Reconstruction of T cell infiltration in an osteosarcoma PDX-organoid interactive biobank for personalized immunotherapy

TL;DR

Researchers created a biobank of osteosarcoma organoids linked to patient tumors and immune cells to test personalized immunotherapies and identify effective drug combinations.

Contribution

A novel PDX-organoid biobank is developed to model immune infiltration and test immunotherapy responses in osteosarcoma.

Findings

PDX-derived organoids maintain spatial and genomic features of osteosarcoma tumors.

Co-culturing with PBMCs enables T cell infiltration and immune response modeling in organoids.

PRMT5MTA inhibition improves immunotherapy response in chromosome 9p21.3-deleted osteosarcoma.

Abstract

Osteosarcoma (OS) is an aggressive malignant bone tumor with limited therapeutic options and low response to immunotherapy. OS rarity slows clinical translation, highlighting the need for models that bridge patient-derived xenograft (PDX) systems and next-generation platforms. Here, we establish an OS PDX-organoid interactive biobank by self-assembling single-cell suspensions into individualized OS organoids (iOSs). iOS models recapitulate OS spatial and architectural features at millimeter scale in vitro and as xenografts and maintain functional pairing with matched PDX models. We validate iOS fidelity using histopathology, spatial features, genomics, transcriptomics, and pharmacogenomics. By reconstructing T cell infiltration in PDX-derived iOS models, we model treatment-associated immune responses and support immunotherapy translational studies. Using paired iOS-PDX models, we show that a PRMT5MTA inhibitor enhances immunotherapy response in chromosome 9p21.3-deleted OS.

•A PDX-organoid biobank captures inter-patient heterogeneity in osteosarcoma•PBMC co-culture rebuilds T cell infiltration and functional immune readouts in organoids•Immune-reconstructed iOS drug testing aligns with patient-specific therapeutic response•Chr9p21.3/MTAP loss marks vulnerability to APRN2169 plus PD-1 blockade

A PDX-organoid biobank captures inter-patient heterogeneity in osteosarcoma

PBMC co-culture rebuilds T cell infiltration and functional immune readouts in organoids

Immune-reconstructed iOS drug testing aligns with patient-specific therapeutic response

Chr9p21.3/MTAP loss marks vulnerability to APRN2169 plus PD-1 blockade

Sun et al. establish a PDX-linked osteosarcoma organoid biobank and rebuild immune infiltration by co-culturing organoids with human PBMCs. This platform connects tumor genotype and immune context to therapeutic response, enabling ex vivo drug testing and supporting a precision combination strategy for Chr9p21.3/MTAP-deleted osteosarcoma.

Figures

Construction and characterization of an osteosarcoma PDX-organoid interactive biobank(A) Illustration of osteosarcoma self-assembling organoid model construction. The process involved 2 steps, single-cell suspension preparation of iOS models and self-assembly of single-cell iOS model suspension.(B) Bright-field images of iOS model during continuous passaging, using iOS_087 as an example, captured on days 3, 7, and 10. The organoids exhibited spherical growth. *n* = 3 biological replicates for each group. Scale bar, 250 μm.(C) Clinical annotation of patient information from osteosarcoma PDX-org

Genomic heterogeneity of iOS models in the biobank(A) Whole-genome analysis of chromosomal CNVs, comparing primary (P0) and passaged iOS models with paired tissue. Genomic deletions are represented in blue, while amplifications are depicted in red.(B) CNVs in genes previously implicated in osteosarcoma pathogenesis were identified. Genomic deletions are represented in blue, while amplifications are depicted in red.(C) Whole-genome analysis of SNVs, comparing primary (P0) and passaged iOS models with paired tissue.(D) Fusion gene alterations using STAR-Fusion analysis, comparing primary (P0) an

Specificity and reproducibility of iOS models in the biobank(A–C) PCA to evaluate the reproducibility of iOS models and paired tissue, including RNA TPM expression profile, ssGSEA score derived from GO genesets, and GDSC drug response prediction score.(D) PCA of transcriptomics to assess the specificity of iOS models and paired tissue. Using TCGA, TARGET-OS, and SGH-OS datasets as references, iOS models, paired tissue, and osteosarcoma samples were clearly distinguishable from other sarcomas.(E) Spearman’s correlation analysis to evaluate the reproducibility of primary (P0) iOS models and pair

Reconstruction of T cell infiltration in PDX-derived iOS models(A) Workflow for T cell infiltration reconstruction in PDX-derived iOS models. Single-cell iOS dissociation was performed followed by PBMC activation and suspension culture for 24 h. The mixture culture system included a 1:1 ratio of PBMCs and iOS, cultured in a medium suitable for immune cell interaction. Scale bar, 250 μm.(B) mIF and statistical analysis of CD8^+^ T cells and HIF-1α expression in iOS models of different diameters (900, 1,900, 1,400, and 2,500 μm). Scale bars are indicated in the figure, statistical significance w

Clinical response validation of T cell-reconstituted iOS models(A) Workflow for clinical validation of T cell-reconstituted iOS models. After osteosarcoma diagnosis, PDX models were established from resectable tumors. For patients with unresectable tumors, PDX models were created directly from biopsy samples. Following standard chemotherapy treatment, corresponding PDX models from the biobank are selected to generate iOS models, which are then reconstituted with PBMCs. These iOS models are treated with doxorubicin, and therapeutic efficacy is assessed through cell viability and bright-field im

Utilization of characteristic Chr9p21.3-deleted PDX-iOS models for personalized immunotherapy development(A) Osteosarcoma-associated genes, *CDKN2A*, and *MTAP*, located at 9p21.3 locus on *Homo sapiens* chromosome 9 (Chr9p21.3).(B) MCP-counter analysis performed to quantify cell proportions in SGH-OS. Relative cell abundance was represented by a color gradient.(C) Mechanistic basis of PRMT5^MTA^ inhibitor (APRN2169) efficacy. MTA and APRN2169 occupy the pocket simultaneously and a “LOCKTAC” mechanism stabilizes the binding to the drug to the PRMT5^MTA^, thereby inhibiting PRMT5 activity.(D) *

Full text

Introduction

Osteosarcoma (OS) is a highly malignant primary bone tumor predominantly affecting children and adolescents.1 It is recognized as one of the most lethal pediatric solid tumors,2^,^3 with an overall survival rate of approximately 70%.4^,^5 Despite advances in treatment, the survival outcomes for OS patients have remained relatively stagnant over the past four decades. Of particular concern is the poor prognosis for patients who fail first-line therapy, where survival rates drop to 20% or lower due to disease recurrence and metastasis, compounded by the lack of effective salvage treatments.6 Immunotherapy, while emerging as a promising approach for solid tumors, has demonstrated limited efficacy in OS.7^,^8^,^9

The rarity of OS, with an incidence rate of 3–5 cases per million people,10 has hindered clinical and translational research on personalized immunotherapy for OS. Current research models for OS immunotherapy often fail to reflect the patient-specific variations critical for effective treatment. While immunocompetent mouse-derived xenograft models (e.g., K7 series and11 DuNN series12) exist, they frequently lack the characteristic mutations and driver features seen in clinical cases, leading to the failure of promising preclinical immunotherapy strategies in subsequent clinical trials.

Although patient-derived xenograft (PDX) models have been developed over nearly a decade, resulting in comprehensive PDX biobanks,13^,^14^,^15 these models in nude mice fail to retain patient-specific immune characteristics. Additionally, humanized PDX models impose excessive economic and physical burdens on pediatric and adolescent patients, making them unsuitable for widespread clinical translation.16 This highlights the urgent need for the development of models that can preserve both the genetic heterogeneity and immune characteristics of OS, in order to accelerate the clinical translation of immunotherapy approaches.

With the increasing adoption of organoid models,17^,^18 traditional PDX models are gradually becoming obsolete.19^,^20^,^21 However, due to the rarity of OS, establishing a comprehensive and mature OS organoid biobank requires significant time. While organoid models can preserve the genetic heterogeneity of patient tumors, they exhibit limitations in immunotherapy research.22^,^23 Tumor-infiltrating lymphocytes (TILs) within these models undergo activation-induced exhaustion over time, leading to the loss of immune characteristics during passaging. Although OS organoid models have been reported, their small size (approximately 100 μm in diameter) fails to fully recapitulate the spatial features of the OS microenvironment.24^,^25^,^26^,^27 Thus, there is an urgent need for a transitional model that bridges the gap between PDX models and next-generation clinical translation models, while accurately representing the OS microenvironment.

In this study, we have established an interactive PDX-organoid biobank for OS using a self-assembly approach. We generated OS organoids with a diameter of approximately 1 mm to better recapitulate the spatial characteristics of the tumor. Virtual drug sensitivity analysis was employed to assess the heterogeneity between organoids and their paired tissue samples. Based on these findings, we reconstructed T cell infiltration in the OS organoids and utilized Chr9p21.3-deleted OS organoid models to evaluate the potential of PRMT5^MTA^ inhibitor combination immunotherapy in preclinical drug development.

Results

Establishment of an osteosarcoma PDX-organoid interactive biobank

Given the rapid proliferation of OS tumors, the generally large volume of diagnostic specimens, and the marked spatial heterogeneity of cellular subpopulations within the tumor microenvironment, we established a self-assembly strategy to generate OS organoids (Figure 1A). Necrosis-free OS tissue specimens were carefully selected and enzymatically dissociated to obtain highly viable single-cell suspensions (Table S1). These single cells were subsequently cultured under optimized conditions to self-assemble into individualized OS organoid (iOS) models. Across all samples, the overall success rate for establishing iOS models was 74.2% (69/93) (Figure S1A). The success rate for iOS models derived directly from patient tumors was 66.7% (24/36), whereas models derived from PDXs reached 78.9% (45/57), likely reflecting the lower extent of necrosis, impurities, blood contamination, and potential infection in PDX tissues compared with primary specimens. Successfully established iOS models could grow to millimeter-scale, macroscopic sizes with consistent dimensions, supporting their utility for high-throughput screening and translational studies (Figure S1B; Table S2). These models were amenable to serial passaging, while maintaining their structural and functional characteristics over multiple generations (Figures 1B and S1C) and collectively encompassed a broad spectrum of clinical OS subtypes with associated multi-omics data, making them highly suitable for translational research (Figure 1C).Figure 1. Construction and characterization of an osteosarcoma PDX-organoid interactive biobank(A) Illustration of osteosarcoma self-assembling organoid model construction. The process involved 2 steps, single-cell suspension preparation of iOS models and self-assembly of single-cell iOS model suspension.(B) Bright-field images of iOS model during continuous passaging, using iOS_087 as an example, captured on days 3, 7, and 10. The organoids exhibited spherical growth. n = 3 biological replicates for each group. Scale bar, 250 μm.(C) Clinical annotation of patient information from osteosarcoma PDX-organoid interactive biobank. Data included age (Age); gender (Gender); postoperative necrosis rate (Necrosis); primary lesions (Primary lesion); pathological characteristics of derived lesions (Derived lesion); key biomarkers, including (KI67, p53, SATB2, and VIM); and multi-omics data (Bright-field image, genome, and transcriptome and pathomics data).(D) mIF and quantitative analysis of the iOS model with diameters of 900, 1,350, 1,800, and 2,300 μm. n = 3 technical replicates for each group. Scale bars are indicated in the figure. Data represented as mean ± SEM. Statistical significance was assessed by one-way ANOVA. The p values are indicated.(E) Comprehensive H&E staining and mIF of the iOS model and paired tissue. The iOS model recapitulated spatial tumor cellular heterogeneity of the paired tissue. Scale bars are indicated in the figure.

