Patient-specific lung cancer tumoroids recapitulate the tumor microenvironments for functional evaluation of therapeutic responses and immune-stromal interactions

Abstract

Background

Tumor heterogeneity and the complexity of the tumor microenvironment (TME) drive therapeutic resistance in lung cancer, highlighting the critical need for experimental models that faithfully recapitulate tumor–stroma–immune interactions.

Methods

We developed patient-specific lung cancer tumoroid (LCT) models by integrating lung cancer organoids (LCOs), cancer-associated fibroblasts (CAFs) and immune cells isolated from single tumor tissue sources using a modular, transwell-based culture platform. The model supports cryopreservation and reproducible reconstruction of cellular components. Multimodal assessments, including histological characterization, dose-response pharmacologic profiling, immune profiling, and transcriptomic analyses, were performed across diverse co-culture configurations to evaluate patient-specific TME features and treatment-associated responses.

Results

The LCT models successfully recapitulated key structural and cellular features of native TMEs, demonstrating high reproducibility across patient samples. Functional analyses revealed tumor-intrinsic heterogeneity in responses to chemotherapy and chemo-immunotherapy. CAF integration altered therapeutic responses and was associated with reduced immune activation and cytotoxic efficacy in selected models, suggesting stromal contributions to treatment resistance. Transcriptomic analyses showed that reconstructed TMEs preserved patient-specific stromal and immune programs associated with therapeutic responsiveness, including transcriptional features linked to clinical outcomes in independent immunotherapy-treated cohorts.

Conclusions

This patient-specific LCT model provides a scalable and translationally relevant approach for the ex vivo reconstruction of native TMEs, facilitating the functional interrogation of tumor–stroma–immune interactions. By capturing heterogeneous stromal responses, this platform offers a valuable tool for investigating therapeutic resistance and supports the further development of precision immune-oncology and patient-tailored therapeutic strategies in lung cancer.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12967-026-08487-2.

Background

Lung cancer remains the leading cause of cancer-related mortality worldwide, largely due to tumor heterogeneity and the complexity of the tumor microenvironment (TME), which contribute to treatment resistance and limited therapeutic efficacy [1, 2]. The TME is a dynamic ecosystem composed of malignant epithelial cells, cancer-associated fibroblasts (CAFs), and immune cells represented by tumor-infiltrating lymphocytes (TILs), whose interactions critically influence tumor progression and immune responses [3, 4]. Notably, substantial interpatient variability in immune infiltration and stromal composition exists even within the same histologic subtype [5–7], underscoring the need for personalized models that accurately recapitulate these features.

Recent studies have focused on integrating tumor, stromal, and immune components into patient-derived cancer models. Organoid-based models replicating this multicellular complexity have been developed for several solid tumors including pancreatic [8], colorectal [9], and, more recently, lung cancers [10]. However, conventional TME- integrated approach typically require separate isolation processes for organoid culture, TIL acquisition and CAF isolation before reassembly, necessitating multiple tissue inputs and substantial experimental workloads. These sequential procedures may fail to preserve native tumor-stroma-immune interactions and limit scalability.

To address these limitations, we developed a patient-specific lung cancer tumoroid (LCT) model coupled with a modular, one-stop transwell-based isolation culture system that enables rapid and reproducible isolation and expansion followed by spatial reassembly of tumor epithelial cells, CAFs, and immune cells from minimal single input of patient-derived NSCLC tissue samples. Each cell population can be cryopreserved and subsequently reassembled into patient-specific LCTs, allowing reconstruction of key features of the TME.

Using this platform, we performed functional assessments of therapeutic responses across multiple co-culture conditions, incorporating multimodal analyses, including therapeutic response evaluation, immune-associated readouts, metabolic characterization, and transcriptomic analyses. These analyses revealed heterogeneous and context-dependent response patterns across patient-derived models.

Overall, this approach provides a scalable and clinically relevant platform for ex vivo modeling of patient-specific TMEs and supports functional evaluation of therapeutic responses.

Methods

Acquisition and preparation of lung tumor tissues

Pathologically confirmed lung cancer tissues were collected from surgically resected specimens (typically ~ 1 cm³ in total volume) through the lung cancer biobank at the Asan Bio-Resource Center (Seoul, Korea) with informed patient consent. This study was approved by the Institutional Review Board of Asan Medical Center (IRB number: 2019–1210) and conducted in accordance with institutional guidelines.

Resected tissues were transported on ice to the laboratory within 1 hour of excision in cold Hank’s Balanced Salt Solution (HBSS; #LB003-04, Welgene) supplemented with antibiotics (Lonza, Basel, Switzerland). For experimental procedures, only a portion of the resected tissue was used for Transwell-based isolation and downstream applications. Upon arrival, samples were washed three times with cold Dulbecco’s Phosphate-Buffered Saline (DPBS; #LB001-02, Welgene) and mechanically sectioned into small fragments using sterile blades. The tumor tissue fragments were incubated at 37 °C in a shaker incubator for 15 to 30 minutes with 0.2 U/µL Collagenase-II (#17010–015, Gibco), undergoing two cycles of tissue dissociation.

Flow cytometry and cell population analysis

Lung tumor-dissociated cell populations were analyzed using flow cytometry to characterize the major cellular compartments within the TME. Single-cell suspensions were incubated with an Fc receptor blocking reagent (#130–059-901, Miltenyi Biotec) to prevent non-specific antibody binding. The cells were then stained with fluorescently labeled antibodies: PE-conjugated EpCAM (#130–091-253, Miltenyi Biotec), APC/Cyanine7-conjugated CD45 (#130–091-253, Miltenyi Biotec), FITC-conjugated CD90 (#561969, BD Biosciences), Brilliant Violet 421-conjugated CD3 (#317344, BioLegend), PerCP/Cyanine5.5-conjugated CD8 (#344710, BioLegend), PE/Cyanine7-conjugated CD14 (#301814, BioLegend), and APC-conjugated CD19 (#302211, BioLegend).

For TIL characterization, we utilized FITC-conjugated CD45RA (#304106, BioLegend) and PE-conjugated CCR7 (#353204, BioLegend), along with the appropriate isotype controls (FITC Mouse IgG2b κ, #400310; PE Mouse IgG2a κ, #400214; BioLegend). Dead cells were excluded using a fixable viability dye (#65–0866-14, Invitrogen). Data were acquired using a BD FACSCanto™ II flow cytometer (BD Biosciences) and analyzed with FlowJo software (BD).

Culture of LCOs from the upper chamber

To establish lung cancer organoids (LCOs), upper chamber components forming organoid-like structures were harvested within 5 days of Transwell separation and re-seeded into Matrigel dome using minimum basal medium (MBM), following a previously described protocol [11]. Briefly, the collected structures were enzymatically dissociated into single cells using TrypLE^TM Express (#11605–010, Gibco) and incubated for 30 minutes at 37 °C in a 5% CO₂ atmosphere. The resulting cell suspension was seeded at a density of approximately 5 × 10⁵ to 1 × 10⁶ viable cells per well and mixed with Matrigel (#356237, Corning) and MBM at a 2:1 ratio. The mixture was plated into pre-warmed culture plates (#30006 and #30024, SPL), and incubated for 30 minutes to allow matrix solidification. LCOs were maintained for approximately three weeks, with MBM refreshed twice weekly, until they reached a diameter of 100–150 μm.

Once mature, LCOs were dissociated using the same protocol and re-seeded at a 1:4 split ratio in fresh Matrigel—MBM mixture. For biobanking, LCOs that consistently reached approximately 150 μm in diameter after at least three passages were cryopreserved in liquid nitrogen, using cells pooled from three wells of a 24-well plate per vial. The freezing medium consisted of a 9:1 mixture of ES grade FBS (#26140079, Gibco) and DMSO (#D1370.0250, Duchefa).

Isolation and culture of TME cells

CAFs were isolated within 14 days using the One-Stop Simultaneous In Vitro Culture System. After transferring LCOs from the upper chamber of the Transwell to culture plates in droplet form, CAFs adhering to the Transwell membranes were harvested by treatment with 0.25% Trypsin-EDTA (#25200–072, Gibco) for 5 minutes, followed by washing and culture in complete DMEM/F12 medium.

