2.1 Transcriptomic and spatial transcriptomic profiling of thyroid cancer

Seven patients underwent thyroidectomy for thyroid tumours with a pathologically confirmed diagnosis of PTC. Following surgery, 12 specimens of primary tumour resections (six regional lymph nodes and six thyroid tumours) were obtained and dissociated for single-cell sequencing. Of these 12 specimens, six (three regional lymph nodes and three thyroid tumours) were sectioned and subjected to spatial transcriptomics (Figure 1A). We detected metastases in the lymph nodes surrounding the thyroid gland, predominantly in the central region, followed by cervical lymph node metastases (LNMs). The number and location of LNMs were significantly correlated with the patient prognosis and the surgical employed for thyroid cancer. Details of the patients and their diagnoses are presented in Table 1.

We combined scRNA-seq data from all samples, treating each sample as a separate batch, using a batch-balanced k-nearest neighbours algorithm. After stringent quality control, 144 746 cells were retained, with a median gene detection count of 1414 per cell. We visualized the cellular landscape using uniform manifold approximation and projection (UMAP), highlighting tissue, sample and patient origins (Figure 1B and Figure S1A,B). Notably, we observed substantial cellular diversity at both the sample and patient levels. To distinguish between tumour and normal cells, we assessed cellular ploidy using CopyKat and InferCNV algorithms on a randomly downsampled subset of the data (Figure 1C and Figure S2). We annotated cell types based on marker gene expression (CD3D and CD3E for T cells, CD79A for B cells, LYZ and CD14 for myeloid cells, COL1A1 and COL1A2 for fibroblast cells, PECAM1 and CD34 for endothelial cells, and KRT18 and KRT19 for thyrocytes) (Figure 1D,E), as described in a previous study.¹³ To isolate the heterogeneity specific to PTC, we generated a separate UMAP plot excluding the medullary thyroid carcinoma (MTC) case, as its tumour origin differs (Figure S3). Interestingly, fibroblasts, thyrocytes and endothelial cells exhibited aneuploidy and a novel cellular state expressing both thyrocyte and fibroblast markers, termed ‘thy_fib.’. At the patient level, heterogeneity was primarily observed in tumour cells, as anticipated from previous studies. To further explore this, we plotted the number and proportions of different cell types (Figure S4) and generated individual UMAP plots (Figure S5).

We performed transcriptomic profiling of the cohort to annotate the collected spatial transcriptomic data. A representative of a histological slide prepared for spatial transcriptomic analysis is depicted in Figure 1F. Employing SpaGCN,¹⁴ we identified distinct domains (Figure 1G). Notably, we successfully delineated tumour regions exhibiting differential EPCAM expression, aligning with the expert pathologist's assessment (Figure 1H and Figure S6).

2.2 Lower APOE gene expression is associated with tumour cell metastasis

Consistent with these findings, lower APOE expression was significantly associated with poorer overall survival in patients with PTC from The Cancer Genome Atlas (TCGA) database (n = 256, Figure 2C). To further investigate the role of APOE in tumour progression, we performed sub-clustering of tumour cells (Figure S8A) and pseudo-time trajectory analysis using Monocle 3 (Figure 2D,E and Figure S8B). This analysis identified nine distinct cellular states based on their position along the trajectory (Figure S8C). When examining APOE expression across these 13 tumour subclusters (Figure S8D), we observed a striking correlation between lower APOE levels and later-stage tumour cell states. This finding suggests that late-stage tumour cells, characterized by lower APOE expression, may contribute more significantly to metastatic progression.

We partitioned tumour cells into two groups based on APOE expression, acknowledging that the zero-inflation inherent to single-cell data might introduce some degree of error. Our goal was to identify major trends. We identified differentially expressed genes (DEGs) between APOE⁺ and APOE⁻ tumour cells in both thyroid and lymph node samples (Figure S9). Subsequently, we performed pathway enrichment analysis on these gene sets (Figure 2F). In thyroid APOE⁻ cells, enriched pathways were associated with enhanced metabolic processes, such as the respiratory electron transport chain and angiogenesis regulation. In contrast, lymph node APOE⁻ cells exhibited enrichment in pathways related to protein synthesis, including peptide biosynthesis and protein targeting to the endoplasmic reticulum.