The entire process from initial self-assembly to maturation was rapid, typically completed within 5–7 days (Figures S2A, S2B, and S3), thereby enabling completion of experimental workflows within 10 days. To assess the presence of hypoxic necrosis within iOS models, we performed multiplex immunofluorescence (mIF) and quantitative analyses on iOS models with diameters of 900, 1,350, 1,800, and 2,300 μm. Proliferation, as indicated by Ki67, did not differ significantly across model sizes. By contrast, HIF-1α and c-caspase-3 expression increased with organoid size, consistent with nutrient and oxygen diffusion limitations in larger iOS models, leading to hypoxia and apoptosis—a pattern that mirrors key microenvironmental features of solid tumors (Figure 1D). To further validate the in vivo tumorigenic capacity and biological activity of the established iOS models, we transplanted them subcutaneously into immunodeficient mice. The resulting iOS xenografts exhibited growth kinetics and histological architectures comparable to their paired PDX tissues (Figure S4). Collectively, these findings demonstrate the establishment of an interactive PDX-organoid biobank for OS, providing a robust platform for mechanistic studies, drug development, and translational research.

Millimeter-scale iOS models recapitulate osteosarcoma tumor spatial architecture

To determine whether iOS models recapitulate the spatial characteristics of OS, we performed comparative analyses of whole-slide images from millimeter-scale iOS models and their paired tumor tissues. Histopathological assessment using hematoxylin and eosin (H&E) staining demonstrated that iOS models faithfully mirrored the spatial hierarchy of their paired tissues. The iOS models preserved the characteristic tissue architecture of the original samples, including intratumoral heterogeneity such as dense tumor regions and osteogenic zones, as indicated by arrows in the representative images (Figures S5A and S5B). These findings were consistent with the cellular composition observed in single-cell suspensions derived from the corresponding tissues. Immunohistochemical (IHC) staining for established OS biomarkers—MKI67 (proliferation), SATB2 (osteoblastic differentiation), and VIM (mesenchymal origin)—further confirmed that iOS models maintained histopathological biomarker profiles comparable to those of the original tumors (Figure S5C). Moreover, comprehensive H&E and Masson staining of all successfully established iOS models from the PDX-organoid interactive biobank and their matched in situ tumor tissues revealed conserved histopathological features and bone matrix characteristics across models (Figures S6 and S7). The preservation of these biomarker and matrix features supports that iOS models reliably capture patient-specific pathological attributes, reinforcing their value for investigating OS biology and therapeutic responses.

To further characterize spatial complexity in iOS models, we applied mIF in combination with H&E staining to examine the distribution of major cellular and extracellular components within the tumor microenvironment. SMA was used to label cancer-associated fibroblasts, COL1A1 to visualize bone-associated collagen matrix, and SATB2 to identify OS tumor cells. Integrated analysis demonstrated that iOS models conserved the spatial complexity of the original tissues, with three anatomically and functionally distinct regions: an inner tumor core region, an outer fibrous tissue-rich active zone, and an intermediate transition zone (Figures 1E, S8, and S9). This organized spatial architecture, closely resembling that of native OS, indicates that iOS models accurately recapitulate the structural and cellular complexity of OS.

Pharmacogenomic recapitulation of iOS models from paired patient tumor/PDX tissues

To further validate the genomic recapitulation of iOS models compared to paired tissues, we performed whole-genome sequencing (WGS) on primary (P0) and passaged iOS models alongside their paired tissues. Analysis of copy number variations (CNVs) across chromosomes revealed a generally high concordance between iOS models and their paired tissues, while also highlighting substantial genomic heterogeneity that emerged during passaging, as well as marked inter-patient heterogeneity among OS cases (Figures 2A and S10A). Further analysis of CNVs in key OS driver genes demonstrated that characteristic CNV alterations were preserved during iOS model establishment and passaging (Figures 2B and S10B). Consistently, correlation analysis of CNV profiles between iOS models and their corresponding PDX or patient-derived in situ tumors showed a high degree of genomic similarity (Figures S11A and S11B; Tables S3 and S4). Critical genomic features such as MYC amplification, RB1 deletion, and CDKN2A deletion were consistently retained. Analysis of single-nucleotide variants (SNVs) revealed that while OS harbors relatively few SNVs, hallmark mutations in TP53, BRCA2, and KRAS were preserved. However, additional SNVs emerged during iOS model establishment and passaging, likely due to the inherent genomic instability of OS (Figures 2C and S11C). Additionally, we detected genomic fusions and breakpoints, which are frequent in OS due to its genomic instability. These features were also retained in iOS models during establishment and passaging (Figures 2D and S12). Overall, genomic analyses confirmed that iOS models maintained overall consistency with paired tissues during both establishment and passaging.Figure 2. Genomic heterogeneity of iOS models in the biobank(A) Whole-genome analysis of chromosomal CNVs, comparing primary (P0) and passaged iOS models with paired tissue. Genomic deletions are represented in blue, while amplifications are depicted in red.(B) CNVs in genes previously implicated in osteosarcoma pathogenesis were identified. Genomic deletions are represented in blue, while amplifications are depicted in red.(C) Whole-genome analysis of SNVs, comparing primary (P0) and passaged iOS models with paired tissue.(D) Fusion gene alterations using STAR-Fusion analysis, comparing primary (P0) and passaged iOS models with paired tissue. Intrachromosomal events are represented in red, and interchromosomal events are represented in blue.

To further validate the transcriptomic consistency between iOS models and paired tissues (Tables S5, S6, and S7), we first performed a correlation analysis of transcript abundance (TPM, transcripts per million) between iOS models and their corresponding PDX or patient-derived in situ tumors. Principal-component analysis (PCA) and Spearman’s correlation revealed high transcriptomic concordance among primary (P0) iOS models, passaged iOS models, and their paired tissues (Figures 3A and S13). Likewise, PCA and correlation analyses of molecular pathway activity—assessed by single-sample gene set enrichment analysis (ssGSEA) using Gene Ontology (GO) terms—demonstrated high consistency in pathway alterations between iOS models and matched tissues (Figures 3A and S14). Furthermore, pharmacogenomic profiling based on the Genomics of Drug Sensitivity in Cancer (GDSC) dataset, using OncoPredict, showed that most iOS models retained drug response signatures comparable to those of their paired OS tissues (Figures 3C and S15).Figure 3. Specificity and reproducibility of iOS models in the biobank(A–C) PCA to evaluate the reproducibility of iOS models and paired tissue, including RNA TPM expression profile, ssGSEA score derived from GO genesets, and GDSC drug response prediction score.(D) PCA of transcriptomics to assess the specificity of iOS models and paired tissue. Using TCGA, TARGET-OS, and SGH-OS datasets as references, iOS models, paired tissue, and osteosarcoma samples were clearly distinguishable from other sarcomas.(E) Spearman’s correlation analysis to evaluate the reproducibility of primary (P0) iOS models and paired tissue, including CNV variation, RNA TPM expression profile, ssGSEA score, and GDSC drug response prediction score (82.93% success rate in RNA-seq analysis and 85.71% success rate in sample with WGS and RNA-seq data).

To verify the OS-specific identity of the iOS models, we next compared their transcriptomic profiles with reference datasets from TCGA, TARGET-OS, and our previously published SGH-OS cohort28 as references. PCA demonstrated that iOS models and PDX tissues clustered closely with OS patient samples (Figure 3D), indicating that the culture conditions preserved OS-specific transcriptomic features without overtly altering tumor identity. To quantify the transcriptomic success rate of iOS model construction, we defined successful transcriptomic recapitulation as a Spearman’s correlation coefficient >0.5 with a p value <0.05 between each iOS model and its paired tissue. Using this criterion, 82.93% (34/41) of iOS models exhibited high transcriptional concordance. However, a subset of models—particularly at the initial establishment stage (P0)—displayed variability in RNA expression, pathway activity, and pharmacogenomic profiles (Figure 3E). Moreover, based on these criteria and the integrated genomic and transcriptomic analyses, the success rate of iOS model establishment reached 85.71% (Figure 3E; Tables S8 and S9). These findings indicate that although the majority of iOS models maintain strong transcriptomic fidelity to their source tissues, discrepancies can arise during early passages, highlighting the need for further optimization of initial culture conditions.

Reconstruction of T cell infiltration in PDX-derived iOS models

Analysis of cellular heterogeneity in iOS models showed that patient-derived iOS models retained diverse cell populations—including T cells and other tumor immune components—at the primary (P0) stage. In contrast, PDX-derived iOS models displayed reduced cellular diversity, consisting predominantly of tumor cells with minimal immune infiltration (Figure S16A). Given the breadth of our previously established OS PDX biobank, which integrates matched patient-PDX multi-omics data and serves as a resource for precision therapy research, we sought to leverage this platform by reconstituting immune components in PDX-derived iOS models. To this end, we introduced peripheral blood mononuclear cells (PBMCs) via co-culture, generating T cell-reconstructed iOS models (iOS + PBMC) (Figure 4A). This approach supported the survival of T cells within the models. Using microenvironment cell populations-counter (MCP-counter) analysis of transcriptomic data, we compared T-cell-reconstructed and non-reconstructed iOS models with paired PDX models in vivo, DuNN OS cell lines in vitro, and DuNN cell-derived xenografts (CDX) in immunocompetent syngeneic mice. The reconstructed iOS + PBMC models closely resembled DuNN CDX in vivo (Figure S16B). mIF further revealed that CD8^+^ T cell infiltration decreased as the diameter of T cell-reconstructed iOS models increased, whereas HIF-1α expression within the microenvironment progressively increased. These findings suggest that iOS models with diameters of approximately 1,000 μm maintain favorable immune infiltration with minimal hypoxia (Figure 4B). Characterization of the immune microenvironment showed uniform expression of TIM-3 and PD-1 and widespread infiltration of CD206^+^ macrophages, with limited FOXP3^+^ Treg accumulation (Figures S17A and S17B). Flow cytometry tracking of CD8^+^ T cell long-term persistence confirmed that infiltration remained stable for up to 21 days in the iOS + PBMC models (Figure S17C).Figure 4. Reconstruction of T cell infiltration in PDX-derived iOS models(A) Workflow for T cell infiltration reconstruction in PDX-derived iOS models. Single-cell iOS dissociation was performed followed by PBMC activation and suspension culture for 24 h. The mixture culture system included a 1:1 ratio of PBMCs and iOS, cultured in a medium suitable for immune cell interaction. Scale bar, 250 μm.(B) mIF and statistical analysis of CD8^+^ T cells and HIF-1α expression in iOS models of different diameters (900, 1,900, 1,400, and 2,500 μm). Scale bars are indicated in the figure, statistical significance was assessed by one-way ANOVA, and the p values are indicated (ns, no significance). Data represented as mean ± SEM. n = 3 technical replicates for each group.(C) UMAP plots showing cell composition of clinical patient samples, PBMC-reconstituted iOS models, paired PDX models, and iOS models.(D) Bar chart showing the ratio of different cell types derived from single-cell RNA sequencing analysis.(E) Flow cytometry validated the lymphocyte distribution in organoids. mIF of PDX-derived iOS model, paired PDX tissue, and iOS model after T cell infiltration reconstruction. Scale bars are indicated in the figure.(F and G) Bright-field imaging and cell viability of T cell-reconstituted iOS models (iOS+PBMC) treated with PD-1 mAb. Untreated and non-reconstituted iOS models served as controls. n = 3 biological replicates for each group. Scale bars, 250 μm. Data represented as mean ± SEM. Two-tailed Student’s t test was performed for statistical analysis.(H) Flow cytometric analysis of PD-1 expression in CD8^+^ T cells in reconstituted iOS models (iOS+PBMC), with PBMCs serving as controls. Data represented as mean ± SEM. Statistical significance was assessed by one-way ANOVA, and the p values are indicated. n = 3 technical replicates for each group.

