Defining the cellular and molecular identities of histologic subtypes in lung adenocarcinoma

Abstract

Background

Tumor histology reflects disease aggressiveness and clinical outcomes in cancer patients. Lung adenocarcinomas (LUADs) are classified based on predominant histologic patterns, including high-grade micropapillary and solid subtypes which portend unfavorable clinical features and prognosis. However, the cellular and molecular characteristics underlying these histologic subtypes remain largely unknown.

Methods

We used scRNA-seq to profile 117,266 cells from 18 treatment-naïve LUADs with heterogeneous histologic patterns and also performed spatial transcriptomic analysis (10x Visium) for representative cases. By integrating single-cell transcriptomics with spatial information, we aimed to characterize the cellular identity and spatial organization driving LUAD heterogeneity.

Results

We demonstrated that histologic subtypes can be distinguished by subtype-specific cancer cell subpopulations and immunosuppressive phenotypes in the tumor microenvironment (TME). Our data reveal how intercellular interactions among cancer cells, macrophages, and CD8⁺ T cells in the prognostically unfavorable solid subtype are associated with cancer cell plasticity and promote an immunosuppressive TME. Additionally, we identify HMGA1 as a potential clinically relevant biomarker and therapeutic target for the solid subtype LUAD.

Conclusions

These findings deepen our understanding of the histologic heterogeneity of LUAD and may facilitate the development of subtype-specific biomarkers and targeted therapeutic strategies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40164-025-00740-6.

Background

Lung adenocarcinoma (LUAD) is the most prevalent subtype of non–small cell lung cancer (NSCLC), accounting for approximately 50% of all lung cancer cases [1, 2]. LUAD is characterized by pronounced histologic and molecular heterogeneity, resulting in substantial differences in clinical characteristics and treatment outcomes. Invasive LUADs, when surgically resected, are classified according to the predominant histologic subtype, which is determined through comprehensive evaluation of all distinct morphologic patterns, including lepidic, acinar, papillary, solid, and micropapillary (MP) [2]. Notably, individual LUADs often contain a mixture of multiple histologic patterns [2]. Tumors with a predominant solid or MP component are strongly associated with earlier postoperative recurrence and poorer survival outcomes. Accordingly, they are classified as high-grade, in contrast to acinar- or papillary-predominant (intermediate-grade) and lepidic-predominant (low-grade) tumors [3, 4]. Similarly, LUADs containing a minor component of solid or MP patterns have also been linked to a poorer prognosis [5, 6]. Our previous work demonstrated that even minimal presence of solid or MP components is strongly correlated with aggressive clinicopathologic features, including larger tumor size, lymph node metastasis, lymphovascular invasion, and earlier recurrence, particularly in the early-stage disease [7].

Recent studies have explored the association between the microscopic morphology and the genomic landscape of LUAD. These studies reveal that high-grade subtype-predominant tumors exhibit higher tumor mutational burden (TMB), greater chromosomal instability, and more frequent oncogenic pathway alterations, while showing fewer targetable alterations than lower-grade subtype-predominant LUADs [8, 9]. Notably, the two high-grade subtypes, solid and MP, display unique biological and clonal evolutionary characteristics that distinguish them from each other [9]. Furthermore, transcriptomic profiles, rather than genomic alterations, show a stronger correlation with intratumor heterogeneity of histologic patterns, as demonstrated by DNA and RNA sequencing data from histopathology-guided multiregion samples of LUAD [10].

Advances in molecular profiling, particularly single-cell RNA sequencing (scRNA-seq), have deepened our understanding of LUAD subtypes, providing new insights into cancer cell phenotypes and the tumor microenvironment (TME) [11–15]. Specifically, multiple studies have demonstrated that the solid subtype is characterized by poorly differentiated cancer cells, heightened oncogenic pathway activity, and a more immunosuppressive TME compared with lower-grade subtypes [10, 14, 15].

Despite these findings, the mechanisms underlying the distinct characteristics of LUAD histologic patterns remain unclear. To address this gap, we performed scRNA-seq of 117,266 cells from 18 treatment-naïve LUADs with heterogeneous histologic patterns, and spatial transcriptomic profiling (10x Visium) for representative cases. By integrating single-cell transcriptomic with spatial data, we aimed to delineate the cellular identities and spatial architectures that drive LUAD heterogeneity. Our analysis identified subtype-specific cancer cell subpopulations and immunosuppressive phenotypes. We uncovered distinct intercellular interactions in the solid subtype that drive cancer cell plasticity and promote an immunosuppressive TME, revealing HMGA1 as a novel solid subtype–specific cancer cell marker. These findings provide a comprehensive framework for understanding LUAD histologic heterogeneity at single-cell resolution, and lay the foundation for improved patient stratification and the development of more effective therapeutic strategies.

Results

A single-cell atlas of LUAD with a heterogeneous mixture of histologic patterns

To characterize the cellular and molecular features of LUAD histologic subtypes, we performed single-cell transcriptomic analysis on surgically resected primary tumor samples from 18 treatment-naïve patients with heterogeneous histologic patterns (Table S1). Typical Histologic patterns observed in tumors from our cohort are presented in Fig. S1 for reference. We prioritized cases with higher clinical T and N stages among patients scheduled for surgery to enrich for solid and MP subtypes, given our previous research [7] linking these subtypes to larger tumor size and lymph node metastasis. To compare solid and MP subtypes with lower-grade subtypes, we categorized samples into three histologic subtype groups based on a 20% threshold for solid or MP component, in accordance with recent LUAD grading system proposal by the International Association for the Study of Lung Cancer (IASLC) pathology committee [3]. These groups were: (1) Acinar/Papillary (A/P) group – tumors predominantly exhibiting acinar or papillary patterns with < 20% solid or MP component (n = 11); (2) MP group – tumors containing ≥ 20% MP component (n = 3); and (3) Solid group – tumors with ≥ 20% solid component (n = 4) (Fig. 1A). The same histologic subtype grouping was used throughout all subsequent analyses, including comparisons of tissue HMGA1 immunohistochemistry (IHC) and mRNA expression, and serum HMGA1 levels among subtypes.

Fig. 1.

Single-cell transcriptomic atlas of 18 treatment-naive LUAD patients across histologic subtype groups. A Schematic illustration of the study design and experimental workflow, including clinicopathologic features and scRNA-seq analysis. B UMAP plot of all 117,266 cells profiled using scRNA-seq, colored by major cell type across three histologic subtype groups: acinar/papillary (A/P), micropapillary (MP) and solid (Solid). C, D UMAP visualizations of lymphoid (C) and myeloid (D) cell compartments, colored according to cell subtype. Insets indicate the position of each compartment within the UMAP of all cells shown in B. E Stacked bar plot depicting the relative proportions of cell types across individual patient samples, ordered by histologic subtype. The illustration in A was created using BioRender (https://biorender.com/)

After quality control and batch correction (Fig. S2A-C), we generated a single-cell atlas comprising 117,266 cells from the three LUAD histologic subtype groups (Fig. 1B and S2D), with an average of 2,470 genes and 10,891 unique molecular identifiers (UMIs) per cell. Major cell lineages, including lymphoid, myeloid, stromal, and epithelial cells, were identified based on the expression of canonical marker genes (Fig. S2E). Subclustering analysis further resolved 39 distinct cell subtypes (Fig. 1C-D and S2F; detailed methodology for subclustering analysis provided in Supplementary Methods, "Cell type annotation and subclustering analysis"). While most cell types were shared across samples, epithelial cell proportions varied considerably, indicating substantial inter-sample heterogeneity in epithelial composition (Fig. 1E).

