Single-cell sequencing profiling of intratumoral heterogeneity and immunosuppressive microenvironment in primary thyroid cancer and lymph node metastases

ABSTRACT

Metastasis is a major determinant of treatment failure and mortality in thyroid cancer, yet the interplay between malignant evolution and the immune microenvironment remains poorly characterized. Immunotherapy offers promise, but its efficacy requires a deeper understanding of tumor-associated immune infiltration and checkpoint regulation. In this study, we constructed a high-resolution transcriptomic atlas of the thyroid cancer ecosystem by analyzing 55,005 single cells from paired primary tumors and lymph node metastases. By integrating chromosomal copy number variation (CNV) inference with consensus nonnegative matrix factorization (cNMF), we deciphered the intrinsic heterogeneity of malignant epithelial cells, revealing distinct transcriptional programs and developmental trajectories driving the metastatic cascade. The metastatic niche exhibited significant reprogramming of the immunosuppressive landscape, characterized by the enrichment of FOXP3⁺ regulatory T (Treg) cells, LAMP3⁺ dendritic cells (DCs), and CCL18⁺ M2-like macrophages. Notably, while canonical checkpoints PD-1 and PD-L1/2 showed minimal expression, ligand-receptor interaction analysis identified the LAG3-LGALS3 axes as dominant immune evasion pathways mediating the crosstalk between CD8⁺ T cells and the tumor stroma. In conclusion, this study comprehensively maps the coevolution of malignant thyrocyte plasticity and the immunosuppressive metastatic niche. By uncovering the specific role of LAMP3⁺ DCs and identifying LAG3/TIGIT as critical alternative checkpoints, our findings challenge the utility of conventional PD-1 blockade in this context and provide a robust molecular rationale for developing next-generation immunotherapeutic strategies tailored to thyroid cancer. Although limited by a modest sample size, these findings provide a foundation for further investigation of the metastatic immune landscape in thyroid cancer.

1. Introduction

Thyroid cancer is the most common malignant tumor of the endocrine system, accounting for approximately 3% of all malignant tumors worldwide. ¹ According to the latest global cancer statistics, there were 586,000 new cases of thyroid cancer in 2020, ranking it among the most prevalent malignant tumors globally. ¹ Notably, the number of new cases in females reached 448,915, which was approximately three times that in males (137,287 cases). In recent years, the incidence of thyroid cancer has been gradually increasing, presenting a still severe clinical situation. ² Thyroid cancer encompasses multiple subtypes, among which thyroid papillary carcinoma (PTC) accounts for 85%–90% of all cases. Approximately 90% of PTC cases are indolent; these tumors exhibit slow progression and high inertness, and most patients can be cured through standard treatment, with a 5-y survival rate exceeding 95% and a favorable prognosis. ³ However, nearly 10% of PTC cases are aggressive, prone to local recurrence or distant metastasis. Once recurrence or metastasis occurs, these patients show poor responses to traditional treatments such as surgery, ¹³¹I therapy, and postoperative thyroid-stimulating hormone (TSH) suppression therapy, leading to worse prognosis and quality of life. ^4-6 Over the past decade, targeted immune checkpoint therapy has emerged as an effective treatment strategy. ^{7 , 8} Although the immune checkpoint therapy for metastatic thyroid cancer is promising, limited patients derive benefits from it. Therefore, there is an urgent need to develop more effective immunotherapeutic strategies for metastatic thyroid cancers.

The tumor microenvironment forms a self-regulating ecosystem, ⁹ the interaction and communication in the tumor microenvironment, along with the diverse cellular phenotypes, configure the site-specific tumor ecosystem. This interplay may also contribute to varying responses to immune checkpoint therapy. ^{10 , 11} In recent years, single-cell RNA sequencing (scRNA-seq) has emerged as a valuable tool in studying tumor cell heterogeneity, identifying new mutation sites, investigating tumor cell cloning and evolution mechanisms, and discovering new therapeutic targets. ^{12 , 13} scRNA-seq allows for the dissection of tumor-infiltrating immune cell heterogeneity and their crosstalk with diverse cellular populations within the tumor microenvironment. ^{14 , 15} Several studies have examined the heterogeneity within various subtypes of primary thyroid cancer and identified cell clusters associated with unfavorable prognosis or treatment response. Pu et al. conducted an analysis of paratumors, localized/advanced tumors in 11 patients and identified a premalignant thyrocyte population that is “cancer-primed” and displays normal morphology but altered transcriptomes. Additionally, they discovered premalignant thyrocyte populations and identified three phenotypes of malignant thyrocytes, which contribute to molecular subtypes, tumor characteristics, and treatment response. ¹⁶ Cao et al. employed single-cell RNA sequencing (scRNA-seq) to investigate intercellular communication networks and predict the effectiveness of immunotherapy. The analysis of ligand-receptor (LR) pairs revealed the communication network between cells and led to the development of a new molecular phenotype for PTC based on LR pairs. ¹⁷ Pan et al. utilized integrated scRNA-seq and single-cell assay for transposase-accessible chromatin using sequencing (scATAC-seq) methods to examine immune cell dynamics in eight PTC patients, three of whom had concurrent Hashimoto thyroiditis. Their findings indicate that the presence of tumor-infiltrating B lymphocytes is primarily associated with concurrent HT origin. ¹⁸ However, while scRNA-seq studies on the primary lesion in thyroid cancer have been conducted, limited studies to date have reported systematic single-cell characterization of metastatic lesions. While paired analyzes of primary tumors and lymph node metastases have recently emerged, the functional divergence of immunosuppressive populations and alternative checkpoint axes in the metastatic ecosystem remain incompletely defined. The present study addresses this by focusing on LAMP3⁺ dendritic cells and noncanonical immune evasion pathways at single‑cell resolution. Recent single-cell RNA sequencing studies in breast cancer have provided critical insights into the remodeling of the tumor microenvironment during lymphatic dissemination. One investigation performed single-cell RNA sequencing on 28 lymph node samples from 23 breast cancer patients and identified GLO1 as a key mediator of paracrine-driven microenvironment remodeling that potentiates lymph node metastasis, revealing an increase in APOE⁺ macrophages and exhausted CD8⁺ T cells within metastatic lymph nodes. ¹⁹ Another study characterized single-cell profiles across primary breast tumors, sentinel lymph nodes, and metastatic lymph nodes, delineating the stepwise immunological changes that accompany nodal dissemination. ²⁰ These findings underscore the power of single-cell resolution approaches in dissecting metastatic niche biology. However, systematic single-cell characterization of lymph node metastases in thyroid cancer remains limited, underscoring the necessity of the present investigation.