To evaluate whether PBMC incorporation could mimic features of OS patient, we performed single-cell RNA sequencing on samples from a chemotherapy-naive OS patient, including original tumor tissue (patient, n = 1), iOS model (iOS, n = 1), paired PDX tissue (paired PDX, n = 1), and T cell-reconstructed iOS model (iOS + PBMC, n = 1). Uniform manifold approximation and projection (UMAP) identified seven major lineages: endothelial cells, myeloid cells, myoblasts, OS_cells1, OS_cells2, pericytes, and TILs (Figures S18A and S18B; Table S10). The OS patient tumor tissue and iOS + PBMC model exhibited similar cellular compositions (Figures 4C and 4D), and Spearman’s correlation analysis confirmed high transcriptomic concordance across groups and cell subsets (Figure S18C). Notably, patient tumor and iOS + PBMC model showed comparable cell-cell communication patterns (Figure S19A). Subclustering of TILs revealed memory T cells, cytotoxic T cells, and γδ T cells within the microenvironment (Figure S19B; Table S10). Reconstructed iOS models also displayed distinct cell subtype distributions relative to standard iOS models and PDX tissues. Pseudotime trajectory analysis showed that OS_cells1 and OS_cells2 followed divergent developmental and functional paths involving nuclear factor-κB signaling, immune response, and lipid metabolism pathways, suggesting complex immune-mediated regulation of tumor cell states within the OS microenvironment (Figures S19C and S19D). Differences in endothelial and pericyte composition may reflect sampling variability during PDX tissue collection for iOS model generation.

Spatial patterns of immune infiltration in T cell-reconstructed iOS models were further validated by mIF (Figure 4E). Functional interrogation using PD-1 monoclonal antibody (mAb) treatment (Figure S20A) indicated that T cells in reconstructed models retained cytotoxic potential. PD-1 mAb enhanced T cell activity but did not achieve complete tumor suppression (Figures 4F and 4G). Although PD-1 expression on CD8^+^ T cells decreased post-treatment, a subset remained PD-1 positive (Figures 4H and S20B). Perforin and granzyme B expression in CD8^+^ T cells showed a modest increase after PD-1 mAb treatment (non-significant for Granzyme B), yet levels remained below those in interleukin-2-stimulated models, indicating partial T cell exhaustion and the need for complementary strategies to fully restore anti-tumor function (Figures S20C and S20D). Together, these findings demonstrate successful reconstruction of T cell infiltration in PDX-derived iOS models and establish iOS + PBMC systems as a functional platform for immuno-oncology studies in OS.

Consistent drug sensitivity in T cell-reconstituted iOS models

To evaluate whether T cell-reconstituted iOS models retain the drug sensitivity profiles of their matched clinical samples, we implemented the following workflow. After OS diagnosis, PDX models were established from resected tumors in operable cases. After patients completed standard chemotherapy, the corresponding biobank PDXs were used to establish iOS models, which were subsequently reconstituted with patient-derived PBMCs. These T cell-reconstituted iOS models were treated with doxorubicin (DOX), and therapeutic responses were assessed by cell viability assays and bright-field imaging. Consistency with clinical outcome was retrospectively evaluated by comparing in vitro responses with longitudinal chest computed tomographic (CT) scans from the corresponding patients to monitor metastatic progression or stability (Figure 5A).Figure 5. Clinical response validation of T cell-reconstituted iOS models(A) Workflow for clinical validation of T cell-reconstituted iOS models. After osteosarcoma diagnosis, PDX models were established from resectable tumors. For patients with unresectable tumors, PDX models were created directly from biopsy samples. Following standard chemotherapy treatment, corresponding PDX models from the biobank are selected to generate iOS models, which are then reconstituted with PBMCs. These iOS models are treated with doxorubicin, and therapeutic efficacy is assessed through cell viability and bright-field imaging. Retrospective evaluation of treatment response consistency is conducted by comparing the effects in iOS models with patient lung CT follow-up images to assess pulmonary metastasis progression or stability.(B and C) Cell viability of T cell-reconstituted iOS models treated with doxorubicin, with retrospective clinical follow-up indicated; non-reconstituted models served as controls. Color-coded circles represent different batches, gray indicates censored data, C: Control. n = 3 biological replicates for each group.(D) Bright-field imaging of T cell-reconstituted iOS models treated with doxorubicin and patient retrospective clinical radiological images, using non-reconstituted models as controls. Scale bar, 250 μm. nMeta, no pulmonary metastasis; SD, stable disease; PD, progressive disease. n = 3 technical replicates for each group.

We first examined the effects of graded DOX concentrations on iOS models derived from patients with distinct clinical statuses: no metastasis (nMeta), stable disease (SD), and progressive disease (PD). Dose-response profiles differed across models. T cell-reconstituted iOS models showed significantly altered treatment responses compared with non-reconstituted controls, underscoring the contribution of immune components to chemosensitivity. Most iOS models from nMeta and SD patients (e.g., iOS_166+PBMC and iOS_072+PBMC) exhibited pronounced DOX sensitivity, with steep declines in cell viability at higher drug concentrations (Figures 5B and 5C, left and middle panels). By contrast, models derived from PD patients (e.g., iOS_4235+PBMC and iOS_111+PBMC) displayed relative resistance, maintaining higher viability even at the maximum DOX dose (Figures 5B and 5C, right panel). Bright-field imaging corroborated these findings, revealing marked morphological changes in organoids following DOX treatment. iOS models from nMeta/SD patients showed substantial reductions in organoid size and structural integrity, whereas PD-derived models exhibited minimal morphological alterations, consistent with reduced drug sensitivity (Figures 5D and S21). This imaging-based readout provides a straightforward and scalable metric for in vitro drug response assessment, supporting its application in high-throughput screening. Comparison of these in vitro responses with clinical follow-up data further validated the model: OS patient CT scans during treatment follow-up demonstrated typical patterns of tumor control or progression, and nMeta-derived T cell-reconstituted iOS models closely mirrored the clinical response to DOX, whereas PD-derived models reflected ongoing tumor progression and drug resistance (Figures 5D and S21).

Overall, in most cases, T cell-reconstituted iOS models faithfully recapitulated the clinical drug sensitivity of their matched patients, supporting their potential utility for personalized response prediction. These results also indicate that the iOS models in our biobank, derived from patients in the SGH-OS cohort,28 capture the molecular and clinical characteristics of their source tumors. In future studies, they may, therefore, serve as representative models for patients with similar molecular and clinical profiles, enhancing the relevance of the iOS biobank for personalized medicine. Nonetheless, some discrepancies were observed. For example, iOS_072—derived from an nMeta patient who had completed standard postoperative chemotherapy—exhibited chemoresistance in vitro and showed low pharmacogenomic concordance with its paired tumor (Figure S14), suggesting that non-genomic factors such as microenvironmental dynamics or epigenetic regulation may contribute to discordant drug responses. In addition, the intrinsic genomic instability of OS may limit the temporal stability of organoid models: iOS_051, for instance, initially reflected an SD state but later aligned more closely with PD. In summary, T cell-reconstituted iOS models generally exhibit robust tumor cell killing and, when supported by pharmacogenomic correlation, reliably emulate the drug responsiveness of corresponding clinical tissues. These exceptions underscore the importance of integrating multi-level profiling and longitudinal monitoring to maximize the reliability and clinical applicability of organoid-based platforms.

Utilization of Chr9p21.3-deleted PDX-organoid model for personalized immunotherapy

As described above, treatment of iOS models with PD-1 mAb alone yielded suboptimal effects, with a subset of T cells remaining in an exhausted state. These findings underscore the need for personalized immunotherapeutic strategies and highlight the importance of combinatorial approaches to enhance the efficacy of immune checkpoint blockade in OS. Our previous work identified chromosome 9p21.3 deletion (Chr9p21.3 deletion) as a genomic alteration associated with poor response to immunotherapy in OS.29 Chr9p21.3 deletion results in the co-deletion of two key genes, CDKN2A and MTAP (Figure 6A), and is associated with reduced infiltration of CD8^+^ T cells and cytotoxic T cells compared to Chr9p21.3 wild-type tumors (Figure 6B), likely contributing to the observed resistance to immunotherapy in Chr9p21.3-deleted OS. However, the loss of MTAP due to Chr9p21.3 deletion creates an individualized vulnerability, as it induces synthetic lethality with PRMT5 inhibition. Specifically, MTAP deficiency leads to the intracellular accumulation of methylthioadenosine (MTA), which binds to PRMT5 to form PRMT5^MTA^ complexes and inhibiting its enzymatic activity.30^,^31 Although PRMT5 inhibitors exhibit preclinical efficacy in Chr9p21.3-deleted tumors, their clinical use is hampered by systemic toxicity. Recent efforts have yielded selective PRMT5^MTA^ complex inhibitors with improved safety. Here, we introduce APRN2169, a selective inhibitor of PRMT5^MTA^ complexes (Figures 6C and S22A–S22C). We validated its preclinical efficacy using Chr9p21.3-deleted iOS_111 models derived from our PDX-organoid interactive biobank. In PDX models, high-dose APRN2169 significantly reduced tumor volume (Figure 6D) with minimal observed toxicity (Figure S22D) and effectively inhibited tumor cell proliferation (Figure S22E). To further mitigate potential toxicity and assess compatibility with immunotherapy, we tested low-dose APRN2169 together with PD-1 mAb in DuNN MTAP^−/−29^ CDX models. The combination significantly suppressed tumor growth while maintaining a favorable safety profile. H&E and IHC staining revealed a significant reduction in MKI67 expression and an increase in CD8^+^ T cell infiltration, further supporting the efficacy of the combination therapy (Figures 6E and S23). In T cell-reconstituted iOS_111 models, the APRN2169 and PD-1 mAb combination led to a pronounced reduction in organoid size and viability (Figures 6F and 6G). mIF and flow cytometry analyses confirmed increased infiltration of CD8^+^ T cells and reduced expression of PD-1 (Figures 6H, 6I, and 6J). Physical contact between tumor cells and CD8^+^ T cells was significantly increased (Figure S24A), and CD8^+^ T cells showed robust upregulation of granzyme B and perforin expression (Figures 6K, 6L, and S24B), indicative of restored T cell cytotoxic activity. Together, these findings demonstrate that our PDX-organoid interactive biobank provides a powerful platform for drug discovery and translational validation. The combination of the PRMT5^MTA^ inhibitor APRN2169 with PD-1 blockade represents a promising personalized therapeutic strategy for Chr9p21.3-deleted OS, simultaneously exploiting tumor-specific genetic vulnerabilities and overcoming the limitations of single-agent immunotherapy.Figure 6. Utilization of characteristic Chr9p21.3-deleted PDX-iOS models for personalized immunotherapy development(A) Osteosarcoma-associated genes, CDKN2A, and MTAP, located at 9p21.3 locus on Homo sapiens chromosome 9 (Chr9p21.3).(B) MCP-counter analysis performed to quantify cell proportions in SGH-OS. Relative cell abundance was represented by a color gradient.(C) Mechanistic basis of PRMT5^MTA^ inhibitor (APRN2169) efficacy. MTA and APRN2169 occupy the pocket simultaneously and a “LOCKTAC” mechanism stabilizes the binding to the drug to the PRMT5^MTA^, thereby inhibiting PRMT5 activity.(D) In vivo experiment of Chr9p21.3-deleted PDX models, treated with APRN2169 and compared with doxorubicin (DOX). Data represented as mean ± SEM. n = 5/group. Two-tailed Student’s t test was performed for statistical analysis.(E) In vivo experiment of MTAP^−/−^ DuNN models, treated with APRN2169 and PD-1 mAb combination. Data represented as mean ± SEM. n = 5/group. Two-tailed Student’s t test was performed for statistical analysis.(F and G) Bright-field imaging and cell viability of Chr9p21.3-deleted T cell-reconstituted iOS models treated with APRN2169 and PD-1 mAb combination. Scale bars, 250 μm. Data represented as mean ± SEM. Statistical significance was assessed by one-way ANOVA. n = 12 technical replicates for each group.(H) mIF of Chr9p21.3-deleted T cell-reconstituted iOS models treated with APRN2169 and PD-1 mAb combination. Scale bars, 100 μm. n = 3 technical replicates for each group.(I) Flow cytometric analysis of CD8^+^ T cell proportions in Chr9p21.3-deleted T cell-reconstituted iOS models. n = 3 technical replicates for each group. Data represented as mean ± SEM. Statistical significance was assessed by one-way ANOVA (ns, no significance).(J) Flow cytometric analysis of PD-1 expression in CD8^+^ T cells in Chr9p21.3-deleted T cell-reconstituted iOS models. Data represented as mean ± SEM. Statistical significance was assessed by one-way ANOVA. MFI, mean fluorescence intensity. n = 3 technical replicates for each group.(K and L) Flow cytometric analysis of granzyme B and perforin expression in CD8^+^ T cell from Chr9p21.3-deleted T cell-reconstituted iOS models. n = 3 technical replicates for each group. Data represented as mean ± SEM. Statistical significance was assessed by one-way ANOVA (ns, no significance).