Cellular and molecular characteristics of the lymphoid compartment across LUAD histologic subtype groups

Given the high intertumor heterogeneity in lymphoid cell type composition (Fig. 1C, S2H, and S3A ), we next sought to characterize the molecular features of lymphoid cells across LUAD histologic subtypes. To achieve this, we inferred the cell-level gene-regulatory network (GRN) activity (hereafter referred to as ‘regulon’) within the lymphoid compartment using SCENIC [16]. Most lymphoid cell types exhibited cell-type-specific regulon activity, however, innate lymphoid cells (ILCs), natural killer (NK) cells, and T cells displayed distinct activation patterns, particularly in the Solid group (Fig. S3B). These cell types were characterized by activated regulons of IRF1, IRF2, IRF9, STAT1/2, MBD2, and IKZF3 (Fig. S3B). Notably, IRF1, IRF2, and IRF9 play pivotal roles in regulating the type I interferon (IFN) response [17]. Among them, IRF2 has emerged as a key driver of IFN-mediated CD8⁺ T cell exhaustion, limiting the anti-tumor immune response [18, 19]. Consistent with this, regulons targeted by PRDM1, IKZF3, MBD2 and ETV1, which are markers of exhausted CD8⁺ T cells [20], were significantly activated in CD8⁺ T cells from the Solid group compared with the other groups (Fig. S3C and D; Table S2). These findings suggest that the solid subtype-enriched TME is immunosuppressive.

To better characterize the cellular and molecular features of lymphoid cells across LUAD histologic subtypes, we classified lymphocytes into 20 cell subtypes based on cell type-specific marker expression and signature scores derived from curated gene sets in prior studies [21–23] (Fig. 1C and S3E-G; see Supplementary Methods and Table S3 for annotation procedures and references for marker genes and gene sets). This classification encompassed four B cell subtypes, five CD4⁺ T cell subtypes, five CD8⁺ T cell subtypes, regulatory T cells (Treg), γδ T cells (Tgd), innate lymphoid cells (ILC), NK cells, natural killer T cells (NKT), and proliferating lymphoid cells (Lymphoid.Prolif). These subtypes exhibited distinct transcriptional profiles and functional states, reflecting the diverse immune landscape within LUAD tumors.

Exhaustion of CD8⁺ T cells in the Solid group

Given the activated regulons associated with the exhausted phenotype of CD8⁺ T cells from the Solid group (Fig. S3B-D), we calculated signature scores using curated gene sets from Chu et al. [21] (Fig. 2A). CD8⁺ T cells in the Solid group exhibited significantly higher signature score of CD8⁺ T cell exhaustion markers, including PDCD1, HAVCR2, LAG3, and CTLA4 compared with other groups (Fig. 2B). These cells also displayed elevated signatures for CD8⁺ T cell activation, positive regulation of cell death, and stress response, whereas anti-apoptotic signature was downregulated (Fig. S3F). Consistent with this pronounced exhausted phenotype, the T.CD8.Exhausted subtype was markedly expanded in the Solid group compared with other groups (Fig. 2C–G and S3A). Flow cytometric analysis confirmed the higher proportion of PD-1⁺CD8⁺ T cells in the Solid group (Fig. 2H and S4). Collectively, these findings suggest that the TME in the Solid group is characterized by a higher abundance of exhausted CD8⁺ T cells compared with other groups.

Fig. 2.

CD8⁺ T cell exhaustion is enriched in the Solid group of LUAD. A Violin plots showing signature scores for curated gene sets related to CD8⁺ T cell functionality. Gene sets are listed in Table S3. B Violin plots displaying T cell exhaustion signature scores across histologic subtype groups: A/P, MP and Solid. C Stacked bar plot illustrating the relative proportions of CD8⁺ T cell subtypes across individual patient samples, ordered by histologic subtype. D Force-directed graph (FDG) plot identifying five CD8⁺ T cell subtypes: naive, pre-dysfunctional, interferon-stimulated (IFN), cytotoxic and exhausted. E–G FDG plots displaying CD8⁺ T cell distribution density by histologic subtype groups: A/P (E), MP (F), and Solid (G). H Box plot comparing the proportion of exhausted CD8⁺ T cells across histologic subtype groups, measured by fluorescence-activated cell sorting (FACS)

Cellular and molecular characteristics of the myeloid compartment across LUAD histologic subtype groups

To characterize cell-state diversity within the myeloid compartment across LUAD histologic subtypes, we defined 14 transcriptionally distinct myeloid subtypes using established cell-type-specific markers (Fig. 1D and S5A; see Supplementary Methods and Table S3 for annotation procedures and references for marker genes and gene sets for molecular phenotyping). This subclustering analysis resolved four dendritic cell subtypes, two monocyte populations, and five macrophage states, along with mast cells, neutrophils, and a proliferating myeloid cluster (Myeloid.Prolif).

Next, we characterized the molecular features of macrophage subtypes using curated macrophage signatures from Luo et al. [24]. Mac.CXCL9 exhibited the M1-like, pro-inflammatory phenotype [25], whereas Mac.SELENOP displayed the M2-like phenotype associated with anti-inflammatory responses, tissue repair, and immunosuppressive functions in the TME [26]. SPP1⁺ macrophages (Mac.SPP1 and Mac.SPP1.GPNMB) showed pro-tumorigenic features characteristic of tumor-associated macrophages (TAMs), consistent with prior reports linking SPP1 expression to TAM polarization and tumor-supporting functions [27]. Notably, these macrophages exhibited increased activation of the angiogenesis signature, with the highest activation in Mac.SPP1.GPNMB, characterized by co-expressing SPP1 and GPNMB (Fig. 3A). Increased angiogenesis in TAM is known to promote tumor growth, invasion, and metastasis by supplying oxygen and nutrients to tumor cells [28]. The increased angiogenic potential, along with the upregulation of GPNMB, a known promoter of tumor progression and immune evasion [29], suggests that Mac.SPP1.GPNMB likely contributes to a more aggressive and immunosuppressive TME. Taken together, these findings suggest that SPP1⁺ macrophages, particularly Mac.SPP1.GPNMB, play a critical role in shaping a pro-tumorigenic and immunosuppressive microenvironment, consequently promoting tumor progression and poor clinical outcomes.

Fig. 3.

Characterization of myeloid cells in the LUAD TME. A Violin plots showing signature scores for gene sets associated with macrophage phenotypes in the TME. Gene sets are listed in Table S3. B Stacked bar plot depicting the relative proportions of myeloid cell subtypes across histologic subtype groups: A/P, MP and Solid. C Differential abundance analysis of myeloid cell subtypes between A/P and Solid groups using Mixed-effects modeling of Associations of Single Cells (MASC) [reference 30]. The x-axis represents log₂(Odds ratio), dot size indicates −log₁₀(P-value) and error bars indicate 95% confidence intervals. D Heatmap displaying signature scores of signaling pathways from MSigDB across macrophage subtypes and histologic subtype groups. E Violin plots comparing cholesterol efflux metabolism scores in macrophages across histologic subtype groups. F Box plot comparing tissue cholesterol levels between A/P and Solid groups. G Violin plots showing cellular senescence signature scores in all macrophages and macrophage subtypes across histologic subtype groups

To analyze the cellular and molecular differences in the myeloid compartment associated across LUAD histologic subtypes, we compared the relative abundance of myeloid subtypes between the Solid and A/P groups using mixed-effects logistic regression [30]. Our analysis revealed that mast cells were significantly enriched in the A/P group, whereas Mac.SPP1.GPNMB cells, characterized as pro-angiogenic TAMs [29, 31], were more abundant in the Solid group (Fig. 3B and C). These findings highlight the unique cellular composition and molecular characteristics of myeloid cells across different histologic subtype groups.