In this study, we employed scRNA-seq to profile intratumoral heterogeneity and the immunosuppressive microenvironment in primary and metastatic lesions. By analyzing the single-cell resolution data, we identified various cell types present in these lesions, including cancer cells, epithelial cells, mural cells, B cells, NK/T cells, neutrophils, mast cells, MPs, and DCs. Notably, the metastatic ecosystem displayed significant reprogramming of immunosuppressive cells, such as FOXP3+ Treg cells, LAMP3+ DCs, and CCL18+ M2-like macrophages. These findings shed light on the most effective immunotherapeutic strategies for lymph node metastatic thyroid cancer.

2. Materials and methods

2.1. Human specimen collection and ethics statement

This study enrolled three patients diagnosed with thyroid carcinoma who underwent surgical resection at Sun Yat-sen University Cancer Center (Table S1). The study was conducted from March 2023 to December 2025. Paired primary tumor tissues and corresponding lymph node metastatic lesions were harvested. Ethical approval was granted by the Institutional Review Board and all participants provided written informed consent prior to inclusion. All identifiable personal information (pathology and medical record numbers) has been removed in accordance with the International Committee of Medical Journal Editors (ICMJE).

2.2. Tissue dissociation and single-cell suspension generation

Immediately following surgical excision, fresh tissue samples were preserved in sCelLive™ Tissue Preservation Solution (Singleron, China) on ice. To generate single-cell suspensions, samples were rinsed thrice with Hank's balanced salt solution (HBSS) and subjected to enzymatic digestion using sCelLive™ Tissue Dissociation Solution (Singleron) at 37 °C for 15 minutes with gentle agitation. Following digestion, erythrocytes were depleted by incubation with red blood cell lysis buffer at 25 °C for 10 minutes. The resulting suspension was centrifuged at 500 × g for 5 minutes and resuspended in PBS. Cellular viability was assessed via Trypan blue exclusion assays, and only samples exceeding 90% viability were processed for sequencing. ¹³

2.3. scRNA-seq library construction and sequencing

Single-cell capture and library preparation were executed utilizing the Chromium Next GEM Single Cell 3ʹ Kit v3.1 (10× Genomics). The cell suspension was loaded onto the Chromium Controller to generate Gel Beads-in-emulsion (GEMs), enabling cell lysis and barcoded reverse transcription.

2.4. Data preprocessing and quality control

Raw BCL files were demultiplexed and converted to FASTQ format. Alignment to the human reference genome (GRCh38) and quantification of gene expression matrices were performed using Cell Ranger (version 7.0.0). Downstream analysis was conducted in R using the Seurat package (version 4.3.0). Cells were retained only if they met the following criteria: (1) detection of >200 unique features (genes); (2) gene counts and UMI counts falling within the biologically reasonable range (excluding the top 2% to remove potential doublets); and (3) mitochondrial gene content < 20%.

2.5. Dimensionality reduction and cell type annotation

Postfiltering, data normalization was performed using the “NormalizeData” function, followed by variance stabilization via “FindVariableFeatures”. Principal component analysis (PCA) was utilized for dimensionality reduction. To visualize the data structure, uniform manifold approximation and projection (UMAP) was generated based on the top significant principal components. Unsupervised clustering was achieved using the shared nearest neighbor (SNN) modularity optimization technique (“FindClusters” function) with a resolution set to 1.2. Differentially expressed genes (DEGs) for each cluster were identified using the Wilcoxon rank-sum test via the “FindMarkers” function (logFC > 0.25, min.pct > 0.1). Cell identities were assigned by cross-referencing identified DEGs with canonical markers cataloged in the SynEcoSys database.

2.6. Inference of malignancy and CNV analysis

To discriminate malignant epithelial cells from their nonneoplastic counterparts, chromosomal copy number variations (CNVs) ²¹ were inferred using the InferCNV R package (version 1.12.0). Immune cells (T and B lymphocytes) from the same microenvironment served as the diploid reference baseline. Relative expression levels were normalized to a center value of 1, with a threshold ceiling set at 1.5 standard deviations derived from residual-normalized expression data. A sliding window of 101 genes was applied to smooth expression signals across chromosomes. Cells exhibiting high CNV scores and consistent chromosomal aberrations were classified as malignant.

2.7. Characterization of transcriptional programs (cNMF)

To dissect the intrinsic heterogeneity of cancer cells, we employed the consensus nonnegative matrix factorization (cNMF) algorithm (https://github.com/dylkot/cNMF). This approach allowed for the extraction of latent transcriptional programs. The top 100 genes contributing to each factor were defined as the meta-signature, and program usage scores were calculated for individual cells to define functional modules.

2.8. Functional enrichment analysis

Biological interpretation of cell subsets was performed using the clusterProfiler R package (version 4.6.0). Gene Ontology (GO) biological process and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyzes were conducted on cluster-specific DEGs. Statistical significance was defined as an adjusted P-value < 0.05. Additionally, gene set variation analysis (GSVA) was utilized to evaluate the relative activity of hallmark pathways (MSigDB) across different cell populations.

