligonucleotides, representing the transcriptomic content of each ROI, were collected via microcapillary aspiration into separate wells. Finally, the collected oligonucleotides were processed for library preparation and subjected to next-generation sequencing (NGS). Details of ROI selection for the samples involved in this study are provided in the Supplementary Material (DSP). Multiplex Immunofluorescence (mIF) Four mIF panels were developed to evaluate macrophages, T lymphocytes, cancer-associated fibroblasts (CAFs), dendritic cells (DCs), and B cells, respectively. The cell subtypes, including M1- and M2-polarized macrophages, effector and exhausted CD8+ cytotoxic T cells (Tc-effector and Tc-exhausted), effector and exhausted regulatory T cells (Treg-effector and Treg-exhausted), CD8+ regulatory T cells (T8reg), type I, II, and III CAFs (CAFI, CAFII, CAFIII) [20], B effector and regulatory B cells (Be and Breg) [21], as well as classical, myeloid, and plasmacytoid DCs (cDCs, mDCs, pDCs) [22], were identified based on the expression profiles across different channels within each panel. Pathway enrichment analysis Gene set enrichment analysis (GSEA) was performed based on the whole DSP data. The reference gene sets were selected as KEGG + GO + Hallmark gene sets. A total of 1,350 gene sets related to tumors and the TIME were analyzed and classified into six major categories based on the Molecular Signatures Database (MSigDB): 421 “immunity” gene sets, 127 “cell cycle” gene sets, 315 “metabolism and energy” gene sets, 286 “genetic and epigenetic information” gene sets, 129 “extracellular matrix (ECM) and metastasis” gene sets, and 72 “cell death” gene sets. We ranked genes by the lnHR from univariate Cox regression for PFS. In GSEA, gene sets with a raw P < 0.1 were considered statistically significant. We adopted a lenient significance threshold to maximize the retention of biologically relevant signals. We also performed enrichment analysis on DEGs using the STRING database. Single-sample GSEA (ssGSEA) was conducted using the Gene Set Variation Analysis (GSVA) package to evaluate activation levels of key biological pathways, including signaling, proliferation, metabolic, and other tumor-related key pathways. Key gene screening and validation To initially screen genes associated with patient prognosis, univariate Cox proportional hazards regression models were employed to analyze each gene in the integrated dataset. Using PFS as the endpoint event, the hazard ratio (HR), P-value, and concordance index (C-index) were calculated for each gene. Genes with a P-value < 0.005 were selected as the candidate gene set for subsequent analysis. The Least Absolute Shrinkage and Selection Operator (LASSO) Cox regression model was then utilized for further analysis. First, the expression values of candidate genes were log2-transformed (log2(x + 1)) to reduce data skewness, followed by Z-score normalization to ensure consistent scaling across genes. To determine the optimal regularization parameter λ, repeated cross-validation (50 repetitions, 5-fold cross-validation per repetition) was performed to evaluate model performance (based on C-index) across different λ values. The average of the minimum λ and λ1se values across all repetitions was calculated, and the more conservative average λ1se was selected as the optimal regularization parameter to balance model complexity and predictive performance. To assess the stability of gene selection, Bootstrap resampling (1000 times with replacement) was applied to validate the LASSO model. The selection frequency of each gene across the 1000 resamplings was calculated and ranked in descending order. Finally, high-frequency genes (with a selection frequency > 10%) were identified as key prognostic genes. Using the median expression of each high-frequency gene as the threshold, patients were divided into high-expression (expression ≥ median) and low-expression (expression < median) groups. Kaplan-Meier survival curves were fitted, and the log-rank test was used to compare differences in survival curves between the two groups. Cox proportional hazards regression models were further employed to calculate the C-index, HR, and 95% confidence interval (95% CI) for each gene. Follow-up and survival analysis Patients were stratified into a Better