Aberrant cholesterol metabolism and senescence signatures in macrophages from the Solid group

To gain a deeper understanding of the role of macrophages in shaping the histologic subtype-specific TME, we compared their signaling pathway activity across three different histologic subtype groups. We found that cholesterol efflux metabolism was significantly elevated in macrophages in the Solid group compared with the A/P group (Fig. 3D and E). Specifically, monocyte-derived macrophages (Mac.Infiltrated) showed a statistically significant increase, while alveolar macrophages (Mac.PPARG) exhibited a similar trend that did not reach statistical significance (Adj P = 0.088). Increased cholesterol efflux metabolism in macrophages is likely to contribute to the immunosuppressive TME in the Solid group, as membrane cholesterol efflux has been shown to drive TAM reprogramming toward an immunosuppressive and tumor-promoting phenotype [32]. In addition, the Solid group displayed a markedly higher abundance of monocyte-derived macrophages —the dominant macrophage population in the TME—compared with other groups (Fig. S2H and S5B-C). Notably, macrophages exhibited substantially higher cholesterol efflux metabolism than any other cell type in TME in our dataset (Fig. S5D). Collectively, these findings suggest that both the expansion of macrophages and their elevated cholesterol-efflux metabolism contribute to shaping the immunosuppressive microenvironment in the Solid group. It has been shown that cholesterol in the TME induces CD8⁺ T cell exhaustion through increased ER stress [33], prompting us to measure cholesterol content in LUAD tissues using archived fresh-frozen samples. We found that the Solid group exhibited significantly higher tissue cholesterol content compared with A/P group (Fig. 3F). Taken together, these results suggest that tumor tissue enriched with cholesterol, along with increased cholesterol efflux from macrophages, is associated with CD8⁺ T cell dysfunction in the Solid group.

To validate the exhausted state of CD8⁺ T cells and their spatial relationship with macrophages involved in cholesterol efflux within the TMEs of the A/P and Solid groups, we used multiplex IHC using specific markers. We used CD8 to identify CD8⁺ T cells, CD68 to label macrophages, ABCA1 as a marker for cholesterol efflux, and PD-1 to indicate exhausted T cells. The inForm software, a digital image analysis program, was used to segment individual cells, quantify marker expression, and classify cell phenotypes based on the combination of markers expressed by each cell. The analysis included 10 cases—all of the 4 cases from the Solid group and 6 out of the 11 cases from the A/P group (Table S1). First, we compared representative images from a Solid group case (KNUCH03) and an A/P group case (KNUCH22) (Fig. 4A–B and S6A–F). In the Solid group case, only a few CD8⁺ T cells infiltrated the tumor (Fig. 4A), indicating exclusion of these cells, whereas substantial CD8⁺ T cell infiltration was observed in the A/P group case (Fig. 4B). Figure 4C shows representative areas in which both PD-1⁺CD8⁺ T cells and ABCA1⁺ macrophages are located within the same local microenvironment.

Fig. 4.

Multiplex IHC images of representative cases of A/P and Solid groups. A, B Representative multiplex IHC images illustrating limited CD8⁺ T cell infiltration in the tumor of KNUCH03 (A) and prominent CD8⁺ T cell infiltration in the tumor of KNUCH22 (B). These images correspond to the regions outlined by yellow boxes in Fig. S6C and F, respectively, which show whole-slide images with delineated tumor areas. C Higher-magnification view of the area indicated by the orange box in A, demonstrating co-expression of CD8 with PD-1 and CD68 with ABCA1 within the tumor microenvironment. Scale bars: 200 μm (A, B), 20 μm (C). D–F Boxplots comparing cell type proportions within the tumor nest between A/P and Solid groups: D CD8⁺  T cells as a proportion of total cells, E exhausted CD8⁺ T cells as a proportion of total CD8⁺ T cells, and F ABCA1⁺ macrophages as a proportion of total macrophages. G–I Boxplots comparing cell type proportions within the peritumoral stroma between A/P and Solid groups: G CD8⁺ T cells as a proportion of total cells, H exhausted CD8⁺ T cells as a proportion of total CD8⁺ T cells, and I ABCA1⁺ macrophages as a proportion of total macrophages

When we analyzed 200 regions of interest (ROIs) from 10 LUAD cases (Fig. S6G and H), we observed a significantly lower density of tumor-infiltrating CD8⁺ T cells within the tumor nest of the Solid group compared with the A/P group (P = 0.031; Fig. 4D). However, no significant differences were observed in peritumoral stroma (P = 0.274; Fig. 4G). Consistent with our scRNA-seq (Fig. 2B) and flow cytometric analysis (Fig. 2H), there was a trend toward a higher proportion of the CD8⁺ cells exhibiting an exhausted phenotype in the Solid group compared with the A/P group, both in the tumor nest (P = 0.118; Fig. 4E) and peritumoral stroma (P = 0.076; Fig. 4H). Although our scRNA-seq analysis indicated increased cholesterol efflux metabolism in macrophages from the Solid group (Fig. 3E), multiplex IHC did not reveal significant differences in the proportion of ABCA1⁺ macrophages between Solid and A/P groups, either in the tumor nest or peritumoral stroma (Fig. 4F and I). However, the density of total CD68⁺ macrophages was markedly higher in the peritumoral stroma of the Solid group compared with the A/P group (Table S4). Consistent with this overall increase in macrophage infiltration, both ABCA1⁺CD68⁺ and ABCA1⁻CD68⁺ subsets showed numerically higher densities in the Solid group, although the difference in ABCA1⁺CD68⁺ macrophages did not reach statistical significance (P = 0.19). This pattern suggests that the macrophage expansion in the Solid group reflects a global increase in stromal infiltration rather than a subset-specific enrichment. Despite the lack of statistical significance for the ABCA1⁺ subset, the greater absolute number of ABCA1⁺ macrophages in the Solid group is consistent with a stromal context exhibiting increased cholesterol-efflux capacity. These results align well with our scRNA-seq analysis, which revealed an increased abundance of macrophages and their elevated cholesterol efflux metabolism in the Solid group. Combined with the observed trend toward a higher proportion of PD-1⁺CD8⁺ T cells in the Solid group, these findings raise the possibility that the macrophage-rich and cholesterol-efflux–oriented TME in the Solid group may be associated with an exhausted CD8⁺ T cell phenotype. Further investigation will be required to clarify this potential functional relationship.

Based on recent studies linking alveolar macrophage senescence to tumor progression [34, 35], we observed a significantly increased senescence-associated secretory phenotype (SASP) in Mac.PPARG (alveolar macrophages) in the Solid group compared with other groups (Fig. 3G; SASP gene set from Saul et al. [36]). Additionally, this pattern was observed in Mac.Infiltrated (monocyte-derived macrophages) and across all macrophages, encompassing both Mac.PPARG and Mac.Infiltrated subsets (Fig. 3G). These findings demonstrate that macrophages from the Solid group exhibit dysregulated cholesterol efflux and enhanced SASP, along with their pro-angiogenic and immunosuppressive characteristics. The distinct characteristics of macrophages observed in the Solid group suggest potential interactions with tumor cells that may shape the TME. Such interactions could drive tumor progression and contribute to the unfavorable prognosis associated with the solid subtype of LUAD.

Enhanced cellular plasticity and loss of AT2 cell identity in cancer cells from solid group

To elucidate the molecular characteristics of cancer cells associated with histologic subtype groups, we identified 34,278 cancer cells using Numbat [37] (Fig. S2G and S7A). Since alveoli are the primary site of LUAD [38], we excluded cancer cells exhibiting transcriptional signatures of ciliated and goblet cells, which are typically located in the airway epithelium, from downstream analyses. The remaining 32,527 cells were subclustered into 6 distinct clusters, further categorized into three groups based on their molecular phenotypes (Fig. 5A). Group 1 exhibited an increased signature score of canonical AT2 cell markers, including ABCA3, ETV5, and SFTPC (Fig. 5A and B). In contrast, Groups 2 and 3, primarily composed of cancer cells from the Solid group, displayed a decrease in AT2 signature scores along with increased glycolysis and MYC target activity. Group 3 exhibited highly proliferative phenotypes, as indicated by the increased mitotic cell cycle and G2M checkpoint activity (Fig. 5A and B).

Fig. 5.