2.9. Trajectory inference and RNA velocity

Developmental trajectories were reconstructed using Monocle 2. Ordering genes were selected based on high dispersion, and dimensionality reduction was performed using the DDRTree method to visualize the pseudotime continuum. To predict the future state of individual cells, RNA velocity analysis was computed using velocyto (Python). Spliced and unspliced transcript abundances were quantified, and velocity streams were projected onto the UMAP embedding to infer directional cellular differentiation.

2.10. Stemness scoring and cell cycle evaluation

Cellular differentiation potential was quantified using the SLICE (Single Cell Lineage Inference using Cell Expression variability) algorithm, which computes single-cell entropy (scEntropy) as a proxy for stemness. Cell cycle phase assignment was performed within Seurat using a predefined list of cell cycle-dependent genes.

2.11. Intercellular communication network analysis

Cell‒cell interactions within the tumor microenvironment were deciphered using CellphoneDB (version 4.0.0). We analyzed the expression of ligand-receptor pairs between cancer cells and stromal/immune subsets. Significance was determined via permutation testing (1,000 iterations). Interactions with a P-value < 0.05 and biologically relevant expression levels were visualized to map the communication landscape.

2.12. Statistical framework

All statistical computations were executed in the R statistical environment (version 4.2.0). Continuous variables were compared using Student's t-test or Wilcoxon rank-sum test, while categorical variables were assessed via Fisher's exact test. Correlation analyzes were performed using Pearson or Spearman coefficients. Data visualization was generated using “ggplot2” and Seurat plotting functions. A P-value < 0.05 was considered statistically significant.

3. Results

3.1. Single-cell transcriptomic profiling unveils the cellular ecosystem of thyroid cancer

To comprehensively characterize the cellular heterogeneity within the tumor microenvironment (TME) of thyroid cancer, we performed single-cell RNA sequencing (scRNA-seq) on matched primary thyroid cancer (TC) tissues and thyroid cancer lymph node metastasis (TLNM) specimens obtained from three patients (Figure 1A). There were 55,005 high-quality single cells were retained for subsequent bioinformatic analyzes. Unsupervised graph-based clustering was implemented using the Seurat pipeline to delineate major cell populations exhibiting similar transcriptomic profiles (Figure 1B). Specifically, we identified the following ten major cell types: epithelial cells (marker gene: EPCAM), endothelial cells (ECs; marker genes: PECAM1, VWF, CDH5), mural cells (marker genes: ACTA2, CALD1, MCAM), proliferating cells (marker genes: MKI67, TOP2A, STMN1), B cells (marker genes: CD79A, MS4A1, CD19), T and natural killer (NK) cells (marker genes: CD3D, CD3E, CD3G), neutrophils (marker genes: FCGR3B, S100A9, S100A8), mast cells (marker genes: TPSAB1, TPSB2, CPA3), mononuclear phagocytes (MPs; marker genes: CD14, CSF1R, ITGAM), and plasmacytoid dendritic cells (pDCs; marker genes: CLEC4C, IL3RA, LILRA4) (Figure 1C and D). Quantitative analysis of cell composition revealed the following distribution: epithelial cells constituted the largest population (n = 22,707), followed by T and NK cells (n = 14,122), MPs (n = 8251), B cells (n = 3349), ECs (n = 3037), mural cells (n = 1586), neutrophils (n = 820), proliferating cells (n = 503), mast cells (n = 473), and pDCs (n = 157) (Table S2). Cluster-specific marker expression was visualized via dot plots, illustrating both scaled expression values and the proportion of expressing cells per subpopulation (Figure 1E). To validate the presence and spatial distribution of these identified cell populations in situ, we performed multiplex immunofluorescence staining on both primary TC tissues and TLNM tissues, which confirmed the existence of the annotated cell subpopulations (Figure 1F and G). Given the limited number of patients (n = 3), these findings represent an initial characterization requiring validation in larger cohorts.

Figure 1.

Single-cell transcriptomic profiling reveals the cellular ecosystem of thyroid cancer. (A) Schematic illustration of the experimental workflow. Matched primary thyroid cancer (TC) and thyroid cancer lymph node metastasis (TLNM) tissues were collected from three patients and subjected to single-cell RNA sequencing (scRNA-seq) followed by comprehensive bioinformatic analyzes. (B) Uniform manifold approximation and projection (UMAP) visualization of 55,005 single cells, displaying major cell clusters identified by unsupervised clustering analysis. (C) Feature plots depicting the expression of canonical marker genes used for cell type annotation. (D) UMAP plot with cells color-coded according to their annotated cell type identities. (E) Dot plot illustrating the scaled expression levels (color intensity) and the percentage of cells expressing (dot size) cluster-specific marker genes across all identified cell subpopulations. (F–G) Representative multiplex immunofluorescence staining images validating the presence of major cell populations in primary TC tissues (F) and TLNM tissues (G).

3.2. Transcriptional heterogeneity and functional modules of malignant cells in thyroid cancer

To further investigate the intrinsic heterogeneity of malignant cells, we analyzed the transcriptomic patterns and performed subpopulation clustering of cancer cells. The InferCNV analysis identified a total of 18,032 malignant epithelial cells, comprising 11,555 cells from primary TC and 6477 cells from TLNM. Based on differentially expressed genes (DEGs), nine major cancer cell subclusters were identified and annotated (Figure 2A). Feature plots and heatmaps were employed to visualize the expression profiles of highly variable genes across cancer cell subclusters (Figure 2B and C).

Figure 2.

