Longitudinal spatial multi-omics delineates tumor microenvironment remodeling across sequential EGFR-TKIs in EGFR-mutant NSCLC

Abstract

Background

Resistance to therapy is a frequent occurrence in patients with epidermal growth factor receptor (EGFR)-mutant non-small-cell lung cancer (NSCLC) who are treated with EGFR-tyrosine kinase inhibitors (EGFR-TKIs). However, spatial information on how the tumor microenvironment (TME) changes within the same patient from baseline to resistance to first- and then third-generation EGFR-TKIs is scarce. Here, we used rare consecutive re-biopsy samples to build an exploratory, longitudinal, and spatially resolved atlas, aiming to describe within-patient TME remodeling patterns as a resource for hypothesis generation.

Methods

A patient-matched cohort with serial samples obtained at baseline (T0), after first-generation EGFR-TKI resistance (T1), and after third-generation EGFR- TKI resistance (T2) was established. Using GeoMx Digital Spatial Profiling (DSP), tumor-enriched and stroma-enriched areas of interest (AOIs) were segmented, and paired RNA and protein profiles were quantified. We tracked temporal changes in compartment-specific heterogeneity and immune remodeling and explored associations between early remodeling and subsequent T790M acquisition.

Results

We analyzed the data from 15 samples of 6 patients. Tumor- and stroma-enriched compartments were not only transcriptionally but also protein-wise consistently distinct. Spatial heterogeneity was relatively stable from T0 to T1 but increased from T1 to T2, with changes in the stroma-enriched compartments accounting for most of the increase. Early remodeling featured loss of T cell activation programs, reduced neutrophil signatures, increased myeloid remodeling, and impairment of antigen presentation. Longitudinal analyses suggested a biphasic immune trajectory, with early myeloid remodeling and late stromal checkpoint reprogramming. Early spatial remodeling patterns showed potential differences between samples that later acquired T790M and those that did not.

Conclusions

This study provides preliminary evidence for dynamic TME changes during sequential EGFR-TKI therapy in EGFR-mutant NSCLC. It suggests a late, stroma-associated expansion of spatial heterogeneity and stage-dependent immune remodeling. Additionally, it offers an initial spatial atlas and generates testable, time-aware hypotheses for future validation.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12967-026-08897-2.

Background

Treatment of epidermal growth factor receptor (EGFR)-mutated non-small-cell lung cancer (NSCLC) has advanced rapidly with successive generations of EGFR-tyrosine kinase inhibitors (TKIs) improving outcomes [1–3]. Yet acquired resistance remains nearly inevitable, and despite substantial progress in characterizing resistance, its dynamic evolution and determinants are still poorly defined [4, 5].

An intractable clinical problem is pronounced inter-patient heterogeneity: progression tempo, lesion behavior, and subsequent therapeutic benefit vary widely across individuals, implying that resistance is often a trajectory rather than a single endpoint [6–8]. Cross-sectional sampling across different patients inevitably conflates baseline differences with treatment-associated remodeling, obscuring true time-dependent trends [9, 10]. Collecting longitudinal, within-patient samples is the best way to address this confounding, but it is inherently difficult to execute because sequential EGFR-TKI courses span long intervals, requiring prolonged follow-up and multiple biopsy windows under real-world constraints.

In parallel, some studies have highlighted that the TME in EGFR-mutant NSCLC is spatially heterogeneous, with compartment-specific cellular and molecular programs that are not captured by bulk profiling [11, 12]. Spatially resolved platforms enable high-plex RNA and protein quantification in morphologically defined compartments from FFPE biopsies, providing a pragmatic approach to interrogate tumor–stroma ecosystems in situ [13]. Emerging Digital Spatial Profiling (DSP) studies further suggest that EGFR-TKI exposure can drive compartment-specific remodeling. However, the avaliable spatial evidence remains cross-sectional or limited to two time points [14]. Thus, the full sequential trajectory remains unclear. To address this gap, we established a rare, patient-matched three-time-point biopsy series and applied compartment-resolved DSP profiling to explore tumor–stroma ecosystem evolution under EGFR-TKI therapy.

Methods

Patient and sample collection

We retrospectively screened patients who were initially diagnosed with EGFR-mutant NSCLC at Shanghai Pulmonary Hospital from 2015 onward who had longitudinal follow-up documenting subsequent systemic therapies. Patients were eligible if they had archived lung biopsy tissues at three matched clinical time points obtained at our institution: baseline biopsy at initial diagnosis, re-biopsy at acquired resistance to first-generation EGFR-TKIs, and re-biopsy again at acquired resistance to third-generation EGFR-TKIs. All three biopsies were obtained from the lung and had sufficient material for spatial profiling.

Using these criteria, three patients met the requirements. All three patients developed the EGFR T790M mutation after first-generation EGFR-TKI resistance, received third-generation EGFR-TKI therapy, and eventually developed resistance, enabling collection of the third time-point biopsy. To explore stratification by T790M acquisition, we additionally included three T790M-negative patients who had paired baseline and post–first-generation EGFR-TKI resistance lung biopsies, but they did not proceed to third-generation EGFR-TKI treatment, therefore, only two time points were available for these patients. In total, 15 lung biopsy specimens from 6 patients were included for subsequent DSP protein and RNA analyses.