TILs were collected within 3 days using the same culture system, concurrently with the transfer of LCOs to the culture plate. For experiments involving TILs, the cells were activated in 6-well culture plates using AIM V™ Medium (research grade) supplemented with AlbuMAX™ (#31035–025, Gibco), 10% human serum (#H3667, Sigma-Aldrich), 50 μM 2-mercaptoethanol (# M3148, Sigma-Aldrich), and 3000 IU/mL recombinant human IL-2 (#200–02, PeproTech), collectively referred to as “TIL culture medium.” Activation was conducted over a period of up to 14 days, with the addition of 1 μg/mL anti-CD3 (#14–0037-82, Invitrogen) and 5 μg/mL anti-CD28 (#16–0289-95, Invitrogen). For biobanking, both CAFs and TILs were cryopreserved using the same freezing medium as LCOs and stored in liquid nitrogen.

TME reconstitution

For histological analysis (Fig. 4), high-density tumoroid constructs were generated by combining matched LCOs, CAFs, and immune cells at a 1:1:1 ratio, with a total of approximately 1 × 10⁶ cells embedded in Matrigel.

Fig. 4.

Spatial reconstitution of patient-derived lung tumor microenvironments. (A) Tumoroids comprising matched tumor cells, cancer-associated fibroblasts (CAFs), and immune cells (ICs) were co-embedded in Matrigel and cultured for 4 days. Scale bar, 200 μm. (B) Tumoroid LCT788, generated from original lung cancer tissue LC788, H&E staining and immunofluorescence staining. Scale bar, 100 μm. (C) Brightfield images of tumoroids within the microfluidic system. Scale bar, 200 μm. (D) Immunofluorescence image at 2 days culture of tumoroids. Scale bar, 100 μm

For microphysiological system (MPS) models, cells were combined at a 1:5:5 ratio, with a total of approximately 1 × 10⁴ cells loaded into each microfluidic device to account for differential growth rates. Culture medium was continuously perfused through the chip at a flow rate of ~1.5 μL/min to support nutrient exchange and dynamic cell-cell interactions.

Tumoroid constructs were cultured for 4 days and subsequently fixed in 10% neutral-buffered formalin (#HT501128, Sigma-Aldrich) for 2 hours at RT. Constructs were gently harvested using a cell scraper (#90020, SPL), embedded in a Histogel base (#HG-4000–012, Epredia), and processed for paraffin embedding. Paraffin sections were used for hematoxylin and eosin (H&E) staining and immunofluorescence analysis. MPS cultures were fixed in situ after 2 days using 10% neutral-buffered formalin, as described above, and processed for immunofluorescence staining.

Immunofluorescence staining

All samples—including cultured CAFs on slides, paraffin-embedded lung cancer tumoroid sections (LCTs), and microchip-based 3D cultures—were fixed in 4% paraformaldehyde (PFA) for 15 min at room temperature (RT), permeabilized with 0.5% Triton X-100 in PBS for 5 min, and blocked with 3% bovine serum albumin (BSA) in PBS for 1 h. Samples were then incubated overnight at 4 °C with the primary antibodies.

For CAFs, anti-vimentin (1:1000; #5741, Cell Signaling Technology), anti-fibroblast activation protein (FAP; 1:50; #66562, Cell Signaling Technology), and anti-paired-related homeobox 1 (Prrx1; 1:50, #PA5-111567, Invitrogen; or 1:100, #H00005396–B01P, Abnova) antibodies. For LCTs, we used anti-EpCAM (1:1000; #ab71916, Abcam), anti-CD45 (1:1000; #M070101-2, Agilent Dako), and anti-vimentin (1:1000; #MAB2105, R&D Systems) antibodies.

After three washes with PBS containing 0.01% Tween-20 (PBST; 10 min each), samples were incubated with Alexa Fluor-conjugated secondary antibodies (1:1000; Invitrogen and Abcam) for 1 h at RT in the dark. CAFs were stained with Alexa Fluor 488- and 594-conjugated antibodies (#A-21202 [anti-mouse] and #A-11012 [anti-rabbit], Invitrogen), while LCTs and microchip-based samples were stained with Alexa Fluor 488-, 594-, and 647-conjugated antibodies (#A-21202 [anti-mouse] and #A-11012 [anti-rabbit], Invitrogen; #ab150155 [anti-rat], Abcam). Nuclei were counterstained with DAPI (1:2000; #D9564, Sigma-Aldrich) for 15 min. Fluorescence images were acquired using a Carl Zeiss fluorescence microscope (Carl Zeiss, Germany).

Histology and H&E staining

Tissue specimens and LCOs were fixed in 4% PFA, whereas LCTs were fixed in 10% neutral-buffered formalin. All samples were subsequently dehydrated through graded ethanol, embedded in paraffin, and sectioned. Hematoxylin and eosin (H&E) staining was performed using Harris hematoxylin according to standard protocols. Bright-field images were acquired using an Olympus BX50 microscope.

3D spheroid TME model and drug treatment

Cancer cells, CAFs, and peripheral blood mononuclear cells (PBMCs) derived from patient-matched non-small cell lung cancer (NSCLC) samples were co-seeded into ultra-low attachment 96-well plates (#7007, Corning) at a ratio of 2:1:6 (Cancer cells:CAFs:PBMCs), with a total of approximately 2,500 viable cells per well to facilitate tumoroidsformation.

On day 1, tumoroids were treated with pemetrexed (Pem; 100 μM; #S1135, Selleckchem), carboplatin (Carbo; 100 μM; #S1215, Selleckchem), or a combination of both agents (Pem+Carbo; 50 μM each), either alone or in combination with an anti-PD-1 antibody (α–PD-1; 100 μg/mL; Keytruda, MSD). Anti-PD-1 monotherapy (100 μg/mL) was also included. Drug concentrations were selected based on preliminary dose–response experiments to ensure comparable exposure levels across conditions. All treatments were performed in three independent wells per condition and maintained for 6 days.

To compare treatment responses across co-culture conditions with different cellular compositions, the treatment-induced reduction in viable tumor organoid area relative to the corresponding untreated control was normalized to the total number of cells in each well and reported as Δ tumor area per total cell. The metric was calculated as: Δ Tumor Area per Total Cell = (Control Tumor Area − Treated Tumor Area) / Total Cell Number.

Cell viability assay

Cell viability was assessed on day 6 following treatment using the LIVE/DEAD™ Viability/Cytotoxicity Kit (#L3224, Invitrogen) according to the manufacturer’s instructions, with minor modifications. Half of the culture medium was aspirated and replaced with a 2× staining solution containing calcein-AM (final dilution, 1:3000) and ethidium homodimer-1 (final dilution, 1:5000), followed by incubation for 30 min at 37 °C. Fluorescence images were acquired using the Operetta HighContent Screening system (PerkinElmer) and analyzed using Harmony software (version 3.5.1, PerkinElmer).

Drug sensitivity assay

Tumoroids were treated with seven concentrations of pemetrexed, carboplatin, or DMSO as a control for 6 days. Following treatment, tumoroids were transferred to white 96-well flat-bottom plates optimized for luminescence detection (#3610, Corning), and 100 μL of CellTiter-Glo reagent (#G7571, Promega) was added to each well. Plates were agitated for 30 min at RT, and luminescence was measured using an EnVision Xcite Multimode Plate Reader (Revvity). IC₅₀ values were calculated by nonlinear regression using GraphPad Prism (GraphPad Software, La Jolla, CA, version 10.5.0).

Enzyme-linked immunosorbent assay (ELISA)

Culture supernatants were collected on day 6 from LCO+IC and LCO+CAF+IC (LCT) tumoroids treated with anti-PD-1 antibody (α–PD-1) alone or in combination with pemetrexed (Pem+α–PD-1) or carboplatin (Carbo+α–PD-1). IFN-γ levels were quantified using the DuoSet Human IFN-γ ELISA kit (#DY285B, R&D Systems) according to the manufacturer’s instructions. Briefly, 96-well plates were coated with capture antibody overnight, followed by blocking and incubation with either culture supernatants or recombinant standards. Detection antibody, streptavidin-HRP, and substrate solution were added sequentially. Absorbance was measured at 450 nm with wavelength correction at 540 nm using a SUNRISE microplate absorbance reader (TECAN), and cytokine concentrations were calculated based on the standard curve.