Molecular characterization of cancer cells across histologic subtype groups. A UMAP plot showing six cancer cell clusters (C0–C5) classified into three subtypes based on transcriptional profiles: Group 1 (C0, C2), Group 2 (C1, C3, C4) and Group 3 (C5). B Heatmap depicting signature scores of cancer cell-related signaling pathways from MSigDB across clusters and subtypes of cancer cells. "H" denotes Hallmark gene sets. Rows represent pathways and columns represent cancer cell clusters. C Violin plots comparing cancer-related pathway scores across histologic subtype groups: A/P, MP and Solid. D UMAP plot displaying cellular differentiation levels computed by CytoTRACE. E UMAP plot showing trajectories of cellular state transitions based on velocities predicted by CellDancer. F UMAP plot showing pseudotime values predicted by Monocle3, with the root set at the cell with the highest AT2 marker expression. G UMAP plots showing the distribution of cancer cells by histologic subtype groups: A/P, MP and Solid

Next, we characterized the molecular differences among cancer cells across the three histologic subtypes of LUAD. The Solid group exhibited a significantly reduced AT2 signature score and upregulation of hypoxia, glycolysis, MYC, and epithelial-to-mesenchymal transition (EMT) signaling compared with the other subtypes (Fig. 5C), suggesting increased cellular plasticity [39–41]. To further investigate cellular plasticity and differentiation across the three LUAD subtypes, we performed RNA velocity analysis [42, 43] and estimated differentiation levels using Cytotrace [44]. These analyses revealed a cellular trajectory from a well-differentiated state, characterized by a high AT2 signature, to poorly differentiated cancer cells (Fig. 5D–G). Differential expression analysis along pseudotime revealed dynamic changes in the expression of genes associated with molecular phenotype of AT2 cells, ribosome synthesis, hypoxia response, cytokeratins, and cell cycling (Fig. S7B). These findings indicate a gradual loss of AT2 identity and the acquisition of a poorly differentiated phenotype along pseudotime, which was most pronounced in the Solid group.

HMGA1 activation and its association with tumor progression and prognosis in LUAD

To dissect the underlying GRN driving this transition across cancer cells with distinct histologic patterns, we performed SCENIC analysis [16, 45] within cancer cells (Table S5). We identified a regulon module (M9) specifically activated in cancer cells from the Solid group (Fig. 6A). Among the transcription factors in the M9 regulons, only HMGA1 was both highly expressed and significantly upregulated in the Solid group compared with the A/P group (Fig. 6A and S7C; Table S6). Notably, HMGA1, a multifaceted gene involved in chromatin organization, cancer cell stemness, and extracellular signaling that promotes tumor growth [46–49], was enriched in cells at the terminal state of the cancer cell trajectory (Fig. S7B).

Fig. 6.

Association of HMGA1 with aggressive phenotype and poor survival in LUAD. A Heatmap showing regulon activities predicted by SCENIC in cancer cells, averaged per sample. B Lollipop plot of GSEA results comparing HMGA1^high and HMGA1^low groups. Dot size indicates −log₁₀(adjusted P-value) and the x-axis represents normalized enrichment score (NES). C UMAP plot showing predicted cell state transitions following in silico HMGA1 knockdown using CellOracle. The perturbation reverses the trajectory observed in Fig. 5E and F, suggesting a shift toward a more differentiated, AT2-like state. D Box plot comparing HMGA1 protein expression levels measured by IHC between A/P and Solid groups. MP group was excluded due to limited sample size. E,F Kaplan–Meier survival curves showing prognostic significance of HMGA1 expression by IHC in the KNUCH cohort: two-group (E) and three-group (F) comparisons. G Box plot comparing Z-scaled HMGA1 expression (TPM) between A/P and Solid groups in TCGA-LUAD bulk RNA-seq data. Statistical significance of differential expression was tested using DESeq2 on raw counts. H Kaplan–Meier survival curves showing prognostic significance of HMGA1 mRNA expression in the TCGA-LUAD cohort. I Box plot comparing serum HMGA1 protein levels between A/P and Solid groups

To further validate the molecular function of HMGA1, we analyzed LUAD cell lines from the Cancer Cell Line Encyclopedia (CCLE) database [50], comparing those with high versus low HMGA1 expression. Gene Set Enrichment Analysis (GSEA) indicated that signaling pathways, including EMT, MYC targets, and the G2M checkpoint, were enriched in HMGA1^high cell lines, further supporting the scRNA-seq results regarding the role of HMGA1 in cellular plasticity and tumor progression (Fig. 6B). Furthermore, in silico knockout of HMGA1 using CellOracle [51] resulted in a transcriptional shift towards a more well-differentiated state (Fig. 6C).

Next, to investigate the clinical relevance of HMGA1, we analyzed its expression by IHC in 83 LUAD samples (71 A/P group, 2 MP group, and 10 Solid group tumors). HMGA1 expression was quantified by the percentage of viable tumor cells showing nuclear staining on each slide (Fig. S8A-I). Solid group had a significantly higher level of HMGA1 expression compared with A/P group (P = 0.011, Fig. 6D). When patients were stratified into high and low HMGA1 expression groups based on the median expression level, high HMGA1 expression was significantly associated with aggressive features, including higher stage, larger tumor size, lymph node metastasis, pleural invasion, and lymphovascular invasion (all P < 0.05) (Table 1). The Solid group showed a significantly higher proportion of high HMGA1 expression compared with other histologic groups (90% vs 53.4%, P = 0.039). Similarly, tumors with any solid component (≥ 1%) exhibited a higher rate of high HMGA1 expression than those without (83.3% vs 50.8%, P = 0.013). EGFR mutations were more frequent in the low HMGA1 group (P = 0.013), whereas HMGA1 expression was not significantly associated with age, sex, or smoking status. Survival analysis revealed a significant association between high HMGA1 expression and poorer survival rate (P = 0.018) (Fig. 6E). When patients were categorized into three subtypes based on the proportion of solid component and HMGA1 expression (solid ≥ 20%, solid < 20% and high HMGA1, and solid < 20% and low HMGA1), survival curves remained well-separated (P = 0.033) (Fig. 6F). To validate the differences in HMGA1 expression across histologic subtype groups by IHC, we examined the archived whole-slide images and the corresponding pathology reports from the TCGA-LUAD dataset (https://cancer.digitalslidearchive.org). A total of 477 cases were selected after excluding mucinous adenocarcinomas and adenosquamous carcinomas based on pathology reports and image review. Because the original pathology reports did not provide quantitative estimation of histologic subtypes, the proportions of solid and MP components were semiquantitatively assessed by experienced pathologists as < 20% or ≥ 20%, taking into account the limitations of image quality. Based on this revised pathologic classification, we further analyzed the TCGA-LUAD RNA-seq data (https://www.cbioportal.org/study/summary?id=luad_tcga_gdc) using samples with both clinical and transcriptomic data. Consistent with our IHC findings, the Solid group exhibited significantly higher HMGA1 mRNA expression compared with the A/P (Fig. 6G). Survival analysis further confirmed that elevated HMGA1 expression was associated with decreased overall survival (Fig. 6H).

Table 1.

Clinicopathologic characteristics of patients with lung adenocarcinoma and HMGA1 expression

Variables	Total	HMGA1 low†	HMGA1 high‡	P-value*
	(N=83)	(N=35)	(N=48)
Age
<65	40	18 (45.0)§	22 (55.0)	0.614
≥65	43	17 (39.5)	26 (60.5)
Sex
Male	35	11 (31.4)	24 (68.6)	0.091
Female	48	24 (50.0)	24 (50.0)
Smoking
Never	45	21 (46.7)	24 (53.3)	0.367
Ever	38	14 (36.8)	24 (63.2)
Pathologic stage
I	66	33 (50.0)	33 (50.0)	0.004
II/III	17	2 (11.8)	15 (88.2)
Tumor size
≤ 2.0 cm	40	23 (57.5)	17 (42.5)	0.006
> 2.0 cm	43	12 (27.9)	31 (72.1)
Mean		2.01 ± 0.92	2.66 ± 1.44	0.014**
Lymph node metastasis
No	70	33 (47.1)	37 (52.9)	0.033
Yes	13	2 (15.4)	11 (84.6)
Pleural invasion
No	56	30 (53.6)	26 (46.4)	0.002
Yes	27	5 (18.5)	22 (81.5)
Lymphovascular invasion
No	66	32 (48.5)	34 (51.5)	0.022
Yes	17	3 (17.6)	14 (82.4)
Solid component
<20%	73	34 (46.6)	39 (53.4)	0.039
≥20%	10	1 (10.0)	9 (90.0)
0%	65	32 (49.2)	33 (50.8)	0.013
≥1%	18	3 (16.7)	15 (83.3)
EGFR
Wild	37	11 (29.7)	26 (70.3)	0.013
Mutant	30	18 (60.0)	12 (40.0)
Intensity of staining
Low	28	23 (82.1)	5 (17.9)	<0.001
High	55	12 (21.8)	43 (78.2)
H score
Low	46	34 (73.9)	12 (26.1)	<0.001
High	37	1 (2.7)	36 (97.3)