Sample preparation for Nanostring GeoMx DSP protein assay

The formalin-fixed paraffin-embedded (FFPE) slides were placed in a 60 °C drying oven and baked for 1.5 h. The slide were then placed in a rack and the following wash steps were performed to deparaffinize and rehydrate the slides: 3 washes (5 min each) in CitriSolv, 2 washes (5 min each) in 100% ethanol, 2 washes (5 min each) in 95% ethanol, and 2 washes (5 min each) in double-distilled water. For antigen retrieval, the slides were placed in a staining jar with 1 × freshly prepared Citrate Buffer (pH = 6) at room temperature. The staining jar was placed into a preheated pressure cooker and run on high pressure and temperature for 15 min. When the timer reached zero, the pressure was released. The staining jar was transferred to a lab bench at room temperature. The lid of the staining jar was removed, and the slides were allowed to stand for 25 min. 1 × Tris-buffered saline with Tween-20 (TBS-T) was used to wash the slides for 5 min. After antigen retrieval, the slides were placed into a humidity chamber and incubated with Buffer W (NanoString) for 1 h at 25 °C to perform blocking. Next, the slides were stained with morphological markers CD45 and panCK as well as a mixture of UV-photocleavable oligo antibody sets (Immune Cell Typing Module, IO Drug Target Module, Immune Cell Profiling Core, and Pan-Tumor Module) (NanoString, 121,300,301) in the humidity chamber at 4 °C and incubated overnight. The complete list of protein markers and their corresponding panels is provided in Table S1. The slides were then washed three times (10 min each) in TBS-T, covered with 4% paraformaldehyde, and incubated for 30 min in a humidity chamber at 25 °C. After incubation, the slides were washed twice (5 min each) in TBS-T. Then the slides were stained with SYTO 13 (NanoString, 121,300,301) at 25 °C for 15 min and rinsed with 1 × TBS-T. After staining, the slides were loaded onto the GeoMx instrument.

Sample preparation for Nanostring GeoMx DSP RNA assay

FFPE slides (4 µm) were placed in a 60 °C drying oven and baked for 1.5 h. The slides were then placed in a rack and underwent the following washes to deparaffinize and rehydrate the slides: 3 washes (5 min each) in CitriSolv, 2 washes (5 min each) in 100% ethanol, 1 wash (5 min) in 95% ethanol, and 1 wash (5 min) in phosphate-buffered saline (PBS). To perform target retrieval, the slides were placed in diethylpyrocarbonate-treated water for 10 s to bring the temperature up to 99 °C. Then the slides were transferred to the preheated 1 × Tris-Ethylene Diamine Tetraacetic Acid at 99 °C and incubated for 20 min. When the retrieval time was up, the slides were immediately moved to 1 × PBS at 25 °C and were washed in 1 × PBS for 5 min. For RNA target exposure, the slides were incubated in proteinase K solution with a concentration of 1 µg/mL at 37 °C for 15 min and then washed in 1 × PBS for 5 min.

When RNA target exposure was completed, the slides were postfixed as follows: 1 wash (5 min) in 10% neutral buffered formalin (NBF), 2 washes (5 min each) in NBF Stop Buffer, and 1 wash (5 min) in 1 × PBS. The slides were removed from 1 × PBS and placed in hybridization chamber to perform in situ hybridization. The hybridization solution containing UV-photocleavable oligo probes (Cancer Transcriptome Atlas, CTA) was added to the slide to detect the target RNA. The full list of RNA targets and their signature classifications is provided in Table S2. The hybridization chamber was closed and inserted into the hybridization oven. Then, the slides were incubated at 37 °C overnight. After incubation, the slides were dipped in 2 × Saline Sodium Citrate (SSC) buffer and then subjected to stringent washes to remove off-target probes as follows: 2 washes (25 min each) in Stringent Wash at 37 °C, 2 washes (2 min each) in 2 × SSC. Then, buffer W was used to cover the tissue at room temperature for 30 min in the dark. After removing the Buffer W, the slides were stained with morphological markers CD45 and panCK as well as nuclear stain SYTO 13 (NanoString, 121,300,310) for 1 h in a humidity chamber at room temperature. After staining, the solution was removed and the slides were washed in 2 × SSC, 2 times for 5 min each. Finally, the slides were loaded onto the GeoMx DSP instrument.

Digital spatial profiling

Prepared slides were scanned to capture fluorescent images. Regions of interest (ROIs) were independently selected by two participating authors and a pathologist based on immunofluorescence images. Prior to DSP, adjacent sections were stained with hematoxylin and eosin (H&E) to evaluate tissue morphology and identify representative regions. These regions were marked as references for subsequent ROI selection on the DSP platform. For each specimen, 1–10 ROIs were selected. ROIs were manually delineated using custom polygons, excluding necrotic, hemorrhagic, or low-cellularity areas. A sequential mask strategy was used to segment the ROI into stromal areas of interest (AOIs) (panCK-negative, CD45-positive staining) and tumor AOIs (panCK-positive staining) using the DSP platform’s integrated algorithm. AOIs were treated as repeated measurements within each specimen; primary inference was performed at the specimen level using aggregation to avoid pseudo-replication. After segmentation, the GeoMx DSP instrument (NanoString) was used to cut and collect the spatially indexed barcode. To quantify protein expression, the collection plate was removed from the instrument. The aspirate containing the barcode in the collection plate was dried and then rehydrated. For hybridization, the samples were transferred to a new plate and combined with GeoMx Hyb Code Pack reagents. The hybridization reaction took place overnight. After hybridization, the products were pooled into a strip tube. Then the strip tubes were loaded on the nCounter Analysis System for counting the probes collected with GeoMx DSP. When counting was completed, the reporter code count (RCC) files were generated by the nCounter system and uploaded onto the GeoMx DSP system.