Metabolic flux analysis (Seahorse XF assay)

Cellular metabolic activity was assessed using the Seahorse XF Cell Mito Stress Test and Glycolysis Stress Test (Agilent Technologies), following the manufacturer’s instructions with modifications optimized for 3D lung cancer tumoroid cultures. Patient-derived LCOs were co-cultured with matched PBMCs, with or without autologous CAFs, in Seahorse XF96 spheroid microplates (Agilent). Tumoroids were treated with α–PD-1 (100 µg/mL), pemetrexed (50 µM), or carboplatin (50 µM), either alone or in combination, for 24 hours prior to the assay.

On the day of measurement, cells were washed and incubated in Seahorse XF assay medium at 37 °C in a non-CO₂ incubator for 1 hour. The oxygen consumption rate (OCR) and extracellular acidification rate (ECAR) were measured following sequential injection of oligomycin (1.5 µM), FCCP (2 µM), and rotenone/antimycin A (0.5 µM each).

Basal respiration was calculated by subtracting non-mitochondrial respiration (post-Rot/AA) from basal OCR (pre-oligomycin). ECAR values were obtained from the steady-state phase (25–75 min). Data were normalized to well-specific background levels and processed using Wave software (Agilent). Each condition was analyzed in three independent biological replicates.

RNA sequencing (RNA-seq) and analysis

Bulk RNA sequencing was performed on LCO, LCO+IC, and LCT samples collected under chemotherapy or immunotherapy treatment conditions. For each treatment condition, three independently prepared culture replicates (six wells per replicate) were pooled prior to RNA extraction, resulting in a single RNA-seq library per treatment condition.

Total RNA was extracted using standard protocols. RNA concentration was quantified using the Quant-iT RiboGreen assay (Invitrogen, #R11490), and RNA integrity was assessed using the Agilent TapeStation RNA ScreenTape system (#5067–5576). Only samples with RIN ≥ 7.0 were included for sequencing. For each RNA-seq library, 0.5 µg of total RNA was used for construction with the Illumina TruSeq Stranded Total RNA Library Prep Gold kit (#20020599), which includes rRNA depletion, fragmentation, cDNA synthesis, end repair, adapter ligation, and PCR enrichment. Libraries were quantified using a KAPA Library Quantification Kit and sequenced as paired-end 100 bp reads on the Illumina NovaSeq X platform. All samples were processed in parallel using the same workflow to minimize technical variability.

Adapter trimming and quality filtering were performed with Trimmomatic v0.38. The sequenced reads were aligned to the GRCh38 reference genome using HISAT2 v2.1.0 and processed with SAMtools v1.9. Transcript assembly and quantification were conducted using StringTie v2.1.3b, resulting in raw counts, fragments per kilobase of transcript per million mapped reads (FPKM), and transcripts per million (TPM). Normalization and differential expression analysis were conducted using edgeR (v3.40.2) with trimmed mean of M values (TMM) normalization and quasi-likelihood F-tests (glmQLFTest); a fixed biological coefficient of variation (BCV) was used for comparisons without replicates. Differentially expressed genes (DEGs) were defined as an FDR of less than 0.05 and logCPM greater than 0. The absolute logFC threshold was adjusted to obtain a manageable number of DEGs for downstream analysis. Log₂-CPM values were used for visualization and signature scoring.

Tumor purity, stromal scores, and immune signature scores were estimated using the ESTIMATE algorithm via immunedeconv (v2.1.0) [12]. Immune cell fractions were inferred using CIBERSORT (LM22, 100 permutations) with log₂-transformed TPM values [13, 14]. Samples with p values greater than 0.05 were excluded. Drug sensitivity was predicted using oncoPredict (v1.2) [15] trained on GDSC v2 data [16], following batch correction and variance filtering.

Responder- and CAF-specific DEGs were identified through pairwise comparisons, excluding overlapping genes to define responder-specific CAF signatures. Highly expressed genes were Z-score normalized and hierarchically clustered using the pheatmap package (v1.0.12). Pathway enrichment analysis was performed with REACTOME gene sets (MSigDB C2, v7.5.1) via the clusterProfiler package (v4.6.2) [17]. CAF subtypes were scored using singscore (v1.18.0) [18] using published markers. T cell exhaustion and functional pathways were assessed using the singscore (v1.18.0) with MSigDB gene sets. Independent Student’s t-tests were applied to all signature comparisons, and biologically relevant results were retained for interpretation.

To assess the clinical relevance of the LC990-derived CAF-specific transcriptional program, four publicly available cohorts of patients with lung adenocarcinoma treated with immune checkpoint inhibitors (GSE126044, GSE135222, GSE161537, and GSE190265) were analyzed. Signature scores for the top 20 LC990 CAF-specific genes were calculated using the singscore (v1.18.0) package and evaluated for associations with treatment response and survival outcomes.

Statistical analysis

Statistical analyses were conducted using GraphPad Prism (v10.5.0) and R (v4.2.2). Data are presented as mean ± standard error of the mean (S.E.M). Group comparisons were performed using appropriate parametric or nonparametric tests based on data distribution, including two-sided Student’s t-tests, Mann–Whitney U tests (Wilcoxon rank-sum tests), Wilcoxon matched-pairs signed-rank tests, Friedman tests for repeated measures, and Kruskal–Wallis tests followed by Dunn’s multiple comparisons. Statistical significance was defined as follows: ns, not significant (p > 0.05), *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.

Results

TME heterogeneity in non-small cell lung cancer tissue

To establish a patient-specific lung cancer model that incorporates the TME, we analyzed cellular compositions of lung tumor tissues from patients histologically (Supplementary Fig. S1A), and by using multiparametric flow cytometry (Supplementary Fig. S1B). There was significant interpatient heterogeneity in the TME across both squamous cell carcinoma (SqCC) and adenocarcinoma (ADC) samples. While some SqCC tumors exhibited immune-enriched profiles, others were predominantly composed of epithelial or stromal components. Similarly, ADC samples also displayed a broad spectrum of immune and stromal infiltration (Fig. 1A).

Fig. 1.

Heterogeneity of tumor microenvironment across lung cancer subtypes. (A) Stacked bar plot showing the normalized proportions of major tumor microenvironment (TME) cell populations in patient-derived lung tumor samples (n = 10), grouped by histological subtype [squamous cell carcinoma (SqCC, n = 4) vs adenocarcinoma (ADC, n = 6)]. Flow cytometry identified EpCAM⁺ tumor cells, CAF-enriched stromal cells, and CD45⁺ immune subsets (CD8⁺, CD8⁻ T cells, CD14⁺ monocytes, CD19⁺ B cells, and others). (B) Boxplots comparing the proportions of each cell type between SqCC (teal) and ADC (salmon). Wilcoxon rank-sum test (*p < 0.05). (C) Volcano plot showing log₂ Fold change (SqCC/ADC) versus –log₁₀(p-value) for major TME cell populations. (D) Hierarchical clustering heatmap of z-score normalized cell type proportions across patient-derived lung tumor samples

Comparing the proportions of individual cell populations between SqCC and ADC samples, CD19⁺ B cells and CD45⁺ immune cells were significantly enriched in SqCC (p = 0.038 for both), consistent with previous observations in lung and pan-cancer immune profiling [19–21]. In contrast, no significant differences were observed for CAF⁺ stromal cells, T cell subsets, CD14⁺ monocytes, or EpCAM⁺ tumor cells [22] (Fig. 1B). These findings were further supported by volcano plot analysis, which showed selective enrichment of CD19⁺ B cells and CD45⁺ immune cells in SqCC (log₂FC > 1 and p < 0.05), aligning with prior reports on immune microenvironmental divergence [23, 24] (Fig. 1C). Despite subtype-associated trends in specific immune subsets, the overall TME composition remained variable within each histological group.

To further evaluate inter-patient heterogeneity in the lung TME, we performed hierarchical clustering of z-score cell-type proportions across all tumor samples. This analysis revealed that the samples did not cluster strictly by histological subtype, with several SqCC and ADC cases exhibiting overlapping immune and stromal profiles (Fig. 1D). Notably, a subset of ADC samples clustered with immune-enriched SqCCs, while others grouped with stromal-dominant tumors, regardless of histology. These findings demonstrate the diversity of the lung TME across histological subtypes and the need for personalized models to represent this heterogeneity.

