Mechanisms and therapeutic targets in distant metastasis of lung adenocarcinoma

Abstract

Distant metastasis, characterized by organotropism, is a major cause of mortality in lung adenocarcinoma (LUAD). In this study, digital spatial profiling (DSP), multiplex immunofluorescence (mIF), and clinical data from 52 LUAD patients were integrated to develop organ-specific metastasis risk models, and the molecular mechanisms underlying metastatic organotropism were investigated. Random forest models based on primary tumor spatial transcriptomics accurately predicted metastasis to the brain (AUC = 0.974), liver (AUC = 0.975), adrenal gland (AUC = 0.929), and bone (AUC = 0.907). Key compartment-specific gene expression signatures associated with organotropic metastasis were identified, including those expressed in tumors (e.g., FKBP1A for the brain and MOCOS for the liver), immune cells (e.g., ADAMTSL2 for the liver), and stromal cells (e.g., CKAP2 for the brain). Pathway analyses revealed distinct biological processes associated with organotropism, such as enriched cell death pathways in brain metastasis and extracellular matrix (ECM) remodeling in liver metastasis. Postmetastasis survival models highlight stromal gene expression (e.g., PKM for OS and VCAM1 for PFS) and immunosuppressive microenvironments (e.g., M2 macrophage infiltration) as critical prognostic factors. The high-precision prediction models and key molecular signatures identified in this study enhance our understanding of “seed–soil” interaction dynamics and offer promising biomarkers and therapeutic targets for future clinical use.

Introduction

Distant metastasis is a major contributor to cancer-related mortality and therapeutic failure. The development of metastases typically signifies progression to advanced-stage disease and is strongly associated with poor prognosis and compounded treatment challenges. Epidemiological evidence has shown that metastatic disease contributes to more than 90% of cancer-related deaths.¹ Metastasis is not a stochastic event but a highly coordinated, multistep process known as the metastatic cascade. This cascade comprises three key stages: local invasion, intravasation into the circulation, and eventual colonization of distant organs.² As early as 1889, the British surgeon Stephen Paget proposed the seminal “seed and soil” hypothesis through his observations of breast cancer metastasis patterns, highlighting that metastatic colonization critically depends on the interaction between tumor cells (seeds) and the microenvironment of specific distant organs (soil).^3,4 This classical theory provides a foundational framework for understanding metastasis and has been continually expanded and refined in the era of modern molecular biology.

From the “seed” perspective, successful metastasis requires that tumor cells exhibit remarkable plasticity and adaptability. During the invasion phase, tumor cells acquire migratory and invasive capabilities through epithelial‒mesenchymal transition (EMT), enabling them to degrade and breach the basement membrane, thereby entering blood or lymphatic vessels.⁵ Once circulating tumor cells (CTCs) enter the circulation, they must resist anoikis⁶ and evade immune surveillance and destruction. Ultimately, a minute fraction of CTCs reach distant organs, where they extravasate and adapt to the new microenvironment through processes such as mesenchymal‒epithelial transition (MET), leading to proliferation and the formation of micrometastases.⁷

On the other hand, the “soil”, which refers to the microenvironment of the target organ, also plays an active role in the entire metastatic process. Recent studies have greatly advanced our understanding of the “soil”, revealing that primary tumors can remotely precondition distant organs via secreted factors (e.g., exosomes and cytokines) to form a premetastatic niche (PMN) that facilitates the colonization of CTCs.^8–10 PMN formation is a critical preparatory event preceding metastasis and involves multiple mechanisms, such as immunosuppressive cell infiltration, extracellular matrix remodeling, enhanced vascular permeability, and metabolic reprogramming,¹¹ thereby creating a favorable ecological niche for metastatic cells.

Moreover, the process of metastasis is characterized by a distinct pattern of organotropism. This refers to the consistent and nonrandom predisposition of particular cancers to metastasize to specific organs. This nonrandom pattern of migration is dictated by intricate interactions between tumor cells and the microenvironment of distant organs, with underlying molecular mechanisms that have been increasingly elucidated.¹² Central to this process are chemokine–receptor signaling axes, which play pivotal roles in directing tumor cell homing. For example, the CXCR4/CXCL12 and CCR7/CCL21 signaling axes promote the directed migration of breast cancer cells toward organs with high local chemokine expression, such as regional lymph nodes and lungs, through chemotactic gradients.¹³ Second, the specific interaction between adhesion molecules on the surface of tumor cells (e.g., integrins and CD44) and their corresponding ligands expressed on the endothelium of blood vessels in distant organs (e.g., VCAM-1 and selectins) facilitates tumor cell adhesion and retention at extravasation sites.¹⁴ Furthermore, the physical anatomical structure of organs contributes to this process. Physical interactions between cancer cells and their microenvironment, along with the regulatory role of mechanical forces, are also key determinants of metastasis.¹⁵ Recent studies have further revealed that metabolic adaptation between the metabolic properties of tumor cells and the unique nutrient microenvironment of distant organs serves as another core mechanism driving organotropic metastasis. For instance, pancreatic cancer cells “decide” whether to metastasize to the liver or the lungs on the basis of the expression level of their PCSK9 gene and the availability of cholesterol in the microenvironment of distant organs.¹⁶ Recent spatial omics studies have demonstrated that the spatial dialog between “seeds” and “soils” constitutes a key mechanism underlying organotropism. For example, spatial transcriptomic analyses of pancreatic cancer metastases revealed that aggressive “basal-like” cancer cells are consistently spatially coupled with TGFB1-expressing myofibroblastic cancer-associated fibroblasts (myCAFs), driving immune exclusion via the CXCR4-CXCL12 signaling axis.¹⁷ Similarly, a recent integrated spatial transcriptomic and metabolomic analysis of non-small cell lung cancer (NSCLC) brain metastases revealed a spatially localized, metastasis-initiating cell cluster (LOX⁺ Malig-5) that colocalizes with a specific neutrophil niche, driving metastasis through a spatially defined metabolic reprogramming circuit.¹⁸

Although these mechanisms offer valuable insights into organotropism, many fundamental questions in this field remain unanswered. For instance, why do distinct subpopulations of cancer cells from the same primary tumor preferentially metastasize to different organs? How is the organ microenvironment dynamically remodeled to facilitate the colonization and subsequent proliferation of metastatic cells? What are the key transcriptomic features that influence patient survival following metastasis? Furthermore, predictive models and therapeutic approaches targeting organ-specific metastasis remain inadequate. There is an urgent need to decode the spatiotemporal regulatory networks of organotropism using integrated multiomics analyses, advanced in vitro models (e.g., organ-on-a-chip), and real-time imaging techniques.^19,20

