Introduction

As one of the most prevalent endocrine malignancies, thyroid cancer often presents diagnostic challenges due to its nonspecific symptoms, which overlap with those of other thyroid disorders1. According to the National Cancer Center of China, thyroid cancer ranks as the third most rapidly increasing cancer in China, with an estimated 466,100 new cases reported in 2022, highlighting its growing public health burden2. The development of thyroid cancer is influenced by multiple risk factors, including family history, radiation exposure, and genetic mutations, which collectively drive carcinogenesis by dysregulating key oncogenic signaling pathways3. Furthermore, thyroid cancer is characterized by the accumulation of diverse genomic, transcriptomic, and epigenetic alterations, leading to significant intertumoral and intratumoral heterogeneity through a stepwise progression4. Given its complex molecular landscape and clinical variability, there is an urgent need to further elucidate the biological mechanisms underlying thyroid cancer to advance precision medicine and develop individualized therapeutic strategies.

The accumulation of diverse transcriptomic heterogeneity and dynamic immunological alterations in thyrocytes significantly influences the malignant progression and clinical outcomes of thyroid cancer5. Notably, the tumor microenvironment (TME) in thyroid cancer is composed of heterogeneous cellular components (including stromal cells and infiltrating immune cells), physical elements (such as the extracellular matrix), and chemical factors (e.g., chemokines)6,7. These intricate cellular interactions play a pivotal role in tumor progression and treatment responsiveness. Collectively, these findings underscore the pressing clinical need to comprehensively map the ecosystem of thyroid cancer, which will be critical for identifying potential therapeutic targets and developing personalized treatment strategies.

Although current genomic and transcriptomic analyses based on bulk samples have identified critical genes and underlying signaling pathways involved in thyroid cancer progression, these approaches are limited in their resolution and accuracy for capturing tumor heterogeneity and molecular diversity. Recently, single-cell RNA sequencing (scRNA-seq) has emerged as a groundbreaking technology, enabling unprecedented insights at the cellular level. This technique has revolutionized the identification of transcriptional profiles, the reconstruction of cellular trajectory lineages, and the discovery of novel cell subpopulations8,9. Furthermore, single-cell genomic analysis in human tumors has opened new avenues for exploring intra-tumor heterogeneity10,11 and deciphering complex cellular cross-talk within the tumor microenvironment12,13. These advancements provide a powerful toolkit for unraveling the intricate molecular landscape of thyroid cancer and advancing precision oncology.

In this study, we performed single-cell RNA sequencing (scRNA-seq) to construct a comprehensive atlas of 405,077 single cells derived from 50 thyroid cancer samples (including papillary thyroid cancer, anaplastic thyroid cancer, metastatic lymph nodes, and subcutaneous metastases) and 14 normal thyroid tissue samples. This approach enabled us to decipher tumor heterogeneity and identify key molecular drivers of tumor progression. Through unbiased clustering, we identified four major cellular clusters and characterized their molecular properties, providing a detailed view of the multicellular ecosystem within the tumor microenvironment (TME). Furthermore, cell-cell communication analysis revealed intricate interactions that contribute to the immunosuppressive reprogramming of the tumor ecosystem. By leveraging the NicheNet algorithm, we highlighted critical signaling pathways and ligand-receptor interactions among hub niche cell types, uncovering a shared hub niche neighborhood that drives tumor progression. These findings were validated using multiplex immunohistochemistry (mIHC), which confirmed the spatial organization and interactions of key cell populations within the TME. Collectively, our scRNA-seq analysis provides a high-resolution atlas that elucidates the complex tumor heterogeneity, delineates the comprehensive landscape of the TME, and uncovers the molecular mechanisms underlying cancer progression. These insights offer promising therapeutic targets and pave the way for the development of novel treatment strategies for thyroid cancer.

Results

A single-cell transcriptome atlas in thyroid cancer lesions

To explore the cellular diversity within thyroid cancer lesions, we conducted single-cell RNA sequencing (scRNA-seq) on 64 thyroid cancer samples. Following rigorous quality control (as detailed in Supplementary Fig. 1), a total of 405,077 high-quality cells were retained for downstream analysis, from which single-cell transcriptomes were successfully obtained. After normalizing gene expression, we performed principal component analysis (PCA) on variably expressed genes across all cell samples, selecting 20 principal components (PCs) for further analysis. Using unbiased clustering based on Uniform Manifold Approximation and Projection (UMAP), we identified four major distinct cellular clusters. To define these clusters, we employed differential expression analysis to identify representative genes for each cluster based on their expression profiles and established cell markers (Fig. 1A).