Transwell-based compartmentalized isolation of patient-specific TMEs

To accurately model the individual TME, we developed a Transwell-based, one-stop isolation culture system that allows spatial separation and simultaneous culture of patient-derived LCOs, CAFs, and TILs over a two-week culture period. The system is assembled within a standard 6-well plate using a Transwell insert coated with a temperature-sensitive ECM component, Matrigel. Upon incubation at 37 °C, the ECM transitions from a liquid to a gel state, providing structural support for organoid formation and serving as a selective barrier to facilitate the differential migration of various cell types (Fig. 2A) [25].

Fig. 2.

One-stop isolation culture of tumor components and intrinsic factors influencing tumor microenvironment recovery. (A) Transwell-based system enabling spatially segregated isolation of lung cancer organoids (LCOs; upper chamber), cancer-associated fibroblasts (CAFs), and tumor-infiltrating lymphocytes (TILs; lower chamber and Matrigel) over two weeks. Scale bar, 200 μm. (B) Histological subtype distribution among 55 lung cancer samples and venn diagram showing overlap in successful isolation of LCOs, CAFs, and TILs. (C) Isolation success rates of individual components and triple co-isolation by major histological subtype. (D) Difference of triple co-isolation across histologic subtypes. (E) Heatmap showing isolation success by histological differentiation. (F) Success rates of LCO, CAF, TIL, and triple isolation stratified by patient sex and age group

Following the enzymatic and mechanical dissociation of resected cancer tissues, the resulting heterogeneous single-cell suspension was seeded into the upper chamber. Cell segregation occurred based on inherent differences in cell size and motility: cancer cells remained in the upper compartment and formed organoid structures [26], while CAFs were retained within the ECM layer [27], and TILs migrated through the porous membrane into the lower chamber. The concentration and thickness of the Matrigel layer were optimized to maximize the selective retention and migration of each cell population. Media conditions were optimized to support each population: organoid medium was used in the upper chamber [11], TIL expansion medium in the lower chamber [28], and a CAF-promoting formulation was introduced after the harvest of TILs.

We applied this platform to 55 lung cancer tissue samples and assessed the recovery rates of LCOs, CAFs, and TILs. Although initial LCO formation was successful in all cases (100%), subsequent passaging and expansion were limited. Overall, LCOs, CAFs, and TILs were successfully recovered and expanded from 21 (38.2%), 25 (45.5%), and 49 (89.1%) of the samples, respectively (Fig. 2B). Triple co-isolation—defined as the concurrent establishment of all three components, was achieved in 10 (18.2%) of the cases (Table 1 and Supplementary Tables 1 and 3). Although the cellular composition appeared to differ, there was no significant difference in recovery rates based on tumor histology (Figs. 1 and 2C). Further stratification revealed substantial heterogeneity within subtypes, particularly concerning the recovery of all three components.

Table 1.

Success rates for isolation of tumor microenvironment (TME) components by histological subtype

Histological subtype	LCOs (n/total, %)	CAFs (n/total, %)	TILs (n/total, %)	Triple (n/total, %)
Adenocarcinoma (n = 40)	16/40 (40.0%)	19/40 (47.5%)	35/40 (87.5%)	8/40 (20.0%)
Squamous cell carcinoma (n = 13)	4/13 (30.8%)	6/13 (46.2%)	13/13 (100.0%)	2/13 (15.4%)
Other NSCLC variants (n = 2)	1/2 (50.0%)	0/2 (0.0%)	1/2 (50.0%)	0/2 (0.0%)
Total (n = 55)	21/55 (38.2%)	25/55 (45.5%)	49/55 (89.1%)	10/55 (18.2)

Aggregated success rates of lung cancer organoids (LCOs), cancer-associated fibroblasts (CAFs), tumor-infiltrating lymphocytes (TILs), and triple co-isolation from 55 patient-derived NSCLC samples, stratified by histological subtype. Values are shown as n/total (%)

LCOs, lung cancer organoids; CAFs, cancer-associated fibroblasts; TILs, tumor-infiltrating lymphocytes; NSCLC, non–small cell lung cancer; Triple, co-isolation of LCOs, CAFs, and TILs

Among adenocarcinomas, triple co-isolation was most frequently observed in acinar tumors (36%), but it was significantly less common in solid (0%), papillary (14%), and mucinous (20%) variants (Fig. 2D). These findings suggest that histological subtypes represented by architectural organization of cancer cells, CAFs and TILs in the tumor play a role in determining the feasibility of isolating complex TMEs, and future reconstruction ex vivo [29].

Stratification by tumor differentiation further underscored this concept. TILs were consistently recovered across all grades, with success rates of 100% in well- and moderately differentiated tumors, and 87.1% in poorly differentiated tumors. In contrast, the recovery of LCOs and CAFs varied significantly. Well-differentiated tumors exhibited low yields of LCOs (33.3%) and entirely failed to produce CAFs (0%). Moderately and poorly differentiated tumors yielded both LCOs (42.9% and 41.9%, respectively) and CAFs (50% and 48.4%, respectively) at higher frequencies. Triple co-isolations were established only in moderately (21.4%) and poorly differentiated tumors (22.6%), but not in well-differentiated tumors (0%) (Fig. 2E and Supplementary Table 2). These results suggest that while immune components remain accessible regardless of tumor grade, the isolation of epithelial and stromal components is significantly limited in well-differentiated tumors [30]. Thus, tumor differentiation may emerge as a key factor influencing the efficiency of ex vivo TME reconstruction [31].

To assess the impact of host-intrinsic variables, we stratified isolation outcomes by patient age and sex. TILs were consistently recovered across all groups, and neither sex nor age significantly affected immune cell isolation. In contrast, the success of CAF and triple co-isolation decreased with advancing age. Patients under 60 years of age demonstrated significantly higher recovery rates for CAF (77.8%) and triple components (57.1%) compared to those aged 60–69 (CAF: 25–33%, Triple: 0–8.3%) and those over 70 (CAF: 21–67%, Triple: 0–11.1%) (Fig. 2F). These findings indicate that age-related declines in stromal viability and immune cell function may hinder comprehensive TME reconstruction in older individuals, aligning with previous reports of fibroblast senescence and diminished T cell migration in aged tissues [32].

Taken together, our results reveal that the feasibility of isolating a complete patient-specific TME is influenced by subtype-specific architectural diversity, differentiation state, and host-intrinsic biology [33].

Ex vivo preservation of phenotypic and functional integrity in lung TME components

We evaluated whether each component of the isolated TME retained its phenotypic identity and function. Initially formed LCOs in upper chamber of the Transwell were subsequently expanded for minimum 2 weeks using droplet-based culture [11]. The LCOs preserved the histological architecture of the original tumors. LCOs derived from ADC samples (LC-975, LC-977, LC-979) exhibited retained glandular or papillary morphology, whereas those derived from SqCC samples (LC-978) retained squamoid features, effectively recapitulating subtype-specific characteristics in vitro (Fig. 3A).

Fig. 3.

Phenotypic features of the patient-derived lung tumor microenvironment components. (A) H&E staining of primary lung cancer tissues (top, original magnification x100) and matched lung cancer organoids cultured for 3 weeks (bottom). LC975, LC977 and LC979, adenocarcinomas; LC978, squamous cell carcinoma. Scale bar, 100 μm. (B) Cultured cancer-associated fibroblasts (CAFs) express vimentin (red) and Prrx1 (green); nuclei stained with DAPI (blue). Scale bar, 100 μm. (C) Flow cytometry of lower chamber-derived cells (n = 14) showing predominant CD45⁺ immune cell composition. (D) CD8⁻ /CD8⁺ ratios in tumor tissue and matched Transwell-derived tumor-infiltrating lymphocytes (TILs) (n = 8; Wilcoxon matched-pairs signed-rank test, p > 0.9999; ns). (E) Time-course of TIL expansion in the Transwell lower chamber (n = 4; Friedman test, p = 0.0174). (F) Upper panel, representative brightfield images of the LC914 lower-chamber culture at days 0, 3, 7 and 14 during ex vivo culture. Lower panels, changes in the proportions of T_EM CD8⁻ and CD8⁺ cells, and T_EMRA CD8⁻ and CD8⁺ cells within the CD3⁺ T cell population during culture period. Wilcoxon matched-pairs signed-rank test. ns, not significant (p > 0.05), *p < 0.05

CAFs isolated from the Matrigel layer expressed canonical mesenchymal and activation markers, including vimentin, Prrx1, and FAP. These cells exhibited an elongated fibroblastic morphology and DAPI-positive nuclei, indicating the preservation of stromal identity and an activated phenotype (Fig. 3B and Supplementary Fig. S2A).