In the GeoMx DSP system, the oligonucleotide counts with spatial data were integrated and then used for data analysis. To quantify gene expression, the collection plate was removed from the instrument and was dried down by incubating on a thermocycler at 65 °C for 1 hour. The sample was then rehydrated by adding nuclease-free water. The polymerase chain reaction (PCR) setup was performed as follows: 5 × PCR Master Mix was aliquoted into each well of a new PCR plate; Primer Mix from each well of the Seq Code Primer Plate was transferred to the corresponding well of the PCR plate; DSP aspirate from each well of the DSP collection plate was transferred to the corresponding well of the PCR plate. When the PCR was completed, the PCR products were pooled and purified. After purification, the library quality and quantity were assessed. Next, the sequencing was performed on an Illumina NextSeq 550AR instrument. After the sequencing, the FASTQ (a sequencing data format) files were generated. The FASTQ files were then converted to digital count conversion (DCC) files using the NanoString GeoMx next-generation sequencing (NGS) Pipeline. The DCC files were then selected and uploaded in the GeoMx DSP system, selecting and uploading the DCC files. Finally, the oligonucleotide counts with spatial data were integrated and used for data analysis.

Spatial proteomic data analysis

The RCC files were uploaded onto the GeoMx DSP system. In the GeoMx DSP Control Center, the Data Analysis module (Version 2.4.0.421) was used to perform quality control (QC) checks and data analysis. There were three steps in the QC process for the nCounter readout: the percentage of fields of view (FOVs), binding density, minimum nuclei and surface area count. At least 75% of FOVs were successfully counted to obtain robust data. The binding density measureed the concentration of barcodes detected by the nCounter instrument. AOIs were retained when the binding density ranged from 0.1 to 2.25. For each AOI, the cut-off values for the number of nuclei and the surface area were set at 20 and 1600 square microns, respectively. In order to avoid the variations across different ROIs, the sizes of these ROIs were adjusted through two steps: area normalization and different cell count. Furthermore, the background of immunoglobulin G was used to normalize all targets that met QC across samples. The normalized counts were used to compare across different groups. The QC parameters for each AOI in the protein and RNA assay are summarized in Table S3A and S3B.

Spatial transcriptomic data analysis

The generated FASTQ files were converted to DCC files with the GeoMx NGS Pipeline Software (version 2.2.0.2). The DCC files were then uploaded onto the GeoMx DSP system. In the GeoMx DSP Control Center, the Data Analysis Module (Version 2.4.0.421) was used to perform QC checks and data analysis. There were four steps in the QC process for the transcriptome data: technical signal QC, technical background QC, probe QC, and normalization QC. Technical signal QC assessed the quality of sequencing for each segment. When the percentage of the raw reads that align to a target sequence was less than 80%, the corresponding ROI was removed. Technical background QC was a GeoMx DSP run control. There were three indicators for technical background QC control: negative probe count, no template control (NTC) count, minimum nuclei and surface area count. The negative probe count established the level of technical noise expected and flags segments with signal below that level. The cut-off value for the negative probe count was 4 counts. The NTC count established the level at which counts in the NTC would be flagged. This QC was used to detect contamination in the library preparation. ROIs were removed when the NTC count was more than 1000. The parameters for the minimum nuclei and surface area count for the ROI were 100 and 8000 square microns, respectively. In order to avoid variations in the different ROIs, the sizes of these ROIs were adjusted through two steps: area normalization and different cell counts. Furthermore, the Quantile 3 (Q3) normalization was used to normalize all targets that met QC across samples [15].

Spatial heterogeneity was quantified using the coefficient of variation (CV). For each time point (T0, T1, T2) and each compartment (tumor vs. stroma), CV was calculated as the standard deviation divided by the mean of expression values for all detected genes across all AOIs within the respective group. A higher CV indicates greater expression variation across AOIs, reflecting increased spatial heterogeneity. Compartment-specific CV values were calculated separately for tumor AOIs and stromal AOIs.

The gene expression matrix was used to compare across groups and calculate signature scores with single sample gene set enrichment analysis (ssGSEA), including immune response, cell function, signaling pathways, and metabolism [16]. The edgeR (3.34.0) package was used to analyze the differences between different groups, and the significant difference was determined with a False Discovery Rate (FDR) value < 0.05 and |log2 (Fold Change)| >1 [17]. KOBAS-i web tool was used to perform Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment with a p < 0.05 [18].

Statistical analysis

The R (version 4.1.0) package of ComplexHeatmap (version 2.8.0) was used for heatmap visualization [19]. Uniform Manifold Approximation and Projection (UMAP) (version 0.2.8.0) was used for dimensionality reduction. The two-tailed Wilcoxon rank-sum test was used for group comparisons, with p < 0.05 considered statistically significant. Pearson’s correlation analysis was employed to assess pairwise correlations. The Benjamini–Hochberg FDR was used to correct for multiple testing in high-dimensional comparisons. Given that multiple AOIs were collected from a single specimen, all analytical outcomes were aggregated at the patient level to avoid pseudo-replication arising from intra-patient correlation.