Given these challenges, this study aims to systematically explore the mechanisms and prediction of distant metastasis in lung adenocarcinoma (LUAD). Common sites of LUAD metastasis include the brain, liver, bones, and adrenal glands.²¹ However, reliable models that can accurately predict organ-specific metastatic risk on the basis of primary tumor characteristics are lacking. Moreover, the heterogeneous responses of LUAD metastases to standard therapies and their underlying mechanisms remain poorly understood. We hypothesize that despite the complexity of the metastatic cascade, organotropism and therapeutic response patterns are embedded within the genomic and tumor microenvironmental landscapes of cancer cells. Thus, systematic multiomics profiling of primary tumors may facilitate the development of clinically useful predictive models for metastasis. In parallel, in-depth molecular analysis of metastatic lesions could reveal critical mechanisms affecting patient survival and treatment susceptibility, potentially revealing novel therapeutic opportunities.

Here, we assembled a cohort of primary and matched distant metastatic lesion tissues from patients with advanced LUAD and employed digital spatial profiling (DSP) technology²² to conduct high-resolution molecular phenotyping across functionally distinct tumor regions—including tumor cell nests, immune cell zones, and stromal compartments. By integrating spatial multiomics profiles with comprehensive clinical data, we developed organ-specific metastasis prediction models and postmetastatic survival prognostic models for LUAD. These models are intended to provide more accurate tools for predicting tumor biological behavior in clinical practice and to support the discovery of novel therapeutic targets, ultimately informing improved treatment strategies for advanced LUAD patients.

Results

Study design

This study included a total of 52 LUAD patients. Matched samples were collected from 23 pairs of brain metastases and corresponding primary tumors, 26 pairs of liver metastases and primary tumors, and 3 pairs of adrenal metastases and primary tumors. Sufficient adjacent normal tissue samples were also collected from each metastatic group to serve as controls. Multiplex immunofluorescence (mIF) staining was performed on each sample, and three distinct regions of interest (ROIs)—tumor, immune, and stromal compartments—were defined for each sample prior to DSP analysis. The data obtained were compiled into an expression matrix for subsequent analysis.

Clinically, for patients without distant metastasis, only primary tumor samples are available for predicting the risk of future metastasis. On the basis of these observations, we hypothesized that the transcriptomic profile of the primary tumor encodes organotropism for metastasis. By integrating transcriptomic data from the primary tumors of all 52 patients with corresponding clinical information, we constructed random forest models to predict the risk of metastasis to the brain, liver, bones, and adrenal glands. We subsequently investigated the key differentially expressed genes (DEGs) and their interrelationships between metastatic and primary foci through the integration of DSP platform data.

Postmetastatic survival represents another critical clinical concern. We hypothesized that genes influencing postmetastatic survival are not randomly distributed but rather are enriched among the DEGs between primary and metastatic tumors. Potential key genes were identified through paired Wilcoxon test analysis of DSP data from matched tumor samples and were further validated using random forest survival models.

Furthermore, on the basis of the results of mIF staining, this study analyzed the correlation between the proportions of various immune cell types within the tumor immune microenvironment (TIME) and the risk scores generated by the predictive models.

Finally, we proposed a novel metric termed the HEindex, derived from H&E staining, defined as the average proportion of nontumor cells surrounding each tumor cell. The association between the HEindex and the model-derived risk scores was also evaluated.

A flowchart summarizing the study design is presented in Fig. 1.

Fig. 1.

Study flow chart

Risk model and identification of key signatures for brain metastasis in LUAD

Using a random forest model, we identified the genes most strongly associated with the risk of brain metastasis in LUAD and determined their cellular localization within the tumor microenvironment. These key genes include FKBP1A and MRPS33, which are expressed in tumor cells; AFAP1L1, ELP6, REC8, PRRG4, MS4A15, CCDC92, and CDCA8, which are expressed in immune cells; and CKAP2, which is expressed in stromal cells. The importance rankings of these genes in the random forest model and their differential expression between the brain metastasis and nonmetastasis groups are presented in Fig. 2a. The brain metastasis risk prediction model based on these genes showed excellent performance in the test cohort, with an area under the curve (AUC) of 0.974 (95%CI: 0.903–1.000), indicating strong discriminative ability (Fig. 2b and Table S1).

Fig. 2.

Brain metastasis risk model and differential gene expression between brain metastases and primary lung tumors. a Feature importance heatmap of the brain metastasis risk model. b ROC curve of the brain metastasis risk model. c GSEA1350 pathway enrichment analysis based on all genes ranked by log2FC between the metastatic and nonmetastatic groups or on all genes ranked by Spearman correlation with the model score. d Comparison of immune cell infiltration proportions in primary lung tumors between the brain metastasis group and the nonmetastasis group (Wilcoxon rank-sum test). e Immune cell subsets in primary lung tumors associated with the model risk score. f Spearman correlation analysis between the HEindex and the model risk score. g Venn diagram identifying key overlapping genes. h Enrichment analysis of key genes using the STRING database. i Validation and spatial localization of key genes using DSP data. j Gene regulatory network. Meta metastasis, NA not available

Next, we conducted gene set enrichment analysis (GSEA) on DSP data from different ROIs in primary lung lesions. GSEA based on actual metastasis status revealed that in the tumor and stromal compartments of patients with brain metastasis, pathways related to ECM and metastasis (47/129, 36.4% in tumor; 40/129, 31.0% in stroma), cell death (21/72, 29.2% in tumor; 19/72, 26.4% in stroma), and immunity (111/418, 26.6% in tumor; 76/418, 18.2% in stroma) were significantly enriched. Significant enrichment of cell death (23/72, 31.9%) and immunity (54/418, 12.9%) was also observed in the immune compartment. Further GSEA based on model prediction scores revealed that a high risk of brain metastasis was significantly associated with enrichment of cell death (41/72, 56.9%) pathways in tumor cells and enrichment of cell death (39/72, 54.2%), immune (154/418, 36.8%), genetic and epigenetic information (104/286, 36.4%), and cell cycle (39/127, 30.7%) pathways in the immune compartment and enrichment of cell death (38/72, 52.8%) and immune (127/418, 30.4%) pathways in the stromal compartment (Fig. 2c).