Transwell lower chamber-derived cells were first characterized using 14 patient samples to confirm their identity as TILs. Flow cytometric analysis revealed that the majority of cells recovered from the lower chamber were CD45⁺ immune cells, which is consistent with the lymphocyte enrichment observed in this compartment (Fig. 3C).

To evaluate whether the composition of TILs mirrors that of the original tumor tissue, we compared the CD8⁻ /CD8⁺ ratio among CD45⁺ immune cells in tumor-dissociated samples and Transwell-derived cells after 3 days of culture. In matched samples (n = 8), no significant difference was observed, indicating that the Transwell system maintains the relative composition of T cell subsets during selective migration (Fig. 3D and Supplementary Fig. S2B).

TIL expansion in the Transwell lower chamber was monitored over a 14-day period in four patient-derived samples. Total lymphocyte counts progressively increased at days 3, 7, and 14 compared to baseline, with consistent trends observed across individuals. These results indicate that TILs isolated from primary tissue retain both viability and proliferative capacity throughout the ex vivo culture period (Fig. 3E).

To assess changes in T cell subset composition during expansion, we performed flow cytometric analysis on TILs derived from three patients on days 7 and 14. Representative brightfield images of the LC914 lower-chamber culture show a progressive increase in size of TIL aggregates over the 14-day culture period (Fig. 3F, upper panel), which correspond to the increase in cell number (Fig. 3E). Effector memory T cells (T_EM; CD45RA⁻ CCR7⁻) represent antigen-experienced T cells, whereas terminally differentiated effector memory T cells (T_EMRA; CD45RA⁺CCR7⁻) represent a more differentiated subset. Notably, CD8⁻ T cells showed a trend toward enrichment for the T_EM phenotype, whereas CD8⁺ T cells exhibited heterogeneous dynamics without consistent enrichment [34, 35]. In both CD8⁻ and CD8⁺ populations, T_EM cells constituted the predominant subset, whereas T_EMRA cells remained minor proportions (Fig. 3F). This trend coincided with a decline in naïve and central memory phenotypes, with near-complete depletion of CCR7⁺ CD8⁺ cells by day 14 [36] (Supplementary Fig. S2C). Although most comparisons between days 7 and 14 fell short of statistical significance, the observed trends suggest a substantial preservation of effector potential with limited progression toward terminal differentiation. While canonical exhaustion markers such as PD-1 and TOX were not assessed, the predominance of the T_EM subset is consistent with preservation of effector-associated characteristics during ex vivo expansion. Together, these findings demonstrate that the Transwell system preserves the phenotypic identity of tumor, stromal, and immune compartments while facilitating the ex vivo expansion of effector memory T cells that may be enriched for tumor-associated reactivity [37].

Reconstruction of the TME in tumoroids

To reconstruct patient-specific lung cancer TMEs ex vivo, we reassembled LCO, CAFs, and autologous immune cells into a three-dimensional lung cancer tumoroid (LCT). These components were isolated from patients’ lung cancer tissues using the Transwell-based platform and cryopreserved prior to use (Supplementary Fig. S3).

The ex vivo expansion and viability of TILs varied across patient samples, limiting their applicability in LCT model for subsequent experiments. To enable consistent and comparative analyses across experimental conditions, we utilized autologous peripheral blood mononuclear cells (PBMCs) as a standardized, patient-matched immune cell source [38, 39]. Live-cell imaging showed treatment-associated caspase activation and morphologic changes in PBMC-containing co-cultures (Supplementary Fig. S7).

Upon co-embedding tumor cells with autologous CAFs, and immune cells in equal ratios within a Matrigel droplet, tumor cells formed compact organoid structures at the center, while CAFs and PBMC-derived immune cells migrated toward the organoid. This spatial arrangement resembled the architecture of native tumor tissue (Fig. 4A). Immunofluorescence staining confirmed distinct organization: EpCAM⁺ tumor cells were located at the core, Vimentin⁺ CAFs at the periphery, and CD45⁺ immune cells infiltrated the stromal zones [40] (Fig. 4B).

To further capture dynamic TME remodeling, we incorporated matched tumor cells, CAFs, and immune cells into a microphysiological system (MPS) consisting of a microfluidic chip containing a perfusable extracellular matrix [41]. The three cell types were co-cultured under continuous medium flow to support multicellular interactions. Brightfield imaging revealed progressive organization of tumoroid structures over two days, and by day 2, immunofluorescence further confirmed the phenomena (Fig. 4C–D). Compared with static culture, the MPS provided a more physiologically relevant environment that facilitated rapid organization into tumoroids.

Together, these LCTs substantially replicate essential structural and cellular characteristics of native lung TMEs, thereby facilitating investigation of tumor-stroma-immune interactions and their impact on therapeutic responses.

Patient-derived LCTs reveal tumor-intrinsic variations in therapeutic responses across reconstituted TMEs

To investigate how LCT reflect tumor-immune-stromal interactions in therapeutic responses, we first established various ex vivo co-culture models incorporating patient-derived lung cancer organoids (LCOs), CAFs or autologous immune cells (ICs) originated from two representative EGFR/ALK-negative tumors LC990 (mucinous ADC) and LC978 (SqCC). Concentrations of pemetrexed (Pem) and carboplatin (Carbo), standard chemotherapeutic agents for NSCLC without targetable mutations [42], were selected based on preliminary dose–response analyses. Across all co-culture conditions, the IC₅₀ for carboplatin ranged from 14.6 to 45.3 μM. In contrast, pemetrexed exhibited higher and more variable IC₅₀ values (56.5 μM to >1,100 μM, with some conditions undetermined), indicating relative resistance in both models (Supplementary Fig. S4A). Consequently, both drugs were applied at a fixed concentration of 100 μM—approximately 2× the highest observed IC₅₀ of carboplatin—to enable standardized comparisons despite the relative resistance to pemetrexed. Tumor responses were evaluated across four configurations: LCO, LCO+CAF, LCO+IC, and LCT. For the immune-relevant models (LCO+IC and LCT), anti-PD-1 or chemo-immunotherapy was administered. To access efficacy, the area of viable tumor organoid (>100 μm) was quantified and normalized to control samples (Fig. 5A–D). Additionally, to account for variations in total cell number across co-culture conditions, the change in Δ tumor area per total cell was calculated as a complementary measure of tumor-suppressive efficiency (Supplementary Fig. S4F–I).

Fig. 5.

