Proteogenomic characterization delineates clinically relevant subtypes of advanced differentiated thyroid cancer

Summary

Advanced differentiated thyroid cancer (DTC) is characterized by limited therapeutic options and unfavorable prognosis. To address this, we conduct proteogenomic analysis of 113 advanced DTCs, identifying three molecularly distinct subtypes: canonical, stromal, and immunogenic. These subtypes exhibit differences in driver mutations, histopathological features, and clinical outcomes. Based on their unique biology, we suggest distinct therapeutic strategies for each subtype. To facilitate clinical application, we develop a machine learning classifier that accurately predicts these subtypes using routinely available gene mutation and digital pathology data. The biological relevance of this classification is further confirmed in an independent cohort analyzed by single-cell and spatial transcriptomics. Moreover, analysis of a real-world cohort of patients receiving various systemic therapies provides preliminary clinical evidence supporting the potential utility of this subtyping framework for informing treatment decisions. Collectively, this study provides a rationale and a practical tool for future exploration of personalized treatment in advanced DTC.

Graphical abstract

Highlights

• •Proteomic and spatiotemporal profiling of advanced DTC
• •Three proteomic subtypes of advanced DTC show distinct molecular and clinical features
• •Establishment of an accurate model to predict proteomic subtypes of advanced DTC
• •Real-world cohort explores clinical relevance of molecular subtypes in advanced DTC

Zhang et al. systematically reveal the heterogeneity of advanced differentiated thyroid cancer (DTC) from a biological perspective and develop a genomics-pathomics model that provides a preliminary basis for a therapeutic framework to advance precision oncology in this disease.

Introduction

Thyroid cancer has a rapidly increasing global incidence, with approximately 600,000 new cases diagnosed annually worldwide. Of these, over 500,000 (representing more than 95% of cases) are differentiated thyroid cancer (DTC), which includes papillary thyroid carcinoma (PTC), follicular thyroid cancer (FTC), and their pathologic variants.^1,2 Despite the overall favorable prognosis of DTC, a proportion of patients may unfortunately progress to advanced stage, presenting with locally advanced or metastatic disease. These patients have limited treatment options and a markedly poorer prognosis compared to those with early-stage disease, imposing a significant burden on society, families, and individuals.

Complete removal of the thyroid gland followed by radioactive iodine (RAI) therapy represents the conventional treatment paradigm for advanced DTC. If the patient is refractory to RAI therapy or ineligible for thyroid resection, targeted therapy may be an alternative option.³ Over the last decade, anti-angiogenic multi-targeted tyrosine kinase inhibitors (mTKIs) have revolutionized the treatment of advanced RAI-refractory (RAIR) DTC, with representative agents—sorafenib, lenvatinib, cabozantinib, donafenib, and anlotinib —approved in the United States or China.^4,5,6,7,8 In recent years, some studies have also attempted to apply combination targeted therapy with immunotherapy in advanced DTC.^9,10,11 Although these therapies have brought clinical benefits to many patients, most patients fail to achieve durable efficacy and some do not respond at all. This highlights an unmet need to better understand the molecular heterogeneity underlying treatment vulnerabilities, which is essential for tailoring optimal therapies or rational combinations.

In this study, we performed a proteogenomic characterization of tumors from 113 patients with advanced DTC (Discovery Cohort). Based on these data, we classified advanced DTC into three consensus clustering (CC) subtypes (CC1: canonical; CC2: stromal; CC3: immunogenic) and proposed tailored treatment strategies corresponding to their distinct biological features. Subsequently, we established a multimodal deep learning framework that integrates gene mutations and pathological features using convolutional neural networks (CNNs) and random forest. This genomics-pathomics (G-P) framework facilitates the integrated analysis of whole-slide pathology images and genomic data to infer the three molecular subtypes with high discriminatory power.