HMGA1, High Mobility Group AT-Hook 1; EGFR, Epidermal Growth Factor Receptor†The percentage of viable tumor cells showing nuclear staining of HMGA1 < 70%. ^‡The percentage of viable tumor cells showing nuclear staining of HMGA1 ≥ 70%. ^§Row percentage *P value by chi-square test or Fisher’s exact test **P value by t-test

The concordant results from both our institutional cohort and the TCGA dataset strongly support the prognostic significance of HMGA1 in LUAD. Additionally, serum HMGA1 protein levels, measured by enzyme-linked immunosorbent assay (ELISA), were significantly elevated in the Solid group compared with other groups (Fig. 6I). The robust association of HMGA1 with aggressive clinicopathological features and poor prognosis underscores its pivotal role in LUAD progression. These results suggest that HMGA1 could serve as a clinically relevant biomarker and therapeutic target, particularly for the solid subtype of LUAD.

Tumors of the A/P group with minor solid component exhibit higher transcriptional and genetic heterogeneity of cancer cells

Given the presence of cancer cells with Solid-like molecular phenotypes within the A/P group (Fig. 5G), we hypothesized that minor solid components (< 20%) in A/P tumors could contribute to transcriptional heterogeneity. Indeed, A/P tumors with minor solid components (solid-present) exhibited higher transcriptional heterogeneity compared with those without any solid component (solid-absent) (Fig. 7A). Notably, cancer cells from solid-absent A/P tumors were predominantly found in the well-differentiated state, whereas those from solid-present A/P tumors spanned well-differentiated to poorly-differentiated states (Fig. S9A). These findings suggest that the increased transcriptional heterogeneity in A/P tumors may be driven by the presence of minor solid components.

Fig. 7.

Solid component contributes to tumor heterogeneity in A/P LUAD. A, B Box plots showing transcriptomic (A) and clonal (B) heterogeneity of cancer cells across histologic subtype groups: A/P (solid-absent), A/P (solid-present), MP and Solid. C Phylogenetic tree of cancer cells in KNUCH22 (solid-absent A/P type). D UMAP plot highlighting clone-2 cell density in KNUCH22. Clone-1 was excluded as it represents normal epithelial cells. E Phylogenetic tree of cancer cells in KNUCH13 (solid-present A/P type). F UMAP plots showing clone-specific cell density in KNUCH13. Clone-1 was excluded as it represents normal epithelial cells. G Spatial plot of tumor sample KNUCH22 showing STAGATE-defined spatial clusters. Each 10x Visium spot is colored by its assigned cluster, illustrating the spatial organization of transcriptionally distinct tumor regions. H Heatmap showing normalized abundance of cancer cell subgroups (Groups 1–3, defined in Fig. 5B) across spatial clusters in KNUCH22. Color intensity represents relative cell type abundance estimated by Cell2location deconvolution. I Spatial plot of KNUCH23, a tumor containing 20% solid component without matched scRNA-seq data. Each 10x Visium spot is colored by its STAGATE-defined spatial cluster (C1–C7). J Heatmap showing normalized abundance of cancer cell subgroups (Groups 1–3) across spatial clusters in KNUCH23. Group 1 (A/P-like) cells are enriched in clusters C2 and C3, whereas Group 2 (Solid-like) cells are enriched in clusters C1 and C5. K Spatial map of CytoTRACE-inferred differentiation states across KNUCH23. Lower scores (dark blue) indicate more differentiated regions, whereas higher scores (yellow) denote less differentiated, Solid-like states. Spatial clusters C1 and C5 show higher CytoTRACE scores, consistent with the enrichment of poorly differentiated Group 2 cells in J

To further validate this observation, we assessed whether cancer cell transcriptomic profiles could discriminate histologic subtypes. PCA of patient-level pseudobulk profiles revealed that A/P and Solid tumors occupy distinct transcriptomic spaces, with PC1 representing the major axis of variation separating these subtypes (Fig. S9B). A binary logistic regression classifier (A/P vs. Solid) achieved complete separation under leave-one-out cross-validation (LOOCV) (macro F1 = 1.00) and showed reasonable transferability to external TCGA-LUAD bulk RNA-seq data (macro F1 = 0.7248) (Fig. S9C; see Supplementary Methods for details), supporting the notion that A/P and Solid subtypes harbor distinct transcriptional programs. Extending this analysis to distinguish three classes (A/P solid-absent; A/P solid-present, referred to as A/P + S; and Solid) achieved a macro F1 score of 0.7897 (LOOCV), suggesting that the presence of a solid component in A/P tumors confers a distinct transcriptomic profile. Together, these findings support our hypothesis that solid components, even at minor levels, introduce transcriptional heterogeneity that distinguishes these tumors from solid-absent A/P cases.

Next, we examined whether this increased transcriptional heterogeneity was accompanied by genetic heterogeneity. Numbat-inferred tumor phylogenies revealed that solid-present A/P tumors displayed more complex, branched subclonal architectures compared with solid-absent cases (Fig. 7B–F and S9D-E). For example, the solid-absent A/P tumor KNUCH22 exhibited an unbranched phylogeny with cancer cells restricted to well-differentiated states (Fig. 7C and D). In contrast, KNUCH13, an A/P tumor containing 10% solid component, displayed a branched evolutionary structure with cancer cells spanning well-differentiated to poorly differentiated states (Fig. 7E and F).

Spatial transcriptomic analysis further validated the subpopulation structure observed in solid-present tumors. In the solid-absent A/P tumor KNUCH22, Group 1 cancer cells (A/P-like, well-differentiated) were uniformly enriched across spatial domains, except for a small edge-localized cluster (C5) likely reflecting an edge-related artifact (Fig. 7G and H). For solid-present A/P tumors, Visium profiling was performed on one additional Solid group case (KNUCH23) that lacked matched scRNA-seq data. Because Visium profiling was initiated only in the latter phase of sample collection and required fresh-frozen tissue, it was not feasible to retrospectively analyze the previously collected specimens. In KNUCH23, cancer cells from Groups 1, 2, and 3 coexisted within the tissue (Fig. 7I and J). Spatial clusters enriched for Group 2 cells (C1 and C5) exhibited elevated CytoTRACE scores, indicating poorly differentiated states (Fig. 7I and K). Moreover, Group 1 (well-differentiated) and Group 2 (poorly differentiated) cancer cells showed distinct regional enrichment, preferentially occupying different tumor domains (Fig. 7I and J).

This intratumoral heterogeneity suggests that even minor solid components can contribute to increased transcriptional diversity in solid-present A/P tumors. The presence of cancer cells with Solid-like molecular phenotypes within A/P tumors provides a plausible explanation for their aggressive clinical behavior, elucidating why LUADs with minimal solid components have prognoses comparable to solid-predominant LUADs. Our spatial transcriptomic analysis, showing mixed cellular composition within individual tumors, offers compelling evidence for the clinical significance of detecting minor solid components. These findings highlight cellular heterogeneity as a potential driver of adverse outcomes, underscoring the need for refined prognostic assessments and tailored treatment strategies in LUAD.