efficacy group and a Worse efficacy group using 6-month PFS as the cutoff. PFS was defined as the time from the initiation of ICI-chemo treatment to disease progression or death from any cause. The last follow-up date was October 10, 2025. The Kaplan-Meier method was used to estimate survival probabilities and generate survival curves, while the log-rank test was applied to assess statistically significant differences in survival distributions between the two groups. Statistics Statistical analyses were performed using R software (version 4.5.0; https://www.r-project.org/). Continuous variables were compared between two independent groups using the Wilcoxon rank-sum test. Spearman rank correlation tests were performed to assess correlations between continuous variables. Statistical significance was defined in each section. Results Study design A total of 12 patients with advanced LUAD who had progressed following treatment with third-generation EGFR-TKIs were enrolled in this study. Of these, 9 patients carried the EGFR L858R mutation and 3 carried the EGFR 19del mutation before the initial treatment. All patients received a combination therapy of PD-1 monoclonal antibody and chemotherapy after developing resistance to third-generation EGFR-TKIs, and disease progression was observed in all cases. The overall median PFS was 7.51 months, and the objective response rate was 25%. Based on the clinical trial results from CheckMate-722 [11] and Keynote-789 [12], patients were stratified into a Better efficacy group (N = 7) and a Worse efficacy group (N = 5) using 6-month PFS as the cutoff. Baseline patient characteristics for both groups are summarized in Table 1. Baseline characteristics and treatment history of all patients before receiving ICI-chemo were summarized in the Supplementary Tables. Table 1Clinical characteristicsCharacteristicsAll casesBetter GroupWorse GroupTotal, n1275Median age, years (range)62.5 (45–73)63 (45–66)62 (53–73)Age group (years), n (%) < 657 (58.3)4 (57.1)3 (60.0) ≥ 655 (41.7)3 (42.9)2 (40.0)Gender, n (%) Male6 (50.0)3 (42.9)3 (60.0) Female6 (50.0)4 (57.1)2 (40.0)Smoking history, n (%) Never smoker8 (66.7)5 (71.4)3 (60.0) Former/current smoker4 (33.3)2 (28.6)2 (40.0)ECOG performance status, n (%) 05 (41.7)3 (42.9)2 (40.0) 17 (58.3)4 (57.1)3 (60.0)EGFR mutation type, n (%) 19del3 (25.0)2 (28.6)1 (20.0) L858R9 (75.0)5 (71.4)4 (80.0)Pathological classification, n (%) Adenocarcinoma12 (100)7 (100)5 (100)Tumor differentiation, n (%) Well-differentiated1 (8.3)0 (0.0)1 (20.0) Moderately differentiated9 (75.0)5 (71.4)4 (80.0) Poorly differentiated2 (16.7)2 (28.6)0 (0.0)Biopsy location, n (%) Lung12 (100)7 (100)5 (100)Previous treatment regimen, n (%) Gefitinib7 (58.3)3 (42.9)4 (80.0) Erlotinib1 (8.3)0 (0.0)1 (20.0) Icotinib3 (25.0)3 (42.9)0 (0.0) Afatinib3 (25.0)2 (28.6)1 (20.0) Dacomitinib1 (8.3)0 (0.0)1 (20.0) Osimertinib12 (100)7 (100)5 (100)ICI-chemo regimen, n (%) Pembrolizumab-chemotherapy3 (25.0)1 (14.3)2 (40.0) Camrelizumab-chemotherapy2 (16.7)2 (28.6)0 (0.0) Tislelizumab-chemotherapy1 (8.3)1 (14.3)0 (0.0) Nivolumab-chemotherapy1 (8.3)1 (14.3)0 (0.0) Toripalimab-chemotherapy1 (8.3)0 (0.0)1 (20.0) Sintilimab-chemotherapy1 (8.3)0 (0.0)1 (20.0) Atezolizumab-chemotherapy3 (25.0)2 (28.6)1 (20.0)PFS for ICI-chemo, months (range)7.5 (0.5–60.7)8.5 (7.5–60.7)2.3 (0.5–5.6)BOR for ICI-chemo, n (%) Partial response3 (25.0)3 (42.9)0 (0.0) Stable disease8 (66.7)4 (57.1)4 (80.0) Progressive disease1 (8.3)0 (0.0)1 (20.0)ICI Immune checkpoint inhibitor, chemo Chemotherapy, PFS Progression-free survival, BOR Best of response Table 1Clinical characteristicsCharacteristicsAll casesBetter GroupWorse GroupTotal, n1275Median age, years (range)62.5 (45–73)63 (45–66)62 (53–73)Age group (years), n (%) < 657 (58.3)4 (57.1)3 (60.0) ≥ 655 (41.7)3 (42.9)2 (40.0)Gender, n (%) Male6 (50.0)3 (42.9)3 (60.0) Female6 (50.0)4 (57.1)2 (40.0)Smoking history, n (%) Never smoker8 (66.7)5 (71.4)3 (60.0) Former/current smoker4 (33.3)2 (28.6)2 (40.0)ECOG performance status, n (%) 05 (41.7)3 (42.9)2 (40.0) 17 (58.3)4 (57.1)3 (60.0)EGFR mutation type, n (%) 19del3 (25.0)2 (28.6)1 (20.0) L858R9 (75.0)5 (71.4)4 (80.0)Pathological classification, n (%) Adenocarcinoma12 (100)7 (100)5 (100)Tumor differentiation, n (%) Well-differentiated1 (8.3)0 (0.0)1 (20.0) Moderately differentiated9 (75.0)5 (71.4)4 (80.0) Poorly differentiated2 (16.7)2 (28.6)0 (0.0)Biopsy location, n (%) Lung12 (100)7 (100)5 (100)Previous treatment regimen, n (%) Gefitinib7 (58.3)3 (42.9)4 (80.0) Erlotinib1 (8.3)0 (0.0)1 (20.0) Icotinib3 (25.0)3 (42.9)0 (0.0) Afatinib3 (25.0)2 (28.6)1 (20.0) Dacomitinib1 (8.3)0 (0.0)1 (20.0) Osimertinib12 (100)7 (100)5 (100)ICI-chemo regimen, n (%) Pembrolizumab-chemotherapy3 (25.0)1 (14.3)2 (40.0) Camrelizumab-chemotherapy2 (16.7)2 (28.6)0 (0.0) Tislelizumab-chemotherapy1 (8.3)1 (14.3)0 (0.0) Nivolumab-chemotherapy1 (8.3)1 (14.3)0 (0.0) Toripalimab-chemotherapy1 (8.3)0 (0.0)1 (20.0) Sintilimab-chemotherapy1 (8.3)0 (0.0)1 (20.0) Atezolizumab-chemotherapy3 (25.0)2 (28.6)1 (20.0)PFS for ICI-chemo, months (range)7.5 (0.5–60.7)8.5 (7.5–60.7)2.3 (0.5–5.6)BOR for ICI-chemo, n (%) Partial response3 (25.0)3 (42.9)0 (0.0) Stable disease8 (66.7)4 (57.1)4 (80.0) Progressive disease1 (8.3)0 (0.0)1 (20.0)ICI Immune checkpoint inhibitor, chemo Chemotherapy, PFS Progression-free survival, BOR Best of response Clinical characteristics ICI Immune checkpoint inhibitor, chemo Chemotherapy, PFS Progression-free survival, BOR Best of response Given the limitations of EGFR-TKI resistant tumor samples—including sampling bias, spatial heterogeneity, and variable tissue quality—we analyzed pretreatment lung tumor specimens from enrolled patients (collected before any systemic therapy) to identify robust, clinically translatable biomarkers. Oligoconjugated antibody staining was performed on each sample, and three distinct ROIs—tumor, immune, and stromal compartments—were annotated for each specimen prior to DSP. The quality-controlled sequencing data were compiled into a gene expression matrix for downstream analysis. We then applied univariate Cox regression, LASSO-Cox regularization, and Kaplan-Meier-based survival stratification sequentially for prognostic feature selection, ultimately identifying eight spatially informative genes. Subsequently, we characterized 19 immune cell populations in the TIME using mIF staining and compared the TIME differences between the two patient groups. A flowchart summarizing the study design is presented in Fig. 1. Fig. 1Study flow chart Study flow chart Transcriptomic differences and functional enrichment Kaplan-Meier curves for PFS in the Better efficacy group and the Worse efficacy group following acquired resistance to EGFR-TKIs and subsequent treatment with ICI-chemo are presented in Fig. 2A. Using the edgeR package [23], we performed a systematic, compartment-specific differential gene expression analysis to compare the two efficacy groups within each of the tumor, immune, and stromal compartments of primary lung lesions (Fig. 2B). Given the distinct expression profiles of different cellular compartments, compartment-specific thresholds for log2 fold change (log2FC) and false discovery rate (FDR) were applied to identify differentially expressed genes (DEGs). In the tumor compartment, genes exhibiting |log2FC| > 1 and FDR < 0.05 were considered statistically significant, yielding 40 DEGs; the top 10 genes with the highest |log2FC| values were SFTPC, PAEP, FGB, DLK1, FGG, NRGN, CLDN10, LAPTM4B, SPANXA1, and MUC3A. In the immune compartment, a threshold of |log2FC| > 1 and FDR < 0.25 was applied to account for higher biological variability, resulting in 20 DEGs; the top 10 genes with the highest |log2FC| values were SFTPC, SFTPA1, SFTPA2, FABP4, AGER, SLC6A4, TMEM100, SCNN1B, USP40, and PDXP. In the stroma compartment, |log2FC| > 1 and FDR < 0.25 were used, identifying 58 DEGs, and the top 10 genes with the highest |log2FC| values were SFTPC, SFTPA2, SFTPA1, NPFF, FGB, AGER, AQP1, CXCL14, MATN3, and TMEM100. Hierarchical clustering heatmap of the key DEGs is presented in Fig. 2C. The Upset plot revealed that SFTPC was consistently differentially expressed across all three compartments and showed significantly higher expression in the Worse group compared to the Better group (Fig. 2D). Fig. 2Differential gene expression analysis between the better efficacy group and the worse efficacy group. A Kaplan-Meier curves for PFS of patients in the Better efficacy group versus Worse efficacy group treated with ICI-chemo. B Volcano plots displaying DEGs in tumor, immune, and stroma compartments identified using the edgeR method. C Heatmap visualization of