Discussion

The interactive PDX-organoid biobank described here provides an OS-centered model system that is both scalable and deeply annotated. Existing sarcoma organoid collections have established the feasibility of patient-derived tumor organoids (PDTOs) across multiple bone and soft tissue sarcomas, but they are not OS-specific and rarely integrate systematically matched multi-omics datasets for molecular validation.26 A recent OS study reported good concordance between normal tissue, tumor, and organoid transcriptomes based on three patients,32 but the field still lacks large, genomically annotated OS organoid cohorts to robustly establish model validity and to support precision medicine efforts.

Our interactive PDX-organoid biobank focuses on OS and is anchored to the SGH-OS PDX cohort, enabling direct linkage of histology, immunohistochemistry, genomics, transcriptomics, and pharmacogenomics between iOS models and their matched patient or PDX tissues. We show that iOS models preserve key architectural, phenotypic, and molecular features of the original tumors, that iOS-derived xenografts retain in vivo tumorigenicity, and that T cell-reconstituted iOS models can be used to interrogate chemotherapy and immunotherapy responses in genomically defined contexts such as Chr9p21.3 deletion.

The efficiency and reproducibility of the iOS platform are among its most distinctive features. Functionally mature organoids can be generated within 7–10 days, yet they retain high concordance with matched patient/PDX tissues at the levels of CNVs, transcriptome, pathway activity, and drug-response signatures. This rapid and standardized workflow is particularly important in OS, which is often aggressive and associated with narrow therapeutic windows. Although PDX models preserve patient-specific genetic alterations, their protracted establishment time substantially limits their use for real-time clinical guidance.33^,^34 By enabling rapid, reproducible modeling of patient-specific responses—such as doxorubicin sensitivity—in a time frame compatible with treatment planning, iOS provides a practical and clinically actionable platform.

Recurrent and metastatic OS patients face poor prognoses and limited treatment options. While immunotherapy has succeeded in other solid tumors, it has fallen short in OS,7^,^8 in part due to low TIL abundance and a poorly understood immune microenvironment.9 Existing OS immunology models (cell line-derived xenografts, explants, and generic 3D co-cultures)35^,^36 have yielded mechanistic insights but are often not patient specific and only partially capture the complexity of the human OS immune microenvironment. In this context, our OS-focused biobank systematically integrated with a large, genomically and clinically annotated PDX resource and provides a patient-tailored platform for linking genotype, phenotype, and treatment response at scale. iOS models can be stably passaged with a short culture cycle, and millimeter-scale organoids recapitulate key features of the OS microenvironment, including reduced T cell infiltration and emerging hypoxia, thereby enhancing the relevance of immune-response studies.

Our study provides several conceptual and technical refinements over previously reported OS organoid and immune co-culture systems. First, unlike mixed sarcoma PDTO collections,26 our resource is OS-focused and anchored in the large OS clinical cohort and corresponding PDX biobank, with systematically paired histology, transcriptomics, genomics, and pharmacogenomics. Second, the stable passaging and rapid cultivation cycle of iOS models make them practical and efficient for large-scale, reproducible studies. Third, we show that increasing organoid size is associated with progressively reduced immune cell infiltration, indicating that larger OS organoids more faithfully reproduce the spatial organization and diffusion constraints of the native tumor microenvironment and thus better model immune-tumor interactions. The extensive integration creates a comprehensive platform for personalized immunotherapy research and drug development. An additional advance is the reconstruction of the OS immune microenvironment by incorporating PBMCs into PDX-derived iOS models. This approach allows tumor cells to shape the activation and exhaustion status of T cells, enabling our models to mimic T cell exhaustion within the OS TME and offering mechanistic insight into immune evasion and therapeutic resistance. Because the immune compartment in reconstructed iOS models is induced by tumor-PBMC interactions, it more closely reflects how tissue-associated immune cells in native OS are reprogrammed into tumor-associated phenotypes. This makes the platform particularly valuable for evaluating immunotherapies, including checkpoint blockade and engineered cell therapies such as CAR-T, and for optimizing their efficacy in OS.

Immunogenomic features such as Chr9p21.3 deletion, which are relatively common in OS,28^,^37 reduce TIL abundance and contribute to resistance to immunotherapy.38^,^39 Using Chr9p21.3-deleted PDX-organoid models, we demonstrated that the PRMT5^MTA^ inhibitor APRN2169, in combination with PD-1 mAb, enhances CD8^+^ T cell infiltration, restores cytotoxic effector functions, and achieves improved tumor control. Notably, APRN2169 derivatives have progressed to phase 1 clinical trials, underscoring the translational potential of this strategy and supporting PRMT5^MTA^ inhibition plus PD-1 blockade as a promising personalized immunotherapy approach for Chr9p21.3-deleted OS.

In summary, our iOS models recapitulate patient-derived heterogeneity and key biological, immunological, and genomic features of OS. Built on a large, multi-omics-annotated PDX-organoid biobank, this platform enables clinical drug testing and the development of genotype- and immune-informed therapeutic strategies to support precision oncology in OS. At the same time, careful attention to temporal constraints, culture-associated genomic drift, and the complexity of immune-tumor-stroma interactions will be essential for responsible use in research and clinical settings. Because PBMC donor identity can influence immune receptor diversity (e.g., human leukocyte antigen [HLA]/T cell receptor [TCR] composition) and baseline immune activation, our immune-reconstituted organoid readouts may vary between donors. While we used a single donor to reduce variability in the present study, future studies will test multiple donors and, where feasible, consider donor-tumor HLA context to better define the applicability of this platform.

Limitations of the study

Our data indicate that iOS models best recapitulate patient-specific treatment responses within a limited temporal window. For example, iOS_051 closely mirrored the patient’s early clinical response, yet the same patient experienced disease progression approximately 6 months later. This observation highlights that OS is a dynamically evolving disease and underscores that a single organoid snapshot cannot indefinitely represent the clinical trajectory. Thus, prospective clinical application will likely require longitudinal sampling and repeated model generation to track tumor evolution over time.

OS is characterized by marked chromosomal instability, and our genomic analyses revealed the emergence of additional somatic variants in a subset of models over serial passages. Some additional SNVs may reflect ongoing tumor evolution resembling in vivo dynamics, whereas others may result from selection of clones better adapted to in vitro conditions or from culture-induced DNA damage.40^,^41^,^42 These observations suggest that long-term culture can introduce genetic drift, emphasizing the need to prioritize low-passage, genomically benchmarked iOS models—particularly when they are used for clinical decision support or mechanistic studies where genomic fidelity is critical.

Several aspects of the iOS model still require refinement. The successful establishment of PDX models in immunodeficient mice is facilitated by the absence of functional tumor-killing T cells, although stromal components such as macrophages and fibroblasts remain active contributors to tumor formation.43 This rationale informed our initial focus on reconstructing lymphocyte populations in iOS models. While we have assessed CD8^+^ T cell functionality, comprehensive characterization of additional immune and stromal cell types remains an important goal. Single-cell RNA sequencing has already revealed substantial divergence in tumor cell subpopulations between PBMC-reconstituted and conventional iOS models, and preliminary analyses point to several candidate signaling pathways involved. Future work will be needed to delineate how specific immune-tumor interactions shape tumor cell differentiation and to integrate a broader spectrum of microenvironmental components into the iOS platform.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Yingqi Hua (yhua@shsmu.edu.cn).

Materials availability

ARPN2169 generated in this study are available from the lead contact with a completed materials transfer agreement.

Data and code availability

• •Data: Single-cell RNA-seq, bulk RNA-seq and WGS data have been deposited at National Genomics Data Center (NGDC) and are publicly available. Accession numbers are listed in the key resources table.
• •Code: This study did not result in any development of original code.
• •Any additional information about this work paper is available from the lead contact upon request.

Acknowledgments

We are deeply grateful to Professor Bing Li (Shanghai Jiao Tong University School of Medicine, China) for his help that greatly improved this work. We are deeply grateful to Jun Zhang (SeekGene, China), Yinghua Gao (Shanghai General Hospital), and Chen Chen (Benagen, China) for their suggestions and great help in improving this work. This study was supported by the National Natural Science Foundation of China (82473491, 32500565, 82272773, 82373177, 82404064, 82203043, and 82172366), STCSM Shanghai Natural Science Grants (23ZR1451200 and 25ZR1402344), Physician-Scientist, Shanghai Jiao Tong University School of Medicine (20240811 and 20240812), and Clinical Cohort Shanghai (no. SHDC2025CCS010).