We examined APOE expression in histological slides of lymph node tissues containing metastatic tumours. Employing a novel algorithm capable of super-resolution analysis of tumour ecosystems,¹⁵we first identified tumour cells based on the expression of EPCAM, TG, KRT18 and KRT19. This automated segmentation closely aligned with the tumour regions delineated by expert pathologists (Figure 2G,H,K,L). Similarly, the main tumour edge was defined using these marker sets (Figure 2I,M). Intriguingly, we observed a higher level of APOE expression within the core of the tumour compared to the tumour edge (Figure 2J,N). This finding suggests that APOE-expressing tumour cells may play a crucial role in facilitating lymph node metastasis.

2.3 Overexpression of APOE inhibits tumour cell proliferation and invasion in vitro

We further validated the role of APOE gene expression in PTC using scRNA-seq data. We established APOE overexpression Hth-7 and TPC-1 cell lines via lentiviral transfection (Figure 3A). To assess the impact of APOE overexpression on cellular behaviour, we performed CCK-8 assays and colony formation assays. These experiments revealed a significant decrease in the proliferation of APOE-overexpressing cells (Figure 3B,C). Additionally, Matrigel invasion chamber assays suggested a substantial reduction in the invasive ability of these cells (Figure 3D). These results collectively indicate a potential role for APOE gene expression in inhibiting PTC progression. To explore the underlying mechanisms, we conducted transcriptome sequencing analysis of APOE-overexpressing Hth-7 and TPC-1 cell lines. This analysis identified 44 overlapping DEGs with an FDR-adjusted p-value < .05, |log2foldchange| > 1 and raw counts > 10 (Figure 3E). Gene set enrichment analysis revealed that genes associated with the ‘Hallmark_TNFa_Signaling_via_NFKB’ pathway were positively correlated with APOE overexpression in both cell lines (Figure 3F). This pathway is closely linked to immune responses and tumour progression. Notably, ABCA1, a gene previously reported as critical for APOE lipidation,¹⁶ was significantly upregulated in both APOE-overexpressing cell lines (Figure 3G). ABCA1, in conjunction with APOE, may contribute to the activation of the LXR transcription factor, thereby potentially restricting immunosuppression.¹⁷ These findings suggest a complex interplay between APOE, ABCA1 and LXR that may influence tumour immunity and progression.

Furthermore, KEGG pathway enrichment analysis revealed significant enrichment of metabolism-related and immune-related pathways in APOE-overexpressing cells (Figure S10A). To gain deeper insights into potential metabolic changes, we performed an untargeted Liquid Chromatograph Mass Spectrometer (LC-MS) metabolomics analysis. Metabolites were analysed for fold change between the groups, and a t-test with criteria of VIP ≥ 1 and p < .05 was conducted to identify significant differences (Figure S10B). A one-way ANOVA was utilized for statistical comparisons, followed by post-hoc pairwise comparisons (Figure S10C). The volcano plot visualized the differential metabolites between the two groups (Figure S10D,E). Additionally, KEGG pathway analysis revealed significant enrichment of metabolites (Figure S10F,G). Collectively, these data suggest that APOE may influence the tumour microenvironment and progression through multiple pathways.

2.4 In vivo validation of APOE function in tumour inhibition and large-scale human validation

Next, we sought to validate these findings in vivo. We utilized APOE-overexpressing cells for subcutaneous xenograft experiments in athymic, female nude mice. Notably, APOE overexpression significantly inhibited the tumorigenicity of PTC cells in vivo. The tumour growth of APOE-expressing cells was substantially slower compared to the control group (Figure 4A,B), while these genetic modifications did not affect mouse body weight (Figure 4C). We further performed immunohistochemical analysis to assess APOE and ABCA1 expression in tumour tissues. Our results demonstrated that APOE overexpression was accompanied by increased ABCA1 expression in tumour tissues (Figure 4D,E).