Using mIF staining of primary tumor samples from patients with and without brain metastasis, we compared the infiltration levels of various immune cells between the two groups. Wilcoxon rank-sum tests revealed no significant differences in immune cell infiltration proportions between the two groups (Fig. 2d). Further correlation analysis revealed that the proportions of cancer-associated fibroblast type I (CAFI) and normal fibroblasts (NF) were significantly positively correlated with the model risk score (Fig. 2e). However, no significant correlation was observed between the HEindex derived from H&E staining and the model risk score (Fig. 2f).

To gain deeper insights into the core molecular differences between brain metastases and primary lung tumors, we applied a multitiered screening strategy. First, by integrating differential gene expression from three bulk-level comparisons—metastatic brain tumor vs. normal brain tissue, primary lung tumor vs. normal lung tissue, and metastatic brain tumor vs. primary lung tumor—and taking their intersection, we identified a core set of 78 DEGs (Fig. 2g). Enrichment analysis was subsequently conducted on these 78 genes through the Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) website (https://cn.string-db.org/), and pathways related to protein folding and cell signal transduction were significantly enriched (Fig. 2h). To further understand the function of the 78 DEGs, spatial information from the DSP platform data was introduced into the analysis, ultimately identifying 100 key DEGs with precise spatial localization (Fig. 2i), and the results of pathway enrichment analysis based on the upregulated and downregulated key DEGs of the tumor, immune and stromal compartments are listed in the Supplementary Material (Comprehensive Data and Analysis Results). In the tumor compartment, pathways related to protein folding were upregulated, and pathways related to interferon were downregulated. In the immune compartment, pathways related to protein folding were also upregulated, and pathways related to ribosomes were downregulated. In the stromal compartment, pathways related to neural development were upregulated, and pathways related to collagen were downregulated. On the basis of this set, we constructed a Spearman correlation coefficient-weighted gene regulatory network to explore the interactions among these key genes during brain metastasis (Fig. 2j). This network elucidates potential gene regulatory interactions across distinct tumor compartments, with altered expression levels observed for key regulatory genes as follows: tumor cells exhibited key changes, such as upregulation of HNRNPA2B1, H3C13 and CYCS and downregulation of BST2 and COL3A1; immune cells exhibited upregulation of HSPA8, HSP90AA1, HNRNPA2B1 and HSP90AB1 and downregulation of RPL37, COL3A1, and RPS2; and stromal cells primarily showed downregulation of RPS2, COL3A1, COL5A2, SERPINH1, RPL37, and IFITM1 and upregulation of CCT6A.

Risk model and identification of key signatures for liver metastasis in LUAD

The genes most strongly linked to liver metastasis risk in LUAD are MOCOS, PDGFRB, PDLIM1, HSPA6, RNF11, and LCN2 (tumor cell-specific), along with ADAMTSL2, SPCS1, IGLL5, and CASZ1 (mainly expressed in immune cells). The importance rankings of these key genes in the random forest model and their differential expression between the liver metastasis group and nonliver metastasis group are presented in Fig. 3a. A liver metastasis risk prediction model based on these genes exhibited excellent performance in the test cohort, with an AUC of 0.975 (95%CI: 0.906–1.000), indicating a high level of discriminative accuracy (Fig. 3b and Table S2).

Fig. 3.

Liver metastasis risk model and differential gene expression between liver metastases and primary lung tumors. a Feature importance heatmap of the liver metastasis risk model. b ROC curve of the liver metastasis risk model. c GSEA1350 pathway enrichment analysis based on all genes ranked by log2FC between the metastatic and nonmetastatic groups or on all genes ranked by Spearman correlation with the model score. d Comparison of immune cell infiltration proportions in primary lung tumors between the liver metastasis group and the nonmetastasis group (Wilcoxon rank-sum test). e Immune cell subsets in primary lung tumors associated with the model risk score. f Spearman correlation analysis between the HEindex and the model risk score. g Venn diagram identifying key overlapping genes. h Enrichment analysis of key genes using the STRING database. i Validation and spatial localization of key genes using DSP data. j Gene regulatory network. Meta metastasis, NA not available

GSEA results of transcriptomic data from primary tumors of patients without liver metastasis revealed significant enrichment of pathways related to cell death (39/72, 54.2%), metabolism and energy (65/315, 20.6%), and genetic and epigenetic information (59/286, 20.6%) within tumor cells. In the immune compartment, pathways associated with cell death (26/72, 36.1%), immunity (81/418, 19.4%), and genetic and epigenetic information (43/286, 15.0%) were notably enriched. Within the stromal compartment, pathways related to cell death (26/72, 36.1%) and immunity (74/418, 17.7%) were most prominently enriched. Further GSEA based on model prediction scores revealed that a low risk of liver metastasis was significantly associated with the enrichment of cell death pathways in all compartments (21/72, 29.2% in tumors; 29/72, 40.3% in the stroma; and 33/72, 45.8% in the immune cells) (Fig. 3c).

mIF revealed that the proportions of M1 macrophages and myeloid dendritic cells (mDCs) in the primary lung lesions of patients with liver metastasis were significantly greater than those in patients without liver metastasis (Fig. 3d). Correlation analysis further revealed that the proportions of both M1 macrophages and mDCs were significantly and positively correlated with the model risk score (Fig. 3e). However, no significant correlation was observed between the HEindex and the model risk score (Fig. 3f).

By integrating DEGs from three bulk-level comparison sets (metastatic liver tumor vs. normal liver tissue, primary lung tumor vs. normal lung tissue, and metastatic liver tumor vs. primary lung tumor) and identifying their intersection, we identified a core set of 72 DEGs (Fig. 3g). Enrichment analysis was subsequently conducted on these 72 genes through the STRING website, and pathways related to DNA replication-dependent chromatin assembly and nucleosomes were significantly enriched (Fig. 3h). To further understand the function of the 72 DEGs, spatial information from the DSP platform data was introduced into the analysis, ultimately identifying 133 key DEGs with precise spatial localization information (Fig. 3i), and the results of pathway enrichment analysis based on the upregulated and downregulated key DEGs in the tumor, immune and stromal compartments are listed in the Supplementary Material (Comprehensive Data and Analysis Results). In the tumor compartment, pathways related to nucleosomes were upregulated, and pathways related to organic acid binding were downregulated. In the immune compartment, pathways related to apoptosis were also upregulated, and pathways related to angiogenesis were downregulated. In the stromal compartment, pathways related to nucleosomes were also upregulated, and pathways related to angiogenesis were also downregulated. Finally, we constructed a Spearman correlation coefficient-weighted gene regulatory network to explore the interactions among these key genes during liver metastasis (Fig. 3j), revealing the mutual regulatory relationships among different compartments. In tumor cells, genes such as H3C15, H2AC19, H3C2, H4C12, H3C13, RPL36A, RPL38, and CCT5 were upregulated following liver metastasis, whereas genes such as LCN12 and SERPINA5 were downregulated. In the immune compartment, characteristic gene expression changes included the upregulation of RPL36A, NPM1, GSTP1, RPL38, UQCRHL, H3C13, and LDHB, whereas genes such as ECSCR and SERPINF1 were downregulated. In the stromal compartment, notable upregulation was observed for NPM1, H3C2, H3C13, H3C15, RPL38, RPL36A, H4C12, GSTP1, and SLC3A2, whereas genes such as RAMP2, ECSCR and SERPINF1 were downregulated.