Leveraging the G-P framework, we independently validated the molecular profiles of our proposed subtypes using single-cell or spatial transcriptomics data (a single-cell RNA [scRNA]/spatial RNA sequencing [spRNA-seq] cohort, External Cohort 1). More importantly, we established a large-scale real-world cohort of advanced DTC receiving systemic therapy (External Cohort 2), which included patients treated with four approved drugs or enrolled in seven clinical trials. This cohort provided preliminary, hypothesis-generating evidence supporting the potential relevance of our subtype-specific treatment strategies (Figure 1). These findings lay the foundation for future prospective validation of precision treatment guided by molecular subtyping in advanced DTC.

Figure 1.

Schematic overview of the study

We performed proteogenomic profiling of a large cohort of patients with advanced DTC (n = 113, Discovery Cohort) and classified them into three CC subtypes (CC1: canonical; CC2: stromal; CC3: immunogenic; based on their respective molecular features). Building upon these characteristics, we proposed tailored subtype-specific treatment paradigms (CC1: RAI monotherapy; CC2: RAI + mTKI targeted therapy; CC3: immunotherapy + mTKI targeted therapy. To facilitate clinical use, we developed a machine learning-based genomics-pathomics framework that predicts subtypes using gene mutation and digital pathology data with high discriminative ability. With the help of this framework, we independently validated the three subtypes’ biological characteristics using scRNA/spRNA-seq data (External Cohort 1). Most importantly, we established a large-scale real-world systemic therapy cohort (External Cohort 2) incorporating data from four approved therapies and seven clinical trials, which effectively validated our subtype-specific treatment paradigm.

Results

Proteomics-based molecular subtypes of advanced DTC

We collected surgically resected tumors from 113 patients with advanced DTC, all of whom underwent RAI therapy after surgery (Discovery Cohort). For all samples, global proteomics was conducted to characterize molecular heterogeneity at the protein level, while next-generation sequencing (NGS) was performed using a targeted panel to identify driver gene mutations and fusions (STAR Methods).¹² The key clinicopathologic characteristics of patients and data flow chart are summarized in Figure 1 and Table S1.

To establish a molecular classification of advanced DTC, we performed CC on the proteomic data from 113 patients to generate cluster assignments (STAR Methods). We then determined the optimal number of clusters. Based on our analyses for k = 2 to 5, the area under the cumulative distribution function curve increased remarkably at k = 3 and k = 4 (Figures S1A and S1B), while the clusters for k = 3 exhibited clearer boundaries than those for k = 4 (Figures S1C–S1F). In addition, k = 3 yielded a lower proportion of ambiguously clustered pairs value compared to k = 4, and an elbow point was observed at k = 3 in the Relative Cluster Stability Index 1 plot from Monte Carlo Reference-based Consensus Clustering analysis (Figures S1G and S1H), supporting the selection of k = 3 as the optimal number of clusters. This resulted in three CC subtypes (CC1, CC2, and CC3), comprising 43 (38.1%), 40 (35.4%), and 30 (26.5%) patients, respectively (Figure 2A).

Figure 2.

Proteomic subtyping of advanced DTC and their genomic and clinical correlations

(A) Heatmap showing CC of global proteome data revealing three CC subtypes of advanced DTC in the Discovery Cohort: CC1 (n = 43), CC2 (n = 40), and CC3 (n = 30).

(B) Bar plot comparing the proportion of RAS mutation across CC subtypes in the Discovery Cohort (n = 113). ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

(C) Bar plot comparing the proportion of BRAF-TERT promoter co-mutation across CC subtypes in the Discovery Cohort (n = 113). ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

(D) Forest plot summarizing the hazard ratios and 95% confidence intervals (CIs) in multivariate Cox proportional hazards model for post-RAI PFS of different gene alteration statuses in the Discovery Cohort (n = 113), adjusted for pathological subtype.