Intercellular interaction analysis reveals how the solid group TME becomes immunosuppressive and enhances cancer cell aggressiveness

To identify histologic subtype-specific ligand-receptor (L-R) interactions that shape the TME and influence the molecular characteristics of cancer cells, we performed MultiNicheNet analysis [48]. By comparing the top 100 prioritized L-R pairs between A/P and Solid groups, we found that immunosuppressive and cholesterol efflux-associated intercellular interactions were significantly activated in the Solid group, including IL-10-IL10RA/B, PLTP-ABCA1, and APOE-SCARB1 (Fig. 8A and B). Furthermore, WNT5A-mediated and TGF-β signaling pathways between macrophages and cancer cells were highly activated in the Solid group (Fig. 8C), suggesting that cancer cells from this subtype have more stem-like properties and higher metastatic potential compared with those from the A/P subtype [52]. Notably, we found that TGF-β signaling from monocyte-derived macrophages (Mac.Infiltrated) to cancer cells can potentially activate HMGA1 in the Solid group cancer cells (Fig. 8C and D) [53, 54]. This finding corroborates our observation that HMGA1 plays a pivotal role in regulating cancer cell plasticity and differentiation in the Solid group. Furthermore, IL-6 signaling from cancer cells to alveolar macrophages (Mac.PPARG), which is known to stimulate cholesterol efflux through the JAK/STAT3 pathway [55], was specifically activated in the Solid group (Fig. 8E). Consistent with this, we observed that the IL-6/JAK/STAT3 signaling activity was higher in alveolar macrophages of the Solid group compared with those of the A/P group (Fig. 8F).

Fig. 8.

Distinct cell–cell interactions in A/P versus Solid groups. A, B Circos plots illustrating representative ligand–receptor pairs for cell–cell interactions in A/P (A) and Solid (B) groups. Colored segments represent different cell types, with colored arrows indicating interactions between them. C Dot plot showing scaled product of ligand and receptor pseudobulk expression for predicted interactions between Mac.Infiltrated (monocyte-derived macrophages, sender) and cancer cells (receiver). D Dot plot showing pseudobulk expression of downstream target genes in cancer cells activated by Mac.Infiltrated-derived TGFBI signaling. TGF-β signaling in C and D was not experimentally validated. E Dot plot showing scaled product of ligand and receptor pseudobulk expression for predicted interactions promoting cholesterol efflux between cancer cells (sender) and alveolar macrophages (Mac.PPARG, receiver). F Violin plot comparing IL6/JAK/STAT3 signaling signature score in Mac.PPARG between A/P and Solid groups

To further validate Solid group-specific L-R interactions inferred from scRNA-seq, we predicted spatially resolved L-R pairs from the Solid group. We observed TGF-β and non-canonical WNT signaling from macrophage- to cancer cell-enriched regions, and IL-6 signaling from cancer-cell to macrophage-enriched regions (Fig. S10A and B), consistent with our scRNA-seq-based predictions. Together, these results suggest that solid subtype-specific intercellular interactions between cancer cells and macrophages contribute to the formation of an immunosuppressive TME and promote the acquisition of an aggressive phenotype by cancer cells.

Discussion

In this study, we employed single-cell and spatial transcriptomics to characterize the cellular and molecular identities of LUAD histologic subtypes. We identified immunosuppressive phenotypes within the TME of the Solid group, characterized by CD8⁺ T cell exhaustion and the enrichment of pro-angiogenic and immunosuppressive macrophage subtypes. Aberrant cholesterol metabolism and senescence signatures in macrophages from the Solid group may further contribute to an immunosuppressive TME. These findings highlight the complex interplay between immune cells and the TME in shaping the aggressive phenotype of the solid-subtype LUAD. Analysis of cancer cells revealed enhanced cellular plasticity and loss of AT2 cell identity in the Solid group, with HMGA1 identified as a key regulator of these processes. Collectively, our findings provide novel insights into the TME and cancer cell biology, shedding light on the mechanisms underlying the aggressive behavior and poor prognosis associated with the solid subtype of LUAD.

Several studies have reported that the solid subtype of LUAD is associated with T cell exhaustion [10, 14, 15]. The extent of T cell exhaustion is directly influenced by both the quantity of antigens and the duration and frequency of T cell receptor stimulation [56, 57]. Therefore, the increased antigenicity associated with higher TMB [8, 9] may be a primary cause of CD8⁺ T cell exhaustion in the Solid group compared with other groups. Beyond antigenic stimulation, tumor microenvironmental factors such as hypoxia, acidity, and metabolic reprogramming also contribute to CD8⁺ T cell dysfunction [15, 58–60]. In our study, we found that elevated cholesterol levels in the TME and increased macrophage cholesterol efflux may further exacerbate CD8⁺ T cell dysfunction in solid-subtype LUAD. The role of cholesterol in T cell activation and function is complex and context-dependent [33, 61–64]. Cholesterol is essential for T cell activation and proliferation, as cholesterol deficiency drives T cell dysfunction [63], whereas increased cholesterol uptake has been associated with enhanced T cell activation in tumors responsive to immune checkpoint inhibitor therapy [64]. Conversely, accumulating evidence indicates that excess cholesterol can drive tumor-infiltrating CD8 + T cells into a dysfunctional and exhausted state [33, 65], which is in agreement with our findings. These observations suggest that context-specific modulation of cholesterol may represent a promising strategy to enhance T cell-based cancer immunotherapy by converting an immunosuppressed tumor into an immunologically responsive one. For example, targeting cholesterol metabolism by reeducating TAMs could alleviate immune suppression and act synergistically with immune checkpoint inhibitors [66]. However, clinical trials of statins in various cancer types have so far failed to demonstrate clear benefit, likely because a systemic approach does not account for the nuanced, cell type-specific roles of cholesterol [67, 68]. Thus, a deeper understanding of cholesterol metabolism and its functional consequences in immune-effector and immunosuppressive cells within the TME will be crucial for designing effective cholesterol-targeted cancer immunotherapies.

The recently introduced IASLC grading system for LUAD [3] defines poorly differentiated tumors as those containing 20% or more high-grade patterns (solid, MP, or complex gland). Although a high proportion of these distinct high-grade patterns defines poor differentiation, it remains unclear whether cancer cells evolve from lower-grade patterns into high-grade patterns during LUAD progression. Our single-cell analysis of cancer cells provided evidence of a transition from a well-differentiated state, characterized by a high AT2 signature, to a poorly differentiated state marked by loss of AT2 cell identity—commonly observed in solid-subtype cancer cells. These findings suggest that lower-grade LUAD subtypes, such as acinar or papillary, may evolve toward the solid subtype over time. Intriguingly, we found that A/P group tumors containing a minor solid component exhibited increased intratumoral genomic and transcriptomic heterogeneity, with a tendency to progress toward a more aggressive phenotype compared with A/P group tumors lacking any solid component. These findings may explain why LUADs with even a minimal solid component have been associated with clinical characteristics and poor prognosis similar to those of solid- predominant LUADs, underscoring the importance of evaluating minor solid component for accurate prognostic stratification in LUAD. Patients with any solid component may warrant more intensive adjuvant therapy and closer surveillance, even in the presence of otherwise favorable clinicopathologic features.

We also identified HMGA1 as a putative key regulator of cancer cell plasticity and differentiation in LUAD. HMGA1 is a chromatin-binding protein that is normally expressed during embryogenesis and in adult stem cells but silenced in most differentiated cells [69–71]. It is overexpressed in many cancers, where elevated HMGA1 expression correlates with poor differentiation, aggressive clinicopathologic features, and unfavorable outcomes [72–74]. Preclinical studies have suggested that HMGA1 represents an attractive therapeutic target, although it has not yet been clinically drugged. Knockdown of HMGA1 in LUAD cells suppresses glycolytic flux and tumor growth [75], highlighting that disrupting HMGA1 function may attenuate tumor aggressiveness. In our study, the clinical relevance of HMGA1 in LUAD was further supported by its strong association with aggressive clinicopathologic features and decreased survival, suggesting its potential utility as both a prognostic biomarker and a therapeutic target. Given the high postsurgical recurrence rate associated with the solid subtype, HMGA1-targeted strategies merit exploration as adjuvant therapies. Our findings also raise the possibility that serum HMGA1 levels could serve as a non-invasive biomarker for detecting aggressive solid components, particularly in inoperable or advanced-stage LUADs where histologic subtype assessment is limited by small biopsy specimens. Therefore, the identification of HMGA1 highlights an opportunity to tailor novel targeted therapeutic strategies according to histologic subtypes in advanced LUAD. Furthermore, intercellular interaction analysis revealed distinct L–R pairs in the Solid group that drive the aggressive phenotype of cancer cells and shape the immunosuppressive TME. Notably, immunosuppressive and cholesterol efflux-associated intercellular interactions were significantly enriched in the Solid group.