To further validate the clinical relevance of APOE expression in metastatic PTC, we investigated APOE expression in a cohort of 41 patients with metastatic disease. Immunofluorescence staining revealed a significant decrease in co-expression of APOE and CK19 in patients with lateral cervical LNM compared to those with central compartment LNMs (Figure 4F,G). This suggests that patients with lateral cervical LNM exhibit lower APOE expression levels in tumour cells relative to normal cells. Considering the link between LXR and APOE, we administered intraperitoneal injections of the LXR-agonist RGX-104 to mice. Treatment with RGX-104 resulted in a suppression of myeloid-derived suppressor cells (MDSCs) and a concomitant increase in APOE expression (Figure 4H−K). Collectively, these findings suggest a potential role for LXR in regulating APOE expression. Moreover, the observed inhibition of tumour growth in vitro and in vivo by APOE overexpression supports the hypothesis that APOE may play a tumour-suppressive role.

2.5 Ligand-receptor pairs driving differential immune−tumour interactions associated with APOE expression profiles

We sought to leverage the comprehensive profiling of immune cells in our cohort by performing a sub-clustering of these cells (Figure 5A). We used established markers to annotate immune cells into 10 distinct types (Figure 5B): naïve CD4⁺ T cells, memory CD4⁺ T cells, CD8⁺ T cells, natural killer (NK) cells, B cells, plasma cells, plasmacytoid dendritic cells, dendritic cells (DCs), CD14⁺ monocytes and CD16⁺ monocytes. A comparative analysis of samples with different metastatic conditions, specifically central, central plus neck (CN) and healthy (none), revealed dramatic shifts in the abundance of certain immune cell types. Notably, the CN group exhibited a clear enrichment of both CD14⁺ and CD16⁺ monocytes, coupled with a depletion of CD8⁺ T cells (Figure S11).

We further investigated the interaction between APOE⁺/APOE⁻ tumour cells with immune cells in lymph nodes. We found that APOE⁻ cells stimulated the immune system to a lesser degree than APOE⁺ cells, regardless of the normalization method employed (Figure 5C). We identified several ligand-receptor pairs that underlie this pattern (Figure 5D), including the CD6- and F11R-driven signalling pathway networks. Specifically, APOE⁺ tumour cells were involved in the CD6-driven network, along with immune cells (Figure 5E). For F11R (JAM1), both APOE⁺ and APOE⁻ cells interacted with immune cells (Figure 5F), with APOE⁺ cells exhibiting stronger interactions. Interestingly, in contrast to APOE⁺ cells, APOE⁻ cells specifically interacted with immune cells, such as DCs, through a CD99-driven network rather than the CD6-driven network. To further explore the role of the CD99-driven network in the interaction between APOE-positive and APOE-negative tumour cells and immune cells, we focused on these cell types at both the primary thyroid site and lymph node metastasis site (Figure S12). Notably, in the APOE-negative group, the ESAM and CD99 pathways exhibited strong interactions at both sites. Furthermore, the CD46 pathway, which is linked to immunosuppression,^{18, 19} showed stronger interactions between APOE-negative tumour cells and immune cells, particularly at the lymph node metastasis site. Using CellTrek,²⁰ we directly visualized the communication networks between APOE⁻ and APOE⁺ tumour cells and immune cells on histological slides. Strikingly, APOE⁻ tumour cells exhibited minimal interaction with NK, B and T cells in both primary tumour tissues (Figure 5G,H) and lymph nodes (Figure 5I,J), while APOE⁺ tumour cells strongly interacted with these cell types.

To further confirm the association of APOE with a suppressive immune microenvironment, we generated APOE knockout mice using a CRISPR-Cas9 system and verified the knockout by PCR (Figure S13A,B). We observed a decrease in mature DCs, NK cells and NKT cells in APOE⁻/⁻ mice (Figure S13C,D). Additionally, the M1/M2 macrophage ratio was also reduced (Figure S13E). However, the numbers of CD4⁺ and CD8⁺ T cells did not differ significantly (Figure S13F). These spatial data suggest that APOE-negative tumour cells may promote tumour growth and invasion by creating an immunosuppressive microenvironment.