(E) Bar plot comparing biochemical responses to RAI across three CC subtypes in the Discovery Cohort (n = 113). Biochemical response was evaluated by comparing post-RAI suppressed thyroglobulin (Sup-Tg) levels vs. baseline Sup-Tg prior to RAI therapy, defined as G1: a decline in Sup-Tg (≥50%); G2: a decline (<50%) or an increase (<10%); G3: an increase (≥10%). ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

(F) Bar plot comparing structural responses to RAI across CC subtypes in the Discovery Cohort (n = 113). Structural response was evaluated by two independent senior physicians according to the RECIST v.1.1 criteria. PR: partial remission; CR: complete remission; SD: stable disease; PD: progression disease. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

(G) Kaplan-Meier survival curves for PFS after RAI treatment across CC subtypes in the Discovery Cohort (n = 113). Also see Figures S1 and S2, Tables S1 and S2.

Distinct clinical features and outcomes across the CC subtypes

We then analyzed the histopathologic and genomic profiles across the CC subtypes. Despite the small number of patients with RAS mutations (n = 12), there was a numerical trend toward a higher prevalence of these mutations in the CC1 subtype than in the CC2 subtype (20.9% vs. 0%, p = 0.007). This finding was concordant with a parallel increase in two well-recognized RAS-driven histotypes—the follicular variant (FV) of PTC and FTC—in the CC1 subtype (CC1 vs. CC2, 41.9% vs. 15.0%, p = 0.021)¹³ (Figures 2B and S2A; Tables S1 and S2).

A progressive increase in the frequencies of both BRAF^V600E and TERT promoter mutations was observed from CC1 to CC3, pronounced for their co-occurrence (Figures 2C, S2B, and S2C), a molecular signature strongly associated with RAIR and worse prognosis, as evidenced both in our data and prior studies (Figures 2D and S2D).^14,15,16 TP53 mutation is a late genetic alteration in DTC dedifferentiation, which is predominantly found in poorly differentiated and anaplastic thyroid cancers and less frequently in high-grade differentiated follicular cell-derived thyroid carcinoma.^17,18 Our cohort demonstrated a TP53 mutation prevalence of 7.1% (8/113) similar to a previous study¹⁹ (Table S2). Although the distribution of TP53 mutations did not differ significantly across the three subtypes, notably, the RAIR-associated TP53/TERT co-mutation pattern was observed in two cases, both of which belonged to the CC3 subtype (Table S2; Figure S2E).

Consistent with these genotypic findings, the CC1-CC3 subtypes showed progressively poorer responses to RAI therapy. Specifically, the proportion of RAIR disease significantly increased from CC1 to CC3 (Figure S2F, p = 1.4e−7 among the three subtypes, Pearson’s chi-squared test), paralleled by a significant decline in both biochemical and structural responses (Figures 2E and 2F, biochemical response: p = 4.5e−8; structural response: p = 5.4e−5 among the three subtypes, Pearson’s chi-squared test). Consequently, progression-free survival (PFS) after RAI administration differed significantly among subtypes (5-year PFS rate: CC1 vs. CC2 vs. CC3: 95.0% vs. 69.3% vs. 39.2%, log rank p = 2.0e−5, Figure 2G).

Collectively, these findings indicate that the pronounced heterogeneity in advanced DTC underlies distinct responses to RAI therapy, highlighting the need to precisely identify RAI-avid tumors and develop tailored therapeutic strategies for RAIR neoplasms.

Clinically applicable genomics-pathomics framework to infer CC subtypes

Nevertheless, the direct clinical application of this proteomic classification faces significant barriers because mass spectrometry-based proteomics involves high operational costs, prolonged analytical timelines, and complex technical requirements. To address these limitations and enhance the clinical applicability of our molecular classification, we have developed a rapid, cost-effective, and convenient workflow that utilizes two types of readily accessible clinical data in advanced DTC: (1) digital pathology whole-slide images (WSIs) and (2) gene mutation data from NGS-based panels, to enable robust prediction of the three proteomic CC subtypes.