Author contributions

Conceptualization, Y.H., W.S., L.Y., H.M., and J.W.; methodology, W.S., Y.T., W.S., Z.Q., X.H., X.Y., H.D., H.W., L.Z., W.C., B.Y., H.M., and Y.G.; investigation, H.M., L.Y., and Y.H.; visualization, Y.T., H.M., J.W., Z.Q., and H.W.; funding acquisition, L.Y., Z.C., H.M., W.S., and Y.G.; project administration, Y.H., W.S., M.L., and Z.C.; supervision, Z.C., M.L., and W.M.; writing – original draft, W.S., H.M., L.Y., and Y.T.; writing – review & editing, Y.H., H.M., L.Y., and J.W.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCESOURCEIDENTIFIERAntibodiesSATB2 Polyclonal AntibodyInvitrogenCat#PA5-83092; RRID:AB_2790248COL1A1 Polyclonal AntibodyInvitrogenCat#PA5-29569; RRID:AB_2547045anti-c-caspase3 antibodyAbcamCat#ab32042; RRID:AB_725947anti-HIF-1α antibodyAbcamCat#ab51608; RRID:AB_880418Goat anti-Rabbit IgG (H + L) Cross-Adsorbed Secondary Antibody, Alexa Fluor™ 594InvitrogenCat#A-11012; RRID:AB_141359Goat anti-Rabbit IgG (H + L) Cross-Adsorbed Secondary Antibody, Alexa Fluor™ 594InvitrogenCat#A-11005; RRID:AB_141372anti-SMA antibodyInvitrogenCat#MA1-06110; RRID:AB_557419anti-CD4 antibodyAbcamCat#ab183685; RRID:AB_2686917anti-CD8 antibodyAbcamCat#ab237709; RRID:AB_2892677anti-PD-1 antibodyInvitrogenCat#14-2798-82; RRID:AB_2572870anti-TIM3 antibodyInvitrogenCat#MA5-32841; RRID:AB_2802488anti-CD206 antibodyInvitrogenCat#PA5-101657; RRID:AB_2851091anti-FOXP3 antibodyInvitrogenCat#41-5773-82; RRID:AB_11219073Goat anti-Mouse IgG (H + L) Cross-Adsorbed Secondary Antibody, Alexa Fluor™ 488InvitrogenCat#A-11001; RRID:AB_2534069Goat anti-Rabbit IgG (H + L) Cross-Adsorbed Secondary Antibody, Alexa Fluor™ 488InvitrogenCat#A-11008; RRID:AB_143165Goat anti-Rabbit IgG (H + L) Cross-Adsorbed Secondary Antibody, Cyanine3InvitrogenCat#A10520; RRID:AB_2770568Goat anti-Mouse IgG (H + L) Cross-Adsorbed Secondary Antibody, Cyanine3InvitrogenCat#A10521; RRID:AB_1500665Goat anti-Rabbit IgG (H + L) Cross-Adsorbed Secondary Antibody, Cyanine5InvitrogenCat#A10523; RRID:AB_2534032Goat anti-Mouse IgG (H + L) Cross-Adsorbed Secondary Antibody, Cyanine5InvitrogenCat#A10524; RRID:AB_10562712PerCP anti-human CD45BioLegendCat#304025; RRID:AB_893341FITC anti-human CD3BioLegendCat#317305; RRID:AB_571906APC anti-human CD4BioLegendCat#317415; RRID:AB_571944PE anti-human CD8BioLegendCat#344705; RRID:AB_1953243BV605 anti-human PD-1BioLegendCat#367425; RRID:AB_2721544Pacific Blue anti-human CD45BioLegendCat#304021; RRID:AB_493654APC anti-human PerforinBioLegendCat#308111; RRID:AB_830872PerCP anti-human Granzyme BBioLegendCat#396415; RRID:AB_2924597Chemicals, peptides, and recombinant proteinsDMEM high GlucoseGibcoCat#11965092DMEM/F-12GibcoCat#11320033McCoy’s 5a mediumGibcoCat#16600082Fetal bovine serumWisentCat#080-150Phophate-Buffered SalineThermo Fisher ScientificCat#J61196.APPhysiological salineShanghai yuanye Bio-Technology Co., LtdCat#R21479Red Blood Cell Lysis SolutionThermo Fisher ScientificCat#00-4333Trypsin-EDTAThermo Fisher ScientificCat#15400054DNase ISigma-AldrichCat#EN0521Triton X-100Sangon BiotechCat#A600198DAPIBeyotimeCat#C1002Trypan blueThermo Fisher ScientificCat#15250061EDTA (0.5 M), pH 8.0, RNase-freeInvitrogenCat#AM9260GBovine Serum Albumin (BSA)Sangon BiotechCat#A500023-0100Nylon cell strainer (40μm)CorningCat#352340Glass bottom culture dishesNESTCat#801001MatrigelCorningCat#356231100mm Culture DishesCorningCat#43016760mm Culture DishesCorningCat#4301666-well platesCorningCat#351696-well platesCorningCat#3798U-bottom 96-well platesCorningCat#7007anti-human PD-1 monoclonal antibodyShlleckCat#A2002ARPN2169APEIRONThis studyFACS Staining BufferInvitrogenCat#00-4222-26Blocking bufferInvitrogenCat#B001T06F01Cell Activation CocktailBioLegendCat#423303GlutaMAX™ supplementGibcoCat#35050061N-2 supplementGibcoCat#17502048MEM Non-Essential Amino AcidsGibcoCat#11140050B-27™ supplementGibcoCat#17504044Insulin, Transferrin, Selenium Solution (ITS-G)GibcoCat#41400045Ascorbic acidMerckCat#200-066-2ROCK inhibitor Y-27632MerckCat#10035-04-8Human recombinant EGF proteinPeproTechCat#AF-100-15Human recombinant bFGF proteinMedChemExpressCat#HY-P7331AHuman recombinant IGF-1 proteinMedChemExpressCat#HY-P70783Human recombinant TGF-β3 proteinMedChemExpressCat#HY-P7120AdenineMedChemExpressCat#HY-B0152ALK inhibitor A 83-01MedChemExpressCat#HY-10432p38 MAPK inhibitor SB 202190MedChemExpressCat#HY-10295DexamethasoneMedChemExpressCat#HY-14648Human recombinant Wnt3a proteinMedChemExpressCat#HY-P70453ASerum-free cell cryopreservation solutionNCM BiotechCat#C40100Anti-human CD3 antibodyBioLegendCat#300306CD28 Monoclonal AntibodyInvitrogenCat#16-0289-85Human Recombinant IL-2PeproTechCat#200-02Human IgG1 isotype controlShlleckCat#A2051Doxorubicin hydrochlorideMedChemExpressCat#HY-15142AgaroseServicebioCat#G5056-5GCellTiter-Lumi™ Luminescent Cell Viability Assay KitBeyotimeCat#C0065SNuclease-free ATPBeyotimeCat#D7378CTAB lysis bufferPromegaCat#MC1411TE bufferThermo Fisher ScientificCat#12090015Agencourt AMPure XP-Medium kitBeckmanCat#A63880Critical commercial assaysChromium Next GEM Single Cell 3ʹ Kit v3.110× GenomicsCat#PN-1000268RNeasy Mini KitQIAGENCat#74104SeekOne DD Single Cell 3′ Transcriptome kitSeekGeneN/ARNA Nano 6000 Assay KitAgilentCat#5067-1511NEBNext® Ultra™ RNA Library Prep Kit for Illumina®NEBCat#E7530Qubit_TM_ dsDNA HS Assay KitThermo Fisher ScientificCat#Q32851Cell Counting Kit-8DojindoCat#CK04Deposited dataWhole genome sequencing (WGS) data of iOSNational Genomics Data Center (This study)HRA008992SGH-OS multi-omics datasetNational Genomics Data CenterHRA003260TCGA-SARC datasetGenomic Data Commons Data Protalhttps://portal.gdc.cancer.gov/TARGET-OS datasetGenomic Data Commons Data Protalhttps://portal.gdc.cancer.gov/Single-cell RNA-seq of OsteosarcomaNational Genomics Data Center - OMIX (This study)OMIX013436Experimental models: Cell linesDuNNDunn LabRRID: CVCL_W629DuNN MTAP^−/−^Shanghai General HospitalN/ASU-DHL-4APEIRONRRID: CVCL_0539AsPC-1APEIRONRRID: CVCL_0152Calu-1APEIRONRRID: CVCL_0608SUDHL-10APEIRONRRID: CVCL_1889CFPAC-1APEIRONRRID: CVCL_1119U87MGAPEIRONRRID: CVCL_D7C9SW780APEIRONRRID: CVCL_1728MDA-MB-231APEIRONRRID: CVCL_0062BxPC-3APEIRONRRID: CVCL_0186SW1573APEIRONRRID: CVCL_1720NCI-H647APEIRONRRID: CVCL_1574MKN-45APEIRONRRID: CVCL_0434LU99APEIRONRRID: CVCL_LU99HT-1080APEIRONRRID: CVCL_0317PSN-1APEIRONRRID: CVCL_1644DOHH-2APEIRONRRID: CVCL_1179A549APEIRONRRID: CVCL_A549LN-18APEIRONRRID: CVCL_0392ACHNAPEIRONRRID: CVCL_1067Experimental models: Organisms/strainsC3H/HeNCrlCharlers RiverRRID: IMSR_CRL:025BALB/cJGpt-Foxn1^nu^/GptCharlers RiverRRID: IMSR_GPT: D000521Software and algorithmsGraphPad Prism (version 8)GraphPadRRID:SCR_002798ImageJImageJRRID:SCR_003070R software (version 4.3.1)The R FoundationRRID:SCR_001905RStudioPositRRID:SCR_000432Leica Application Suite X (LAS X)LeicaRRID:SCR_013673FlowJo (version 10)FlowJoRRID:SCR_008520GRCh38 genomeNIHHomo sapiens genome assembly GRCh38 - NCBI - NLM (nih.gov)mm10 genomeNIHMus musculus genome assembly GRCm38 - NCBI - NLM (nih.gov)FastQC (version 0.11.9)Babraham BioinformaticsRRID:SCR_014583Cell Ranger (version 3.0.1)10× GenomicsRRID:SCR_017344Harmony packageKorsunsky et al.44RRID:SCR_022206clusterProfiler (version 3.14.0)Yu et al.45RRID:SCR_016884DoubletFinder package (version 2.0.2)McGinnis et al.46RRID:SCR_018771inferCNV package (version 1.12.0)Tickle et al.47N/AbiomaRt package (version 2.42.1)Durinck et al.48RRID:SCR_019214STAR (version 2.7.10a)Dobin et al.49RRID:SCR_004463subread (version 2.0.1)http://subread.sourceforge.net/RRID:SCR_009803SAMTOOLS (version v1.16.1)http://htslib.org/RRID:SCR_002105Gene Set Enrichment Analyiss (GSEA, version 4.1.0)Subramanian et al.50RRID:SCR_003199BioRenderBioRenderRRID:SCR_018361CaseViewer (version 2.4)3DHISTECH Ltd.RRID:SCR_017654DESeq2 (version v1.44.0)https://bioconductor.org/packages/release/bioc/html/DESeq2.htmlRRID:SCR_015687Monocle (version 2.28.0)http://cole-trapnell-lab.github.io/monocle-release/docs/RRID:SCR_016339GATK (version 4.2.1.0)DePristo et al.51RRID:SCR_001876OtherIllumina NovaSeq 6000 Sequencing SystemIlluminaRRID:SCR_016387Agilent 2100 Bioanalyzer InstrumentAgilent TechnologiesRRID:SCR_018043Invitrogen Countess II Automated Cell CounterInvitrogenRRID:SCR_025370Leica SP8 LIGHTNING confocal microscopeLeicaRRID:SCR_018169SpectraMax M3 Microplate ReaderMolecular DevicesCat#14327BD LSRFortessa Cell AnalyzerBD BiosciencesRRID:SCR_018655Fresco™ 17 MicrocentrifugeThermo Fisher ScientificCat#75002402

Experimental model and study participant details

Patient information and human specimens

Our study received ethical approval from the Institutional Research Ethics Committee of Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine (Approval No. 2021KY103), Shanghai, China. Specimens, including patient tissues and patient-derived xenografts (PDX), were obtained from newly diagnosed osteosarcoma (OS) patients at Shanghai General Hospital, Shanghai, China, with written informed consent provided by all participants. Comprehensive clinical information associated with these specimens was collected and analyzed (Table S2). Pathological diagnoses for all OS patients enrolled in this study were independently confirmed by three experienced pathologists at Shanghai General Hospital to ensure diagnostic accuracy and consistency.

Animals

All animal experiments were conducted in compliance with the guidelines provided by the Laboratory Animal Center of Shanghai General Hospital. The Clinical Center Laboratory Animal Welfare & Ethics Committee of Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine, approved all animal study protocols used in this study (IACUC: 2023AW067). The establishment and passaging of PDX models were performed following the protocols described in our previous study.28 Briefly, approximately 100 mg of tumor tissue was placed in a 15 mL polypropylene tube containing serum-free DMEM and transported to the laboratory on wet ice. After thorough washing and sectioning, tumor fragments measuring 3–5 mm were implanted into the flanks of NOD-scid IL2Rγnull (NSG) mice. Tumor size and body weight were monitored twice weekly. Tumor volume was determined using the formula: Volume = Length × Width^2^ × 0.5 mm^3^, with measurements taken using a vernier caliper for precision. All animal experimental procedures were conducted in strict compliance with ethical guidelines and were approved by the Animal Care and Use Committee of Shanghai General Hospital.

In cases where models were treated with doxorubicin (DOX) or the PRMT5^MTA^ inhibitor ARPN2169 (this study, APEIRON, US012173002B), DOX was administered via intraperitoneal injection at a dose of 4 mg/kg once daily (QD), while ARPN2169 was administered orally at doses of 30 or 60 mg/kg once daily (QD). N = 5 for each group.

To establish Orthotopic tibia mouse xenografts, this study utilized 6–8-week-old female C3H mice (Strain: C3H/HeNCrl; Charles River, USA/China). DuNN MTAP^−/−^ cells and its culture conditions were established as described in our previous study.29 All mice were housed under specific pathogen-free (SPF) conditions in a controlled environment maintained at 20°C–22°C, with a 12/12-h light/dark cycle and 50–70% humidity. All animal experiments were conducted in strict adherence to the protocols and guidelines of the Laboratory Animal Center of Shanghai General Hospital. The Clinical Center Laboratory Animal Welfare & Ethics Committee of Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine, approved all animal protocols used in this study.