the top ten DEGs ranked by |log2FC| in each compartment. D UpSet plot showing overlapping DEGs among distinct compartments. E Functional enrichment analysis of DEGs using the STRING database. F GSEA for the tumor, immune and stroma compartments. PFS, progression-free survival; ICI, immune checkpoint inhibitors; chemo, chemotherapy; DEGs, differentially expressed genes Differential gene expression analysis between the better efficacy group and the worse efficacy group. A Kaplan-Meier curves for PFS of patients in the Better efficacy group versus Worse efficacy group treated with ICI-chemo. B Volcano plots displaying DEGs in tumor, immune, and stroma compartments identified using the edgeR method. C Heatmap visualization of the top ten DEGs ranked by |log2FC| in each compartment. D UpSet plot showing overlapping DEGs among distinct compartments. E Functional enrichment analysis of DEGs using the STRING database. F GSEA for the tumor, immune and stroma compartments. PFS, progression-free survival; ICI, immune checkpoint inhibitors; chemo, chemotherapy; DEGs, differentially expressed genes Among the 118 DEGs identified in the comparison between the Worse group and the Better group, 101 genes were upregulated and 17 genes were downregulated. The results of the differential gene expression analysis have been included in the Supplementary Tables. Functional enrichment analysis via the STRING database indicated significant associations between the DEGs and pathways related to idiopathic pulmonary fibrosis, surfactant metabolism disorders, and Secretoglobin family 3 A (Fig. 2E). Subsequently, we performed GSEA using all genes (Fig. 2F). GSEA based on PFS revealed that pathways related to extracellular matrix (ECM) and metastasis, metabolism and energy in the tumor compartment were more inclined to be enriched in patients with worse PFS. In the immune compartment, pathways associated with cell death, immunity, genetic and epigenetic information were more likely enriched in patients with better PFS. For the stroma compartment, pathways related to cell cycle and immunity were more prone to enrichment in patients with better PFS. These findings suggest that the efficacy of ICI-chemo in EGFR-mutant LUAD patients may result from the interaction between tumor cells and the TIME. Patients with longer PFS may possess a more activated anti-tumor immune microenvironment, whereas those with shorter PFS may exhibit higher tumor metabolic activity and metastatic potential. Notably, due to the limited sample size, we did not further apply statistical thresholds for filtering; thus, the results represent tendencies. Details of the GSEA are provided in Supplementary Tables. GSVA for the pathway of interest To further explore differences in key pathways, we performed GSVA. In the tumor compartment, patients with longer PFS exhibited higher enrichment scores in pathways including PI3K-AKT-mTOR, TGF β-signaling, mitotic spindle, p53-related, glycolysis, heme metabolism, and protein secretion (Fig. 3A). In the immune compartment, patients with longer PFS showed higher enrichment scores in the apoptosis pathway (Fig. 3B). In the stroma compartment, patients with longer PFS had higher enrichment scores in both the apoptosis and protein secretion pathways (Fig. 3C). Collectively, these findings suggest that the efficacy of ICI-chemo in EGFR-mutant LUAD patients may arise from the synergistic interplay between proliferative tumor cells (susceptible to chemo) and an activated TIME (supportive of immune-mediated cell death), thereby underscoring the functional interdependence of tumor-intrinsic and extrinsic factors in therapeutic response. Fig. 3GSVA for the pathway of interest. A GSVA for the tumor compartment. B GSVA for the immune compartment. C GSVA for the stroma compartment. GSVA, gene set variation analysis GSVA for the pathway of interest. A GSVA for the tumor compartment. B GSVA for the immune compartment. C GSVA for the stroma compartment. GSVA, gene set variation analysis LASSO-Cox proportional hazards regression for key prognostic gene identification To further identify key prognostic genes for patients with advanced LUAD receiving sequential ICI-chemo after EGFR-TKI resistance, we performed LASSO-Cox regression analysis on 22 genes with a P-value < 0.005 from PFS-based univariate Cox