First, we asked whether tumors of the three CC subtypes exhibited distinct pathologic patterns and whether CNN models could differentiate them through digital pathology analysis. Using WSIs from all 113 enrolled patients, we developed a deep learning pipeline to train CNN models for classifying the three CC subtypes. We then integrated the three most prevalent pathogenic gene alterations (BRAF^V600E mutation, RAS mutations, and TERT promoter mutations) with the CNN-based pathomics models using a random forest algorithm to construct a classifier within the G-P framework (Figure 3A; STAR Methods). The G-P framework demonstrated strong predictive performance for all three subtypes in cross-validation, with the area under curve (AUC) values of 0.754 for CC1, 0.804 for CC2, and 0.895 for CC3, significantly outperforming models using only genomic or pathomic data (Figures 3B, S3A, and S3B).

Figure 3.

Construction of the G-P framework inferring CC subtypes of advanced DTC

(A) Schematic diagram illustrating the development of the G-P integration framework to infer CC subtypes of advanced DTC.

(B) Receiver operating characteristic curves for using our G-P framework to infer CC subtypes of advanced DTC with robust discriminability (CC1: AUC = 0.754; CC2: AUC = 0.804; CC3: AUC = 0.895).

(C) Representative tiles from the cases of each CC subtype with the highest prediction score. Class activation maps highlight the subtype-specific discriminative subregions. Scale bar, 20 μm. All tiles presented were among the top 100 tiles according to the prediction score. Also see Figure S3.

Using our G-P framework, we retrospectively examined whether the pathological characteristics of the three G-P-inferred CC subtypes aligned with their proteomic features through visualization of representative image tiles. The morphological features of the tiles corresponding to each CC subtype were as follows: CC1 showed partially conserved thyroid gland morphology; CC2 exhibited abundant tumor stroma; and CC3 demonstrated enriched immune cell infiltration (Figure 3C). Collectively, these findings reveal distinct pathological patterns across CC subtypes that reflect the heterogeneous molecular signatures of advanced DTC.

The CC1 subtype corresponds to canonical tumors with better differentiation

Next, we delved into the intrinsic molecular underpinnings of each subtype. As shown in Figure 2A, we found that the thyroid hormone synthesis and metabolism-related pathways were significantly enriched in the CC1 subtype, suggesting that CC1 tumors may overall retain more functions of normal thyroid follicular cells. To quantify the degree of differentiation, we calculated the well-established thyroid differentiation score (TDS) for each subtype.¹³ The results revealed that CC1 tumors exhibited significantly higher TDS scores compared to the CC3 subtype (Figure 4A, p < 0.0001, ANOVA), particularly demonstrating elevated expression levels of two proteins closely associated with thyroid hormone synthesis: TG (thyroglobulin) and TPO (thyroid peroxidase) (Figures 4B–4D). Additionally, most key proteins in MAPK and PI3K/AKT/mTOR pathways were downregulated in CC1 tumors, accompanied by reduced levels of cell cycle- and cell proliferation-related proteins (Figures 4E, 4F, and S4).

Figure 4.

The canonical CC1 subtype exhibits higher degree of differentiation and RAI susceptibility

(A) Boxplot comparing TDS across CC subtypes in the Discovery Cohort (n = 113). The data are presented as the interquartile range (IQR) and median of TDS for each CC subtype; the two ends represent the maximum and minimum values. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

(B) Heatmap showing the proteomic expression level of TDS-related genes in the Discovery Cohort (n = 113).

(C and D) Line charts comparing the protein expression of thyroid differentiation markers TG (C) and TPO (D) among the three CC subtypes in the Discovery Cohort (n = 113). Error bars denote standard deviation (SD). ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

(E) Heatmap showing the abundance of key proteins in the MAPK pathway and/or PI3K/Akt/mTOR pathway in the Discovery Cohort (n = 113).

(F) Heatmap showing differential protein abundance of cell cycle-related and cell proliferation-related proteins among the three CC subtypes in the Discovery Cohort (n = 113).