Drug response profiling in patient-specific lung cancer tumoroids. (A–D) quantification of normalized tumor area in viable spheroids (>100 μm) derived from LC990 and LC978 models treated with chemotherapy [pemetrexed (Pem), carboplatin (Carbo), or their combination (Pem + Carbo)] (A, C) or chemo–immunotherapy [anti–PD-1 antibody (α–PD-1), Pem + α–PD-1, Carbo + α–PD-1] (B, D). Analyses were performed across four configurations: lung cancer organoid (LCO), LCO with cancer-associated fibroblasts (LCO+CAF), LCO with immune cells (LCO+IC), and lung cancer tumoroid (LCT). Carboplatin-only controls (A, C) were performed in parallel and used as the baseline for the corresponding combination panels (B, D). (E) Interferon-γ (IFN-γ) secretion measured by ELISA in immune cell-containing co-cultures (LCO+IC, LCT) under α–PD-1 or combination therapy for LC990, LC978, and LC993 models. (F) Seahorse metabolic analysis of LC993 models showing oxygen consumption rate (OCR) and extracellular acidification rate (ECAR) under chemo-immunotherapy conditions. Plotted values represent the average OCR and ECAR across three independent biological replicates. Data are presented as mean ± SEM of biological replicates (n = 3), unless otherwise indicated. Each dot represents an individual biological replicate. Statistical analyses were performed for all panels. Statistical significance was assessed using the Kruskal–Wallis test followed by Dunn’s multiple comparisons test (panels A-D, right) or the mann-whitney U test (panel E). Non-significant comparisons are shown in the right panels to indicate the full set of comparisons performed. ns, not significant (p > 0.05); *p < 0.05; **p < 0.01; ***p < 0.001; ****p < 0.0001. Scale bar, 200 μm. Pem, pemetrexed; Carbo, carboplatin; α–PD-1, anti–PD-1 antibody. Chemo, chemotherapy. α–PD-1 ± chemo indicates anti–PD-1 monotherapy or its combination with chemotherapy (pemetrexed or carboplatin)

In LC990, carboplatin and combination chemotherapy (Pem + Carbo) induced significant tumor regression (p < 0.05), particularly under immune cell-containing conditions (LCO+IC, LCT). The inclusion of CAFs attenuated this effect, suggesting partial stromal-mediated resistance (Fig. 5A and Supplementary Fig. S4B). A similar trend was observed with chemo-immunotherapy (Carbo + α–PD-1), which was most effective in LCO+IC (p = 0.0395), but attenuated and not statistically significant in CAF-containing LCT (p = 0.4202) (Fig. 5B and Supplementary Fig. S4D).

Notably, LCT990 showed a trend toward increased tumor reduction than LCT978 across chemo-immunotherapy conditions. In LCT990, this response was accompanied by a modest increase in IFN-γ secretion under anti-PD-1 (α–PD-1)–based treatment (Fig. 5E). Despite this elevated cytokine production, the antitumor effect was attenuated in CAF-containing conditions (Fig. 5A–B and Supplementary Fig. S4B and S4D), indicating a functional dissociation between immune activation and cytotoxic efficacy. In contrast, LCT978 showed limited tumor reduction (Fig. 5C–D and Supplementary Fig. S4C and S4E) and persistently low levels of IFN-γ production across all treatment and co-culture conditions (Fig. 5E). These findings suggest that CAFs modulate immune activation differently across individual patient tumors, and tumor-intrinsic factors more critically impact therapeutic outcomes [43].

To further dissect CAF-driven immunomodulation, we evaluated the LC993 models (LCO+IC and LCT) under α–PD-1 monotherapy and combination treatments (Pem+α–PD-1, Carbo+α–PD-1). Seahorse metabolic assays demonstrated increased mitochondrial respiration and glycolysis in CAF-containing LCT cultures compared to LCO+IC, particularly under Carbo+α–PD-1 treatment, indicating enhanced cellular activation [44] (Fig. 5F). However, this metabolic stimulation was not accompanied by increased IFN-γ production, which generally remained reduced in LCT compared to LCO+IC [45] (Fig. 5E), demonstrating a partial dissociation between immune effector function and metabolic activation. The LC993 model derived from a different patient functionally supports the results of LC990, preserved immune responsiveness despite partial CAF-mediated resistance. These results underscore the context-dependent capacity of CAFs to selectively modulate cytokine production [46].

Clinical information for the patients from whom LCT990, LCT978, and LCT993 were derived is summarized in Supplementary Table 4. Direct comparison between ex vivo drug response and clinical outcome was limited in the LCT978 case, as the patient did not receive α–PD-1 therapy. Similarly, assessment of treatment response in the LCT990 case was challenging because adjuvant α–PD-1 therapy was administered only after completion of adjuvant combination chemotherapy following lobectomy. Both patients currently remain disease-free. In contrast, the patient from whom LCT993 was derived developed recurrence at the surgical stump after adjuvant combination chemotherapy and subsequently received pembrolizumab treatment. However, treatment was discontinued due to grade 5 pneumonitis, and radiologic evaluation demonstrated partial response (Supplementary Fig. S9).

Although direct comparison between patient clinical outcomes and ex vivo treatment responses was therefore limited, our findings collectively support the notion that the patient-specific LCO+CAF+IC (LCT) model not only preserves immune responses within the tumor microenvironment but also captures inter-patient variability in therapeutic response.

By integrating autologous tumor, stromal, and immune components, this ex vivo platform offers a physiologically relevant system for functionally stratifying therapeutic responses [37].

Reconstitution of patient-specific TMEs reveals tumor-intrinsic differences in immune responsiveness and CAF function

Bulk RNA-seq was performed on independently processed LCO, LCO+IC, and LCT samples from LC990 and LC978 using only high-quality RNA (RIN ≥ 7.0). Treatment conditions for RNA-seq were selected based on the functional drug-response results shown in Fig. 5. For LC990 (ADC), both chemotherapy (CM) and immunotherapy (IM) conditions were analyzed. In the CM setting, LCO samples were included as a tumor-only reference to enable comparison with LCO+IC and LCT samples, as anti–PD-1 treatment requires immune cells to exert its biological effects and was not expected to induce substantial transcriptional changes in LCO-only cultures. For LC978 (SqCC), functional assays demonstrated limited tumor reduction under CM and low IFN-γ production following anti–PD-1 treatment. Accordingly, transcriptomic analysis was restricted to the IM condition, comparing LCO+IC and LCT samples under immune-stimulated conditions, while optimizing the use of limited patient-derived material. All libraries were prepared using the same workflow and sequenced on the same platform to minimize technical variability. Differential expression analyses were conducted as pairwise comparisons within each patient model, applying a fixed BCV when replicates were unavailable. This framework ensures that the stromal, immune, and CAF-associated differences described below reflect biological rather than technical variation.

To evaluate the fidelity of TME reconstitution in LCT, we first assessed stromal and immune cell signature scores, which represent their abundance, using the ESTIMATE algorithm and tumor purity scores derived from transcriptome data across co-culture models [47] (Fig. 6A). In LC990, the sequential addition of immune cells (LCO+IC) and CAFs (LCT) progressively increased stromal and immune cell scores while reducing cancer cell purity, with fully reconstituted LCT990 (IM) model exhibiting the highest stromal/immune cell scores. These characteristics persisted under α–PD-1 treatment, indicating a stable and functionally engaged immune architecture. In contrast, LC978-derived models—including LCT978 (IM)—exhibited minimal changes in stromal/immune signatures and consistently high tumor purity, reflecting an immune-excluded phenotype. CIBERSORT-based immune cell deconvolution further revealed a diverse immune repertoire in LC990, including CD8⁺ T cells, memory CD4⁺ T cells, macrophages, and NK cells [13] (Fig. 6B), particularly enriched under LCT and α–PD-1 conditions. In comparison, LC978 exhibited a relatively immune-cold profile across all co-culture conditions (Supplementary Fig. S5D).

Fig. 6.

Divergent CAF-immune interactions shape TME composition and therapy response. (A) ESTIMATE stromal, immune, and tumor purity scores across co-culture conditions in LC990 and LC978 under chemotherapy (CM) and immunotherapy (IM) treatments. (B) CIBERSORT-based immune cell deconvolution of LC990 and LC978. Immune subsets, including CD8⁺ T cells, CD4⁺ memory T cells, macrophages and natural killer (NK) cells are shown. Samples with p values > 0.05 were excluded; fewer LC978 samples met this criterion. (C) Tumor-intrinsic transcriptomic profiles under immune stimulation in LC990 and LC978. (D) Predicted drug sensitivity across 198 compounds for LC990 and LC978 using OncoPredict, assessed by half-maximal inhibitory concentration (IC₅₀) values (Wilcoxon signed-rank test, ***p < 0.001). (E) Top 20 upregulated CAF-associated genes identified in LC990 and reactome pathway enrichment of CAF-specific transcriptional programs. (F) CAF subtype scores across co-cultures in LC990 and LC978 under chemotherapy (CM) and immunotherapy (IM). Two-sided Student’s t-tests; *p < 0.05; **p < 0.01; ***p < 0.001; ****p < 0.0001. (G) Functional pathway activity across models. Singscore-based signatures for IFN-γ, T cell exhaustion, transforming growth factor-β (TGF-β), and epithelial-mesenchymal transition (EMT) were quantified across treatment conditions in LC990 and LC978. LCO, lung cancer organoid; LCO+IC, LCO with immune cells; LCT, lung cancer tumoroid; CM, chemotherapy (pemetrexed [Pem], carboplatin [Carbo], or their combination); IM, immunotherapy (anti–PD-1 [α–PD-1] alone or in combination with Pem or Carbo). Ctrl, control; untreated indicates samples that did not receive either the therapeutic antibody or chemotherapeutic agents