regression. Through repeated cross-validation, the optimal λ was determined as the average λ1se (λ = 0.597), and the average cross-validation curve is shown in Fig. 4A. Subsequently, Bootstrap resampling (1000 times with replacement) was applied to validate the LASSO-Cox results, and the selection frequency of each gene is presented in Fig. 4B. A total of 8 genes with a selection frequency > 10% were identified: DNM1L_Immune (32.2%), ADPGK_Immune (21.9%), SLC12A4_Stroma (18.4%), PLA2G6_Immune (17.8%), FKBP1A_Immune (17.3%), CHST15_Tumor (16.6%), TRMT5_Immune (12.7%), TAX1BP1_Immune (10.3%). Fig. 4LASSO-Cox proportional hazards regression to identify the key prognostic genes. A The average cross-validation curve. B The selection frequency of each gene. C Kaplan-Meier curves for PFS of patients in the High expression group versus Low expression group. PFS, progression-free survival LASSO-Cox proportional hazards regression to identify the key prognostic genes. A The average cross-validation curve. B The selection frequency of each gene. C Kaplan-Meier curves for PFS of patients in the High expression group versus Low expression group. PFS, progression-free survival Next, patients were stratified into high- and low-expression groups based on the median expression of each of the 8 genes, followed by Kaplan-Meier survival analysis. The results revealed distinct separation of survival curves between the two groups for all genes, with the C-index for individual gene prediction all exceeding 0.8. Notably, the C-index for ADPGK_Immune prediction was greater than 0.9 (Fig. 4C). Details of key prognostic gene identification and validation are provided in the Supplementary Tables. TIME and immune checkpoint analysis The infiltration levels of 19 immune cell types in the primary lung tumor tissues from each patient were assessed using mIF technology, with infiltration quantified per unit area for subsequent analysis. A heatmap illustrating the comparison of immune cell infiltration between the Better and Worse groups is presented in Fig. 5A. Wilcoxon rank-sum test analysis revealed no statistically significant differences in overall immune cell infiltration levels between the two groups. Given the limited sample size of this study, which may render the results more susceptible to outlier influence, we further examined the three immune cell types exhibiting the largest magnitude of difference (Fig. 5B). Notably, patients in the Better group tended to exhibit higher infiltration levels of cDCs and mDCs, along with lower infiltration of CAFI, suggesting a potentially more immunologically active TIME. Additionally, when comparing the relative proportions of the 19 immune cell types, the Worse group demonstrated a significantly higher proportion of CAFI infiltration compared to the Better group (Fig. 5C). Fig. 5Comparative analysis of the TIME and immune checkpoints. A Heatmap illustrating the infiltration levels of 19 immune cell types. B Comparison of infiltration levels of the three most significantly differentially abundant immune cells between the Better and Worse efficacy groups. C Comparative analysis of immune cell infiltration proportions. D Expression levels of 15 immune checkpoints in the tumor compartment. E Expression levels of 15 immune checkpoints in the immune compartment. TIME, tumor immune microenvironment Comparative analysis of the TIME and immune checkpoints. A Heatmap illustrating the infiltration levels of 19 immune cell types. B Comparison of infiltration levels of the three most significantly differentially abundant immune cells between the Better and Worse efficacy groups. C Comparative analysis of immune cell infiltration proportions. D Expression levels of 15 immune checkpoints in the tumor compartment. E Expression levels of 15 immune checkpoints in the immune compartment. TIME, tumor immune microenvironment Immune checkpoints are critical regulatory components of the immune system that modulate the intensity and duration of immune responses through activating or inhibitory signaling pathways, thereby maintaining systemic immune homeostasis. While their primary physiological role is to prevent autoimmunity and tissue damage caused by excessive immune activation, tumors frequently co-opt these pathways to evade immune surveillance and establish an immunosuppressive