Results

Establishment of a spatially resolved, patient-matched longitudinal cohort to track microenvironment remodeling during EGFR-TKI therapy

We constructed a patient-matched longitudinal cohort specifically designed along the therapeutic timeline and key molecular evolutionary landmarks (Fig. 1A). All enrolled patients had lung adenocarcinoma harboring classical sensitizing EGFR mutations (L858R, exon 19 deletion, or G719X) and received first-generation EGFR-TKIs (Gefitinib or Icotinib) as first-line treatment. Upon disease progression, patients were stratified according to T790M mutation status. T790M-positive patients subsequently received the third-generation EGFR-TKI (Osimertinib). The baseline characteristics of the enrolled patients are summarized in Table 1, and the detailed clinical information of each individual are provided in Table S4. Tumor samples were categorized into three clinically defined time points: baseline before treatment (T0), post-first-generation EGFR-TKI at first progression (T1), and post-third-generation EGFR-TKI at second progression (T2). Importantly, longitudinal sampling was performed on lung tissues from the same individuals across time points, enabling within-patient comparisons under a consistent host background.

Fig. 1.

A spatially resolved, patient-matched longitudinal cohort to track microenvironment remodeling during EGFR-TKI therapy. (A) Study design of the longitudinal cohort. (B) Overview of GeoMx DSP for parallel high-plex protein and RNA profiling. (C) Representative segmentation of ROI into tumor-enriched and stroma-enriched AOIs. (D) AOI numbers profiled for protein and RNA and the corresponding module coverage

Table 1.

Clinical characteristics

Characteristics	All cases (n = 6)	T790M (-) (n = 3)	T790M (+) (n = 3)
Median age(range), y		66 (57–73)	67 (56–67)
Age group, n (%)
<65 years	2 (33.3)	1 (33.3)	1 (33.3)
≥65 years	4 (66.7)	2 (66.7)	2 (66.7)
Gender, n (%)
Male	5 (83.3)	3 (100.0)	2 (66.7)
Female	1 (16.7)	0 (0.0)	1 (33.3)
Smoking history, n (%)
Never-smoker	3 (50.0)	2 (66.7)	1 (33.3)
Former/current smoker	3 (50.0)	1 (33.3)	2 (66.7)
TNM stage, n (%)
I- II	0 (0.0)	0 (0.0)	0 (0.0)
III- IV	6 (100.0)	3 (100.0)	3 (100.0)
Distant metastasis, n (%)
Yes	6 (100.0)	3 (100.0)	3 (100.0)
No	0 (0.0)	0 (0.0)	0 (0.0)
EGFR mutation type, n (%)
Exon19 19-del	1 (16.7)	1 (33.3)	0 (0.0)
Exon21 L858R	4 (66.7)	2 (66.7)	2 (66.7)
Exon19 19-del & Exon18 G719X	1 (16.7)	0 (0.0)	1 (33.3)

To preserve tissue architecture while quantifying immune and tumor programs, we performed GeoMx DSP for both proteins and transcripts at each time point (Fig. 1B). Each section was segmented into tumor-enriched and stroma-enriched AOIs (Fig. 1C) for compartment-resolved analyses. In total, we profiled 60 tumor and 43 stromal AOIs for proteins, 60 tumor and 46 stromal AOIs for RNA (Fig. 1D). In addition, QC showed comparable AOI area, cellularity, detection metrics, and background across samples (Figs. S1, S2), supporting consistent segmentation and assay performance across the longitudinal dataset.

Longitudinal spatial dynamics reveal a late, stroma-driven increase in heterogeneity during EGFR-TKI treatment

To place subsequent analyses in a global spatial context, we first quantified compartment structure and spatial similarity across the longitudinal series. The same AOIs were profiled at both the protein and RNA levels (Fig. 2A). In UMAP, tumor- and stromal AOIs formed two well-separated clusters in the protein data (Fig. 2B), and the same separation was seen in the RNA data (Fig. 2C), showing that tumor and stroma remained distinct in both readouts.

Fig. 2.

Longitudinal spatial dynamics reveal a late, stroma-driven increase in heterogeneity during EGFR-TKI treatment. (A) The representative images of AOIs selected from the same location for the DSP protein and RNA assay. (B-C) UMAP for dimensionality reduction analysis based on protein (B) and RNA (C) expression data. (D-E) Comparison of stromal or tumor AOIs from the same or different ROIs performed with DSP protein (D) and RNA (E) assay. (F) Heatmap of protein data in different regions for each patient at different timepoint. The top track showed the AOIs region for each sample. The central heatmap displayed the expression of proteins for each AOI. The region and sample information corresponding to AOIs are displayed at the bottom. (G) The change of heterogeneity pre- and post-first and thied generation EGFR-TKIs treatment. *Wilcoxon rank-sum test, p < 0.05

Next, protein and RNA correlations across different spatial distances were compared. At protein level, within-compartment correlations (tumor–tumor and stroma–stroma) were higher than cross-compartment correlations (tumor–stroma; p < 0.001). Cross-compartment similarity also showed a clear gradient: tumor–stroma correlations within the same ROI were higher than tumor–stroma correlations across different ROIs (Fig. 2D). This compartment-separated correlation pattern was further corroborated at the RNA level (Fig. 2E), and also consistent with visual inspection of the tissue images (Fig. S3A, S3B), indicating shared local programs at short distances despite overall compartment separation.