Fig. 1: Single-cell transcriptomic analysis of thyroid cancer lesions.
figure 1

A UMAP plot illustrating the four major cell types identified in thyroid cancer lesions. B UMAP projection of the main cell types, color-coded by sample origin. C Heatmap displaying the expression levels of representative novel markers across the four main cell types. D Stacked barplots depicting the proportional distribution of major cell types within each sample. E Boxplot showing the diversity distribution, quantified using the Shannon equitability index based on the composition of the four main cellular clusters, across normal tissue, papillary thyroid cancer (PTC), anaplastic thyroid cancer (ATC), metastatic lymph node, and subcutaneous metastasis samples.

Full size image

Specifically, the four cellular lineages were characterized as follows: (1) thyrocytes, marked by high expression of EPCAM14, TG (Thyroglobulin), and TPO (Thyroid peroxidase); (2) endothelial cells, exhibiting elevated expression of EMCN and VWF15; (3) mesenchymal cells, expressing PDGFRB16 and COL1A217; and (4) immune cells, specifically expressing PTPRC (encoding the CD45 antigen). The expression patterns of these cluster-specific marker genes across the cellular populations are illustrated in Fig. 1B. Additionally, we observed significant heterogeneity in cellular composition among different patients (Fig. 1C) and across various tumor tissue types (Fig. 1D), highlighting the diverse and complex nature of thyroid cancer at the single-cell level.

Blueprint of tumor microenvironment in thyroid cancer

To further elucidate the dynamic changes in the tumor microenvironment (TME) cellular composition, we identified and analyzed cell subsets to reveal alterations in endothelial cells, mesenchymal cells, T cells, myeloid cells, and B cells across primary tumors, metastatic lesions, and adjacent normal tissues. Using UMAP analysis, endothelial cells were clustered into eight distinct subtypes (Fig. 2A, B), which exhibited tumor subtype-specific distributions (Fig. 2C). Consistent with previous findings18, the End_PLVAP cluster was notably enriched in papillary thyroid tumors. Based on canonical cellular markers (Fig. 2E), unsupervised clustering analysis further delineated nine mesenchymal cell clusters (Fig. 2D), revealing strong intertumoral heterogeneity among different tumor subtypes (Fig. 2F).

Fig. 2: Distinct clusters of endothelial and mesenchymocyte cells in thyroid cancer lesions.
figure 2

A UMAP reduction analysis identified eight distinct endothelial cell subclusters. B Heatmap illustrating the expression levels of representative novel markers across each endothelial cellular subtype. C Boxplot displaying the diversity distribution, calculated based on the composition of endothelial cellular clusters across different tissue origins. D UMAP reduction analysis revealed nine mesenchymocyte subclusters. E Heatmap depicting the expression levels of representative novel markers across each mesenchymocyte cellular subtype. F Boxplot showing the diversity distribution, quantified based on the composition of mesenchymocyte cellular clusters from different tissue origins.

Full size image

To comprehensively characterize the immune landscape of thyroid cancer, immune compartments were reclustered separately (Fig. S2). T cells were partitioned into nine sub-clusters (Fig. 3A), with the expression profiles of cluster-specific marker genes illustrated in Fig. 3B. Notably, CD8 + PDCD1 + T cells, exhibiting high expression of exhaustion-related genes, were predominantly localized in anaplastic thyroid tumors, while CD4 + CCR7 + T cells, displaying a naive phenotype, were significantly upregulated in papillary thyroid cancer (Fig. 3C). Intriguingly, we identified a novel T cell subset, CD4 + HSPA1A7 + T cells, characterized by a stress response state and high expression of heat shock protein genes (such as HSPA1B and HSPA1A), which was specifically enriched in anaplastic thyroid tumors. This unique T cell phenotype may contribute to the poor prognosis associated with anaplastic thyroid carcinoma (ATC) (Fig. 3D).