Risk model for adrenal gland metastasis in LUAD

The genes most strongly associated with the risk of adrenal metastasis in LUAD include PTGR2 and IGKC, which are localized to tumor cells; ADIG, ASB3, SP5, and MLLT1, which are expressed in immune cells; and IRF3, MSLN, RPS26 and C19orf33, which are present in the stromal compartment. The importance rankings of these key genes in the random forest model and their differential expression between the adrenal metastasis and nonmetastatic groups are summarized in Fig. 4a. An adrenal metastasis risk prediction model constructed on the basis of these genes demonstrated excellent performance in the test cohort, with an AUC value of 0.929 (95%CI: 0.789–1.000), indicating high discriminative ability (Fig. 4b and Table S3).

Fig. 4.

Adrenal metastasis risk model and differential gene expression between adrenal metastases and primary lung tumors. a Feature importance heatmap of the adrenal metastasis risk model. b ROC curve of the adrenal metastasis risk model. c GSEA1350 pathway enrichment analysis based on all genes ranked by log2FC between the metastatic and nonmetastatic groups or on all genes ranked by Spearman correlation with the model score. d Comparison of immune cell infiltration proportions in primary lung tumors between the adrenal metastasis group and the nonmetastasis group (Wilcoxon rank-sum test). e Immune cell subsets in primary lung tumors associated with the model risk score. f Spearman correlation analysis between the HEindex and the model risk score. g Venn diagram identifying key overlapping genes. h Enrichment analysis of key genes using the STRING database. i Validation and spatial localization of key genes using DSP data. Meta metastasis, NA not available

The results of the GSEA of the transcriptomic data from the primary tumors of patients with adrenal metastasis revealed no significantly enriched pathways in tumor cells or immune compartments. However, within the stromal compartment of the metastasis group, pathways related to cell death (11/72, 15.3%) were notably enriched. Further GSEA based on model prediction scores also revealed that a high risk of adrenal metastasis was significantly associated with the enrichment of cell death (29/72, 40.3%)-related pathways in the stroma (Fig. 4c).

mIF analysis revealed that the proportion of mDCs in the primary lung lesions of patients with adrenal metastasis was significantly greater than that in patients without metastasis (Fig. 4d). Subsequent correlation analysis revealed that the proportion of plasmacytoid dendritic cells (pDCs) was significantly negatively correlated with the model risk score (Fig. 4e). Moreover, a significant positive correlation was observed between the HEindex and the model risk score (Fig. 4f).

Owing to the unavailability of normal lung tissue from patients with adrenal metastasis, we focused on the intersection of the two gene sets, which yielded 296 key DEGs (Fig. 4g). Enrichment analysis was subsequently conducted on these 296 genes through the STRING website, which revealed that pathways related to apoptosis and oxidoreductase were significantly enriched (Fig. 4h). Although DSP data were incorporated, resulting in the identification of 318 genes with precise spatial mapping (Fig. 4i), the small sample size prevented us from performing weighted regulatory network analysis.

Risk model for bone metastasis in LUAD

The genes most strongly associated with the risk of bone metastasis in LUAD include ARMCX6 and PLPP1, which are localized to tumor cells; DDIAS, POLE, C2CD5, and THAP7, which are expressed in immune cells; and B3GNT5, which is present in the stromal compartment. The importance rankings of these key genes in the random forest model and their differential expression between the bone metastasis and nonmetastasis groups are summarized in Fig. 5a. A bone metastasis risk prediction model constructed on the basis of these genes demonstrated excellent performance in the test cohort, with an AUC value of 0.907 (95%CI: 0.755–1.000), indicating strong discriminative ability (Fig. 5b and Table S4).

Fig. 5.

Bone metastasis risk model. a Feature importance heatmap of the bone metastasis risk model. b ROC curve of the bone metastasis risk model. c GSEA1350 pathway enrichment analysis based on all genes ranked by log2FC between the metastatic and nonmetastatic groups or on all genes ranked by Spearman correlation with the model score. d Comparison of immune cell infiltration proportions in primary lung tumors between the bone metastasis group and the nonmetastasis group (Wilcoxon rank-sum test). e Immune cell subsets in primary lung tumors associated with the model risk score. f Spearman correlation analysis between the HEindex and the model risk score. Meta metastasis, NA not available

GSEA of transcriptomic data from primary tumors revealed no significant differences in pathway enrichment between the bone metastasis group and the nonmetastasis group (Fig. 5c). These findings suggest that the propensity for bone metastasis may not be driven by the activation of specific pathways but rather by broader, more fundamental biological processes.

The mIF results revealed that the proportion of M1 macrophages in the primary lung lesions of patients in the bone metastasis group was significantly greater than that in the nonmetastasis group (Fig. 5d). However, subsequent correlation analysis revealed no significant association between the proportion of M1 macrophages and the model risk index score (Fig. 5e). Similarly, the HEindex was not significantly correlated with the model risk score (Fig. 5f).

Owing to the unavailability of bone metastasis lesion tissues from patients, we were unable to investigate differences in gene expression between metastatic and primary lesions.

Survival prediction models for LUAD patients following metastasis

Distant metastasis is often associated with a shorter survival period in LUAD patients; however, effective postmetastatic survival prediction models are lacking. We hypothesize that the key genes determining postmetastasis survival must be enriched among the DEGs between primary and metastatic lesions, as this would provide a biological explanation for how the metastatic event influences patient survival. Using the Wilcoxon signed-rank test on transcriptomic data from matched primary and metastatic tumor samples, we identified dysregulated genes whose expression significantly changed during metastatic transition and generated a corresponding volcano plot (Fig. 6a). The top 50 genes with the most statistically pronounced expression changes after metastasis are summarized in Fig. 6b.