To investigate the role of CAFs in modulating immune responses, we compared gene expression profiles between CAF⁺ and CAF⁻ models under immune-stimulated conditions (Supplementary Fig. S5A). In the LC990 model (LCT990, IM), CAFs induced both immune-stimulatory genes (e.g., CXCL9-11) [48] and ECM-remodeling genes (e.g., SULF1, DCN) [49]. In contrast, CAFs in the LC978 model (LCT978, IM) upregulated structural ECM programs (e.g., COL1A2, PDGFRA) [50] without activating immune-related transcripts, highlighting divergent functional states. Consistently, tumor-intrinsic transcriptomic profiles under immune stimulation revealed strong interferon signaling and antigen presentation in LC990 (e.g., CXCL9–11, GBP5, IDO1, HLA-DRB6), whereas LC978 showed ECM-associated gene enrichment (e.g., CPXM2, CTSK, SERPINF1) and minimal immune activation (Fig. 6C). In silico drug response predictions using OncoPredict [15, 16] further differentiated the models, with LC990 displaying significantly lower predicted IC₅₀ values across multiple agents compared to LC978, indicating enhanced drug sensitivity (Fig. 6D).

Considering the distinct immune engagement and drug sensitivity observed in LC990 compared to LC978, we hypothesized that qualitative differences in CAF function may underlie this divergence. To investigate this, we conducted a two-step differential expression analysis to define CAF-specific transcriptional programs in the responder context. By excluding immune-induced genes (LCO+IC) from those upregulated in LCT, we identified 137 CAF-enriched transcripts that were uniquely elevated in LC990 (Supplementary Fig. S5B). The top 20 genes included interferon-inducible and immune-regulatory transcripts (EREG, SERPINB2, TNFSF13B, CSF1, TAP1/2, and IFI30) (Fig. 6E, top), which are associated with cytokine signaling and antigen presentation [51]. Pathway analysis using the Reactome database revealed enrichment in immune-relevant pathways such as PD-1 signaling, CD28 co-stimulation, PI3K/AKT, TNFR2, and immune synapse formation (Fig. 6E, bottom).

To validate the clinical relevance of the top 20 CAF-specific genes identified in LC990, we analyzed four public cohorts of lung adenocarcinoma patients treated with immune checkpoint inhibitors (ICIs): GSE126044 [52], GSE135222 [53], GSE161537 [54], and GSE190265 [55]. Enrichment scores for these 20 genes were calculated from patient RNA-seq data using the Singscore algorithm. When patients were stratified into high and low groups based on the median value, the high-score group demonstrated a significantly better response to ICI therapy and a favorable prognosis (Supplementary Fig. S8). These clinical findings strongly align with our observations in the organoid model, where an immunogenic microenvironment was induced, thereby closely mirroring real-world patient responses to immunotherapy.

Collectively, these findings suggest that CAFs in LC990 foster an immune-supportive microenvironment, whereas those in LC978 impose an immune-restrictive state.

We next assessed whether the heterogeneity of CAF subtypes accounted for functional divergence. Singscore-based analysis revealed distinct CAF subtype profiles: LCT990 (IM) was enriched for inflammatory CAFs (iCAF, iCAF.2), while LCT978 (IM) showed higher scores for structural and matrix-remodeling CAFs (myCAF, dCAF, pCAF), along with limited inflammatory activity (Fig. 6F and Supplementary Fig. S5C). Although both models exhibited global CAF activation, LC990-derived CAFs were associated with immune-supportive phenotypes, whereas LC978 CAFs were characterized by structural and proliferative remodeling.

Finally, we examined the functional outcomes associated with CAF heterogeneity by quantifying singscore-based signatures across various treatments (Fig. 6G). In LC990, the addition of CAFs to chemotherapy (LCT990, CM) modestly increased T cell exhaustion scores compared to tumor-only conditions. Under immune stimulation (IM), both LCO990+IC and LCT990 exhibited near-zero T cell exhaustion scores, reflecting active T cell engagement. In contrast, LC978-based models maintained low T cell exhaustion levels, with LCT978 (IM) showing the most suppressed score (~–0.4), consistent with sustained immune exclusion. The TGF-β and EMT pathways were most activated in LCT990 (IM), followed by LCT978 (IM), while remaining low in CAF⁻ models, highlighting CAF-driven mesenchymal activation [56]. Notably, this mesenchymal signaling in LC990 coincided with robust IFN-γ and T cell activation, whereas LC978 lacked such immune signatures. These findings suggest that CAFs in LC990 are associated with coordinated immune-stromal activation, while CAFs in LC978 maintain an immune-silent, structurally reprogrammed niche.

Collectively, these findings suggest that CAFs modulate immune engagement and therapeutic responses in a tumor-intrinsic, context-dependent manner [43]. Specifically, CAFs in LC990, fostered enhanced immune infiltration, IFN-γ–related activation, and mesenchymal remodeling, which correlated with augmented drug sensitivity. In contrast, CAFs in LC978 promoted ECM enrichment and limited immune activation, establishing an immune-restrictive niche with attenuated therapeutic responsiveness. These results highlight the value of patient-specific tumoroid models for reconstituting the TME and dissecting stromal-tumor interactions, suggesting that CAF heterogeneity could serve as a promising axis for lung cancer treatment stratification [57].

Discussion

The TME in lung cancer exhibits significant heterogeneity, characterized by patient-specific variations in stromal and immune cell composition that critically influence therapeutic response [58]. Flow cytometric analyses of NSCLC tissues verified significant interpatient variability in immune and stromal cell composition, independent of histological subtype. These results underscore the need for personalized models that capture the diversity of TME architecture and reflect the biological complexity of the native tumors.

To capture the functional complexity of tumor-stroma-immune interactions in vitro in a personalized model, we developed a modular Transwell-based platform to simultaneously isolate and expand cancer organoids, CAFs, and TILs from a single input of patient lung cancer tissue, which overcomes tissue limitation and intra-tumoral heterogeneity. With this approach, we demonstrated the feasibility of a practically applicable personalized TME- integrated model. While initial LCO formation and TIL acquisition were highly successful, CAF isolation and long-term LCO culture showed markedly lower success rates. The results suggest that although CAF overgrowth have been reported as a major hindering factor for success of LCO formation [11], loss of TME components may influence LCO growth by altering the balance of proliferative and suppressive cues during culture. Analysis of isolation efficiency across diverse tumors revealed notable differences according to tumor differentiation and patient age. These findings may reflect differences in extent of desmoplastic reaction producing CAF in the tumor tissue according to tumor differentiation and host intrinsic biology.

Importantly, the TME components acquired by the Transwell platform preserved the phenotypic and functional integrity of their origin. LCOs retained the histological features of the primary tumor, while CAFs maintained mesenchymal markers, including Vimentin and Prrx1 [59]. TILs preserved their subset distributions and expanded over time. Notably, CD8⁺ and CD8⁻ T cells differentiated toward an effector memory phenotype during culture, indicating sustained proliferative potential and partial activation. To fully validate tumor organoids, in vivo tumorigenicity assays will be ideal. However, their tumor-forming capacity of these organoids has been demonstrated in our previous study [11], supporting the intrinsic viability and functional integrity of the tumor compartment.

After a period of biobanking up to 12 months, the three TME components were reassembled into tumoroids, in which tumor cells, CAFs, and immune cells organized into three-dimensional structures resembling native histologic tissue architecture. These findings may provide a structural basis for functional assessment of tumor-stroma-immune interactions. Successful incorporation of all three cell types into a microphysiological system (MPS) further expanded the utility of our system. Given its compatibility with multiplexed readouts, this system has the potential to serve as clinically relevant tools for personalized immune-oncology research. Additionally, transcriptomic analysis of LCTs revealed heterogeneous tumor-stroma-immune cell compositions across models, further supporting personalized models with inter-patient heterogeneity.