Despite the well-established association between the MP subtype of LUAD and poor prognosis, identifying MP-specific transcriptional patterns in scRNA-seq data was challenging. This is partly due to the limited presence of the MP pattern, which comprises only 20–30% of the tumor area even within the MP group. However, our study identified a distinct regulon specifically activated in MP group cancer cells, prominently featuring the transcription factor HOXB7 and its target genes (Fig. S11A and B). HOXB7 has been linked to poor prognosis in various tumor types, consistent with our findings [76, 77]. Despite our limited sample size, we identified HOXB7 as a potential MP-specific cancer cell marker and further validated its association with poor survival outcomes in LUAD patients using TCGA data (Fig. S11C). These findings underscore the need for further exploration of HOXB7 in MP subtype of LUAD. Given the localized nature of MP pattern within the LUAD tumors, high-resolution spatial transcriptomics technologies integrated with scRNA-seq should be employed across a larger cohort of patients to achieve a more comprehensive understanding of the MP subtype.

This study has several limitations. First, the cohort size was modest and derived from a single institution, which may reduce statistical power and limit the generalizability of the results. Second, as an observational study based primarily on transcriptomic profiling, our analyses identify correlative relationship rather than direct causal mechanisms; therefore, functional validation will be required to establish direct biological roles. Third, intrinsic limitations of scRNA-seq, Visium, and multiplex IHC—including dissociation-related bias, spot-level spatial resolution, and manual ROI selection from a limited number of samples—may restrict complete characterization of tumor heterogeneity. Finally, LUAD exhibits substantial spatial heterogeneity, and the discrete tissue regions used for single-cell and spatial profiling may not fully overlap with the entire tumor area evaluated for histologic subtype assessment on whole-slide pathology, introducing inevitable sampling bias. Despite these limitations, the consistency of results across multiple platforms supports the robustness of our findings. In summary, our study provides a comprehensive characterization of LUAD histologic subtypes, revealing immunosuppressive phenotypes, enhanced cellular plasticity, and distinct intercellular interactions that contribute to the aggressiveness and poor prognosis of the solid subtype. These findings underscore the importance of evaluating even minor solid components and identify potential therapeutic targets for subtype-specific treatment strategies. Future studies should aim to validate these findings in larger cohorts and to investigate the therapeutic efficacy of targeting identified pathways in preclinical and clinical settings.

Materials and methods

Human sample collection

Human lung tumor specimens and clinical information were provided by the National Biobank of Korea—Kyungpook National University Chilgok Hospital (NBK-KNUCH). A collection of 18 fresh LUAD tissue samples for scRNA-seq analysis and one additional sample for Visium profiling, as well as 83 archived formalin-fixed paraffin-embedded samples from surgically resected LUAD for IHC were obtained. In addition, 24 fresh and 78 frozen LUAD tissue samples were also obtained for flow cytometry analysis and cholesterol content measurement, respectively. Comprehensive histologic subtyping based on the 2011 IASLC/ATS/ERS classification was performed at the time of pathological diagnosis of surgical specimens (JYJ) and reviewed by two experienced pathologists (JYJ and MK) prior to analysis. Eighteen serum samples from LUAD patients were also provided by the NBK-KNUCH for the measurement of HMGA1 protein level. All materials derived from the NBK-KNUCH were obtained from patients at the time of surgery under institutional review board-approved protocols. Informed written consent was obtained from all patients and the study protocol was approved by the institutional review board of KNUCH (KNUCH 2020–07-025).

Preparation of single-cell suspensions

The tumor tissues were collected at operation rooms and transported in a cryotube on ice to the research facility within an hour. Fresh tissue samples were cut into small slices and enzymatically digested in the RPMI-1640 (Corning, NY, USA) with Liberase™ (Roche Diagnostics, Indianapolis, IN, USA) at 37 °C for 1 h. Samples were partially dissociated by running the ‘h_tumor’ program on a gentleMACS Dissociator (Miltenyi Biotec, Bergisch Gladbach, Germany). The digested tissue homogenates were filtered through a 500-μm, 250-μm, 70-μm, 40-μm strainer. Following centrifugation at 300 × g for 7 min, the cell pellet was resuspended in red blood cell lysis buffer (Miltenyi Biotec, Bergisch Gladbach, Germany) and allowed to sit at 4℃ for 10 min. After washing with chilled wash buffer (1 × DPBS containing 0.04% BSA), 10 µL of this cell suspension was counted using an automated cell counter (Luna, Logos Biosystems, Gyeonggi-do, South Korea) to determine the concentration of live cells. The cell suspension was transported to the POSTECH for library preparation on the same day that each sample was acquired at the operation room.

Library preparation for single-cell RNA sequencing

The libraries for the single-cell transcriptome were prepared using the Chromium Next GEM Single Cell 3′ Reagent kits (10x Genomics, Pleasanton, CA, USA). Briefly, qualified cell suspensions were washed twice with 0.04% BSA in PBS and concentrated to a density of 7 × 105 ~ 1.6 × 106/mL. The cells were filtered through a 70 µm strainer, counted and combined with a reverse transcription enzyme mixture to capture 10,000–12,000 cells per channel. Single-cell gel bead-in-emulsions (GEMs) were generated by loading the cell-enzyme mixture, gel beads, and partitioning oil onto the Chromium Next GEM Chip G (10x Genomics, PN-1000127) using the Chromium Controller (10x Genomics). Single-cell RNA barcoding was performed through reverse transcription within each GEM, followed by cDNA amplification and library construction using the Chromium Next GEM Single Cell 3’Kit (10x Genomics, PN-1000269) and Dual Index Kit TT Set A (10x Genomics, PN-1000215) according to the manufacturer’s instructions. Libraries were sequenced on the HiSeq X platform (Illumina, San Diego, CA, USA) with 100 bp paired-end reads, targeting a depth of minimum 20,000 read pairs per cell.

Spatial transcriptomics (10x Visium)

Freshly obtained tumor tissue samples were snap frozen in isopentane and liquid nitrogen bath, then embedded in Tissue-Tek OCT Compound (Sakura Finetek, Tokyo, Japan) on dry ice, then stored at − 80 °C in a sealed container for later use. For gene expression, VISIUM Spatial Tissue Optimization and Visium spatial gene expression was processed using a kit (10 × Genomics, Stoneridge Mall Road Pleasanton, CA, USA). The OCT-embedded tissue was cryosectioned to a thickness of 10 μm and mounted on the slide. Tissues were permeabilized for 24 min, based on tissue optimization time course experiments. The resulting barcoded cDNA obtained from mRNA was transferred from the slides, and amplified. cDNA quantification was performed using the Agilent Bioanalyzer High Sensitivity Kit on an Agilent Bioanalyzer 2100 (Agilent Technologies, CA, USA).

Immunohistochemistry(IHC) assay

Immunohistochemical staining for HMGA1 (1:1000; Abcam, Cambridge, UK) was conducted on the formalin-fixed paraffin-embedded tumor sections using Ventana BenchMark XT autoimmunostainer (Roche Ventana, Tucson, AZ, USA) according to the manufacturer’s instruction. Each section was cut from paraffin blocks in 4-μm thickness. Tissue sections were deparaffinized in xylene and pretreated with conditioning solution for antigen retrieval. The incubation time for HMGA1 antibody was 30 min. Standard Ventana signal amplification, counterstaining with hematoxylin, and staining with a bluing reagent were followed. Lastly, stained slides were mounted with cover slides and examined by microscope.