2.6 Machine learning-based model predicts metastasis patterns

Given the significant morbidity and mortality associated with metastatic thyroid cancer, early prediction of metastatic potential is crucial for improved patient outcomes. We leveraged our comprehensive single-cell atlas to develop a predictive model capable of effectively classifying patients into high- and low-risk groups for metastasis. We used only tumour cells to build a predictive model to distinguish between the two groups (Figure 6A). These cells did not clearly differ in their patterns of metastasis in the latent space constructed by principal component analysis (PCA). While incorporating all genes into the model was theoretically possible, we aimed to simplify the model and facilitate the development of practical diagnostic assays by limiting the number of genes. To achieve this, we ranked genes based on their mutual information scores. We constructed random forest classifiers using varying numbers of top-ranked genes and evaluated their performance. The model's performance plateaued after incorporating the top 13 genes with the highest mutual information scores (Figure 6B). This 13-gene model demonstrated excellent predictive power, with an area under the receiver operating characteristic curve (AUROC) of .929 (Figure 6C). We visualized the expression patterns of these 13 genes in tumour cells from the high- and low-risk metastasis groups (Figure 6D), revealing distinct expression profiles between the two groups. Notably, nine of these 13 genes were upregulated in the APOE-negative group, suggesting that their association with metastasis is not directly linked to APOE metabolism. To validate the model's clinical utility, we collected an additional 203 clinical samples obtained through surgery or biopsy and performed qPCR for the 13-gene panel. Recognizing the potential differences between qPCR and scRNA-seq data, we retrained the model separately for surgical and biopsy samples, using 70% of the data for training and the remaining 30% for validation. Both models exhibited excellent performance, with high scores for precision, recall and AUROC (Figure 6E,F). Detailed patient information and diagnostic outcomes are presented in Table 3.

In addition to the presence or absence of metastasis, the pattern of metastasis (central or CN) is also clinically significant. To address this, we employed a similar machine-learning approach to distinguish between these two conditions. Consistently, only tumour cells were utilized to build the model (Figure S13A). We initially developed a model based on single-cell data, incorporating 10 genes as features (Figure S13B−D). To further validate and refine the model, we collected additional clinical samples through biopsy. Unfortunately, we did not obtain a sufficient number of surgical samples for this purpose. Using qPCR data from the biopsy samples, we retrained the model and achieved satisfactory performance (Figure S13E). We are currently developing formal, rapid-diagnostic approaches based on these trained models.

3.1 Human participants

A total of eight thyroid cancer patients were enrolled in this study. Fourteen samples were selected for single-cell transcriptome sequencing, and eight samples were included for spatial sequencing. Among these patients, seven had papillary thyroid carcinoma, and one had medullary thyroid carcinoma. Detailed clinical information for these patients is provided in Table 1. Tissue sections from 41 PTC surgical specimens were selected for immunofluorescence staining. Clinical information for these patients is available in Table 2. A total of 203 thyroid cancer tissue DNA samples were obtained from punch biopsies of patients undergoing thyroidectomy. Clinical information for these patients is described in Table 3. All human tissue samples used in this study were obtained with approval from the Human Ethics Committee of Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China (2023LLSDN11).

3.2 Haematoxylin and eosin staining

Haematoxylin and eosin (H&E) staining was performed using standard protocols. Following deparaffinization and rehydration, sections were stained with haematoxylin solution for 5 min, followed by five consecutive washes with 1% acid ethanol (1% HCl in 70% ethanol). After rinsing with water, sections were stained with eosin for 3 min. Finally, the slides were dehydrated with a graded ethanol series and mounted with a coverslip. The stained slides were examined under a light microscope (Leica DM3000; Leica Microsystems).