Transcriptional heterogeneity and functional modules of malignant cells in thyroid cancer. (A) Nine cancer cell subclusters, identified based on differential gene expression profiles, visualized on a UMAP plot. (B) Feature plot visualization depicting the distribution of highly variable gene expression among cancer cell subclusters. (C) Heatmap showing the top DEGs defining each cancer cell subcluster. (D) Identification of six distinct gene expression modules using meta-cluster algorithm analysis. Representative genes for each module are indicated. (E) Comparative unsupervised clustering analysis of cancer cells from primary TC (seven subclusters) and TLNM (five subclusters). (F) The potential biological functions and signaling pathways of primary TC (seven subclusters) and TLNM (five subclusters) were evaluated by GSVA with hallmark gene sets.

To delineate the functional heterogeneity of cancer cells between primary thyroid tumors and lymph node metastases, we applied a meta-cluster algorithm to the scRNA-seq data and identified six distinct gene expression modules (Figure 2D). Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis was subsequently performed to characterize the predominant biological pathways enriched within each coexpression module (Figure S1). Among these modules, Modules 1 and 2 represented the differentiated thyroid phenotype. Module 1 contained canonical thyroid differentiation markers, including thyroid peroxidase (TPO) and trefoil factor 3 (TFF3), whereas Module 2 was characterized by the expression of thyroglobulin (TG) and thyroid-stimulating hormone receptor (TSHR), with KEGG analysis demonstrating significant enrichment in the “thyroid hormone synthesis” pathway, thereby confirming the retention of thyroid-specific differentiation features in these cell subsets. In contrast, Module 6 was identified as the signaling-driven proliferative module, characterized by elevated expression of immediate-early response genes including FOS, JUN, and early growth response 1 (EGR1), and showed significant enrichment in the “mitogen-activated protein kinase (MAPK) signaling pathway”, suggesting that this module represents a subpopulation of cancer cells with hyperactivated MAPK signaling, which is a well-established driver of thyroid tumorigenesis. Modules 3 and 5 reflected tumor-stroma interactions and invasive potential. Specifically, Module 5, which expressed keratin 19 (KRT19), epithelial cell adhesion molecule (EPCAM), and midkine (MDK), was enriched in “focal adhesion” and “proteoglycans in cancer” pathways, while Module 3 displayed features of epithelial-mesenchymal transition (EMT), with elevated expression of mesenchymal markers such as fibronectin 1 (FN1) and vimentin (VIM), and pathway analysis highlighted enrichment in “cell adhesion molecules”, collectively suggesting that these cells have acquired invasive properties that may promote metastatic dissemination. Notably, Module 4 revealed a unique immunogenic state, as evidenced by high expression of major histocompatibility complex (MHC) Class II genes, including human leukocyte antigen-DRA (HLA-DRA) and CD74, as well as inflammatory mediators such as chitinase-3-like protein 1 (CHI3L1), which correlated with enrichment in “antigen processing and presentation” and “allograft rejection” pathways, indicating extensive crosstalk between this subset of tumor cells and the immune microenvironment.

Based on DEGs, seven cancer cell subclusters were identified in primary tumors, whereas five subclusters were annotated in lymph node metastases (Figure 2E). Hallmark pathway enrichment was assessed via GSVA to delineate the functional signatures and signaling programs underlying each cell type (Figure 2F). The analysis revealed distinct pathway activation patterns between primary and metastatic sites. In primary TC tissues, the subclusters exhibited enrichment in cell cycle-related pathways such as G2/M checkpoint and mitotic spindle, immune regulation pathways including interferon-gamma (IFN-γ) response, interferon-alpha (IFN-α) response, and allograft rejection, as well as metabolic pathways encompassing oxidative phosphorylation and glycolysis. In contrast, TLNM subclusters were characterized by enhanced activation of oncogenic signaling pathways, notably PI3K-AKT-mTOR signaling, transforming growth factor-beta (TGF-β) signaling, Hedgehog signaling, and KRAS signaling upregulation, along with pro-metastatic processes including EMT and angiogenesis, and immune and inflammatory response pathways such as inflammatory response and interleukin-6 (IL-6)-JAK-STAT3 signaling. These findings indicate that metastatic cancer cells undergo significant transcriptional reprogramming to facilitate survival and colonization in lymph nodes, while maintaining dynamic interactions with the lymph node immune microenvironment.

3.3. Genomic instability and developmental trajectories underlying thyroid cancer evolution

Thyroid cancer development is substantially driven by gene copy number alterations. To explore the genomic landscape of thyroid cancer at single-cell resolution, scRNA-seq-based CNV analysis was performed on malignant cells, revealing marked chromosomal heterogeneity in tumor cells. (Figure 3A). We clustered cancer cells harboring CNVs and compared these clusters with T cells and B cells (Figure 3B). Recurrent copy number gains on chromosomes 8 and 20, along with losses on chromosomes 6 and 10, were observed in the majority of cancer cells (Figure 3C).

Figure 3.

Genomic instability and developmental trajectories underlying thyroid cancer evolution. (A) Heatmap displaying inferred copy number variation (CNV) profiles across primary thyroid cancer (TC) and thyroid cancer lymph node metastasis (TLNM) samples. (B) Clustering of cancer cells based on CNV scores, with T cells and B cells serving as diploid reference populations. (C) Summary of recurrent chromosomal amplifications (chromosomes 8 and 20) and deletions (chromosomes 6 and 10) observed in cancer cells. (D) Gene set variation analysis (GSVA) comparing pathway enrichment between high-CNV and low-CNV cancer cell groups. (E) Gene expression dynamics along the pseudotime trajectory. (F–G) Cell cycle phase distribution (F) and proliferation scores (G) across cancer cell subclusters, identifying CancerCells_6 as the most proliferative population. (H) Pseudotime analysis reveals cancer cell maturation trajectories. (I) Heatmap of genes with dynamic expression patterns during cell state transitions along the pseudotime axis. (J) RNA velocity analysis projected onto the UMAP plot, showing directional differentiation trajectories from subclusters c3, c5, c8, c9 toward c2, c1, c4, c6. (K–L) SLICE algorithm analysis showing entropy distribution (K) and stemness scores (L) for each cancer cell subcluster. (M–N) Expression profiles of immune checkpoint ligands (CD274, PDCD1LG2, LGALS3, CD47) across cancer cell subclusters, visualized by UMAP feature plots (M) and dot plots (N).