Fig. 6.

Differential gene expression between paired primary and metastatic tumors and survival prediction models for LUAD patients after distant metastasis. a Wilcoxon signed-rank test results for paired comparisons of gene expression levels between metastatic and matched primary tumors. b Bar plot of the top 50 significant DEGs. c GSEA1350 pathway enrichment analysis based on all genes ranked by lnHR from univariate Cox regression for PFS and OS or on all genes ranked by Spearman correlation with the model score. d Immune cell subsets in metastatic tumors significantly associated with OS following the first distant metastasis. e Immune cell subsets in metastatic tumors significantly associated with PFS following the first distant metastasis. f Spearman correlation analysis between the HEindex and the survival model risk score. Note: For correlation analyses, “yes” indicates patients who reached the PFS or OS endpoint, “no” indicates patients who did not reach the endpoint, and “NA” indicates patients who were lost to follow-up

DEGs for which the raw P value was < 0.05 and the FDR was < 0.01 were considered to be statistically significant and were selected to construct subsequent random forest survival models. Next, we performed univariate Cox regression analysis to evaluate the association of each DEG with survival outcomes. We then constructed separate random forest models to predict overall survival (OS) and progression-free survival (PFS) in LUAD patients from the initiation of first-line therapy after distant metastasis. Ultimately, we identified key genes significantly associated with OS and PFS. The genes significantly linked to OS included ATP8A1_Immune, EPC1_Stroma, PKM_Stroma, DYNLL1_Stroma, FAM83H_Stroma, and EIF4H_Stroma. Those significantly associated with PFS included MDH1_Tumor, PSMA2_Tumor, RIPOR2_Tumor, VCAM1_Stroma, CYCS_Stroma, EPRS1_Stroma, TCF21_Stroma, and MFAP4_Stroma (Tables S5 and S6).

We subsequently conducted GSEA on the basis of the lnHR rankings of all genes in the Cox regression analysis to explore the relevant pathways that affect prognosis (Fig. 6c). GSEA based on OS revealed that pathways related to cell cycle regulation (55/127, 43.3% in tumors; 51/127, 40.2% in stroma) and genetic and epigenetic information (57/286, 19.9% in tumors; 89/286, 31.1% in stroma) in both tumor and stromal compartments were more enriched in patients with poor prognosis. Further GSEA using the OS risk model prediction score indicated that worse OS outcomes were significantly associated with the upregulation of pathways related to the cell cycle (82/127, 64.6% in tumor; 73/127, 57.5% in stroma) and genetic and epigenetic information (130/286, 45.5% in tumor; 139/286, 48.6% in stroma) in both tumor and stromal compartments.

Similarly, GSEA based on PFS revealed that pathways related to genetic and epigenetic information (42/286, 14.7%) in the stromal compartment were more enriched in patients with worse PFS. Further GSEA based on the PFS risk model prediction score revealed that worse PFS outcomes were significantly associated with the upregulation of pathways related to the cell cycle (64/127, 50.4% in tumor; 50/127, 39.4% in stroma), genetic and epigenetic information (103/286, 36.0% in tumor; 119/286, 41.6% in stroma), and cell death (19/72, 26.4% in tumor; 29/72, 40.3% in stroma) in both tumor cells and stromal compartments, as well as with the upregulation of pathways related to cell death (18/72, 25.0%), metabolism and energy (66/315, 21.0%), and genetic and epigenetic information (54/286, 18.9%) in immune compartments.

Furthermore, using mIF staining of metastatic lesions, we investigated the correlation between the infiltration of various immune cell types in the metastatic TIME and survival risk in LUAD patients. Correlation analysis revealed that the risk score of OS from the initiation of first-line therapy after distant metastasis was negatively correlated with the infiltration of M2 macrophages and effector T cells in metastatic lesions (Fig. 6d). The PFS risk score of the first-line treatment after metastasis was negatively correlated with the infiltration of M2 macrophages, effector T cells, and regulatory T cells in metastatic lesions (Fig. 6e). Additionally, the HEindex was negatively correlated with the risk scores derived from both the OS and PFS models (Fig. 6f).

Discussion

Distant metastasis of malignant tumors is characterized by distinct organotropism, which is not a random event but an active and highly selective biological process shaped by both tumor cell characteristics and the microenvironment of target organs.^23,24 Despite significant advances in understanding tumor metastasis mechanisms, the ability to accurately predict organ-specific metastasis patterns in individual patients remains limited. In this study, by integrating DSP technology, mIF, and HE staining, we systematically constructed organ-specific risk prediction models for common distant metastatic sites of LUAD, including the brain, liver, adrenal gland, and bone, on the basis of spatially resolved molecular phenotypes. We also conducted an in-depth analysis of key molecular differences between metastatic and primary lesions, offering novel insights into the mechanisms underlying organ-specific metastasis and its clinical prediction. The tumor ecosystem operates as a whole across multiple biological scales, and risk scores as well as key molecular differences derived from compartmental transcriptional data can capture systemic tumor microenvironment properties through the lens of multidirectional intercompartmental crosstalk.

One of the most significant findings of this study is that random forest models based on the spatial transcriptomic features of primary tumors can predict organ-specific metastasis risk with high accuracy (AUC > 0.9). This strongly supports our central hypothesis: although metastasis is a multistep cascade, the organ-specific metastatic “programme” is already encoded within the molecular blueprint of the primary tumor. This encompasses both the intrinsic characteristics of tumor cells (“seeds”), such as key genes identified in this study to be specifically localized in tumor cells (e.g., FKBP1A in brain metastasis, MOCOS in liver metastasis, IGKC in adrenal metastasis, and PLPP1 in bone metastasis), particularly the critical signaling cues provided by immune and stromal compartments (e.g., AFAP1L1 and CKAP2 in brain metastasis, ADAMTSL2 in liver metastasis, ASB3 and IRF3 in adrenal metastasis, and DDIAS and B3GNT5 in bone metastasis), as well as key cellular compartments (e.g., CAFI and NF positively correlated with risk scores in brain metastasis, M1 macrophages and mDC positively correlated in liver metastasis, and pDC negatively correlated in adrenal metastasis). These findings support the classical “seed and soil” hypothesis by demonstrating that the composition and state of the primary tumor microenvironment (soil) are also critical determinants of metastatic dissemination patterns. Our results indicate that the metastatic program is spatially encoded within the primary tumor ecosystem. This notion is supported by emerging evidence in NSCLC, where a preexisting subpopulation of cancer cells exhibiting neural-like transcriptional programs is specifically enriched within brain metastases, underscoring how organ-specific adaptive traits are already encoded within the spatial architecture of the primary tumor.²⁵ Targeting the spatial metastatic program may therefore represent a promising therapeutic strategy that extends beyond individual gene-based approaches.