Orthotopic tibia mouse xenografts were established via intramedullary injection of OS cells (5 × 10ˆ5 cells in 25 μL PBS supplemented with 10% FBS) into the marrow space of the proximal tibia of C3H mice using a 27-gauge needle, as detailed in our previous studies.52^,^53 Tumor volume was determined using the formula: Volume = Length × Width^2^ × 0.5 mm^3^, with measurements taken using a vernier caliper for precision. Tumor burden was assessed by tracking tumor volume, with a maximum permitted tumor size of 2 cm, as authorized by the Clinical Center Laboratory Animal Welfare & Ethics Committee of Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine.

In cases where models were treated with anti-human PD-1 monoclonal antibody (mAb) (#A2002, Shlleck) or the PRMT5/MTA inhibitor ARPN2169 (this study, APEIRON), the PD-1 mAb was administered at a dose of 5 mg/kg on days 5, 7, 9, and 13 post-model establishment. ARPN2169 was administered orally at a dose of 1 mg/kg once daily (QD). N = 5 for each group.

Cell lines

The mouse OS cell lines DuNN MTAP^−/−29^ were cultured in high glucose Dulbecco’s Modified Eagle Medium (DMEM; Gibco, USA) supplemented with 10% fetal bovine serum (FBS; Wisent, Canada). Human cell lines SU-DHL-4, AsPC-1, Calu-1, SU-DHL-10, CFPAC-1, U87MG, SW780, MDA-MB-231, BxPC-3, SW1573, NCI-H647, MKN-45, LU99, HT-1080, PSN-1, DOHH-2, A549, LN-18, and ACHN were provided by APEIRON. SU-DHL-4, SU-DHL-10, DOHH-2, AsPC-1, BxPC-3, NCI-H647, PSN-1, MKN-45, and LU99 cells were cultured in RPMI-1640 medium (Gibco, USA) supplemented with 10% fetal bovine serum (FBS; Gibco, USA) at 37°C in a humidified incubator with 5% CO_2_. Calu-1 cells were maintained in McCoy’s 5A (modified) medium (Gibco, USA) supplemented with 10% FBS at 37°C with 5% CO_2_. CFPAC-1 cells were cultured in Iscove’s Modified Dulbecco’s Medium (IMDM; Gibco, USA) supplemented with 10% FBS at 37°C with 5% CO_2_. U87MG, HT-1080, and ACHN cells were maintained in Eagle’s Minimum Essential Medium (EMEM; Gibco, USA) supplemented with 10% FBS at 37°C with 5% CO_2_. LN-18 cells were cultured in Dulbecco’s Modified Eagle Medium (DMEM; Gibco, USA) supplemented with 5% FBS at 37°C with 5% CO_2_. A549 cells were cultured in F-12K medium (Kaighn’s modification of Ham’s F-12; Gibco, USA) supplemented with 10% FBS at 37°C with 5% CO_2_. SW780, SW1573, and MDA-MB-231 cells were maintained in Leibovitz’s L-15 medium (Gibco, USA) supplemented with 10% FBS at 37°C in a CO_2_-free incubator. Cell line authentication was performed on cells that used for in vitro and in vivo studies using Short Tandem Repeat (STR) DNA profiling and all cell lines were preserved at Shanghai Bone Tumor Institute (Shanghai, China). All cell lines were routinely tested for Mycoplasma monthly using the Mycoplasma Detection Kit (catalog no. rep-mys-20, InvivoGen) and used within 10 passages.

Method details

Organoid storage and washing solution preparation

The storage solution was prepared by combining the following components to a final volume of 50 mL: fetal bovine serum (FBS) (2%, #10099141C, Gibco), penicillin-streptomycin solution (1%, #15140122, Gibco), and HEPES (10 mmol/L, #15630080, Gibco). The mixture was then adjusted to 50 mL using HBSS (#14170161, Gibco). After preparation, the storage solution was aliquoted into 15 mL sterile centrifuge tubes (#430053, Corning), with 5 mL per tube, and stored at 4°C. The storage solution remained stable for up to 1 month.

The washing solution was prepared by adding penicillin-streptomycin solution (1%) to PBS (#21-040-CVR, Gibco) and adjusting the final volume to 50 mL. The washing solution was prepared immediately prior to use to ensure optimal efficacy.

Organoid dissociation and dissociation termination solution preparation

The dissociation solution was prepared by combining the following components to a final volume of 5 mL: collagenase II (125 CDU/mL, #V900892, Sigma-Aldrich), DNase I (40 Kunitz units/mL, #DN25, Sigma-Aldrich), penicillin-streptomycin solution (0.1%), and FBS (5%). The mixture was then adjusted to 5 mL using RPMI 1640 medium (#SH30809.01B, Hyclone). The dissociation solution was prepared immediately prior to use to ensure optimal enzymatic activity. The activity of collagenase II (CDU) was defined as the protease activity required to release 1 μmol of L-leucine from collagen at 37°C and pH 7.5 over a 5-h period. The activity of DNase I (Kunitz unit) was defined as the amount of enzyme that induces an increase in absorbance of 0.001 A260nm/min/mL at 25°C (in 0.1 M NaOAc, pH 5.0) due to the degradation of highly polymerized DNA.

The dissociation termination solution was prepared by combining penicillin-streptomycin solution (1%) and FBS (10%) and adjusting the final volume to 50 mL with Dulbecco’s Modified Eagle Medium F12 (DMEM/F12) (#11320033, ThermoFisher). The dissociation termination solution was stored at 4°C and remained stable for up to 1 month. Both the dissociation solution and dissociation termination solution were prepared by SeekGene Corporation (SeekGene Corporation, China) under standardized protocols to ensure consistency and reproducibility.

Organoid culture medium and cryopreservation solution preparation

The culture medium was formulated by combining the following components to a final volume of 50 mL: penicillin-streptomycin solution (1%), FBS (10%), GlutaMAX supplement (1%, #35050061, Gibco), N-2 supplement (1.2%, #17502048, ThermoFisher), MEM Non-Essential Amino Acids (NEAA) (1.2%, #11140050, ThermoFisher), B-27 supplement (1.2%, #17504044, ThermoFisher), Insulin, Transferrin, Selenium Solution (ITS-G) (1.2%, #41400045, Gibco), ascorbic acid (100 μmol/L, #200-066-2, Merck), ROCK inhibitor Y-27632 (10 μmol/L, #10035-04-8, Merck), human recombinant EGF protein (20 ng/mL, #AF-100-15, PeproTech), human recombinant bFGF protein (40 ng/mL, #HY-P7331A, MedChemExpress), human recombinant IGF-1 protein (20 ng/mL, #HY-P70783, MedChemExpress), human recombinant TGF-β3 protein (10 ng/mL, #HY-P7120, MedChemExpress), adenine (188 μmol/L, #HY-B0152, MedChemExpress), ALK inhibitor A 83-01 (0.5 μmol/L, #HY-10432, MedChemExpress), selective p38 MAPK inhibitor SB 202190 (10 μmol/L, #HY-10295, MedChemExpress), dexamethasone (10ˆ-5 mol/L, #HY-14648, MedChemExpress), and human recombinant Wnt3a protein (500 ng/mL, #HY-P70453A, MedChemExpress). The mixture was adjusted to 50 mL with DMEM and sterilized by filtration through a 0.22 μm syringe filter (#SLGP033RS, Millipore). The prepared culture medium was stored at 4°C and remained stable for up to 3 weeks.

The cryopreservation solution was prepared by adding Y-27632 (10 μmol/L) to serum-free cell cryopreservation solution (#C40100, NCM Biotech) and adjusting the final volume to 10 mL. The cryopreservation solution was prepared immediately prior to use to ensure optimal cell viability during cryopreservation.

Single-cell osteosarcoma patient/PDX-derived organoid suspension preparation

Under sterile conditions, tumor tissue was harvested from osteosarcoma patients or patient-derived xenografts (PDXs). The tissue was immediately preserved at 4°C using storage solution and transported to the laboratory within 2 h to maintain tissue viability. The firmness of the tissue was assessed to determine the appropriate dilution ratio with dissociation solution. The tissue was meticulously cleaned using washing solution to remove non-tumor components, necrotic tissue, blood clots, and other contaminants. Subsequently, the tissue was cut into small fragments in a clean environment using washing solution, ensuring each fragment measured approximately 1.5–3 mm^3^ to minimize the production of bone shards, which could compromise experimental integrity. The tissue fragments were then dissociated using dissociation solution at 37°C for approximately 15 min. The dissociation process was monitored under a microscope until a significant number of single cells were observed. To halt the dissociation, dissociation solution was mixed with at least three times its volume of dissociation termination solution, followed by gentle pipetting to ensure homogeneous mixing. The resulting cell suspension was filtered through a sterile 40 μm cell strainer (#431750, Corning) to remove residual tissue debris. The suspension was centrifuged at 1500 × g to remove the supernatant, and the cell pellet was carefully resuspended in PBS. A second centrifugation step (1500 × g) was performed to eliminate any remaining dissociation termination solution, yielding a purified single-cell suspension ready for downstream applications.

Self-assembly and cultivation of osteosarcoma patient/PDX-derived organoids

The harvested single-cell OS PD(X)O suspension was counted, and the cell concentration was adjusted to 2 × 10ˆ5 cells/50 μL using culture medium at room temperature (∼25°C). Aliquots of 50 μL were dispensed into each well of a 96-well clear round-bottom ultra-low attachment microplate (#7007, Corning). The plate was centrifuged at 500 × g for 3 min at room temperature, carefully removed, and gently placed on the workstation. A self-assembling matrix was prepared by mixing Matrigel (#356231, Corning) with culture medium in a 3:2 ratio at 4°C or on ice. Subsequently, 75 μL of the matrix was gently overlaid onto the centrifuged single-cell suspension in each well under cold conditions. The 96-well plate was then transferred to a cell culture incubator (37°C, 5% CO2) and left undisturbed for 30–60 min to allow thorough mixing and cross-linking of the matrix with the OS PD(X)O cells. After cross-linking was complete, 100–150 μL of culture medium was gently added to each well at room temperature. The plate was returned to the incubator, and the culture medium was replaced every 1–2 days. The growth of OS PD(X)O was monitored using bright-field microscopy until organoid-like structures with diameters of 1000–1500 μm were observed.

Passaging, cryopreservation, and resuscitation of osteosarcoma patient/PDX-derived organoids

For passaging, cold sterile PBS (200 μL per well) was gently added to the OS PD(X)O organoid models to dissolve the Matrigel. The dissolved organoid structures were transferred to a 15 mL centrifuge tube and incubated at 4°C for 15 min. Subsequently, 1–2 mL of sterile PBS was added, and the organoid-like structures were gently mixed by pipetting. The suspension was centrifuged at 1500 × g for 5 min at room temperature, and the supernatant was carefully removed. Next, 1–2 mL of dissociation solution was added, and the mixture was gently pipetted to ensure thorough dissociation. The tube was incubated in a cell culture incubator (37°C, 5% CO2) for 10 min or until the organoid-like structures dissociated into single cells. The digestion reaction was halted by adding dissociation termination solution at three times the volume of dissociation solution used. The single-cell suspension was collected by centrifugation at 1500 × g for 5 min at room temperature. The cell pellet was resuspended in dissociation termination solution, and the OS PD(X)O organoids were cultured for 2–3 generations to achieve expansion.

For cryopreservation, cold sterile PBS (200 μL per well) was gently added to the OS PD(X)O organoids to dissolve the Matrigel. The organoids were transferred to a 15 mL centrifuge tube and incubated at 4°C for 15 min. After adding 1–2 mL of sterile PBS and gently mixing by pipetting, the suspension was centrifuged at 1500 × g for 5 min at room temperature. The organoid cells were resuspended in cryopreservation solution (0.5–1 mL per well) and gradually cooled to 4°C (∼1–2 h) before transferring to liquid nitrogen for long-term storage.

For resuscitation, sterile 37°C water was prepared in advance. The cryopreserved OS PD(X)O organoids were rapidly thawed in a 37°C water bath. After centrifugation at 1500 × g for 5 min at room temperature, the supernatant was carefully removed. The organoid single-cell suspension was resuspended in culture medium, and cell morphology was observed under a microscope. The cells were counted and subsequently cultured for further use.