microenvironment [24]. The previous Phase III clinical trial, ATTLAS, demonstrated that higher PD-L1 expression levels in patients with driver gene-positive NSCLC were associated with longer PFS with immunotherapy combination [14]. A retrospective study from Japan, the STEP study, showed that only EGFR-mutant NSCLC patients with PD-L1 ≥ 50% could derive PFS benefit from immunotherapy combination [25]. A clinicopathological study revealed that tumors in EGFR-mutant NSCLC patients with PD-L1 ≥ 50% were typically more aggressive and exhibited higher response rates to immunotherapy combination after resistance to EGFR-TKIs [26]. These findings suggest that immune checkpoint expression levels are closely correlated with the efficacy of ICI combination therapy. In this study, we investigated the differences in common immune checkpoints across various efficacy groups. We utilized CD274 transcription levels to indirectly reflect tumor PD-L1 expression and PDCD1 transcription levels to indirectly reflect immune cell PD-1 expression. Our results demonstrated that CTLA4 and HAVCR2 expression was significantly elevated in tumor cells of the Worse group relative to the Better group (Fig. 5D); and PDCD1 expression was significantly elevated in immune cells of the Worse group (Fig. 5E). Correlation analysis of key prognostic genes with immune checkpoints In the preceding analyses, we identified 8 potential efficacy biomarkers for patients with advanced LUAD receiving sequential ICI-chemo after EGFR-TKI resistance. While we compared differences in immune checkpoint expression on tumor cells and immune cells between the Better efficacy group and Worse efficacy group, the potential associations between these key genes and immune checkpoints remained unclear. Herein, we investigated the expression correlations between these key genes and immune checkpoints using Spearman rank correlation tests, with detailed results provided in Supplementary Tables. Our findings revealed that the expression level of PD-1 on immune cells was positively correlated with SLC12A4_Stroma (r = 0.826, p < 0.001) and TRMT5_Immune (r = 0.716, p = 0.009), while negatively correlated with PLA2G6_Immune (r=-0.743, p = 0.006), ADPGK_Immune (r=-0.695, p = 0.012), and CHST15_Tumor (r=-0.658, p = 0.020). The expression level of PD-L1 on immune cells was positively correlated with FKBP1A_Immune (r=-0.597, p = 0.040). The expression level of OX40 on immune cells was positively correlated with FKBP1A_Immune (r = 0.692, p = 0.013) and negatively correlated with DNM1L_Immune (r=-0.676, p = 0.016). The expression level of TNFRSF9 on immune cells was positively correlated with FKBP1A_Immune (r = 0.632, p = 0.028). The expression level of CD47 on tumor cells was positively correlated with PLA2G6_Immune (r = 0.604, p = 0.038) and negatively correlated with DNM1L_Immune (r=-0.596, p = 0.041). The expression level of IDO1 on immune cells was positively correlated with FKBP1A_Immune (r = 0.594, p = 0.042). The expression level of CD276 on immune cells was negatively correlated with SLC12A4_Stroma (r=-0.578, p = 0.049). Pairwise correlations with a correlation coefficient (r) > 0.6 were visualized as scatter plots with regression lines in Fig. 6. Fig. 6Correlation analysis of key prognostic genes with immune checkpoints. TIME, tumor immune microenvironment Correlation analysis of key prognostic genes with immune checkpoints. TIME, tumor immune microenvironment Discussion Whether sequential ICI-chemo represents a viable treatment strategy following resistance to third-generation EGFR-TKIs remains an unresolved clinical challenge in patients with EGFR-mutated NSCLC. Recent evidence [27] demonstrates that upon the development of resistance, 27.3% (9/33) of NSCLC patients exhibited a transition from immune-cold to immune-hot TIME, whereas 12.1% (4/33) shifted from immune-hot to immune-cold states. Notably, 54.5% (18/33) of patients maintained an immune-cold phenotype throughout treatment, and 6.1% (2/33) retained a stable immune-hot status. These findings underscore the limited capacity of EGFR-TKIs to remodel the TIME, with over 60% of patients showing no substantial change in immune contexture before and after therapy. Given this restricted immune plasticity, leveraging intrinsic transcriptomic profiles and baseline TIME features of