By comparing metastasis mechanisms across different target organs, we elucidated the molecular basis for the heterogeneity of organ-specific metastatic pathways. With respect to brain metastasis, GSEA revealed significant enrichment of cell death-related pathways in the tumor, immune, and stromal compartments of high-risk primary tumors. The key DEGs we identified (such as upregulated HNRNPA2B1 and HSP90AA1 and downregulated COL3A1) and the constructed gene regulatory network indicated that metastatic cells undergo profound transcriptional reorganization when adapting to the new environment, involving adjustments in RNA processing, stress response and ECM interactions.^26–29 Concurrently, GO analysis revealed the upregulation of genes involved in protein folding/refolding and telomere maintenance in the tumor compartment during metastasis. These findings point to the preactivation of specific survival strategies, such as enhanced stress resistance and proliferative capacity, which facilitate tumor cell survival in the circulation and successful colonization of the brain.^30,31 With respect to liver metastasis, GSEA revealed significant downregulation of genetic and epigenetic information-related pathways within both the tumor and immune compartments of high-risk primary tumors. Consistently, widespread upregulation of the expression of genes encoding histones (e.g., H3C13 and H4C12) and ribosomal proteins (e.g., RPL36A and RPL38) indicates that transcriptional repression and cancer cell dormancy are key adaptive mechanisms for successful liver metastasis.³² GO analysis revealed the upregulation of genes related to the negative regulation of apoptosis in the immune compartment and nucleosome assembly in the tumor compartment during metastasis. These findings suggest that premetastatic microenvironmental remodeling is a key facilitator of liver colonization.^33,34 With respect to adrenal metastasis, GSEA revealed significant downregulation of immune-related pathways in the stromal compartments of high-risk primary tumors. GO analysis revealed that pathways related to apoptosis and oxidoreductase were significantly enriched. These results suggest the complex role of immunity in the process of adrenal metastasis. With respect to bone metastasis, GSEA did not reveal any key pathways, suggesting that its propensity may stem from broader biological processes rather than specific pathway activation. Collectively, while certain steps in the metastatic cascade are shared, the molecular pathways leading to different target organs are distinct, providing a foundation for the development of organ-specific therapeutic strategies. Our study not only developed risk prediction models for distant metastasis but also identified genes significantly associated with metastatic risk. Following rigorous validation, these genes hold promise as immunohistochemical (IHC) markers for predicting patients’ likelihood of developing distant metastasis at initial diagnosis.

In terms of clinical translation, the postmetastasis survival prediction model developed in this study has significant application potential. By sequentially applying the Wilcoxon signed-rank test, univariate Cox regression, and random forest models, we identified key genes significantly associated with PFS and OS. Our findings indicate that these key genes are primarily localized in the stromal compartment—for instance, PKM for OS and VCAM1 for PFS. These findings further underscore the pivotal role of the tumor microenvironment, particularly stromal components, in driving disease progression following metastasis.^35,36 Interestingly, analysis of the immune microenvironment revealed that M2 macrophage infiltration in metastatic lesions correlated with better survival outcomes. This observation contradicts the established view that M2 macrophages are primarily immunosuppressive, suggesting a more nuanced and complex impact of the immune microenvironment on long-term patient survival.^37,38 Moreover, these observations are influenced by multiple factors, including the inherent TIME characteristics of the metastatic organ, the choice of therapeutic regimens before and after metastasis, the patient’s postmetastatic tumor burden, and the ECOG performance status score following metastasis, among others. Given the numerous contributing factors and substantial heterogeneity in treatment patterns across patients, further quantification of these confounding variables remains challenging. Beyond confounding, one plausible biological explanation worth exploring in future studies is that within the specific context of established metastases receiving systemic therapy, a subset of M2-like macrophages might be involved in tissue repair and remodeling responses that inadvertently stabilize the tumor microenvironment or mitigate treatment-related toxicity, indirectly influencing survival metrics. This hypothesis, however, requires functional validation. Therefore, the conclusions presented in this section should be interpreted with caution.

Building upon these findings, our study ultimately aims to establish a comprehensive, two-stage predictive framework to guide clinical management across the disease continuum of LUAD. 1) Risk stratification at initial diagnosis (Premetastatic Stage): At this stage, we utilized primary tumor specimens to generate four organ-specific metastasis risk scores. The intended clinical actions are as follows: ① General surveillance strategy: Patients identified as high risk for any metastasis are recommended for more intensive and frequent surveillance. Surveillance can be further tailored; for instance, patients at high risk for brain metastasis may benefit from prioritized or more frequent cranial MRI. ② Organ-specific prophylactic or supportive measures: For patients at very high risk for specific sites, clinicians could consider organ-directed preventive strategies. With respect to the brain, the use of prophylactic cranial irradiation or intrathecal chemotherapy in select high-risk populations, similar to strategies used in other cancers, should be discussed; for bone, early initiation of bone-modifying agents and regular bone density monitoring to prevent skeletal-related events should be performed; and for liver and adrenal cancer, while direct prophylaxis is less standardized, high-risk identification would mandate enhanced abdominal imaging surveillance for early detection. This allows for timely consideration of local therapies while lesions are still limited and asymptomatic, potentially improving outcomes. 2) Treatment Guidance After Metastasis Occurs (Postmetastatic Stage): Once biopsy-proven metastasis is obtained, the two postmetastatic survival prediction models (for PFS and OS) are constructed. These models are intended to stratify patients on the basis of their likely response to conventional first-line therapies: For low-risk prognosis, continue with standard-of-care treatment protocols; for high-risk prognosis, signal potential resistance to conventional therapy. For these patients, our findings advocate for a more aggressive and personalized approach early on. This could include enrollment in clinical trials or the use of functional precision medicine platforms, such as deriving patient-derived organoid xenograft models from metastatic biopsy for ex vivo drug sensitivity testing to guide the selection of potentially effective therapies.

Furthermore, we defined a new parameter, the HEindex, which represents the average proportion of nontumor cells surrounding each tumor cell, to quantitatively assess the infiltration of nontumor cells in the tumor microenvironment. Our results showed that the HEindex in primary lesions positively correlated with the risk of adrenal metastasis, suggesting a strong association between adrenal metastasis in LUAD and the presence of immune or stromal cells within the TIME. Conversely, the HEindex in metastatic lesions is negatively correlated with mortality and disease progression risk, indicating that immune or stromal infiltration levels in metastatic sites are crucial for postmetastatic survival.