T cell immune reconstitution in osteosarcoma PDX-derived organoids

Peripheral blood mononuclear cells (PBMCs) were obtained from Milestone (#PB025C, TPCS, Milestone Biotechnologies) with informed consent from donor (Age, 32; Gender, male; Nationality, China; Smoking, yes; Drinking, no; Blood Type: B+), surface Marker Summary (%)——(1) The rate of Leukocyte: CD45^+^ (99.13%), CD14^+^ (21.64%), CD16^+^ (10.85), CD3^+^ (58.20), CD19^+^ (6.62), CD56^+^ (11.88), CD3^+^CD56^+^ (3.81), CD56^+^CD16^+^ (7.79), CD14^+^CD16^+^ (2.57), CD4^+^ (32.89), CD8^+^ (16.69), CD25^+^ (5.20); (2) The rate of CD3^+^T cells(%): CD4^−^CD8^+^ (28.51), CD4^+^CD8^+^ (0.15), CD4^+^CD8^−^ (58.16), CD4^−^CD8^−^ (13.17); (3) The rate of CD3^+^T Cells: CCR7^+^CD45RA^+^(26.84), CCR7^+^CD45RA^−^(30.64), CCR7^−^CD45RA^−^(27.27), CCR7^−^CD45RA^+^(15.26); (4) The rate of CD4^+^T Cells: CCR7^+^CD45RA^+^(34.96), CCR7^+^CD45RA^−^(47.61), CCR7^−^CD45RA^−^(17.04), CCR7^−^CD45RA^+^(0.39); (5) The rate of CD8^+^T Cells: CCR7^+^CD45RA^+^(21.45), CCR7^+^CD45RA^−^(7.67), CCR7^−^CD45RA^−^(36.50), CCR7^−^CD45RA^+^(34.38); (6) HLA: HLA-A (03:303, 24:370N), HLA-B (15:01, 52:01), HLA-C (03:03, 12:02),HLA-DRB1 (07:01, 13:01), HLA-DQB1 (06:03, 03:03).

For PBMC pre-activation, cells were treated with anti-human CD3 antibody (10 μg/mL, #300306, BioLegend) in a 96-well plate, followed by incubation with CD28 Monoclonal Antibody (5 μg/mL, #16-0289-85, Invitrogen) and Human Recombinant IL-2 (1 μL/10 mL, #200-02, PeproTech) at 37°C for 2 h. The dissociated single-cell OS PD(X)O organoid suspension was mixed with the pre-activated PBMCs at a 1:1 cell ratio to ensure adequate cell contact and immune microenvironment reconstruction. Following the self-assembly protocol for OS PD(X)O organoid models, the mixture was embedded in a self-assembling matrix and cultivated. The co-cultivation medium consisted of RPMI 1640 mixed with 2×culture medium in a specified ratio. The medium was replaced every two days.

In cases where anti-human PD-1 monoclonal antibody (mAb) (#A2002, Shlleck) or the PRMT5^MTA^ inhibitor ARPN2169 (this study, APEIRON) and Human IgG1 isotype control (#A2051, Shlleck) was added to the culture conditions, the medium was also refreshed every two days. The co-cultivation period lasted 7–10 days to ensure optimal T cell integration and organoid functionality.

T cell infiltration “success” scoring criteria: (1) significant Infiltration of CD8^+^ T cells: We assessed the infiltration of CD8^+^ T cells using both multiplex immunofluorescence (mIF) and flow cytometry. The mIF images clearly show the distribution of CD8^+^ T cells in the tumor tissue, marking the infiltrating regions. For flow cytometry, we used the following criterion: the proportion of CD8+/CD45+ cells should be greater than or equal to 2% (i.e., at least 2% of the cells should be CD8^+^ T cells). This criterion ensures the presence of significant T cell infiltration and confirms its authenticity. (2) Expression of T cell Functional Markers: To assess the functional status of T cells, we also examined the expression of T cell functional markers, specifically Granzyme B (GZMB) and Perforin. These markers effectively reflect the cytotoxic activity of T cells. The expression levels of GZMB and Perforin are key indicators of T cell functionality.

Agar embedding

The organoid was centrifuged at 5000 rpm for 3 min in a 1.5 mL microcentrifuge tube. The supernatant was carefully removed using a pipette. A 2% molten agarose (#G5056-5G, Servicebio) solution was added to the tube containing the organoid, and the mixture was gently stirred with forceps to ensure uniform encapsulation of the cells. Once the agarose solidified, the agar block was removed, and the region containing the organoid was excised and placed into a dehydration processor (TP1020, Leica). Dehydration was performed using a gradient alcohol series in a tissue processor: 75% alcohol for 4 h, 85% alcohol for 2 h, 90% alcohol for 2 h, 95% alcohol for 1 h, absolute ethanol I for 30 min, absolute ethanol II for 30 min, absolute ethanol III for 30 min, xylene I for 5–10 min, and xylene II for 5–10 min. Following dehydration, the tissue was infiltrated with molten paraffin wax at 65°C in three steps, each lasting 1 h. For embedding (KD-BM IV, KEDEE), the tissue was transferred from the dehydration processor to a pre-labeled embedding mold filled with molten paraffin wax, ensuring proper orientation of the sample. The mold was then placed on a −20°C freezing stage (HistoCore Arcadia C, Leica) to solidify the wax. Once solidified, the paraffin block was removed from the mold and trimmed for further processing.

Hematoxylin and eosin (H&E) staining

Hematoxylin and Eosin (H&E) staining was performed as our previous study.54 Sections were deparaffinized in xylene (twice, 5 min each) and rehydrated through a graded ethanol series (100%, 95%, 70%, and 50%, 2 min each), followed by a distilled water rinse. Nuclei were stained with hematoxylin for 5 min, rinsed in tap water, and differentiated in 1% acid alcohol. Sections were blued in 0.2% ammonia water for 1 min and counterstained with eosin for 2 min. Dehydration was achieved through a graded ethanol series, and sections were cleared in xylene before mounting with a xylene-based medium.

Immunohistochemical (IHC) staining

Immunohistochemical (IHC) staining was conducted on formalin-fixed, paraffin-embedded tissue sections to detect protein expression. Sections were deparaffinized in xylene, rehydrated through graded ethanol, and rinsed in distilled water. Antigen retrieval was performed using preheated citrate or EDTA buffer at sub-boiling temperature for 10–15 min, followed by cooling and PBS washing. Endogenous peroxidase activity was blocked with 3% hydrogen peroxide for 10 min, and nonspecific binding was minimized by incubating with 5% BSA or 10% normal goat serum for 30 min. Primary antibodies were applied overnight at 4°C, followed by incubation with an HRP-conjugated secondary antibody for 1 h. Signal detection was achieved using DAB, and sections were counterstained with hematoxylin, dehydrated, cleared in xylene, and mounted for analysis. The antibodies were including MKI67 (1:500, #GB151499-100, Servicebio), SATB2 (1:500, #GB111449-100, Servicebio) and VIM (1:500, #GB121308-100, Servicebio).

Immunofluorescence staining

Paraffin sections were deparaffinized and rehydrated, while frozen sections were equilibrated to room temperature and rinsed with PBS. Antigen retrieval was performed using citrate or EDTA buffer by heating to boiling and maintaining for 10–15 min. After cooling, sections were blocked with 5% BSA or 10% normal goat serum for 30 min. Primary antibodies were applied and incubated at 4°C overnight, followed by species-specific fluorescent secondary antibodies at room temperature for 1 h. Nuclear staining was performed with DAPI (1:1000) for 5 min. Sections were mounted using antifade medium and imaged using a fluorescence or confocal microscope, with separate channels acquired to avoid crosstalk. The antibodies were including SMA (1:500, # MA1-06110, Invitrogen), SATB2 (1:1000, #PA5-83092, Invitrogen), COL1A1 (1:500, #PA5-29569, Invitrogen), c-caspaes3 (1:200, #ab32042, abcam), HIF-1α (1:500, #ab51608, abcam), CD4 (1:200, #ab183685, abcam), CD8 (1:200, #ab237709, abcam), PD-1 (1:100, #14-2798-82, Invitrogen), TIM3 (1:100, #MA5-32841, Invitrogen), CD206 (1:200, #PA5-101657, Invitrogen). FOXP3 (5 μg/mL, #41-5773-82, Invitrogen). The secondary antibodies were including Alexa Fluor 488 (#A-11001 and # A-11008, Invitrogen), Alexa Fluor 594 (#A-11012 and #A-11005, Invitrogen), Cyanine3 (#A10520 and #A10521, Invitrogen) and Cyanine5 (#A10523 and #A10524, Invitrogen).

3D cell viability assay

Cell viability was assessed using the CellTiter-Lumi Luminescent Cell Viability Assay Kit (#C0065S, Beyotime), following the manufacturer’s instructions. Briefly, the cell culture plate was equilibrated at room temperature for 10 min (not exceeding 30 min). Subsequently, 100 μL of the CellTiter-Lumi luminescent assay reagent was added to each well of a 96-well plate. The plate was gently shaken at room temperature for 2 min to ensure complete cell lysis, followed by a 10-min incubation at room temperature to stabilize the luminescent signal. Luminescence was measured using the SpectraMax M3 Microplate Reader (Molecular Devices). Relative cell viability was calculated directly from the luminescent readings or determined based on ATP content using a standard curve generated with nuclease-free ATP (#D7378, Beyotime). This method provided a reliable quantification of cell viability in 3D cultures based on ATP levels.

Flow cytometry

For flow cytometry analysis, the single-cell suspension was centrifuged to collect the cell pellet. Cells were resuspended in 500 μL of FACS staining buffer (#00-4222-26, Invitrogen) and incubated with 1 μL of blocking buffer (#B001T06F01, Invitrogen) with gentle vortexing to ensure proper mixing. Surface staining was conducted by incubating cells with 1 μL of each antibody (1:500 dilution; antibodies were combined according to the experimental design) for 15 min at 4°C. After thorough vortexing, cells were washed with PBS and fixed with 4% paraformaldehyde (PFA). Samples were stored protected from light at 4°C until analysis on LSR Fortessa flow cytometer (BD Biosciences, USA). The following antibodies were used: PerCP anti-human CD45 (#304025, BioLegend), FITC anti-human CD3 (#317305, BioLegend), APC anti-human CD4 (#317415, BioLegend), PE anti-human CD8 (#344705, BioLegend) and BV605 anti-human PD-1 (#367425, BioLegend).

For the detection of T cells cytotoxic molecules, cells were pre-incubated for 6 h in complete medium supplemented with Cell Activation Cocktail (#423303, BioLegend). Surface staining was performed using Pacific Blue anti-human CD45 (#304021, BioLegend), FITC anti-human CD3 (#317305, BioLegend) and PE anti-human CD8 (#344705, BioLegend) for 15 min at 4°C. Cells were then washed with PBS and fixed/permeabilized using Intracellular Fixation & Permeabilization Buffer (#88-8823-88, eBioscience) for 60 min at room temperature. Intracellular staining was carried out with APC anti-human Perforin (#308111, BioLegend) and PerCP anti-human Granzyme B (#396415, BioLegend) in permeabilization buffer for 60 min at room temperature. After washing with permeabilization buffer, cells were analyzed on LSR Fortessa flow cytometer (BD Biosciences, USA).

For T cell gating, single cells were first selected from total events, followed by sequential gating using the strategy CD45^+^→CD3^+^→CD4^+^/CD8^+^. To assess the cytotoxic function of CD8^+^ T cell, the expression levels of Granzyme B and Perforin were measured within the gated CD8^+^ T cell population. For macrophage gating, single cells were selected and subsequently gated as CD45^+^→CD11b^+^→F4/80^+^.