To investigate the functional implications of CNV burden, we stratified cancer cells into high-CNV and low-CNV groups. GSVA demonstrated that pathways including the reactive oxygen species (ROS) pathway, DNA repair, fatty acid metabolism, and MYC targets V2 were enriched in the low-CNV group, whereas Hedgehog signaling and mitotic spindle pathways were significantly downregulated in this group (Figure 3D). Using canonical cell cycle markers, we calculated cell cycle phase scores and found that cells in G1, G2/M, and S phases were intermixed throughout cancer cell subclusters, indicating that cell cycle status is not the primary determinant of subcluster differentiation (Figure 3F). Instead, subclusters are likely defined by functional specialization, with CancerCells_6 representing the most actively proliferating population (Figure 3G).

To characterize cancer cell lineage trajectories, we performed pseudotime analysis using the Monocle 2 algorithm (Figure 3H) and dissected dynamic gene expression patterns associated with cell state transitions (Figure 3E and I). To further characterize the temporal dynamics of tumor cell evolution, we employed RNA velocity analysis to map cellular developmental lineages. Notably, four cancer cell subclusters (c3, c5, c8, c9) exhibited differentiation trajectories toward another four subclusters (c2, c1, c4, c6), suggesting directional clonal evolution within the tumor (Figure 3J). To assess the stemness potential of each subcluster, we calculated gene expression entropy using the SLICE algorithm (Figure 3K). The analysis revealed that CancerCells_9 exhibited the highest entropy, indicating stronger stemness and progenitor-like characteristics, while CancerCells_7 demonstrated the lowest entropy, suggesting a more differentiated state (Figure 3L).

To evaluate tumor-intrinsic immunosuppressive mechanisms, we profiled immune checkpoint ligand expression across cancer cell subclusters. Notably, the canonical immunotherapy targets CD274 (programmed death-ligand 1, PD-L1) and PDCD1LG2 (programmed death-ligand 2, PD-L2) showed low expression levels across all subclusters (Figure 3M and N). In contrast, the immune checkpoint ligand LGALS3 (galectin-3) was highly expressed across nearly all subclusters. Intriguingly, CD47, an antiphagocytic “don't eat me” signal molecule, also demonstrated robust expression (Figure 3M and N). These molecules represent potential alternative immunotherapeutic targets for thyroid cancer.

3.4. Single-cell profiling of primary thyroid tumors and paired lymph node metastases reveals lymphocyte divergence

To comprehensively delineate the immune microenvironment across thyroid cancer progression, we performed high-resolution subclustering of T and B cells. A total of ten T and innate lymphoid subclusters were identified (Figure 4A). These populations spanned the developmental spectrum, including naive states (Tnaive_JUNB, Tnaive_CCR7), specialized CD4+ subsets (CD4Treg_FOXP3, CD4Tfh_CXCL13), and various CD8+ effector/memory states (CD8Teff_GZMK, CD8Tem_FGFBP2, CD8Trm_ZNF683). The distinctive marker expression for each lineage, such as FOXP3 for Tregs and NK_XCL1 for NK cells, was validated across the integrated dataset (Figure 4B). Similarly, reclustering of B cells from both anatomical sites yielded five major clusters: NaiveB_IGHD, MemoryB_BANK1, MemoryB_TNFSF9, GCB_RGS13, and Plasma_IGHG1 (Figure 4C and D). The distribution and canonical markers of these lymphoid populations are summarized in Figure 4E.

Figure 4.

Single-cell profiling of primary thyroid tumors and paired lymph node metastases reveals lymphocyte divergence. (A) UMAP visualization of integrated T and NK cell subclusters from 3 pairs of primary tumors (PT) and lymph node metastases (LNM). (B) Feature plots demonstrating lineage-specific markers (CCL3, ZNF683, FGFBP2, GZMK, FOXP3, CXCL13) across the joint dataset. (C) UMAP visualization of B cell subclustering into five major subpopulations. (D) Feature plots of key B cell marker genes (IGHD, TNFSF9, BANK1, JCHAIN, RGS13). (E) Dot plot showing the expression levels of top markers across lymphoid clusters; size indicates expression frequency, and color indicates average intensity. (F–G) Heatmaps displaying functional signature scores and relative expression of genes related to cytotoxicity, MHC molecules, and immune checkpoints across clusters. (H) Detailed comparison of cytotoxicity and exhaustion scores within the CD8+ and NK cell compartments. (I) Violin plots comparing “Exhaustion scores” specifically between Primary Tumors (TC) and Lymph Node Metastases (TLNM) for CD4Treg_FOXP3 and Tnaive_JUNB clusters. Statistical significance was determined using Student's t-test. **p<0.01, ***p<0.001. (J–K) Enrichment heatmaps (GO/KEGG) for T, NK, and B cell subpopulations reflecting their biological roles in the tumor-metastasis axis. (L-M) Pseudotime trajectory models for T cells (L) and B cells (M), indicating the transition from naive states to terminal differentiated states (colored by cluster and pseudotime). (N–P) Expression analysis of immune checkpoint receptors (CD96, LAG3, TIGIT, PDCD1) across CD8+ T-cell clusters, visualized by violin plots (N), dot plots (O), and UMAP feature plots (P).