3.3 Spatial transcriptome and data analysis

Spatial transcriptomic slides were printed with capture areas from four patients with lymph node metastatic thyroid cancer, the same patients from whom scRNA-seq data was obtained. Gene expression information for these spatial transcriptomic slides was captured using the Visium Spatial Gene Expression platform (10x Genomics), employing spatially barcoded mRNA-binding oligonucleotides, as per the manufacturer's instructions. Raw sequencing reads from the spatial transcriptomes were subjected to quality control and mapping. The demultiplexed clean reads were aligned against the UCSC human GRCh38 reference genome using Space Ranger (v. 1.3; 10x Genomics). After obtaining the single-cell gene expression count matrix, we utilized Seurat v. 4.0 for downstream analysis on the R platform (v. 4.1.2; The R Foundation for Statistical Computing). In addition, we employed SpaGCN, a recently developed graph convolutional network, to differentiate spatial patterns and domains based on DEGs.¹⁴ In the SpaGCN analysis, the default mode was used to autonomously define domains within the spatial data. Super-resolution tumour-edge maps were generated using TESLA.¹⁵ Finally, integrated single-cell and spatial data analyses were performed using CellTrek, a computational framework that facilitates the direct mapping of single cells to their spatial coordinates in tissue sections by integrating scRNA-seq and spatial transcriptomic data. This integration enabled the generation of a cell−cell communication map.²⁰

3.4 ScRNA-seq analysis

Single-cell sequencing data were aligned and barcode-demultiplexed using the Cell Ranger v. 3.0.2 vdj pipeline (10x Genomics). The data were filtered based on quality control criteria: cells with fewer than 800 detected genes, genes detected in fewer than 5 cells per dataset and cells with mitochondrial gene expression exceeding 10% of the total expression level were excluded. Potential doublets were removed by excluding cells with the top 5% of total transcript unique molecular identifiers (UMIs). Scrublet was also employed to further identify and remove doublets. All datasets were combined, and the top 2000 highly variable genes were selected. To account for technical variations, the total UMI counts and mitochondrial gene expression proportions were regressed out for each cell. The data was scaled, and PCA was performed to reduce dimensionality. Batch correction was applied using the BBKNN method with ‘donor’ as the batch key. These preprocessing steps were conducted using the Scanpy framework.

Seurat was employed for downstream analysis. Initial clustering was performed using the FindClusters function with a resolution of .1 to identify major cell types. For subsequent sub-clustering of immune cells, the resolution was set to .3. Known marker genes were utilized to annotate the clusters. CopyKat was employed to infer cell ploidy on a randomly down-sampled dataset.²¹ To manage computational demands, cells were down-sampled to 2000 cells per cluster before running this analysis. Default parameters were used for CopyKat (ngene.chr = 5, win.size = 25, KS.cut = .1). To validate CopyKatt results, we also applied the InferCNVsoftware. Using InferCNV version 1.12.0, we down-sampled the data to select 7000 cells, ensuring equal representation of 1000 cells from each primary cell type. This process resulted in the retention of 26 428 features for subsequent analysis with InferCNV. Differential gene expression was determined using the FindMarkers or FindAllMarkers function with a fold-change threshold of .25 and gene detection rate threshold (min.pct) set to .25. Only genes with an adjusted p-value less than .05 were considered differentially expressed. Gene set enrichment analysis was conducted using databases incorporated in the enrich R package. Pseudo-time analysis was performed using Monocle with default settings.²² The log fold change threshold was set to .5. Cell−cell interaction analysis was performed with the R package CellChat.²³

3.5 Immunofluorescence

Standard immunostaining techniques were employed. Briefly, sections were permeabilized with .5% Triton X-100 in phosphate-buffered saline (PBS) for 10 min at room temperature and blocked with serum-free protein-blocking solution (Dako; Agilent Technologies) for 60 min. For colocalization staining, slides were sequentially incubated with primary antibodies. After incubation with the secondary antibody, sections were stained with 4',6-diamidino-2-phenylindole (DAPI) for nuclear visualization. The sections were imaged using a confocal laser-scanning fluorescence microscope (ZEISS). ImageJ software was used for quantitative analysis. The following antibodies were utilized: anti-CK19 (MAB-0829, MXB), anti-APOE (Abcam Cat#ab51015) and anti-Gr-1 (Abcam Cat#ab238132).