Fig. 3: Distinct clusters of T cells in thyroid cancer lesions.
figure 3

A UMAP reduction analysis identified nine distinct T cell subclusters. B Heatmap displaying the expression levels of representative novel markers across each T cell subtype. C Boxplot illustrating the diversity distribution, calculated based on the composition of T cell clusters across different tissue origins. D Heatmap showing the expression of 19 curated gene signatures across T cell clusters, with the heatmap generated based on scaled gene signature scores.

Full size image

Myeloid cells were further classified into ten subtypes (Fig. 4A) based on differential gene expression patterns (Fig. 4B). Among these, APOE+ macrophages, identified as M2-like tumor-associated macrophages (TAMs), were predominantly enriched in tumor samples compared to normal tissues, suggesting their role in promoting an immunosuppressive TME (Fig. 4C). Additionally, myeloid subpopulations exhibited significant diversity across different cancer subtypes and metastatic tumors (Fig. 4D). Within the B cell compartment, a total of 38,855 cells were divided into two discrete subpopulations: B cells and plasma cells, both of which showed dysregulation in different tumor samples (Figure S2).

Fig. 4: Distinct clusters of myeloid cells in thyroid cancer lesions.
figure 4

A UMAP reduction analysis identified nine distinct myeloid cell subclusters. B Heatmap displaying the expression levels of representative novel markers across each myeloid cellular subtype. C Boxplot illustrating the diversity distribution, calculated based on the composition of myeloid cellular clusters across different tissue origins. D Heatmap showing the expression of M1-like, M2-like, angiogenesis, and phagocytosis-associated gene signatures across myeloid cell clusters, with the heatmap generated based on scaled gene signature scores. E Bar plot highlighting the top five functional pathways enriched in APOE+ macrophages, as determined by Gene Ontology (GO) analysis.

Full size image

In summary, our analysis revealed heterogeneous immune cell subtypes with unique molecular states, highlighting their specific functional roles in the tumorigenesis of thyroid cancer. These findings provide a deeper understanding of the cellular dynamics within the TME and underscore the potential for targeting specific immune cell populations in therapeutic strategies.

Comprehensive cell-cell communication networks in the TME

Of particular interest, APOE+ macrophages were found to upregulate collagen development processes and vasculature activation programs, underscoring their role as a hub niche within the tumor microenvironment (Fig. 4E). This observation prompted us to further investigate the cell-cell interaction network within the diverse tumor ecosystem. To characterize the tumor microenvironment of thyroid cancer lesions, we employed CellChat to analyze intercellular communication among thyrocytes and immune components.

First, we deciphered the cell-cell crosstalk among cellular subpopulations in papillary thyroid cancer (PTC) and anaplastic thyroid cancer (ATC) separately (Fig. 5A, B). Overall, the total number and strength of potential interactions differed significantly between these two tumor subtypes (Fig. 5C). The inferred incoming and outgoing interaction strengths for each cell cluster were visualized to further explore intercellular communication dynamics (Fig. 5D). Differential interaction analysis revealed that exhausted CD8 + PDCD1 + T cells and immunosuppressive APOE+ macrophages were highly active and engaged in extensive interactions with other cell types (Fig. 5E). Given the hub niche function of APOE+ macrophages, we focused on their differential interactions (Fig. 5F). Intriguingly, inhibitory interactions between APOE+ macrophages and CD8 + PDCD1 + T cells were prominently observed in ATC (Fig. 5G), with specific ligand-receptor complexes (e.g., THBS1-CD47, PECAM1-CD38, LGALS9-HAVCR2, and ICOSL-CD28) identified as key mediators. To validate these findings, we performed multiplex immunohistochemistry (mIHC) to visualize these proteins, directly confirming the interaction between CD47 + T cells and THBS1 + APOE+ macrophages (Fig. 5H). These results provide critical insights into the immunosuppressive mechanisms within the TME and highlight potential targets for enhancing immunotherapeutic efficacy.

Fig. 5: Cell-to-cell communications among thyrocyte and immune cells predicted by the CellChat software in PTC and ATC.
figure 5

A Circle plot summarizing the maximum number of interactions among individual cell types in PTC. The thickness of the lines connecting cells indicates the interaction strength. B The same plot as in (A) but applied to ATC. C Differences of number and strength overall information flow within the inferred networks between PTC and ATC. D The potential outgoing and incoming interaction strength of each cellular cluster in PTC and ATC. Heatmap showing the relative signaling contribution of each cell group based on the number and strength compared PTC with ATC. E Heatmap showing the relative signaling contribution of each cell group based on the number and strength compared PTC with ATC. F Differential strength of network centrality based on APOE+ macrophages (PTC vs. ATC). G Bubble heatmap showing the cell–cell communication of selected ligand–receptor pairs between APOE+ macrophage and CD8+ PDCD1+ T cells. Dot size indicates P value, colored by communication probability. H Multiplex IHC shows the cross-talk between APOE+ macrophages and T cells via the ligand–receptor of THBS1-CD47.