Functional characterization using gene set scoring demonstrated a diverse distribution of MHCI, MHCII, cytotoxicity, and exhaustion signatures across the T and NK cell populations (Figure 4F). Detailed examination of functional genes showed that while CD4+ T-cell clusters expressed costimulatory molecules like TNFRSF4 (OX40) and ICOS, CD8+ T-cell subsets were enriched for cytotoxic effectors such as GZMA and GZMK (Figure 4G). Among the CD8+ compartment, the CD8Trm_ZNF683 and Pro_CD8Teff_STMN1 clusters displayed markedly higher exhaustion scores compared to the CD8Teff_GZMK and CD8Tem_FGFBP2 groups (Figure 4H). Notably, a site-specific comparison revealed that the CD4Treg_FOXP3 and Tnaive_JUNB clusters harbored significantly higher exhaustion scores in the lymph node metastasis (TLNM) samples than in the primary tumor (TC) samples (Figure 4I), suggesting a more suppressive niche in the metastatic compartment.

To further understand the biological roles of these cells, pathway enrichment analysis was performed. The CD8Trm_ZNF683 cluster was associated with T-cell proliferation, response to interferon-gamma, and immune checkpoint pathways, while the NK_CCL3 cluster showed enrichment in necroptosis and the RIG-I-like receptor signaling pathway (Figure 4J). For B cells, the NaiveB_IGHD cluster showed heightened activity in antigen processing and presentation, whereas the Plasma_IGHG1 cluster was primarily involved in protein targeting to the ER and immunoglobulin complex circulation (Figure 4K). Pseudotime trajectory analysis reconstructed the developmental transition of T cells from a naive state toward a terminal effector status, with the CD8Tem_FGFBP2 cluster positioned at the trajectory terminus (Figure 4L). Similarly, the B cell lineage demonstrated an evolutionary path ending at the Plasma_IGHG1 cluster, which exhibited the highest pseudotime score (Figure 4M).

Finally, we assessed the immune checkpoint profile of the CD8+ T-cell subpopulations. Inhibitory receptors, including CD96, LAG3, and TIGIT, exhibited higher expression levels across CD8+ clusters compared to PDCD1 (PD-1) (Figure 4N). This pattern was consistent across the lymphoid landscape, as shown by the high-resolution dot plot of coinhibitory and costimulatory receptors such as TIGIT, LAG3, ICOS, and TNFRSF18 (Figure 4O). The spatial distribution of key checkpoint-related genes, such as CD28, BTLA, CD96, and CD226, was further confirmed on the UMAP embedding, highlighting the specific activation and exhaustion states across the primary and metastatic niche (Figure 4P).

3.5. Myeloid cell diversity and functional polarization within the thyroid cancer niche

Through unsupervised clustering analysis, we identified 12 distinct myeloid cell subsets (Figure 5A). Myeloid signature markers were visualized via feature plots (Figure 5B), with subtype proportions shown in histograms (Figure 5C) and cluster-specific marker profiles summarized in dot plots (Figure 5D). To further dissect functional polarization of the myeloid compartment, we examined representative immune-regulatory and phenotypic genes across the six TAM subsets using violin plots (Figure 5E). Pro-inflammatory TNF was enriched in TAM_c1_FCGBP, TAM_c3_CCL8, and TAM_c4_CXCL8 but largely absent in the others, indicating a more inflammatory M1-like state in these clusters. The phagocytosis checkpoint receptor SIRPA was broadly expressed across all subsets, whereas NECTIN2 showed only modest expression, mainly in TAM_c1_FCGBP and TAM_c5_SPP1. Notably, the scavenger receptor MARCO, together with LGALS3 and APOE, was preferentially enriched in TAM_c5_SPP1, identifying this subset as a lipid-laden, scavenging, immunosuppressive population distinct from the inflammatory TNF⁺ subsets, underscoring the functional heterogeneity of tumor-associated macrophages within the thyroid cancer microenvironment. In addition, to trace the maturation and differentiation trajectory of TAMs, we employed the Monocle 2 algorithm (Figure 5F) and identified alterations in gene expression profiles during TAM phenotypic transitions (Figure 5H). Additionally, RNA velocity was applied to dissect the developmental trajectories and cellular dynamics underlying TAM evolution (Figure 5G). The DC population comprised four distinct clusters: cDC2_c1_FCER1A, cDC2_c2_CD207, cDC2_c3_AREG, and mDCs_c1_LAMP3. The conventional DC type 2 (cDC2) subsets represent cells capable of activating CD8⁺ T cells and orchestrating antitumor immune responses. High expression of costimulatory factors, MHC class II molecules, and proinflammatory cytokines was observed in these cDC2 clusters (Figure 5I).

Figure 5.

Myeloid cell diversity and functional polarization within the thyroid cancer niche. (A) UMAP visualization of 12 distinct myeloid cell subsets identified through unsupervised clustering analysis. (B) Feature plots confirming the myeloid identity of the annotated clusters. (C) Stacked bar plot showing the proportional distribution of each myeloid cell subset across samples. (D) Dot plot illustrating the scaled expression levels (color intensity) and the percentage of cells expressing (dot size) signature marker genes across all myeloid cell subsets. (E) Violin plots displaying the expression of immunomodulatory molecules (TNF, SIRPA, NECTIN2, MARCO, LGALS3, APOE) across TAM subsets. (F) Monocle 2 pseudotime trajectory analysis revealing the differentiation continuum of myeloid cells from monocytes to terminally differentiated TAM subsets. (G) RNA velocity analysis projected onto the UMAP plot, demonstrating that directional differentiation flows within the myeloid compartment. (H) Heatmap displaying differentially expressed genes across all myeloid subsets, highlighting distinct transcriptomic signatures for each population. (I) Heatmap comparing the expression of costimulatory molecules, MHC class II genes, and immunosuppressive mediators across DC subsets. (J) Gene Ontology (GO) and KEGG pathway enrichment analysis comparing the biological functions of DCs. (K) Violin plots showing the expression of immunosuppressive genes (IDO1, NECTIN2, LGALS3, LGALS9) across DC subsets. (L) CellphoneDB ligand-receptor interaction analysis displaying the communication networks between myeloid cell subsets and cancer cells in primary thyroid cancer (TC) and thyroid cancer lymph node metastasis (TLNM) microenvironments.