Under therapeutic conditions, these three-dimensional tumoroids provided insights into the role of CAFs in modulating immune activation and suppression, as seen in the fundamental differences between the therapy-sensitive LC990 and therapy-resistant LC978 models, which exhibited distinct immune and stromal microenvironmental states. These findings suggest that while CAFs can influence immune dynamics, the direction and magnitude of their effects are shaped by tumor-intrinsic properties [60]. Transcriptomic profiling further supported this context dependency. LC990-derived CAFs expressed immune-related programs, including interferon signaling and antigen-processing pathways [61], whereas LC978-derived CAFs were enriched for extracellular matrix remodeling and contractility signatures. CAF subtype classification aligned with these observations, showing predominance of inflammatory CAF subsets in LC990 and structural subsets in LC978 [4]. Collectively, these results suggest that CAF subtype may provide an additional axis for patient stratification in immunotherapy. However, these CAF subtype assignments were inferred from bulk RNA-seq using published gene signatures rather than from the more reliable single cell RNA-seq. Consequently, the findings should be interpreted cautiously as transcriptional associations rather than definitive functional classifications. Additionally, future imaging-based approaches such as direct spatial co-localization of CAFs and immune cells after anti-PD-1 treatment may serve as valuable complementary strategies to visualize CAF–immune interactions in situ. However, this was not technically feasible on our platform due to the microscale and fragility of LCTs.

We additionally observed a CAF-associated dissociation between metabolic activation and cytokine production in the independent patient-derived model LCT993, consistent with previous reports of CAF-mediated metabolic–immune uncoupling [62]. Given the limited number of patient-derived models examined, these observations warrant further validation. Nevertheless, they suggest that additional layers of stromal regulation may emerge in a patient-specific context. Such variability may underlie inconsistent outcomes of stromal-targeting strategies and highlights the potential value of individualized TME models in refining therapeutic prediction and informing precision immuno-oncology approaches. Further validation across larger patient cohorts will be required to determine the broader applicability of these findings.

Although autologous TILs enabled TME reconstruction, their yield and ex vivo expandability varied substantially across patient samples, limiting their use in comprehensive multi-condition analyses. Given the minimal tissue input, PBMCs were therefore employed as a standardized immune input to enable consistent comparison of functional responses across experimental conditions. PBMC-derived T cells do not fully recapitulate the tumor-primed, exhausted, and antigen-experienced states of endogenous TILs and may include bystander immune populations that respond to non-tumor antigens. Nevertheless, previous studies have demonstrated that co-culture of peripheral blood lymphocytes or PBMCs with autologous tumor organoids can generate tumor-reactive T cells under defined conditions and capture patient-specific responses to PD-1 blockade, supporting the use of PBMCs as a practical immune component for functional immunotherapy modeling [38, 39, 63–65]. In our study, live-cell imaging revealed treatment-associated caspase activation and morphologic changes in PBMC-containing co-cultures, and matched imaging in the LC990 model showed comparable qualitative responses in PBMC- and TIL-containing co-cultures following anti–PD-1 treatment. Although these imaging experiments were performed under conditions distinct from the main functional assays and were intended as qualitative observations, together with the IFN-γ ELISA results they suggest that PBMC-derived immune cells in the LCT model retain functional properties relevant to anti-tumor responses.

Nonetheless, our study did not include direct cytotoxicity assays or comprehensive profiling of exhaustion markers, which could further delineate functional immune responses. Future work incorporating TCR clonotype tracking, single-cell transcriptomics, and spatial proteomics will be essential to distinguish tumor-specific immune cells from bystander cells and to further validate the physiological relevance of the reconstructed immune compartment.

Despite these limitations, our findings have significant translational implications. The ability to reproducibly reconstitute individualized TMEs ex vivo, demonstrated here in both immunologically active and cold tumors, provides a physiologically relevant framework for dissecting tumor-immune-stroma interactions and supports functional assessment of therapy-induced changes in situ [66, 67]. These tumoroids, coupled with transwell-based isolation culture system, are scalable for further subsequent applications, including drug screening, immune profiling, transcriptomic analysis, and studies of stromal remodeling and immune modulation [68]. Moreover, evaluating therapeutic responses across heterogeneous microenvironments underscores the necessity of incorporating stromal and immune components into precision treatment strategies [69].

Our findings further suggest that stromal phenotyping, particularly the distinction between IFN-γ–responsive and immunosuppressive subsets, may serve as a framework for stratifying patients based on their likelihood of responding to immunotherapy [70]. This provides a rationale for selectively targeting pro-tumorigenic fibroblasts while preserving or reprogramming those with immune-supportive functions [71]. Current study had limited availability of clinically matched treatment response data for direct comparison with ex vivo findings. Because some patients did not receive immune checkpoint inhibitor therapy, while others underwent multimodal treatment or discontinued therapy due to treatment-related adverse events, accurate assessment of concordance between clinical outcomes and ex vivo drug responses were challenging. In addition, the small number of patient-derived models limits the generalizability of our observations regarding inter-patient heterogeneity in therapeutic response. Further validation using larger clinically annotated cohorts with standardized treatment settings will therefore be necessary to establish the predictive utility of the LCT platform for personalized immunotherapy response assessment and may underscore its potential to capture patient-specific TME dynamics.

Building on this foundation, future extensions of the platform may incorporate additional components of the TME, such as tumor-associated endothelial cells or tertiary lymphoid structures, thereby improving physiological fidelity and enabling a more comprehensive exploration of stromal-vascular-immune crosstalk in personalized immuno-oncology [72].

Conclusions

In conclusion, we developed a patient-specific lung cancer tumoroid (LCT) model that enables the reproducible ex vivo reconstruction of the tumor microenvironment by integrating tumor, stromal, and immune components from a single tissue source. This model successfully recapitulates key structural and functional features of native TMEs while captures interpatient heterogeneity in therapeutic responses.

Importantly, the LCT model reveals context-dependent roles of cancer-associated fibroblasts in modulating immune activation and therapeutic efficacy, highlighting stromal heterogeneity as a critical determinant of treatment response. By enabling functional interrogation of tumor–stroma–immune interactions under clinically relevant conditions, this model provides a scalable and translationally relevant framework for modeling therapeutic responses and investigating resistance mechanisms.

Collectively, these findings support the clinical translation of patient-specific TME models in precision immuno-oncology and the development of personalized treatment strategies for lung cancer.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

ADC

Adenocarcinoma

α-PD-1

Anti-programmed cell death protein 1 antibody

CAFs

Cancer-associated fibroblasts

Carbo

Carboplatin

CM

Chemotherapy

IC

Immune cells

IM

Immunotherapy

LCOs

Lung cancer organoids

LCT

Lung cancer tumoroid

NSCLC

Non-small cell lung cancer

PBMCs

Peripheral blood mononuclear cells

Pem

Pemetrexed

SqCC

Squamous cell carcinoma

TILs

Tumor-infiltrating lymphocytes

TME

Tumor microenvironment

Author contributions

S.J.J. and Y.R.K. designed the study; Y.R.K., S.M.K., and H.S.P. performed the experiments; B.A., H.S.H. resected and provided patient-derived lung cancer tissues; C.O.S., C.W.P. analyzed the RNA-seq data; S.J.J., Y.R.K., J.K. and D.J.J. conducted data analysis; and S.J.J., Y.R.K. drafted the manuscript with contributions from all co-authors.

Funding

This research was supported by the Bio & Medical Technology Development Program funded by Ministry of Science and ICT (MSIT) (2019M3E5D3071926) and Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (RS-2024-00408858).

Data availability

The RNA-seq data generated in this study have been deposited in the NCBI Gene Expression Omnibus (GEO) database under accession number GSE311792 and will be publicly available on December 31, 2026. Prior to public release, these data and all other data supporting the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethical approval and consent to participate

This study was conducted in accordance with the principles outlined in the Declaration of Helsinki and received approval from the Institutional Review Board of Asan Medical Center (IRB no. 2019–1210). Written informed consent was obtained from all participants prior to sample collection.

Consent for publication

Not applicable.

Competing interest

All authors declare no competing interests.