Full size image

Additionally, we mapped the landscape of cellular communication in PTC compared to metastatic lesions to uncover key signaling pathways driving tumor metastasis (Fig. 6A). Notably, both the number and strength of interactions exhibited a declining trend in metastatic lesions relative to primary tumors (Fig. 6B). We analyzed the broadcast cell-cell communication network and established the incoming and outgoing signaling pathways of cellular subsets in metastatic lymph node (MLN) samples (Fig. 6C). Strikingly, the interaction strength between EMT-like cancer-associated fibroblasts (CAFs) and CD8 + PDCD1 + T cells was significantly upregulated in MLN samples (Fig. 6D). An overview of interaction changes involving EMT-like CAFs was compared between PTC and MLN (Fig. 6E). To further elucidate the key signaling pathways driving cancer metastasis, we investigated the extensive ligand-receptor pairs between EMT-like CAFs and CD8 + PDCD1 + T cells. In addition to the well-known LAMC1-CD44 interaction, we identified other molecular interactions rarely reported in thyroid cancer, such as PPIA-BSG and CLEC2C-KLRB1 complexes (Fig. 6F), suggesting their potential roles in cancer invasiveness. These findings were validated using mIHC, which confirmed the PPIA-BSG interaction between BSG + T cells and PPIA + SAA1+ CAFs in PTC lymph node metastases (Fig. 6G).

Fig. 6: Cell-to-cell communications among thyrocyte and immune cells predicted by the CellChat software in PTC and Metastase.
figure 6

A Circle plot summarizing the maximum number of interactions among individual cell types in metastatic lesions. The thickness of the connecting lines represents the interaction strength. B Comparison of the overall information flow, including the number and strength of interactions, within the inferred networks between PTC and metastatic lesions. C Heatmap displaying the potential outgoing and incoming interaction strength of each cellular cluster in PTC and metastatic lesions. D Heatmap illustrating the relative signaling contribution of each cell group based on the number and strength of interactions, comparing PTC with metastatic lesions. E Differential network centrality analysis based on EMT-like cancer-associated fibroblasts (CAFs), comparing PTC and metastatic lesions. F Bubble heatmap showing the cell-cell communication of selected ligand-receptor pairs between EMT-like CAFs and CD8 + PDCD1 + T cells. Dot size indicates the P-value, while color represents the communication probability. G Multiplex immunohistochemistry (mIHC) validation of the cross-talk between SAA1+ CAFs and T cells via the PPIA-BSG ligand-receptor interaction.

Full size image

Collectively, these findings provide a comprehensive understanding of the cellular communication networks within the thyroid cancer TME, revealing novel mechanisms of immune suppression and metastasis. This work lays the foundation for developing targeted therapeutic strategies to disrupt these interactions and improve clinical outcomes.

Hub niche interaction of oncogenesis and invasiveness

In addition to the hub niche properties of APOE+ macrophages, our cellular communication analysis revealed that EMT-like cancer-associated fibroblasts (CAFs) play a pivotal role in driving tumor metastasis. Furthermore, RBP7+ endothelial cells were significantly enriched in metastatic tumor lesions (Fig. 2C), suggesting their potential contribution to the metastatic niche. We hypothesize that the complex interactions among these three cell types—APOE+ macrophages, EMT-like CAFs, and RBP7+ endothelial cells—collectively contribute to the remodeling of the tumor ecosystem during cancer progression. Intriguingly, the strength of both outgoing and incoming signaling pathways was progressively dominated by these hub niche cell types, transitioning from normal tissue to primary PTC tumors and further to metastatic lymph node (MLN) lesions (Fig. 7A). This observation underscores the central role of these hub niche cell types in orchestrating cellular communication networks critical for cancer progression.