In stark contrast, LAMP3⁺ mature DCs (mDCs_c1_LAMP3) exhibited an immunosuppressive phenotype, with elevated expression of several negative immune regulatory molecules, including indoleamine 2,3-dioxygenase 1 (IDO1), galectin-3 (LGALS3), galectin-9 (LGALS9), and nectin cell adhesion molecule 2 (NECTIN2) (Figure 5I and K). To explore the functional characteristics and signaling pathways of each DC subset, Gene Ontology (GO) and KEGG pathway analyzes were performed (Figure 5J). The mDCs_c1_LAMP3 cluster was significantly enriched in pathways related to IFN-γ response, negative regulation of T-cell activation, and negative regulation of immune response. Conversely, the cDC2_c1_FCER1A cluster was enriched in pathways associated with T-cell activation, interleukin-2 (IL-2) production, and T helper 17 (Th17) cell differentiation, consistent with an immunostimulatory phenotype. To investigate intercellular communication patterns, we performed ligand-receptor interaction analysis using the CellphoneDB algorithm (Figure S2, S3). The analysis revealed that mDCs_c1_LAMP3 exhibited the most extensive interactions with cancer cells and other myeloid cells in both primary TC and TLNM tumor microenvironments (Figure 5L). These findings suggest that LAMP3⁺ mDCs serve as central hubs for maintaining the immunosuppressive niche and may represent potential therapeutic targets for immune modulation in thyroid cancer.

4. Discussion

Thyroid cancer represents the most prevalent endocrine malignancy worldwide, with lymph node metastasis serving as a critical determinant of disease recurrence and patient prognosis. ^{22 , 23} Although the majority of differentiated thyroid cancers exhibit favorable outcomes following conventional therapies including surgery, radioactive iodine ablation, and TSH suppression, a subset of patients develops metastatic disease refractory to standard treatments. ^{24 , 25} The emergence of immune checkpoint inhibitors has revolutionized therapeutic paradigms across multiple malignancies; however, clinical trials evaluating PD-1 blockade in thyroid cancer have demonstrated limited efficacy, suggesting that distinct immunobiological mechanisms may govern this disease. ²⁶ Metastatic tumors are known to exhibit more profound immunosuppressive microenvironments than their primary counterparts, attributed to clonal evolution of cancer cells, phenotypic shifts of infiltrating immune populations, and the influence of organ-specific niches. ^{27 , 28} With the advancement of single-cell sequencing technologies, extensive studies have characterized the cellular components and immunophenotypic features of primary thyroid cancer ecosystems. ^{18 , 29} However, the key determinants underlying the immune microenvironment of lymph node-metastatic thyroid cancer, and the mechanisms by which metastatic cells establish immunosuppressive niches, remain poorly defined. In the present study, we constructed a comprehensive single-cell transcriptomic atlas of paired primary thyroid tumors and lymph node metastases, elucidating the coevolutionary dynamics between malignant cell plasticity and immunosuppressive niche remodeling during metastatic progression.

Recent pan-cancer analyzes have demonstrated that intratumoral heterogeneity constitutes a fundamental determinant of therapeutic resistance and disease progression. ^{30 , 31} In thyroid cancer, integrated multiomics studies have identified distinct molecular subtypes characterized by differential MAPK pathway activation and thyroid differentiation scores. ^{32 , 33} Our single-cell resolution analysis extends these bulk-level observations by revealing nine transcriptionally distinct cancer cell subclusters and six functional gene expression modules. The coexistence of differentiated modules retaining canonical thyroid markers (TPO, TG, TSHR) alongside proliferative modules exhibiting MAPK hyperactivation underscores the phenotypic plasticity inherent to thyroid malignancies. Trajectory and RNA velocity analyzes further delineated directional differentiation flows from progenitor-like populations toward more differentiated states, with entropy-based stemness scoring identifying specific subclusters harboring stem-like characteristics. These progenitor populations may serve as reservoirs for tumor regeneration and therapeutic resistance, consistent with the cancer stem cell paradigm observed in other epithelial malignancies. ^{34 , 35}

These transcriptional programs and clonal dynamics establish the intrinsic capacity for metastatic dissemination; however, successful colonization of lymph nodes also requires extrinsic adaptation to the new microenvironment to facilitate invasion, survival, and colonization of secondary sites. ³⁶ Comparative pathway analysis between primary and metastatic lesions in our cohort revealed striking divergence in signaling pathway activation. While primary tumor cells exhibited enrichment for cell cycle regulation and interferon response signatures, metastatic cells demonstrated significant upregulation of PI3K-AKT-mTOR, TGF-β, and Hedgehog signaling cascades, accompanied by enhanced EMT and angiogenesis programs. The TGF-β pathway, a master regulator of EMT, has been implicated in conferring metastatic competence across multiple cancer types. ³⁷ The concurrent activation of IL-6-JAK-STAT3 signaling in metastatic subclusters suggests an inflammatory-metastatic axis that may represent a targetable vulnerability.