Whole genome sequencing (WGS)

Genomic DNA was extracted using the CTAB method. Approximately 200 mg of tissue, ground to a fine powder using liquid nitrogen, was transferred to a preheated (65°C) 2.0 mL tube containing an appropriate volume of CTAB lysis buffer (#MC1411, Promega) and mixed thoroughly by vortexing. The tube was incubated at 65°C for 60 min, followed by centrifugation at 10,000 rpm at room temperature (RT) for 5 min. The supernatant was extracted with an equal volume of phenol/chloroform/isopentanol (25:24:1) and centrifuged at 10,000 rpm for 10 min in a fresh tube. Two-thirds volume of pre-cooled (−20°C) isopropanol was added to the supernatant, and the mixture was incubated at −20°C for at least 2 h to precipitate DNA. The DNA was pelleted by centrifugation at 12,000 rpm for 15 min at RT, washed with 75% ethanol, and air-dried for 3–5 min. The DNA pellet was resuspended in 30–200 μL of TE buffer (#12090015, Thermo Fisher) for downstream applications.

For WGS library preparation, 1 μg of genomic DNA was randomly fragmented using a Covaris Focused-ultrasonicator (Covaris). Fragments of 200–400 bp were selected using the Agencourt AMPure XP-Medium kit. Selected fragments were subjected to end repair, 3′ adenylation, and adaptor ligation. The ligated products were amplified by PCR and purified using the Agencourt AMPure XP-Medium kit (#A63882, Beckman). The purified PCR products were heat-denatured into single-stranded DNA and circularized using a splint oligonucleotide. The resulting single-stranded circular DNA (ssCir DNA) was formatted into the final library and subjected to quality control (QC). Sequencing was performed on the BGISEQ-500 platform (BGI Genomics), where ssCir DNA molecules were amplified into DNA nanoballs (DNBs) containing over 300 copies via rolling-circle replication. DNBs were loaded onto a patterned nanoarray using high-density DNA nanochip technology, and paired-end 100 bp reads were generated through combinatorial Probe-Anchor Synthesis (cPAS). WGS library preparation and sequencing were conducted at Benagen (Wuhan Benagen Technology Co., Ltd.).

WGS data analysis

For data analysis, Xgenome (v1.0.0) was used to remove potential mouse-derived reads from the WGS data following quality control. Filtered reads were aligned to the UCSC hg19 reference genome using Burrows-Wheeler Aligner (bwa mem, v0.7.17). PCR duplicates were removed using Picard (v2.27.1), and the resulting BAM files were indexed using Samtools (v1.10). Base quality score recalibration was performed using the BaseRecalibrator and ApplyBQSR tools from the Genome Analysis Toolkit (GATK, v4.2.6.1) following GATK best practices. Somatic variants, including single nucleotide variants (SNVs) and small indels, were identified using Mutect2 in GATK with a tumor-only model. Variants were annotated using Annovar (v2022/08) based on the RefSeq gene model, and non-coding variants (e.g., upstream, downstream, intergenic, intronic, ncRNA, UTR5, UTR3) were excluded. Germline variants were filtered using databases such as the 1000 Genomes, Exome Aggregation Consortium, NHLBI Exome Sequencing Project (ESP6500), and Genome Aggregation Database (gnomAD). Mutation landscapes were visualized using maftools (v2.24.0). Somatic copy number alterations (SCNAs) were calculated using CNVkit (v0.9.9) with default parameters, and SCNAs were clustered by chromosome in R based on Euclidean distance using Ward’s complete method within the pheatmap (v1.0.13) package.

Bulk RNA sequencing (RNA-seq)

RNA purity was assessed using the kaiaoK5500 Spectrophotometer (Kaiao), while RNA integrity and concentration were evaluated using the RNA Nano 6000 Assay Kit of the Bioanalyzer 2100 system (#5067-1511, Agilent Technologies). For library preparation, 2 μg of total RNA per sample was used as input material. Sequencing libraries were constructed using the NEBNext Ultra RNA Library Prep Kit for Illumina (#7770, New England Biolabs) following the manufacturer’s protocol, with index codes incorporated to distinguish individual samples. Index-coded samples were then clustered on a cBot cluster generation system using the HiSeq PE Cluster Kit v4-cBot-HS (#PE-401-4001, Illumina). Final sequencing was performed on the DNBSEQ-T7 platform (BGI Genomics) at Benagen (Wuhan Benagen Technology Co., Ltd.), generating 150 bp paired-end reads.

RNA-seq analysis

Paired-end 150-bp RNA-seq reads with an additional 6-bp barcode were generated for each sample. Reads were aligned to the human genome assembly hg38 (GRCh38) using STAR (v2.7.10a; RRID:SCR_015899) with the following key parameters: outFilterMismatchNmax 2, outFilterMultimapScoreRange 0, and alignIntronMax 500000. Aligned BAM files were filtered using SAMtools (v1.16.1; RRID:SCR_002105) to remove low-quality alignments (MAPQ <7) and retain only properly paired reads (samtools view -q 7 -f 2). PCR duplicate reads were marked and removed using Picard (v3.1.0; RRID:SCR_006525). To reduce alignment artifacts, reads overlapping the hg38 genomic blacklist regions were removed using bedtools. Gene-level read counts were computed from the filtered BAM files using featureCounts (Subread v2.0.1; RRID:SCR_012919) with a GRCh38/hg38 gene annotation. For genome browser visualization, strand-specific BigWig files were generated using bamCoverage from deepTools (v3.5.1; RRID:SCR_016366), with strand filtering enabled and the resulting tracks were visualized in IGV.

Downstream analyses were performed in the R/Bioconductor environment (R v4.5.1. and Bioconductor v3.22). Genes were considered expressed if they had >10 read counts in at least two summarized replicates, consistent with the filtering strategy used in our pipeline. Differential expression analysis was conducted using edgeR (please specify edgeR v4.6.3). Briefly, library sizes were normalized using the trimmed mean of M-values (TMM) method, dispersions were estimated within the edgeR framework, and differential expression was tested using an appropriate generalized linear model (GLM) approach. Genes were defined as differentially expressed with |log2 fold-change| > 1 (log2FC > 1 or < −1) and adjusted p value <0.01 (Benjamini–Hochberg FDR). To support downstream modeling and correlation analyses, gene expression was additionally represented as TPM. TPM values were calculated from gene-level read counts using gene lengths derived from the same GRCh38 gene annotation used for counting.

Principal Component Analysis (PCA) was conducted using the R package stats. Differential gene expression analysis was performed using DESeq2 (R package, v1.44.0) with the IfcShrink function, where genes meeting the following criteria were considered significantly differentially expressed: fold-change ≥2.00, probability ≥0.80, and false discovery rate (FDR) < 0.05. Functional enrichment analyses were conducted using the R package clusterProfiler. Gene set enrichment analysis (GSEA) was performed using the GSEAPreRanked function with the following parameters: enrichment statistic = classic, 1000 permutations, and normalized p value <0.05. Single sample GSEA (ssGSEA) was implemented using the R package GSVA (v2.2.0), with gene sets curated from the Molecular Signatures Database (MSigDB, https://www.gsea-msigdb.org/gsea/msigdb/).

Drug sensitivity was inferred from RNA-seq expression profiles using the OncoPredict R package (v1.2). Briefly, gene-level TPM expression matrices were used as input for the oncopredict function to generate drug response predictions based on pharmacogenomic models trained on the GDSC2 dataset. Gene identifiers were harmonized between the study TPM matrix and the GDSC2 training expression matrix. Model-based prediction was then performed using oncopredict, leveraging the GDSC2 reference expression and drug response profiles provided by the package (or user-supplied GDSC2 objects, as implemented in our analysis). Predicted drug sensitivity values (predicted IC50) were obtained for each sample across the tested compounds and used for downstream association analyses.

To evaluate associations among transcriptomic features and predicted drug response, Spearman rank correlation was performed using gene expression TPM values, ssGSEA scores, and OncoPredict-predicted drug sensitivity results across samples. Correlation coefficients and p values were computed in R using cor.test(method = “spearman”). Multiple testing correction was applied where appropriate using the Benjamini–Hochberg method, and significant associations were defined based on an adjusted p value (FDR) threshold (FDR <0.05). Correlation results were summarized and visualized as scatterplots and/or correlation heatmaps.

For fusion gene detection, STAR-Fusion (v1.10.0) was utilized. High-quality RNA-seq reads were aligned to the GRCh38 reference genome using STAR aligner (v2.7.10a) with default parameters. The resulting BAM files were processed by STAR-Fusion, which integrates alignment evidence and transcriptome annotations, incorporating information from GENCODE reference annotation (v35) and the FusionCatcher database. Only high-confidence fusion events supported by a minimum of five spanning reads and/or junction reads were retained. Predicted fusions were annotated with detailed gene information, including breakpoint positions, fusion partners, and functional domains. To ensure reliability, fusion candidates were filtered against a panel of normal tissues and known artifacts. Finally, fusion events were visualized and validated using Integrative Genomics Viewer (IGV).

Preparation and analysis of single-cell RNA sequencing (scRNA-seq)

Tissue transportation and dissociation procedures were conducted as described by Mu et al.29 Freshly prepared cell suspensions were immediately processed according to the manufacturer’s instructions provided in the SeekOne DD Single Cell 3′ Transcriptome kit (SeekGene, Beijing, China). The resulting libraries were assessed for quality, with the main peak fragment size required to fall within 350–750 bp. Libraries containing small fragments underwent additional purification until no small fragments were detected, as verified using the Agilent 4200 TapeStation. Library concentration was determined by Qubit 4.0 and was required to be no less than 1 ng/μL. Sequencing was performed on the NovaSeq6000 platform (Illumina, USA).

For quality control and normalization, raw sequencing reads were converted to FASTQ format using Illumina bcl2fastq2 Conversion Software v2.20 and assessed with FastQC v0.11.9. Subsequent processing was carried out using Cell Ranger (v3.0.1), in which reads were aligned to the GRCh38 genome with STAR under default parameters. Gene-barcode matrices were generated for each sample by counting distinct molecular identifiers (UMIs) and filtering out barcodes not associated with cells.

Downstream scRNA-seq analysis was performed in R (v4.3.1) using the Seurat package (v5.3.0). Expression matrices from individual samples were imported via the Read10X function. Cells were filtered based on the following criteria: expressed genes <300 or >5000, or mitochondrial gene content >10%. The filtered data were log-normalized using a default scaling factor. The top 2000 highly variable genes (HVGs) were selected, centered, and scaled prior to principal component analysis (PCA). Clustering was conducted on the integrated joint embedding, and the resulting clusters were visualized in two dimensions using UMAP. Differentially expressed genes (DEGs) between clusters were identified using the non-parametric Wilcoxon rank-sum test with Bonferroni correction in Seurat. Cell clusters were annotated based on these DEGs and classical cell markers. Pseudotime analysis was performed using the Monocle package (v2.28.0). The trajectory was inferred based on the top 2,000 highly variable genes, utilizing the DDRTree reduction method and log normalization. Cell-cell communication analysis was conducted using the CellChat package (v1.6.1). The human-specific ligand–receptor interaction database (CellChatDB.human) covering secreted signaling, ECM-receptor, and cell–cell contact pathways was applied. Interaction probabilities were inferred using the “triMean” method. Unless otherwise stated, all analytical steps employed default parameters.

Quantification and statistical analysis

Detailed descriptions of statistical analyses for individual experiments are provided in the corresponding methods sections. Unless otherwise specified, statistical analyses were performed using GraphPad Prism 8 software (GraphPad Software). Normality of data distribution was assessed prior to analysis. For comparisons between two independent groups, a two-tailed Student’s t test was applied to determine statistical significance. Spearman’s correlation analysis was used to evaluate the relationship between two sets of variables. A p-value <0.05 was considered statistically significant. Specific details regarding the number of replicates, data points, and experimental conditions are outlined in the respective figure legends. Flow cytometry data were analyzed using FlowJo v10 software (Tree Star). All schematic illustrations were created with BioRender (BioRender.com).