Fig. 7: Cell-cell crosstalk among three hub niche cell types from Normal to PTC then to Metastase.
figure 7

A The potential outgoing and incoming interaction strength of each cellular cluster in normal tissue, PTC, and metastatic lesions. Hub niche cell types are highlighted in magenta, while other clusters are denoted by blue points. The center line represents the median value, the lower and upper hinges indicate the 25th and 75th percentiles, respectively, and the whiskers extend to 1.5× the interquartile range. B Ranking of all significant signaling pathways based on differences in overall information flow within the hub niche among normal tissue, PTC, and metastatic lesions. C Bubble heatmap illustrating the cell-cell communication of selected ligand-receptor pairs among hub niche cell types. Dot size indicates the P-value, while color represents the communication probability.

Full size image

To further elucidate the mechanisms underlying these interactions, we deciphered reciprocal ligand-receptor pairs among the hub niche cell types, uncovering key signaling pathways involved in tumorigenesis (Fig. 7B). Among these, the MIF–ACKR3 interaction, which was significantly enriched in RBP7+ endothelial cells, validated the upregulated AUCell score of angiogenesis (Fig. 7C). Additionally, RBP7+ endothelial cells expressed high levels of SELE, while its ligands GLG1 and CD44 were overexpressed by EMT-like CAFs, suggesting that RBP7+ endothelial cells play a critical role in promoting tumor epithelial-mesenchymal transition (EMT). Notably, the ligand ANGPT2, expressed by EMT-like CAFs, interacted with the receptor complex ITGA5 + ITGB1 on APOE+ macrophages, indicating that EMT-like CAFs may facilitate the recruitment of immunosuppressive M2-like macrophages. Collectively, these findings establish a comprehensive map of cellular crosstalk within the tumor microenvironment (TME) of PTC, highlighting APOE+ macrophages, EMT-like CAFs, and RBP7+ endothelial cells as central regulators reshaping the tumor ecosystem during progression.

NicheNet analysis further emphasized the overexpression of predicted ligands within hub niche cells, underscoring their role as crucial drivers of tumorigenesis (Fig. 8A). These ligands mediate broad interactions among hub niche cell types, reinforcing their central role in the TME. Interestingly, many of these prioritized hub niche ligands were also expressed by other bystander cells, such as CD207+ dendritic cells, PLVAP+ endothelial cells, inflammatory CAFs, and thyrocytes (Fig. 8B). This suggests that these bystander cells, by expressing hub niche-prioritized ligands, constitute a nonnegligible component of the TME and contribute to tumor development. Together, these findings provide a holistic view of the cellular interactions within the TME, revealing the multifaceted roles of hub niche and bystander cells in driving tumor progression and metastasis.

Fig. 8: Predicted cell–cell interactions of onco-fetal reshaping in Normal, PTC and Metastase.
figure 8

A Expression of top 20 predicted ligands identified by NicheNet analysis based on the differentially expressed signature genes of RBP7+ endothelial cells in Normal, PTC and Metastase. B Expression of top 20 predicted ligands identified by NicheNet analysis based on the differentially expressed signature genes of EMT-like CAFs in Normal, PTC and Metastase. C Expression of top 20 predicted ligands identified by NicheNet analysis based on the differentially expressed signature genes of APOE+ macrophages in Normal, PTC and Metastase. D Heatmap depicting the expression of shared potential ligands driving hub niche reprogramming in Normal, PTC and Metastase.

Full size image

Discussion

Given the fundamentally distinct carcinogenic mechanisms underlying anaplastic thyroid cancer (ATC) and papillary thyroid cancer (PTC), their clinical behaviors exhibit significant differences19. ATC, characterized by aggressive growth and rapid progression, contrasts sharply with the typically indolent nature of PTC. Furthermore, the malignant tumor microenvironment of metastatic lesions is shaped by a complex sequence of events, including tumor cell dissemination, dormancy, and colonization. These processes are orchestrated by the dysregulation of tissue-resident stromal cells and the recruitment of immune cells into organ-specific protumor niches, which collectively foster a permissive environment for metastatic outgrowth20. This highlights the critical role of the tumor microenvironment in driving metastatic progression and underscores the need for tailored therapeutic strategies to address the unique biological features of different thyroid cancer subtypes.