Despite these findings, this study has several limitations. First, the sample size was relatively small, particularly for adrenal and bone metastasis cases, which may affect the generalizability and robustness of the models. In the risk prediction models for metastasis, we included all primary lesions for each type of distant metastasis (brain, liver, bone, and adrenal gland). Nevertheless, our findings require further validation in larger and more diverse populations. With respect to the postmetastasis survival prediction model, we analyzed all pairs of matched specimens with complete survival and RNA expression data from three compartments (tumor, stroma and immune); however, the distribution of these paired samples across organ-specific metastatic sites was uneven. Specifically, only three matched pairs were available for adrenal metastasis, and no matched pairs were included for bone metastasis. The limited sample size precluded the development of a survival prediction model tailored to specific distant metastatic sites. While this represents a clear limitation, it is worth noting that our pooled analysis across metastatic sites was also motivated by clinical considerations: in real-world practice, patients with multifocal progression typically undergo biopsy of only a single accessible lesion to guide subsequent therapy. An organ-specific model, while biologically more precise, would be challenging to apply in this common scenario, as it would require predicting overall survival from a single biopsy when multiple metastatic sites are present. Nevertheless, this approach inherently limits the model’s applicability to patients with distinct metastatic patterns—particularly those with solitary bone metastasis, for whom no matched specimens were available in our cohort. For such patients, the current model may not adequately capture site-specific prognostic determinants, and future studies with dedicated bone metastasis cohorts are needed to address this gap. Second, the study primarily relied on spatial transcriptomic data and mIF without incorporating multiomics data such as genomics and epigenomics data. Future studies should integrate these additional layers of information to construct more comprehensive predictive frameworks. Finally, functional validation of the identified key genes and pathways remains limited, and further mechanistic studies using in vitro and in vivo models—such as gene editing and organoid coculture systems—are urgently needed. In the future, we will further validate the findings of this study in multicenter patient cohorts, train additional models to predict metastasis to other clinically relevant sites and construct models to predict the time to organ-specific metastasis. Moreover, cellular and animal experiments will be conducted to explore the underlying organ-specific metastatic mechanisms.

In conclusion, from a unique spatial transcriptomic perspective, this study revealed that the molecular features of primary LUAD lesions are related to organ-specific metastatic potential, whereas the molecular and immune characteristics of metastatic lesions influence postmetastatic survival. The high-precision prediction models and key molecular signatures identified in this work not only enhance our understanding of the spatiotemporal dynamics of “seed–soil” interactions but also provide promising biomarkers and therapeutic targets for future clinical applications.

Materials and methods

Subject investigated

This study retrospectively included 52 patients with advanced LUAD and distant metastases, from whom we collected a total of 52 paired tissue samples: 23 pairs of brain metastases with primary tumors, 26 pairs of liver metastases with primary tumors, and 3 pairs of adrenal metastases with primary tumors. Clinical and pathological data were obtained from the Electronic Medical Record system. This study was approved by the Ethics Committee of the National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College (24/111-4391) and was conducted in accordance with the Declaration of Helsinki.

DSP spatial transcriptomic analysis

First, formalin-fixed, paraffin-embedded (FFPE) tissue sections were hybridized with a panel of mRNA-targeting oligonucleotide probes equipped with photocleavable linkers. After hybridization, the sections were subjected to multiplex fluorescent morphological staining (including SYTO13, panCK, and CD45) to enable precise histological annotation (e.g., tumor, immune, and stroma types). On the basis of high-resolution fluorescence images combined with H&E staining, the ROIs were selected by the researchers. Using the GeoMx^® DSP instrument,³⁹ each predefined ROI was irradiated with ultraviolet (UV) light to selectively cleave and release the index oligonucleotides from the probes bound specifically within the illuminated area. The released oligonucleotides, representing the transcriptomic content of each ROI, were collected via microcapillary aspiration into separate wells. Finally, the collected oligonucleotides were processed for library preparation and subjected to next-generation sequencing (NGS). All samples run on the NanoString GeoMx® DSP platform passed standard QC metrics, including positive control probe performance, sequencing saturation, and negative control backgrounds. ROIs were prospectively annotated by a pathologist prior to DSP. The sole inclusion criterion was a minimum of 100 cells within the annotated morphology-guided circle. No ROIs were excluded postprofiling because of quality issues. Details of HE staining and DSP data ROI selection for the samples involved in this study are provided in the Supplementary Material (HE and DSP).

mIF

Four mIF panels were developed to evaluate macrophages, T lymphocytes, cancer-associated fibroblasts (CAFs), dendritic cells (DCs), and B cells. The cell subtypes, including M1- and M2-polarized macrophages, effector and exhausted CD8⁺ cytotoxic T cells (Tc-effector and Tc-exhausted), effector and exhausted regulatory T cells (Treg-effector and Treg-exhausted), CD8⁺ regulatory T cells (T8reg), type I, II, and III CAFs (CAFI, CAFII, and CAFIII),⁴⁰ B effector and regulatory B cells (Be and Breg),⁴¹ and classical, myeloid, and plasmacytoid DCs (cDCs, mDCs, and pDCs),⁴² were identified on the basis of the expression profiles across different channels within each panel. The mIF panels used in this study are described in the Supplementary Material (mIF panel information).

Pathway enrichment analysis

GSEA was performed on the basis of the whole DSP data. The reference gene sets were selected as KEGG + GO + hallmark gene sets. A total of 1,350 gene sets related to tumors and the TIME were analyzed and classified into six major categories on the basis of the Molecular Signatures Database (MSigDB): 421 “immunity” gene sets, 127 “cell cycle” gene sets, 315 “metabolism and energy” gene sets, 286 “genetic and epigenetic information” gene sets, 129 “extracellular matrix (ECM) and metastasis” gene sets, and 72 “cell death” gene sets. In the metastasis risk model, we ranked genes either by the logFC between the metastatic and nonmetastatic groups or by the Spearman correlation coefficient between each gene and the model risk score. In the survival prediction model, we ranked genes either by the lnHR from univariate Cox regression for PFS or OS or by the Spearman correlation coefficient between each gene and the model risk score. In GSEA, gene sets with a raw P < 0.05 and FDR < 0.25 were considered to be statistically significant. We employed a dual threshold to identify robust yet biologically informative signals. The raw P filter ensures a baseline level of unadjusted statistical evidence, while the lenient FDR cutoff is consistent with common practice in exploratory GSEA, aiming to reduce false negatives and capture broader functional themes without compromising interpretability.