Multiplex IHC

Multiplex IHC staining, scaning and analysis were performed at prismCDX Co.,Ltd. (Gyeonggi-do, Korea). 4-μm sections of specimens were cut from formalin-fixed paraffin-embedded (FFPE) blocks. Slides were heated for at least one hour in a dry oven at 60℃, then followed by multiplex IHC staining with a Leica Bond Rx™ Automated Stainer (Leica Biosystems, Wetzlar, Germany). The list of the antibody and fluorophore used is summarized in the Table S7. Briefly, the slides were dewaxed with Leica Bond Dewax solution (Leica Biosystems, Wetzlar, Germany), followed by antigen retrieval with Bond Epitope Retrieval 2 (Leica Biosystems, Wetzlar, Germany) for 30 min. The staining proceeds in sequential rounds of blocking buffer (TheraNovis, Bingen a. Rhein, Germany), followed by primary antibody incubation for 30 min and Mouse/Rabbit HRP secondary antibody (TheraNovis, Bingen a. Rhein, Germany) incubation for 10 min. Visualization of antigen was accomplished using Astra-dye (TheraNovis, Bingen a. Rhein, Germany) for 10 min, after which the slide was treated with Bond Epitope Retrieval 1 (Leica Biosystems, Wetzlar, Germany) for 20 min to remove bound antibodies before the next step in the sequence. The process from the blocking step to the antigen retrieval step is repeated for every antibody staining. Nuclei were stained with DAPI (Thermo Scientific, Waltham, MA, USA) for counterstaining after the last round of antigen retrieval. The slides were coverslipped using ProLong Gold antifade reagent (Invitrogen, Carlsbad, CA, USA).

Multispectral imaging and analysis

Multiplex stained slides were scanned using the PhenoImager™ HT (Akoya Biosciences, Marlborough, MA, USA) at 20 × magnification. Representative images for training were selected in Phenochart™ Whole Slide Viewer (1.1.0 version, Akoya Biosciences, Marlborough, MA, USA), and an algorithm was created in the inForm® Tissue Analysis software (2.6 version, Akoya Biosciences, Marlborough, MA, U SA). Multispectral images were unmixed using the spectral library in inForm software, and tumor tissues were segmented according to the presence or absence of CK antibody expression. Based on DAPI staining, each single cell was segmented and phenotyping was performed according to the expression compartment and intensity of each marker. After designating the ROIs to be analyzed on the tissue slide, the same algorithm created in this way was applied and batch-running. A total of 20 ROIs (10 in tumor nest and 10 in peritumoral stroma) per sample were manually selected using Phenochart software by the pathologists who were blinded to downstream measurements and outcomes. In the tumor nest, a fixed-size stamp (931 × 698 μm) with an area of 0.65 mm² was used and 10 ROIs were selected to evenly cover the whole tumor area. Given the uneven stromal distribution of immune cells, 10 ROIs in the peritumoral stroma were intentionally sampled from immune-cell–dense hotspots by drawing (average area of 0.59 mm² per ROI) to capture functional immune niches relevant to antitumor immunity. The number of individual cell phenotypes in each ROI was expressed as density per mm². The exported data is consolidated and analyzed in R studio (4.2.1 version) using the phenoptr (Akoya Biosciences, Marlborough, MA, USA) and phenoptrReport (Akoya Biosciences, Marlborough, MA, USA) packages.

ELISA assay

Serum HMGA1 levels were quantified using a commercial ELISA kit (MyBioSource, MBS2024012) with a detection range of 0.312–20 ng/mL and a sensitivity (lower limit of detection, LLD) < 0.112 ng/mL. Values between 0.112 and 0.312 ng/mL were included in statistical analyses, while a value below 0.112 ng/mL was considered below the limit of detection and excluded from quantitative analysis. Enzyme immunoassay was carried out following manufacturer’s instructions. HMGA1 level was measured using Multiskan SkyHigh Microplate Reader (Thermo Scientific, Waltham, MA, USA) at 450 nm. All analyses were performed using SPSS (version 25.0).

Flow cytometry analysis

Single-cell suspensions from lung tumor tissues were washed in PBS and then stained with FcR blocking reagent (Miltenyi Biotec, Bergisch Gladbach, Germany). Cells stained for 30 min at 4 °C in the dark with conjugated antibodies specific for the cell surface antigens. The cell populations were defined as follows: Live/Dead – APC Cy7 (Invitrogen, Carlsbad, CA, USA), CD3-FITC, CD8-PerCP-Cy5.5, CD279-APC (BioLegned, San Diego, CA, USA). The stained cells were washed twice with FACS buffer, resuspended in 300μL of buffer, and detected for surface markers using flow cytometry (BD FACS Aria III, BD Biosciences, Franklin Lakes, NJ, USA). Data were analyzed with FlowJo software.

Cholesterol content measurement

For quantification, lysates from tumor tissues were prepared with a TissueLyser II (Qiagen, Maryland, MD, USA) and cholesterols were extracted using a commercially available kit (Abcam, Cambridge, MA, USA). Cholesterols were resuspended in 50 μl of reaction buffer and measured as recommended by the Amplex® Red Cholesterol Assay kit (Invitrogen, Carlsbad, CA, USA). Cholesterol levels were normalized to tissue weight and expressed as cholesterol content per milligram of tissue.

Bioinformatics analysis

All bioinformatics analyses are described in detail in the Supplementary Methods.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

SYL and JKK conceived the study and designed the experiments. JL, JYJ, SYL, and JKK analyzed the scRNA-seq and spatial transcriptomics data. MJH, YHC, JYK, JHL, JEC performed the experiments. JYJ, MK, YWD, EBL, SHC, SSY, JYP, SYL contributed to acquisition of tumor samples and clinicopathological data analysis. JL, JYJ, MJH, SYL, and JKK wrote the original draft, with discussion and feedback from all co-authors. JL, SYL, and JKK reviewed and edited the manuscript.

Funding

This work was supported by the National Research Foundation of Korea (NRF) grants funded by the Korean government (MSIT) (Grant No.: RS-2023-NR077247 and RS-2023-00221112 to JKK, RS-2023-NR076546 and RS-2020-NR049556 to SYL), and by the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (Grant No. RS-2025-25410994 to SYL).

Data availability

The raw scRNA-seq and spatial transcriptomics data for all samples generated in this study were deposited in the Korea Sequence Read Archive (KRA, https://kbds.re.kr/KRA) at the Korea Bioinformation Center, Korea Research Institute of Bioscience and Biotechnology, under the accession number KAP241022. The gene expression count matrices with associated metadata are available via CELLxGENE (https://cellxgene.cziscience.com/collections/0bebef1a-4607-4584-9070-dacf89a0d635), and all reproducible code is provided at GitHub (https://github.com/CB-postech/luad_histologic_subtypes).

Code availability

The raw scRNA-seq and spatial transcriptomics data for all samples generated in this study were deposited in the Korea Sequence Read Archive (KRA, https://kbds.re.kr/KRA) at the Korea Bioinformation Center, Korea Research Institute of Bioscience and Biotechnology, under the accession number KAP241022. The gene expression count matrices with associated metadata are available via CELLxGENE (https://cellxgene.cziscience.com/collections/0bebef1a-4607-4584-9070-dacf89a0d635), and all reproducible code is provided at GitHub (https://github.com/CB-postech/luad_histologic_subtypes).

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with ethical standards and national and international guidelines. This study was approved by the institutional review board of Kyungpook National University Chilgok Hospital (KNUCH 2020-07-025). The biospecimens and data used for this study were provided by the Biobank of Korea-Kyungpook National University Hospital (KNUH), a member of the Korea Biobank Network. All materials derived from the National Biobank of Korea-KNUH were obtained with informed consent under institutional review board (IRB)-approved protocols.

Competing interests

The authors declare no competing interests.