In this study, we identified four predominant cell populations through UMAP clustering analysis of 64 thyroid cancer biopsies, and further characterized their molecular and cellular features to elucidate their potential roles in thyroid cancer tumorigenesis. Notably, the proportions and functional phenotypes of non-tumor cell subtypes—such as endothelial cells, immune cells, and fibroblasts—varied significantly across different sample sources, highlighting their heterogeneity and prompting a focused investigation into these subsets. Given the central role of myeloid cells in regulating inflammation and tumorigenesis, we identified ten distinct myeloid cell subsets, among which APOE+ macrophages emerged as key players in reprogramming the tumor microenvironment (TME) of thyroid cancer. These APOE+ macrophages exhibited high expression of established M2-like macrophage markers, consistent with their role in modulating T cell states toward an immunosuppressive phenotype, as previously reported21. Furthermore, while angiogenesis-related signature genes were universally expressed across macrophage subsets, APOE+ macrophages uniquely displayed collagen metabolic characteristics, suggesting their involvement in extracellular matrix remodeling and TME modulation. These findings underscore the multifaceted roles of myeloid cells, particularly APOE+ macrophages, in shaping the immunosuppressive and pro-tumorigenic landscape of thyroid cancer.

Anaplastic thyroid cancer (ATC) is widely recognized for its poor prognosis and resistance to conventional therapies, and effective immunotherapeutic strategies for ATC remain limited. Therefore, there is an urgent need to unravel the molecular characteristics of the immune landscape in ATC at high cellular resolution. Our cell-cell communication analysis revealed distinct ligand-receptor interactions between PTC and ATC, highlighting the critical role of APOE+ macrophage–CD8 + PDCD1 + T cell interactions in ATC immunotherapy. Specifically, the THSB1-CD47 axis emerged as a potential novel target for enhancing therapeutic efficacy. Furthermore, compared to primary PTC tumors, metastatic lesions exhibited upregulated interactions between EMT-like cancer-associated fibroblasts (CAFs) and CD8 + PDCD1 + T cells, particularly through ligand-receptor pairs such as PPIA-BSG. These findings provide valuable insights into the underlying biological mechanisms driving metastatic progression and offer a foundation for developing targeted therapeutic strategies to disrupt these interactions. By elucidating the cellular and molecular dynamics of ATC and metastatic thyroid cancer, this study paves the way for advancing immunotherapy and improving clinical outcomes for patients with aggressive thyroid cancer subtypes.

Metastasis represents a critical stage in tumor progression and is a major contributor to treatment failure20. However, the molecular mechanisms driving lymph node metastasis and local invasion in papillary thyroid cancer (PTC) at single-cell resolution remain poorly understood. In this study, we identified APOE+ macrophages as a central player in the hub niche ecosystem of thyroid cancer, orchestrating key interactions within the tumor microenvironment (Fig. 7g). Through cell-cell crosstalk analysis, we revealed that APOE+ macrophages, together with EMT-like cancer-associated fibroblasts (CAFs) and RBP7+ endothelial cells, form a multicellular hub niche. These interactions are mediated by signaling pathways such as ANGPT, SELE, and MIF, which collectively reinforce the core remodeling processes in PTC tumor metastases.

Based on our single-cell RNA sequencing (scRNA-seq) data and cell-cell communication analysis, we have identified compelling evidence that the complex interactions among APOE+ macrophages, EMT-like cancer-associated fibroblasts (CAFs), and RBP7+ endothelial cells play a central role in remodeling the tumor ecosystem during thyroid cancer development. Specifically, these interactions involve key signaling pathways such as ANGPT, SELE, and MIF, which collectively promote immune suppression, extracellular matrix (ECM) remodeling, and angiogenesis. For example, APOE+ macrophages modulate T cell states toward an immunosuppressive phenotype, EMT-like CAFs drive epithelial-mesenchymal transition (EMT) and metastatic progression, and RBP7+ endothelial cells facilitate vascular remodeling. Together, these cell types form a hub niche that orchestrates critical processes in tumor progression and immune evasion.

NicheNet analysis predicted that hub niche programming ligands are predominantly enriched in these three hub niche cell types. However, some of these ligands are also expressed by bystander cells that are not traditionally classified as hub niche components. This finding underscores the potential role of bystander cells in modulating the tumor ecosystem during the metastatic process of PTC. Together, these results provide a deeper understanding of the complex cellular and molecular interactions driving PTC metastasis and highlight potential therapeutic targets for disrupting metastatic progression.