Beyond malignant cell-intrinsic alterations, the composition and functional state of tumor-infiltrating immune cells critically influence disease progression and therapeutic response. ³⁸ Our profiling unveiled profound remodeling of the myeloid compartment during metastatic progression, most notably the emergence of LAMP3⁺ tolerogenic dendritic cells expressing elevated levels of immunosuppressive mediators including IDO1, LGALS3, LGALS9, and NECTIN2. Pathway enrichment analysis confirmed significant association with negative regulation of T-cell activation and immune response. Ligand-receptor interaction analysis positioned LAMP3⁺ DCs as central communication hubs within both primary and metastatic microenvironments, suggesting their critical role in establishing and maintaining immune tolerance. ³⁹ Furthermore, the organ-specific adaptations observed in thyroid cancer lymph node metastases, including the emergence of LAMP3⁺ tolerogenic DCs and the upregulation of LAG3/TIGIT checkpoint axes, complement recent single-cell characterizations of breast cancer brain metastases, which identified ILF2 as a potential therapeutic target and revealed distinct cellular states and therapeutic vulnerabilities shaped by the central nervous system microenvironment. ⁴⁰ Consistent with this notion, LAMP3⁺ DCs have been increasingly recognized as immunoregulatory cells across multiple cancer types. In cervical cancer, LAMP3⁺ DCs expressing IDO1 interact with exhausted CD8⁺ T cells and regulatory T cells to form an immunosuppressive cycle. ⁴¹ In osteosarcoma, LAMP3⁺ DCs have similarly been identified as key components of the immunosuppressive microenvironment. ⁴² Importantly, a recent single-cell study in papillary thyroid carcinoma demonstrated that LAMP3⁺ DCs promote CD8⁺ T-cell exhaustion via NECTIN2-TIGIT interactions and increase regulatory T-cell infiltration through CCL17-CCR4 signaling. ⁴³ These convergent observations across tumor types support the hypothesis that LAMP3⁺ DCs may serve as central immunosuppressive hubs within the metastatic niche, although direct functional evidence in thyroid cancer remains to be established. The lymphocyte compartment exhibited parallel dysfunction, with tissue-resident memory and proliferating effector CD8⁺ T cells demonstrating elevated exhaustion signatures. Notably, FOXP3⁺ regulatory T cells in metastatic lesions harbored significantly higher exhaustion scores than their primary tumor counterparts, indicating a more profoundly suppressive niche in the metastatic compartment. ⁴⁴

Regarding immune checkpoint expression, our analysis revealed minimal expression of the canonical PD-1/PD-L1 axis across both lymphoid and tumor/stromal populations, while LAG3, TIGIT, and CD96 emerged as the dominant inhibitory receptors on CD8⁺ T cells, with cognate ligands abundantly expressed on cancer and myeloid cells. ⁴⁵ Additionally, CD47 demonstrated robust expression across malignant subclusters. ⁴⁶ The functional relevance of the LGALS3-LAG3 axis has been experimentally validated in several cancer types. In a mechanistic study, galectin-3 (encoded by LGALS3) was shown to bind LAG-3 on activated CD8⁺ T cells specifically within the tumor microenvironment, and LAG-3 expression was necessary for galectin-3-mediated suppression of CD8⁺ T cells in vitro. ⁴⁷ In ovarian cancer, mesenchymal cancer cells were shown to promote CD8⁺ T-cell exhaustion through the LGALS3-LAG3 axis. ⁴⁸ These findings provide a molecular rationale for exploring alternative checkpoint inhibitors in metastatic thyroid cancer. Taken together, these findings delineate a coherent pathogenic sequence: chromosomal instability generates heterogeneous malignant populations; upon reaching the lymph node, these cells encounter and shape an immunosuppressive niche dominated by LAMP3⁺ DCs and exhausted T cells; and within this niche, the LGALS3-LAG3 and TIGIT-NECTIN2 axes emerge as the principal mediators of immune evasion. This framework explains the limited efficacy of PD-1 blockade in thyroid cancer and directs future therapeutic efforts toward alternative targets specific to the metastatic ecosystem.

Several limitations warrant consideration. First, the modest sample size (n = 3 patients) limits the statistical power and generalizability of our findings; although the paired design enhances internal validity, the identified cellular states and signaling axes require validation in larger, independent cohorts. Additionally, spatial transcriptomic approaches would provide valuable insights into the architectural organization of immune-tumor interactions. Functional validation of identified therapeutic targets in preclinical models remains essential. Based on the expression landscape and cell‒cell communication networks delineated in this study, we propose that the LAG3-LGALS3 and TIGIT-NECTIN2 axes represent the most compelling high-priority candidates for such validation. Additionally, the broad expression of CD47 across malignant subclusters warrants further exploration as a potential target for macrophage-mediated clearance strategies. In conclusion, this study presents a comprehensive single-cell atlas of paired primary and lymph node-metastatic thyroid cancer, revealing the coordinated evolution of malignant cell plasticity and immunosuppressive microenvironment remodeling. The identification of LAMP3⁺ tolerogenic DCs as potential central orchestrators and LAG3/TIGIT as prominent checkpoint axes offers a foundation for future investigation into tailored immunotherapeutic strategies for metastatic thyroid cancer. Nevertheless, it should be emphasized that the proposed immunosuppressive functions of LAMP3⁺ DCs and the LAG3/TIGIT axes are inferred from transcriptomic signatures and computational ligand‑receptor analyzes, rather than direct functional evidence. These findings are therefore presented as hypothesis‑generating and require experimental validation in future studies.

Supplementary Material

Supplementary Material

Supplementary_Figure.docx

Disclosure of potential conflicts of interest

No potential conflicts of interest were disclosed.

Funding

No funding was received for this work.

Data availability statement

Data and materials associated with this study will be available upon reasonable request from the corresponding author.

Ethical approval

This study was approved by the Medical Ethics Committee of Sun Yat-sen University Cancer Center (G2022-066-01). All patients provided written informed consent prior to enrollment. This study was conducted in accordance with and the principles of the Declaration of Helsinki.

Supplementary material

Supplemental data for this article can be accessed at https://doi.org/10.1080/2162402X.2026.2701504.