primary tumors may provide a rational and predictive framework for assessing response to ICI-chemo after EGFR-TKI failure. In this study, we collected treatment-naïve primary tumor specimens from patients with advanced LUAD who had progressed on third-generation EGFR-TKIs and subsequently received ICI-chemo treatment, for exploratory analysis. Using the DSP platform, tumor tissues were spatially segmented into tumor, immune, and stroma compartments for region-specific transcriptomic sequencing. Compared with the Better efficacy group, patients in the Worse efficacy group exhibited significantly elevated SFTPC gene expression across all three compartments. Prior whole-genome sequencing study has identified indel hotspots in surfactant protein genes—including SFTPA1, SFTPB, and SFTPC—in patients with LUAD [28], and carriers of surfactant-related gene variants are at increased risk of developing lung cancer [29]. Recent evidence indicates that SFTPC suppresses epithelial-mesenchymal transition in LUAD by upregulating SOX7 and inhibiting the WNT/β-catenin signaling pathway [30], while miR-629-3p-mediated downregulation of SFTPC promotes LUAD cell proliferation and is associated with poor survival [31]. SFTPC is a marker of AT2 cell differentiation, but it is also transcribed at low levels in various cell types, including lymphocytes. Previous studies reported tumor‑suppressive roles for SFTPC, but our finding that high SFTPC expression correlates with poor ICI‑chemo efficacy is not contradictory. Those studies focused on intrinsic gene functions, whereas we assessed treatment response involving multi‑mechanism interactions. Elevated SFTPC expression is known to cause idiopathic pulmonary fibrosis with collagen deposition [32, 33]. We thus hypothesize that SFTPC may affect ICI‑chemo efficacy by modulating the TIME, particularly CAFs. Indeed, we found that higher SFTPC expression significantly correlated with increased CAFI abundance and decreased dendritic and effector T cells, explaining the poor response. Correlation results are shown in the Supplementary Tables. These findings indicate that SFTPC holds promise as a predictive biomarker of therapeutic response; however, validation in a larger, independent cohort is warranted to confirm its clinical utility. Enrichment analysis results revealed that the efficacy of ICI-chemo in EGFR-TKI resistant LUAD is a collective outcome of the interplay between tumor cells and the TIME. Activation of cell death pathways within tumor cells and the presence of an anti-tumor immune microenvironment serve as the basis for sensitivity to ICI-chemo, whereas the metabolic and energetic status of tumor cells profoundly influences tumor responses to ICI-chemo. Quantitative analysis of immune cell infiltration in the TIME revealed that CAFI—a CAF subset robustly implicated in tumor protection and characterized by high expression of hepatocyte growth factor (HGF) and fibroblast growth factor 7 (FGF7) [20]—exhibited significantly elevated infiltration levels and proportional abundance in tumors from the Worse efficacy group. These observations align with prior evidence demonstrating that high density of matrix CAFs correlates with suppressed intratumoral immune infiltration and is negatively correlated with patient survival [34]. Mechanistically, collagen secreted by CAFs deposits extensively around tumor cells and forming a collagen-rich “capsule” around and within tumors, compartmentalizing cancer cells into small clusters and blocking direct contact between anti-tumor immune cells and tumor cells [35]. Recently, emerging preclinical data suggest that proteolysis-targeting chimera (PROTAC)-based nanotherapeutics may overcome this stromal resistance by simultaneously eliminating both lung cancer cells and CAFs, thereby potentiating antitumor immunity and therapeutic efficacy [36]. Through stepwise regression, we identified eight prognostic genes with spatial information. The present study found that elevated expression of DNM1L and TRMT5 in the immune compartment may indicate prolonged PFS in LUAD patients. Correlation analysis further revealed that DNM1L expression was negatively correlated with that of the costimulatory immune checkpoint OX40 (TNFRSF4), while TRMT5 expression was positively correlated with that of the inhibitory immune checkpoint