Compared to existing studies exploring single-cell RNA sequencing (scRNA-seq) data in thyroid cancer, our research demonstrates several notable advancements. First, our analysis encompasses a substantial dataset of 405,077 single cells derived from 50 thyroid cancer samples, including papillary, anaplastic, and metastatic tumors, as well as 14 normal tissue samples, providing a robust and comprehensive resource for understanding thyroid cancer heterogeneity. Second, we identified and characterized the orchestrated roles of a hub niche comprising APOE+ macrophages, EMT-like cancer-associated fibroblasts (CAFs), and RBP7+ endothelial cells in thyroid cancer progression. This contrasts with previous studies, which primarily focused on thyroid epithelial cells or individual immune cell types. Third, to the best of our knowledge, this study is the first to report the presence of CD4 + HSPA1A1 + T cells in anaplastic thyroid cancer (ATC) and to investigate their potential biological roles in immunotherapy. These findings provide novel insights into the tumor microenvironment and highlight potential therapeutic targets for aggressive thyroid cancer subtypes.

In this study, our single-cell RNA sequencing (scRNA-seq) analysis offers a unique and comprehensive resource for dissecting the molecular landscape of papillary thyroid cancer (PTC), anaplastic thyroid cancer (ATC), and metastatic thyroid cancer at an unprecedented single-cell resolution. By analyzing a large-scale dataset encompassing diverse thyroid cancer subtypes and normal tissues, we unveiled the intricate intratumoral heterogeneity of thyroid cancer cells, uncovering their distinct molecular signatures and signaling pathway activities. Furthermore, we meticulously characterized the cellular composition and functional properties of immune cell subtypes within the tumor microenvironment (TME), shedding light on their roles in tumor progression and immune evasion. Notably, this study is the first to propose APOE+ macrophages, EMT-like cancer-associated fibroblasts (CAFs), and RBP7+ endothelial cells as a coordinated hub niche that drives the remodeling of the tumor ecosystem in thyroid cancer. These hub niche cells orchestrate critical interactions through key signaling pathways, such as ANGPT, SELE, and MIF, which collectively promote tumor growth, immune suppression, and metastatic dissemination. Our findings highlight the central role of this multicellular hub in shaping the TME and provide a deeper understanding of the cellular and molecular dynamics underlying thyroid cancer progression. In conclusion, this work establishes a novel foundation for identifying therapeutic targets and advancing precision medicine strategies in thyroid cancer. By elucidating the complex cellular interactions and molecular mechanisms that drive tumor progression and metastasis, our findings offer critical insights for the development of targeted therapies. These advancements hold promise for improving clinical outcomes and addressing the unmet therapeutic needs of patients with aggressive and metastatic thyroid cancer.

Methods

Patient data and tissue specimens

PTC tissues, metastatic lesions, and ATC samples were collected from patients who underwent surgical resection in May 2024. Immediately after surgery, tissue specimens were preserved in formalin for subsequent multiplex immunohistochemistry (mIHC) analysis. To minimize potential confounding factors, all patients included in the study were diagnosed through histopathological examination, and those who had previously received chemotherapy or radiotherapy were excluded. Prior to participation, all patients provided written informed consent. This study was conducted in accordance the guidelines of the Declaration of Helsinki (2013 amendment) and was approved by the Ethics Committee of the Second Affiliated Hospital, School of Medicine, Zhejiang University (Approval No. 2024-0548).

Sources of datasets

The single-cell RNA sequencing data were obtained from the Gene Expression Omnibus (GEO) database, including datasets GSE23223722, GSE19358123, GSE19128824, GSE18436225, and GSE241184, which collectively comprise 50 thyroid cancer tumor samples and 14 normal thyroid biopsies. Since all data were publicly available and open-access, no additional approval from the Ethics Committee was required for their use in this study.