On this basis, we then explored whether spatial heterogeneity changed with treatment stage. Global expression patterns became more dispersed at T2 in both protein and RNA spaces (Fig. 2F, S3C). Statistical analysis found that heterogeneity did not increase from T0 to T1 (p = 0.5) but increased from T1 to T2 (p = 0.021) (Fig. 2G). This suggests a trend that the increase in spatial complexity was not evident during the early period but was observed in later stages under sustained accumulation of therapeutic pressure. When assessed by compartment, this increase was confined to stromal AOIs (T1 to T2, p = 0.008) (Fig. S4A), with no significant change in tumor AOIs (p = 0.5) (Fig. S4B). Consistent with these results, stromal AOIs were more similar to each other than tumor AOIs in the baseline correlation analysis (p < 0.0001) (Fig. S3D), and this relative stromal uniformity was progressively lost at later stages.

Compartment identity was supported by expected molecular features. Tumor AOIs showed higher epithelial markers, whereas stromal AOIs were enriched for extracellular matrix and immune-related molecules (Fig. S4C), with concordant trends at the RNA level (Fig. S4D). Pathway enrichment further showed that tumor regions were more biased toward tumor signaling and resistance-associated modules, whereas stromal regions were significantly enriched for immune–stromal response networks such as chemokine signaling, complement and coagulation cascades, and cytokine–cytokine receptor interactions (Fig. S4E). Together, these data indicate that the increase in spatial complexity emerges late during sequential EGFR-TKI therapy and appears to be primarily associated with stromal remodeling.

Early EGFR-TKI exposure induces stromal immune reprogramming and tumor intrinsic adaptation

To identify early events that precede the later rise in heterogeneity, we compared T0 with T1. In DSP protein data, T1 showed lower T-cell activation and checkpoint markers [tumor necrosis factor receptor superfamily member 9 (TNFRSF9/4-1BB), programmed cell death protein 1 (PD-1), cytotoxic T-lymphocyte–associated protein 4 (CTLA4)], reduced CD66b, increased CD163, and decreased β2-microglobulin (B2M) with a lower major histocompatibility complex class II (MHC-II) antigen presentation score (Figs. 3A, B, and S5A). In contrast, V-domain Ig suppressor of T cell activation (VISTA) was increased at T1 (p = 0.035) (Fig. 3B). These findings indicate a shift toward myeloid remodeling and impaired antigen presentation, with concurrent upregulation of a myeloid associated checkpoint. To clarify the spatial origin of these immune signals, we found that PD-1 declined significantly in stromal areas (p = 0.035), whereas VISTA increased (p = 0.044). No significant changes were observed in tumor areas (Fig. S5B), indicating that early immune remodeling may be primarily associated with the stromal compartment.

Fig. 3.

Early EGFR-TKI exposure induces stromal immune reprogramming and tumor intrinsic adaptation. (A) Volcano map of differentially expressed proteins pre- and post-treatment of first-generation EGFR-TKIs. (B) Box plot of differentially expressed proteins pre- and post-treatment of first-generation EGFR-TKIs. (C) Comparison of the transcript signature scores of core metabolic and biosynthetic programs pre- and post-treatment of first-generation EGFR-TKIs in the tumor region. (D) Comparison the scores of cell state and microenvironment-associated programs. (E) Comparison the scores of oncogenic bypass and plasticity-related signaling pathways. *Wilcoxon rank-sum test, p < 0.05

Tumor-side transcriptomic signatures at T1 pointed to a stress-and-repair response, with higher cell cycle (p < 0.001), nucleotide synthesis (p = 0.022), and DNA damage repair scores (p < 0.001), no overall increase in energy modules, and decreased lipid metabolism (p = 0.045) (Fig. 3C). In parallel, pathways linked to outward signaling and microenvironment communication were dampened, including vascular endothelial growth factor (VEGF) (p = 0.010), platelet-derived growth factor (PDGF) (p = 0.028), and G protein coupled receptor (GPCR) signaling (p = 0.030) (Fig. 3D). Among major bypass programs, only Notch signaling rose selectively (p = 0.017) (Fig. 3E). Therefore, first-generation EGFR-TKI exposure might already be associated with measurable stromal immune changes and tumor intrinsic adaptation, providing an early background for later spatial evolution.

Longitudinal profiling reveals a biphasic shift from early innate remodeling to late stromal checkpoint reprogramming

To test whether the early changes observed after first-generation EGFR-TKI evolve into a more stable state under continued selective pressure, we extended the analysis to T2 and examined the immune cell composition. The markers used to define each immune cell type are listed in Table S5. The abundances of immune cells showed stage-dependent shifts, with greater overall variability at T2 (Fig. 4A). From T0 to T1, the dominant change involved innate immunity. Neutrophil abundance decreased (p < 0.001) and did not decrease further at T2 (Fig. 4B). In contrast, fibroblasts and natural killer (NK) cells increased from T1 to T2 (p = 0.004 and p = 0.001, respectively), consistent with late stromal expansion and reconfiguration of innate components (Fig. 4B). Notably, the overall cytotoxicity score did not increase, suggesting a dissociation between cell composition and functional state at late stage.

Fig. 4.