3.6 Cell culture and transfection

Hth-7 and TPC-1 cell lines were purchased from Procell Company (BCRJ Cat# 0397, RRID:CVCL_6298). Cells were cultured in DMEM supplemented with 10% foetal bovine serum (FBS) and 1% penicillin-streptomycin at 37°C in a 5% CO₂ atmosphere. APOE overexpression control and lentiviral vectors were purchased from GenePharma. TPC-1 and Hth-7 cells were seeded in six-well plates 24 h prior to transfection with viruses and transfection reagent (RNAi-Mate, GenePharma). APOE-overexpressing cell lines were selected and cultured in a medium containing 5 µg/mL puromycin.

3.7 RNA isolation and qPCR

Total cellular RNA was extracted using an RNA extraction kit (Takara Bio). The extracted RNA was reverse transcribed into cDNA using the Vazyme reverse transcription kit. The resulting cDNA was subjected to quantitative-PCR (qPCR) using the SYBR Green kit (Vazyme). qPCR reactions were performed on a CFX96 Touch Real-Time PCR Detection System (Bio-Rad Laboratories, RRID:SCR_008426). The primers used for qPCR were as follows: βactin-forward: 5′-CATGTACGTTGCTATCCAGGC-3′; βactin-reverse: 5′-CTCCTTAATGTCACGCACGAT-3′. APOE-forward: 5′- GTTGCTGGTCACATTCCTGG-3′; APOE-reverse: 5′-GCAGGTAATCCCAAAAGCGAC-3′. The relative expression levels were calculated using the 2^‒ΔΔCt method.

3.8 Cellular proliferative ability

The TPC-1 and Hth-7 cells were seeded in 96-well plates at a density of 2000 cells per well and incubated for 24 h. A CCK-8 assay (Vazyme) was performed according to the manufacturer's instructions at 37°C in 5% CO₂. The CCK-8 reagent was added to each well and incubated for 2 h at 37°C. Optical density was measured at 450 nm using a SpectraMax 190 microplate reader (Molecular Devices).

3.9 Colony formation assays

We seeded 200 TPC-1 and Hth-7 cells per well into 6-well plates and replaced the medium every 3–4 days. After a 2-week incubation period, cells were fixed with 4% paraformaldehyde, washed three times with PBS and stained with .05% crystal violet for 30 min. The plates were then washed, and colonies containing more than 50 cells were photographed and counted.

3.10 Transwell migration and invasion assays

Transwell chambers (8-µm pores; Corning) pre-coated with Matrigel were employed for the invasion assay. Cells (2 × 10⁵ cells/well) were cultured in the upper chamber in 200 µL of DMEM supplemented with 5% FBS. In the lower compartment, 700 µL of DMEM containing 20% FBS was added. After a 24-h incubation period, cells were fixed with 4% paraformaldehyde, washed thrice with PBS and stained with .05% crystal violet for 30 min. Cells on the inner surface of the chamber were removed using sterilized cotton swabs. The cells were subsequently examined under a light microscope (Leica DM3000; Leica Microsystems). ImageJ software (ImageJ, RRID:SCR_003070) was utilized for quantification.

3.11 Xenotransplantation model

Female BALB/c nude mice, aged 3–4 weeks and weighing 13–15 g, were obtained from Shanghai Phenotek Laboratory Animal Co., Ltd. Tumour models were established by subcutaneously injecting1 × 10⁷TPC-1 and Hth-7 cells, suspended in Matrigel/PBS, into the left hind limbs of nude mice. Tumour size and mouse weight were monitored every 3–4 days. Tumour volume was calculated using the formula: .5 × length × width × width. After 2 weeks, mice were euthanized, and tumours were harvested, photographed and fixed with 4% paraformaldehyde. All animal studies were approved by the Animal Use Committee of Shanghai Jiao Tong University (SYXK 2018-0027).

3.12 Immunohistochemistry

After dewaxing, samples were microwaved in ethylene diamine tetra acetic acid buffer and blocked for 60 min with a serum-free protein-blocking Solution (Agilent Technologies). Primary antibodies against APOE and ABCA1 (Abcam; 1:500 dilution) were applied to the sections and incubated overnight at 4°C. Subsequently, the tissues were incubated with secondary antibodies for 1 h at room temperature. The sections were then imaged using a confocal laser-scanning fluorescence microscope (ZEISS). ImageJ software (RRID:SCR_003070) was employed for quantification.