Quality control and the dimensionality reduction

The Seurat object containing gene expression data was imported into the Seurat (v2.3.0) R toolkit using the Read10×() function26. The gene-cell matrices were filtered to remove low-quality cells (those with <200 transcripts/cell, >6000 transcripts/cell, or >25% mitochondrial genes) and genes expressed in fewer than three cells. For each sample, gene expression values were normalized by calculating the fraction of each gene’s expression relative to the total expression per cell, multiplied by 10,000, and then log-transformed after adding a pseudocount of 1 to avoid taking the logarithm of zero. The top 1000 highly variable genes (HVGs) were identified from the normalized expression matrix and used as input for principal component analysis (PCA). Significant principal components were determined using Jackstraw analysis, and heatmaps were generated to visualize the contribution of genes to the top 40 principal components. Based on this analysis, the first 20 principal components (PCs) were selected for downstream graph-based clustering to identify distinct cell populations. The Seurat functions FindNeighbors() and FindClusters() were applied to compute dimension-reduction coordinates, and the resulting single-cell clusters were visualized using Uniform Manifold Approximation and Projection (UMAP).

Cell-clustering and annotation

Cluster-specific marker genes were identified using the FindAllMarkers function in the Seurat package, which employs a non-parametric Wilcoxon rank sum test with Bonferroni correction as the default statistical method. To detect differentially expressed genes (DEGs) between two specific clusters, the find.markers function was utilized. Cell groups were annotated based on the identified DEGs and well-established cellular markers referenced from the literature.

DEGs identification and gene set variation analysis

We identified differentially overexpressed genes in specific clusters compared to all other clusters using the FindMarkers function in Seurat, applying the Wilcoxon Rank-Sum Test with the following parameters: adjusted P-value < 0.05, only.pos = TRUE, and logfc.threshold = 0.25. To evaluate the activation of hallmark pathways and metabolic pathways, gene set variation analysis (GSVA) was performed using gene sets downloaded from the MSigDB database (https://www.gsea-msigdb.org/gsea/msigdb)27. Subsequently, the AddModuleScore function in the Seurat package was employed to assign pathway activity scores, enabling the assessment of relative pathway activities within specific cell types based on their active gene signatures.

Cell–cell interactions analysis with CellChat

Cell-cell communication was analyzed using the CellChat package28, which evaluates the expression of known ligand-receptor pairs. Briefly, the count data were imported into CellChat with default parameters. The “Secreted Signaling” pathways were selected as the database, incorporating pre-compiled human protein-protein interactions (PPIs) as prior network knowledge. Core analyses were performed using the functions computeCommunProbPathway, computeCommunProb, and aggregateNet with standard settings and fixed randomization. Differences in cell-cell interactions between groups were considered significant at p < 0.05.

To identify potential ligands driving the phenotype of specific clusters, we employed the NicheNet algorithm29. First, differentially expressed genes (DEGs) for each cluster were calculated. The top 350 DEGs with an AUC > 0.1 were selected as the gene sets of interest, while all expressed genes in the scRNA-seq dataset served as the background. Potential ligands included all expressed ligands in the dataset, and only receptors expressed in the indicated cluster were used to construct ligand-receptor interactions and compute ligand activity. The top 20 predicted ligands, ranked by ligand activity, were further analyzed to calculate their mean expression levels across each cluster.

multiplex IHC

We performed multiplex immunofluorescence staining using the following primary antibodies: THBS1 rabbit anti-human antibody (Affinity Biosciences; catalog no. DF6848), CD47 rabbit anti-human antibody (Affinity Biosciences; catalog no. DF6649), CD68 rabbit anti-human antibody (Affinity Biosciences; catalog no. DF7518), APOE rabbit anti-human antibody (Affinity Biosciences; catalog no. AF5178), CD3 rabbit anti-human antibody (Affinity Biosciences; catalog no. DF6848), THBS1 rabbit anti-human antibody (Affinity Biosciences; catalog no. DF6594), BSG rabbit anti-human antibody (Affinity Biosciences; catalog no. AF5221), COL3A1 rabbit anti-human antibody (Affinity Biosciences; catalog no. AF5457), PPIA rabbit anti-human antibody (Proteintech Group; catalog no. 10720-1-AP), and SAA1 rabbit anti-human antibody (Proteintech Group; catalog no. 16721-1-AP). Following the manufacturer’s protocol (PerkinElmer Opal Kit), the stained slides were scanned using the PerkinElmer Vectra3 platform, and the results were quantified using the same platform for accurate and reproducible analysis.

Statistical analysis

All statistical analyses were conducted using R (http://www.r-project.org). Depending on the experimental design, either a two-sided paired or unpaired Student’s t-test, or an unpaired Wilcoxon rank-sum test, was applied as appropriate. A P-value < 0.05 was considered statistically significant.