(G) Uniform manifold approximation and projection (UMAP) visualization of 95,578 cells from 10 advanced DTC samples in the scRNA/spRNA-seq cohort (External Cohort 1, n = 10), colored by cell cluster.

(H) Density plot visualization illustrating the abundance of seven major cell types in two CC1 samples in External Cohort 1 (n = 10).

(I) Violin plot comparing TDS scores of malignant thyrocytes across CC subtypes in External Cohort 1 (n = 10). The data are presented as the IQR and median of TDS for each CC subtype.

(J) Violin plot comparing TDS scores using spRNA-seq data from four patients in External Cohort 1: two with CC1 tumors (P1, P2), one with a CC2 tumor (P3), and one with a CC3 tumor (P9). The data are presented as the IQR and median of TDS for each CC subtype.

(K) Spatial transcriptomics mapping of TDS scores using spRNA-seq data from four patients in External Cohort 1: two with CC1 tumors (P1, P2), one with a CC2 tumor (P3), and one with a CC3 tumor (P9).

(L) Pseudotime trajectory of malignant thyrocytes from 10 advanced DTC samples in External Cohort 1, colored by CC subtypes.

(M) Scatterplot showing a significantly negative correlation between TDS scores and pseudotime values inferred from the trajectory analysis for all malignant thyrocytes in External Cohort 1.

(N) Post-therapeutic whole-body scans (Rx-WBS) and post-RAI chest computed tomography (CT) scans obtained after the first and third courses of RAI treatment for a representative patient (P1) with CC1 tumor in External Cohort 1, demonstrating strong I¹³¹ uptake and favorable therapeutic efficacy in the pulmonary metastatic lesions (red arrows). The images clearly show significant remission of lung metastases after three courses of RAI therapy (200 mCi each). Also see Figures S4, S5, and S6 and Tables S3.

We next used an scRNA sequencing (scRNA-seq) cohort of 10 advanced DTC tumors for validation, 4 of which also underwent spRNA-seq profiling (scRNA/spRNA-seq cohort, External Cohort 1; Table S3). In total, 95,578 cells from 10 tumors were analyzed and 7 distinct major clusters were identified, including thyrocytes (also called thyroid follicular cells), B cells, T cells, macrophages, mast cells, endothelial cells, and fibroblasts (Figure 4G and S5A). Cell types were assigned based on consensus among expression-based clustering, copy number alteration inference, and canonical marker annotation (Figure S5B, STAR Methods). The G-P framework inferred that two, four, and four patients in this scRNA/spRNA-seq cohort were classified as CC1, CC2, and CC3 subtypes, respectively (Figure S5C). Within this cohort, CC1 tumors contained a higher proportion of thyrocytes compared to the other two subtypes (Figures 4H, S5D, and S5E). Using a previously developed classifier for thyrocyte malignancy classification on our scRNA-seq data,²⁰ we successfully distinguished malignant from benign thyrocytes, as confirmed by a difference in copy number variation (CNV) levels between the two groups (Figures S5F and S5G; STAR Methods). In scRNA-seq data, we observed that CC1 malignant thyrocytes displayed higher TDS scores than their counterparts in the other two subtypes (Figures 4I and 4J), which was further supported by spRNA-seq profiling in four patients from this cohort (Figures 4K and S5H). Furthermore, pseudotime analysis of malignant thyrocytes further demonstrated that CC1 tumor cells showed a tendency to be positioned at the starting point of tumor evolutionary dynamics (Figures 4L and 4M). These findings also aligned with our clinical data, that is, both of the CC1 patients in this cohort exhibited strong avidity for RAI therapy (Figures 4N and S5I).

Collectively, these findings validate that the CC1 subtype represents a “canonical” category of DTC characterized by better differentiation and heightened sensitivity to RAI therapy.