Longitudinal profiling reveals a biphasic shift from early innate remodeling to late stromal checkpoint reprogramming. (A) The lymphocyte infiltration abundance estimated by protein expression of AOIs at different timepoint. (B) The longitudinal changes of lymphocyte infiltration with treatment stage. (C) Bar chart of the protein Fold change pre- and post-first-generation as well as third-generation EGFR-TKIs treatment. (D) The longitudinal changes of the expression of CTLA4, VISTA, and tim-3 as the treatment progresses. (E) Comparison of some signature score for the growth, invasion, metastasis of tumor cells and the related cellular signal regulation mechanisms. *Wilcoxon rank-sum test, p < 0.05

Protein-level longitudinal comparisons further indicated a directional switch in immune regulation. Most T cell and checkpoint markers declined from T0 to T1, whereas several exhaustion-associated axes rebounded at T2 (Fig. 4C). CTLA4 and VISTA increased after an initial decline, and Tim-3 also rose significantly from T1 to T2 (p = 0.02) (Fig. 4D). This down-then-up trajectory suggests that T1 represents a transitional reset, followed by renewed checkpoint remodeling under sustained treatment pressure, in line with the more complex immune landscape at T2.

Transcriptomic signatures supported this staged remodeling at the functional program level. Nucleotide synthesis was already increased from T0 to T1 (p = 0.022) (Fig. S5C), while GPCR signaling decreased from T0 to T1 (p = 0.029) and remained low thereafter (Fig. 4E). At T2, tumor state modules shifted further. Differentiation signatures decreased from T1 to T2 (p = 0.006) (Fig. 4E), mitochondrial metabolism and tricarboxylic acid cycle (TCA) also declined (p = 0.026) (Fig. S5C). In contrast, canonical bypass and growth pathways showed no consistent cohort-level increase (Fig. S5D), supporting a late-stage phenotype dominated by reorganization of the microenvironment and immune regulatory networks rather than broad activation of classic bypass programs.

T790M stratification reveals distinct early remodeling trajectories

To assess whether early remodeling already diverges into distinct evolutionary directions, we stratified patients by T790M acquisition, a hallmark mechanism after first-generation EGFR-TKI resistance. From T0 to T1, T790M-positive patients showed a predominant downward shift in immune regulation and antigen presentation signals, because CD66b, CTLA4, VISTA, and B2M all trended lower (Fig. 5A left). In contrast, T790M-negative patients showed a more prominent increase in effector and tissue-response signals, including Granzyme B (GZMB) and fibronectin (Fig. 5A right). These results suggest that early stage remodeling may differ by resistance background, though this observation is exploratory.

Fig. 5.

Early on-treatment immune remodeling diverges by subsequent T790M acquisition. (A) Volcano map of differentially expressed proteins for EGFR T790M positive (left) and negative (right) NSCLC patients pre- and post-first-generation EGFR-TKIs treatment. (B) Heatmap of estimated abundance of infiltrating lymphocytes for EGFR T790M positive (top) and negative (bottom) NSCLC patients. (C) Box plot of major infiltrating lymphocytes for EGFR T790M positive and negative NSCLC patients. (D) Box plot of antigen presentation-related markers for EGFR T790M positive and negative NSCLC patients. *Wilcoxon rank-sum test, p < 0.05

Cell abundance estimates and functional scores aligned with these patterns. In the T790M-positive group, the most pronounced change from T0 to T1 was a sharp decline in neutrophil abundance (p < 0.001), consistent with reduced CD66b. In the T790M-negative group, the dominant change was an increase in the cytotoxicity signature (p = 0.0051) (Fig. 5B, C). Along antigen presentation and immune regulation axes, B2M decreased with treatment in both groups, suggesting a shared impairment in a key antigen presentation component. However, the more specific MHC class II antigen presentation module showed stronger reduction in the T790M-negative group (p < 0.001), with no clear change in the T790M-positive group (Fig. 5D). Checkpoint proteins like CTLA4 and VISTA decreased significantly in T790M-positive cases (p = 0.00034 and p = 0.01, respectively) (Fig. S5E), whereas they remained largely stable in T790M-negative cases (Fig. S5F).

We next compared the two T790M strata at the group level, independent of treatment time. Differential features showed consistent between-group shifts in representative immune and stromal molecules (Fig. 6A). The T790M-negative group had higher cytotoxicity scores (p = 0.0014), higher estimated fibroblast (p = 0.012) and neutrophil (p = 0.027) abundance (Fig. 6B). In contrast, antigen presentation modules were higher overall in the T790M-positive group (Fig. 6C). Compartment-resolved comparisons of checkpoint expression indicated that most checkpoints were higher in the T790M-negative group (Fig. 6D). At the program level, metabolic modules particularly mitochondrial metabolism/TCA and fatty acid oxidation, differed between groups (Fig. 6E), whereas migration/stromal and growth factor–related modules showed weaker separation; among canonical bypass pathways, only Wnt signaling differed significantly (p = 0.042) (Fig. 6F, G). Together, T790M stratification revealed broad, coordinated differences in immune–stromal state rather than isolated marker changes, suggesting that early remodeling patterns may be associated with resistance context, although this finding requires validation.

Fig. 6.