PD-1 (PDCD1). These correlations suggest potential links between these genes and immune checkpoint regulation, but their functional implications require further investigation. DNM1L encodes a member of the dynamin superfamily of GTPases, and the encoded protein mediates mitochondrial and peroxisomal division and is involved in developmentally regulated apoptosis and programmed necrosis [37, 38]. TRMT5 is involved in mitochondrial tRNA methylation [39], but its functional role in tumorigenesis and antitumor immunity remains incompletely characterized and warrants systematic investigation. Decreased expression of ADPGK, PLA2G6, FKBP1A, and TAX1BP1 in the immune compartment may indicate prolonged PFS in LUAD patients. ADPGK and PLA2G6 were negatively correlated with PD-1 (PDCD1) expression in immune cells, while PLA2G6 was positively correlated with the expression of the inhibitory immune checkpoint CD47, suggesting that ADPGK and PLA2G6 are biomarkers of insensitivity to PD-1 antibody therapy. ADPGK has been found to be associated with poorer survival in prostate cancer patients [40], and PLA2G6-mediated reprogramming of phospholipid metabolism is associated with the development of hepatocellular carcinoma [41]. Although FKBP1A expression was positively correlated with that of the costimulatory immune checkpoints OX40 (TNFRSF4) and 4-1BB (TNFRSF9) in immune cells, previous study has shown that FKBP1A expression correlates with poor overall survival, disease-specific survival, disease-free interval, and progression-free interval in several cancers [42]. TAX1BP1 is a selective autophagy receptor that mediates autophagic clearance in lung cancer cells [43]. Additionally, CHST15 encodes a glycosyltransferase and has been found to promote PD-L1-mediated immune evasion in lung cancer cells [44]. Therefore, the eight candidate prognostic genes identified in this study represent biologically plausible and statistically supported candidates warranting independent experimental and clinical validation. In summary, through an exploratory analysis of a small-sample cohort of patients with EGFR-mutated LUAD, we identified candidate biomarkers predictive of response to sequential ICI-chemo following disease progression on third-generation EGFR-TKIs, and comprehensively characterized the TIME and immune checkpoint expression. This study has several limitations. First, the sample size is small and the LASSO-Cox model is at risk of overfitting, and all findings are derived from exploratory analyses; therefore, the results require validation in larger and more diverse cohorts. Second, only treatment-naïve primary tumor specimens were used, without paired specimens collected after the development of EGFR-TKI resistance, limiting our ability to assess dynamic changes in tumors and the TIME induced by EGFR-TKI therapy. Third, the complex tumor architecture was divided into three spatial compartments (tumor, immune, and stroma), a pragmatic yet inherently reductionist framework that may fail to capture fine-scale spatial interactions within the TIME. Fourth, recent studies have demonstrated that among NSCLC patients, those carrying the EGFR L858R mutation experience a more pronounced PFS benefit from immunotherapy combination [45, 46]. Nevertheless, our study did not allow for subgroup analyses stratified by EGFR mutation subtypes. Fifth, our study did not include a control group of patients who received chemotherapy alone after resistance to EGFR-TKIs. Therefore, it remains unclear whether these biomarkers are truly effective predictors of response to ICI-chemo or merely reflect the impact of tumor aggressiveness or host-related factors. Conclusions Baseline upregulation of SFTPC, enrichment of CAFI, and specific immune checkpoint expression patterns are associated with poor response to ICI-chemo following EGFR-TKI resistance in EGFR-mutant LUAD. The eight spatially resolved genes identified may serve as potential predictive biomarkers for patient stratification. These findings underscore the functional interdependence between tumor-intrinsic properties and TIME characteristics in determining immunotherapy outcomes, and provide biologically plausible candidates warranting independent validation in larger cohorts. Supplementary Information Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2