For brain and liver metastases, we obtained the final location-specific DEGs by taking the intersection of DEGs between primary and metastatic lesions at the bulk level and DEGs at the DSP level, which were then subjected to clusterProfiler analysis (GO and KEGG enrichment analyses were performed separately for upregulated and downregulated DEGs). For the brain, liver, and adrenal metastases, we performed enrichment analysis on DEGs between primary and metastatic lesions obtained at the bulk level using the STRING database.

Weighted regulatory network analysis during metastasis

Spearman’s rank correlation coefficients were calculated between genes in both primary and metastatic lesion samples. Only when the correlation coefficients of the same gene pair showed consistent directions (i.e., both positive or both negative) in primary and metastatic lesions was the average value of these correlation coefficients computed. This average value was used to evaluate the strength of the potential regulatory relationship between the gene pair during tumor metastasis and to construct a weighted regulatory network.

Follow-up and survival analysis

OS for patients was defined as the period from the initiation of first-line treatment following metastasis to death due to any cause. PFS was defined as the period from the initiation of first-line treatment following metastasis to either tumor progression or death due to any cause. The last follow-up date was June 30, 2025. Survival analysis was performed with a Cox proportional hazards model.

Construction of risk models for distant metastasis of LUAD

To predict distant metastasis risk, we developed a random forest model using primary lung tumor specimens. We integrated DSP data from distinct tumor regions (tumor, immune, and stroma) with clinical metastasis data (brain, liver, bone, and adrenal) and pathological information. Metastatic status was defined as follows: ① Metastasis confirmed (status = 1): Radiographic or pathological evidence of metastasis to the target organ. ② No metastasis was confirmed with sufficient follow-up (status = 0): The follow-up duration exceeded the median time from primary diagnosis to first metastasis in that organ within our cohort. ③ Indeterminate status (excluded, coded as NA): Follow-up was shorter than the median interval mentioned above, preventing reliable classification. Only samples with a status of 1 or 0 were used to train the corresponding organ-specific prediction model. For each organ-specific metastasis risk model, we independently partitioned the cohort using a stratified 60/40 training/test split. Stratification was based on the metastatic status of the target organ, which was ensured via the “createDataPartition” function from the R “caret” package, to maintain consistent class proportions across sets. Using the training set, we identified the top 1000 genes ranked by the absolute log2-fold change (logFC) in expression between metastatic and nonmetastatic samples. This candidate set was further filtered by receiver operating characteristic (ROC) analysis, which revealed an AUC > 0.9 or <0.1 and a lower confidence interval (LCI) > 0.5 or an upper confidence interval (UCI) < 0.5 in the training cohort to ensure statistical significance. We trained random forest models with selected features and applied additional selection by feature importance (IncNodePurity) when the number of features exceeded 10 to simplify the model. Model performance was evaluated using ROC curves and AUC values in the training and test cohorts.

Construction of survival prediction models for LUAD patients following metastasis

To develop survival prediction models, we applied a random forest algorithm combined with Cox proportional hazards regression. The dataset was randomly split into training (60%) and test (40%) cohorts. We selected features through a two-step process: identifying DEGs between primary and metastatic lesions via the Wilcoxon signed-rank test (raw P < 0.05, FDR < 0.01), followed by univariate Cox regression, which retained genes with P < 0.05 and a C-index > 0.75. A random forest survival model was then trained on these features, and if their number exceeded 10, we retained only the top 10 by importance. Model performance was assessed by evaluating the association between the predicted scores and the outcome using Cox regression, which provided the C-index, HR, and corresponding P in both the training and test cohorts.

HEindex

Whole-slide H&E-stained images were digitized and segmented at the single-cell level using QuPath software (version 0.6.0; https://qupath.github.io/). The tumor, stroma, immune, and necrotic components were automatically identified and classified across the entire slide utilizing the machine learning algorithm of the software. For quantitative analysis of the tumor microenvironment, we calculated the proportion of nontumor cells within a 20-micrometer radius from the center of each tumor cell. This proportion, termed the local nontumor ratio, defines the spatial context of individual tumor cells: a ratio of 0 indicates a “core” region (exclusively surrounded by other tumor cells), whereas a ratio greater than 0 and less than or equal to 1 indicates a “border” region (in contact with nontumor cells). The HEindex for each sample was then derived by averaging this ratio across all its tumor cells. A higher HEindex signifies greater tumor–stroma contact and a more dispersed tumor growth pattern, whereas a lower index indicates a more compact, solid tumor mass. Notably, the calculation of the HEindex itself is a fixed spatial algorithm based on geometric relationships and cell type proportions; it does not involve any machine learning algorithms or fitted coefficients.

Statistics

Statistical analyses were performed using R software (version 4.5.0; https://www.r-project.org/). Continuous variables were compared between two independent groups using the Wilcoxon rank-sum test, while the Wilcoxon signed-rank test was applied for paired comparisons. Correlations between continuous variables were assessed using Spearman’s method. Model performance was assessed using ROC curves. Statistical significance was defined in each section.

All the models and computational results presented in this study are included in the Supplementary Material (Comprehensive Data and Analysis Results).

Author contributions

Tongji Xie: Conceptualization, Data curation, Formal analysis, Methodology, Visualization, Software, Writing - original draft. Lige Wu: Conceptualization, Data curation, Formal analysis, Methodology, Visualization, Software, Writing - original draft. Yan Li: Data curation, Software. Mengxing You: Data curation. Zihe Wang: Data curation. Ran Gao: Data curation. Xuezhi Hao: Resources, Writing – review & editing. Jianming Ying: Conceptualization, Resources, Project administration, Supervision, Writing – review & editing. Junling Li: Conceptualization, Resources, Project administration, Supervision, Writing – review & editing. Puyuan Xing: Conceptualization, Resources, Project administration, Supervision, Writing – review & editing. All the authors have read and approved the article.

Data availability

The data reported in this paper have been deposited in the OMIX, China National Center for Bioinformation/Beijing Institute of Genomics, Chinese Academy of Sciences (https://ngdc.cncb.ac.cn/omix: accession no. OMIX015799). Any information required to reanalyze the data reported in this article is available from Puyuan Xing (xingpuyuan@cicams.ac.cn) upon request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41392-026-02817-y.