T790M-associated immune–stromal features at the comparable timepoint complement early on-treatment remodeling. (A) Volcano map of differentially expressed proteins for EGFR T790M positive and negative NSCLC patients. (B) Heatmap of estimated abundance of infiltrating lymphocytes for EGFR T790M positive and negative NSCLC patients post-first-generation EGFR-TKIs treatment. (C) Box plot of antigen presentation for EGFR T790M positive and negative NSCLC patients. (D) Expression of major immune checkpoints in all AOIs, stromal AOIs and tumor AOIs. (E) Comparison of metabolic and biosynthetic program signature scores between EGFR T790M–negative and positive groups. (F) Comparison of differentiation, EMT/motility, and growth factor related signaling signature scores between EGFR T790M–negative positive groups. (G) Comparison of canonical oncogenic/adaptive signaling pathway signature scores between EGFR T790M–negative and positive groups. *Wilcoxon rank-sum test, p < 0.05

Discussion

In EGFR-mutated NSCLC patients, resistance to EGFR-TKI is not a one-time point event but rather an adaptive process under the treatment pressure [20]. Identifying when this process begins to shift organizationally and distinguishing treatment-induced changes from patient-specific variation in a real disease course remains a significant clinical challenge. Most previous evidence of EGFR-TKI resistance comes from cross-sectional comparisons of different patients or re-biopsy at a single progression time point [21, 22]. These studies have revealed various resistance mechanisms but are inevitably affected by inter-patient differences, sampling site, and intratumoral heterogeneity, making it difficult to capture the longitudinal dynamic changes of patients from baseline to the final resistance. Spatial omics studies further suggest that the TME has significant regional characteristics, with many key changes confined to specific compartments rather than being globally consistent [23, 24]. Ignoring compartment discrimination may cause bulk sequencing to dilute or mask these important local signals. To address this gap, we integrated spatial transcriptomics and proteomics to construct a longitudinal cohort spanning the entire clinical process from baseline to T790M-mediated resistance after first-generation EGFR-TKI and subsequent re-resistance after third-generation EGFR-TKI. Within a unified spatio-temporal framework, we explored the dynamic trajectory of the TME under therapeutic pressure. Our study suggests that early adaptive reprogramming may lay the foundation for the subsequent expansion of spatial heterogeneity, with co-evolution between tumor genotype evolution and microenvironment phenotypic remodeling jointly driing end-point drug resistance.

In our study, tumor and stromal regions remained clearly separated at both protein and RNA levels, suggesting that compartmental differences are stable and that longitudinal changes should be interprated within a compartmentalization framework. In the dynamic observation of evolution under EGFR-TKI treatment, bulk sequencing may confound compartmental differences with temporal changes, potentially diluting or masking the true signals [25]. Many previous spatial studies have shown that the TME is not a homogeneous mixture but contains diverse cellular compositions, signaling networks, and immune-related pathways [13, 23, 24, 26]. For example, in a spatial single-cell map of lung adenocarcinoma, different immune cell activation states exhibit reproducible spatial organization associated with clinical outcomes, indicating that spatial information itself is part of the biological characteristics [27]. Similarly, a study on cancer-associated fibroblasts (CAFs) demonstrated that such effects often become non-significant after overall averaging [28]. Therefore, spatial information is essential for dynamia tracking and is a necessary prerequisite for our subsequent temporal comparisons.

Next, we observed that spatial heterogeneity increased significantly after third-generation EGFR-TKI treatment, and this increase was primarily associated with stromal region. A more plausible explanation is that under early treatment pressure, the overall TME may remain relatively consistent in its adaptive direction. However, under prolonged and more complex drug selection pressures, the stromal region is more prone to form diverse niches, driving increased spatial complexity. This finding is consistent with recent evidence showing that in EGFR-mutated NSCLC, the intrinsic tumor cells differences, along with transforming growth factor beta (TGFβ) - related secretory programs, CAF recruitment, and immunosuppression, collectively contribute to a drug-tolerant TME [22]. Previous clinical trials also suggest that resistance mechanism to Osimertinib are highly diverse, with many cases unexplained by single classical bypass pathway, providing clinical support for our conclusion [21]. Therefore, we view the stroma-associated heterogeneous expansion in the late stage of EGFR-TKI treatment as a framework conclusion that required further testing. The key feature of the terminal progression may not be general pathway activation in the tumor cells but rather increased spatial differentiation of the microenvironment. This stroma-associated heterogeneity has potential therapeutic implications. When resistance to third-generation TKIs develops, treatment strategies may need to shift from directly targeting tumor cells to modulating the stromal microenvironment. For example, targeting CAFs or stromal immune checkpoints could represent a complementary approach to overcome late-stage resistance.

Previous re-biopsy studies after EGFR-TKI resistance have shown that targeted therapy induces changes in PD-L1, tumor-infiltrating lymphocyte (TIL) density, and other immune-related markers, indicating that EGFR-TKI not only affect tumor cells intrinsic properties but also reshape the TME [29, 30]. Existing evidence indicates that patients with EGFR mutations generally have a weak response to immune checkpoint inhibitors (ICIs), with the immunosuppressive microenvironment after EGFR-TKI resistance considered a key contributing factor [31]. Single-cell data further show that post-resistance samples exhibit an immunosuppressive microenvironment dominated by tumor-associated macrophages [32]. A meta-analysis of the EGFR-mutated subgroup also suggested that the overall survival of ICIs is more limited in the mutant population than in wild-type ones, with low PD-L1 expression and CD8+TIL infiltration partly explaining this phenomenon [33]. Closer to spatial partitioning, a recent DSP study by Kim et al. used spatial analysis of pre- and post-treatment samples to demonstrate that EGFR-mutated NSCLC has an immunosuppressive TME, with changes in immune cells proportions and related pathways at the spatial level after treatment [12]. In addition to stromal remodeling, we observed stage-dependent immune evolution. Our study provides further evidence that immune-related remodeling is a phased transition throughout treatment. The axis of immunosuppression and exhaustion became apparent after third generation EGFR-TKI resistance, with changes concentrated in the stroma, offering more specific spatiotemporal dynamic evidence. From a translational perspective, this stage-dependent immune evolution suggests that the timing of immunotherapy may be critical. The late upregulation of stromal checkpoints raises the possibility that immune checkpoint blockade could be more effective when administered after resistance to first-generation TKIs but before progression on third-generation TKIs. This window may represent a time when the TME is more immunologically responsive.