The CC2 subtype corresponds to stroma-enriched tumors with higher expression of anti-angiogenic mTKI target

As shown in Figure 2A, integrin-related pathways were significantly enriched in the CC2 subtype, suggesting abundant extracellular matrix-tumor cell interactions in CC2 tumors. We further investigated protein expression patterns across the three subtypes and found that multiple categories of stromal proteins exhibited elevated expression in the CC2 subtype, including collagens, integrins, laminins, and thrombospondins (Figure 5A). Furthermore, well-known anti-angiogenic mTKI targets—PDGFRB (platelet-derived growth factor receptor β), SRC, and FGF (fibroblast growth factor) receptor 2—exhibited elevated expression in the CC2 subtype, with intermediate expression in CC3 (Figures 5B, S6A, and S6B).

Figure 5.

The stromal CC2 subtype reveals enriched tumor stroma and elevated expression of mTKI target

(A) Abundance of signature protein families (including collagens, laminins, thrombospondins, integrins, fibronectins, and others) across CC subtypes in the Discovery Cohort.

(B) Violin plot comparing the protein expression of PDGFRB across CC subtypes in the Discovery Cohort. The data are presented as the IQR and median of the protein expression of PDGFRB for each CC subtype. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

(C) Violin plot comparing stromal scores across CC subtypes at the single-cell level in scRNA-seq data in External Cohort 1 (n = 10). The data are presented as the IQR and median of stromal scores for each CC subtype.

(D) Violin plot comparing stromal scores across CC subtypes at the sample level (pseudobulk) in scRNA-seq data in External Cohort 1 (n = 10). The data are presented as the IQR and median of stromal scores for each CC subtype.

(E) Density plot visualization illustrating the abundance of seven major cell types in four CC2 samples in External Cohort 1.

(F) UMAP visualization illustrating PDGFRB gene expression across seven major cell types in four CC2 samples in External Cohort 1.

(G) H&E-stained section (left, scale bar, 1 mm) and spatial transcriptomics mapping of stromal score (calculated by the ESTIMATE algorithm, middle) and PDGFRB gene expression (right) in a representative CC2 tumor (P3). Also see Figure S6 and Table S3.

We next characterized the CC2 subtype in the scRNA/spRNA-seq cohort (External Cohort 1). At single-cell resolution, CC2 tumors showed elevated stromal scores and greater fibroblast infiltration compared to the other subtypes (Figures 5C–5E and S6C), with the characteristic target PDGFRB predominantly expressed on fibroblasts in this subtype (Figure 5F). Consistent with these findings, spatial transcriptomic mapping of a representative CC2 tumor revealed increased stromal abundance (Figure 5G and S6D). Moreover, regions expressing PDGFRB showed marked spatial overlap with these stroma-abundant areas (Figure 5G). Clinically, a representative male patient with CC2 tumor in this cohort received three courses of RAI therapy. Although the lesions remained I¹³¹-avid, his pulmonary metastases consistently showed no remission (Figure S6E). This finding echoes the suboptimal therapeutic effect of RAI on CC2 tumors observed in the Discovery Cohort (Figures 2E–2G), prompting us to explore anti-angiogenic targeted therapies (particularly those targeting PDGFR), either in combination with RAI or as second-line therapy, for precision treatment of this subtype.

The CC3 subtype is characterized by a more immunogenic phenotype

As illustrated in Figure 2A, the CC3 subtype demonstrated significant enrichment of immune-related pathways. We performed immunohistochemical (IHC) staining on all 113 patients in the proteomic cohort. The results revealed that CC3 tumors exhibited significantly higher CD4⁺ and CD8⁺ T cell infiltration compared to non-CC3 subtypes, along with a significantly elevated expression of the immune checkpoint protein PD-L1 (Figures 6A, 6B, and S7A). These observations suggest that the CC3 subtype may possess greater immunogenicity than other subtypes, indicating potential susceptibility to immunotherapy.

Figure 6.