Different resistance pathways lead to distince immune ecologies, with T790M stratification being the most clinically significant explanation. Haratani et al. compared T790M status in NSCLC patients receiving Nivolumab after EGFR-TKI progression and found that T790M-negative patients were more likely to benefit, suggesting a link to immune-related factors such as PD-L1 [34]. This provides circumstantial evidence supporting our observation of divergent microenvironment remodeling in the early treatment window. However, most previous studies were retrospective, with non-uniform sampling of individuals and biopsy locations. Early stage differentiation trends may accompany subsequent molecular outcomes of drug resistance. This association between early immune remodeling and subsequent T790M status, while exploratory, raises the possibility of early predictive biomarkers. However, given the very small sample sizes and the lack of independent validation, these findings should be considered hypothesis-generating rather than definitive. If validated in larger cohorts, dynamic changes in myeloid signatures, T cell activation, and antigen presentation features during the first few weeks of first-generation TKI treatment could help identify patients at high risk for T790M-mediated resistance. Such patients might benefit from earlier transition to alternative strategies, although this hypothesis requires prospective validation.

Our study has several limitations. First, complete three-stage paired re-biopsy chains are exceptionally rare, so our cohort is small. This limits statistical power, especially for subgroup comparisons. Therefore, our findings should be interpreted as a description of longitudinal trends within this limited cohort rather than population-level estimates. Second, despite our efforts to obtain orthogonal validation, we were unable to perform immunohistochemistry or multiplex immunofluorescence due to insufficient residual tissue after DSP profiling, and no suitable publicly available longitudinal dataset was identified for external validation. The extreme difficulty of collecting comparable longitudinal biopsy chains also precluded independent cohort analysis Third, we acknowledge technical limitations of AOI-based spatial segmentation. Although ROI selection was guided by histopathological assessment and manual delineation, and AOI segmentation was performed using the DSP platform’s integrated algorithm, this approach cannot achieve completely pure cell populations. Therefore, our conclusion that late-stage heterogeneity is primarily associated with the stromal compartment should be interpreted with this consideration in mind. Finally, the study is retrospective, observational, and exploratory. All findings are correlative, and we acknowledge the absence of mechanistic validation. The conclusions should be considered hypothesis-generating rather than definitive. Independent replication in future prospective cohorts will be required to confirm their wider applicability.

Conclusions

This study offers an initial, exploratory framework for understanding the dynamic evolution of EGFR-mutated NSCLC patients during targeted therapy, based on rare consecutive biopsies from the same patient combined with spatial multi-omics technology. In our limited cohort, tumor and the stroma regions remained distinctly separated at both protein and RNA levels, highlighting the potential value of spatial information in such longitudinal analyses. The observed increase in spatial heterogeneity was more pronounced in the late stage and appeared to be primarily associated with the stromal compartment, raising the possibility that late-stage evolution may reflect the reorganization and diversification of TME rather than synchronous amplification in cancer cells. In addition, our data suggest a trend towards non-linear and stage-dependent immune remodeling. Early T790M status appeared to be associated with potentially different spatial ecological trajectories, though this observation requires validation (Fig. 7). Although this study was observational and based on a small sample size, the longitudinal spatial evidence provides preliminary clues and generates testable hypotheses for understanding the evolution under EGFR-TKI pressure, serving as a data resource and conceptual reference for future validation in independent cohorts.

Fig. 7.

Graphical summary of compartment-resolved, longitudinal TME remodeling across EGFR-TKI treatment

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

XS was responsible for designing the experiments, analyzing the data, and writing the paper, under the overall coordination of HL and CZ. PH and XZ assisted with the analysis of DSP data. The interpretation of tissue sections was independently guided by XS, ZL and pathologist ZZ. XS and AX assisted with the collection of clinical samples. All authors reviewed the paper. All authors read and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (82403395), the Natural Science Foundation of Shandong Province (ZR2024QH219), China Postdoctoral Science Foundation (Special Fund, 18th Batch, 2025T180618) and Shandong Provincial Government Publicly Funded Overseas Study Project.

Data availability

The processed GeoMx DSP CTA and protein outputs supporting the findingsof this study have been deposited in Zenodo (DOI: 10.5281/zenodo.18603267).

Declarations

Ethics approval and consent to participate

This study was approved by the Institutional Review Board of Shanghai Pulmonary Hospital (Approval No.L24-404–1). The study was conducted inaccordance with the Declaration of Helsinki (as revised in 2013). Informed consent was obtained from all participants.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.