The immunogenic CC3 tumor subtype exemplifies prototypical features of immune-inflamed tumors characterized by an immunosuppressive TME and highly invasive behavior

(A) Representative IHC-stained sections of CD4, CD8, and PD-L1 for each CC subtype (left). Boxplots comparing the abundance of CD4⁺ T cells (upper right), CD8⁺ T cells (middle right), and PD-L1 expression (lower right) in IHC staining of all cases in the Discovery Cohort (n = 113), measured by average optical density (AOD) value. Scale bars, 100 μm. Data are presented as the IQR and median of AOD for each CC subtype; the two ends represent the maximum and minimum values. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

(B) Multiplex immunofluorescence staining of a representative CC3 tumor from a patient with widely invasive FTC in the Discovery Cohort. The results demonstrate positivity for CD56 and pan-CK, supporting the diagnosis of widely invasive FTC, while concurrent positivity for CD4, CD8, and CD134 supports the high immunogenicity of this tumor. Scale bars, 100 μm.

(C) Density plot visualization illustrating the abundance of seven major cell types in four CC3 samples in External Cohort 1.

(D) Density plot visualization illustrating the single-cell expression profiles of PD-L1 and PD-1 across CC subtypes samples in External Cohort 1 (n = 10).

(E) Spatial transcriptomics mapping of immune scores using spRNA-seq data from four patients in External Cohort 1: two with CC1 tumors (P1, P2), one with a CC2 tumor (P3), and one with a CC3 tumor (P9).

(F) H&E-stained section (left, scale bar, 1 mm) and spatial transcriptomics mapping of gene expression profiles inferring the distribution of macrophages (middle) and T cells (right) in a representative CC3 tumor (P9).

(G) Preoperative contrast-enhanced neck CT images for the third, fifth, and sixth neck surgeries in a representative CC3 tumor patient (P8) from External Cohort 1. This patient was radioiodine refractory and experienced repeated recurrences despite undergoing multiple extensive neck dissection surgeries, indicating a highly aggressive disease.

(H) CT image of the primary tumor in a representative CC3 patient (P10) from External Cohort 1. The tumor measures approximately 10 cm in maximum dimension and shows invasion into the skin.

(I) H&E staining and CD8, CD68, and PD-L1 IHC images from a representative CC3 tumor patient (P9) in External Cohort 1 demonstrate relatively abundant infiltration of CD8⁺ T cells and macrophages in the tumor microenvironment, along with PD-L1-positive expression. Scale bar, 50 μm.

(J) Multiplex immunofluorescence in a representative CC3 tumor patient (P7) from External Cohort 1 demonstrates relatively abundant infiltration of CD8⁺ T cells and macrophage infiltration in the tumor microenvironment, along with PD-L1-positive expression. Scale bar, 100 μm. Also see Figure S7 and Table S3.

We characterized CC3 tumors in the scRNA/spRNA-seq cohort (External Cohort 1). At single-cell resolution, our data suggested that CC3 tumor cells localized toward the terminal end of the pseudotime trajectory map, and malignant thyrocytes within this subtype exhibit the lowest TDS score at the single-cell level, consistent with their poor differentiation in bulk data (Figures 4I, 4J, and 4L). CC3 tumors exhibited enhanced infiltration of T cells and macrophages, consistent with their highest immune score compared to the other two subtypes (Figure 6C and S7B–S7D). In contrast to the other two subtypes, CC3 tumors exhibited predominant PD-L1 expression on tumor cells and macrophages. The interaction between PD-L1 and PD-1 on T cells provides a rationale for pursuing immune checkpoint blockade (Figure 6D). These observations were corroborated by the spatial transcriptomic profile of CC3 tumors (Figures 6E, 6F, and S7E–S7G). We reviewed the clinical data of all four CC3 patients in this cohort and found that all had intractable RAIR disease. One patient had undergone six neck surgeries due to recurrent neck lesions (Figure 6G). The other three patients presented with extensive primary tumor invasion, involving the skin, hypopharynx, and laryngeal cavity, respectively (Figures 6H, S7H, and S7I). We performed IHC and multiplex immunofluorescence staining on primary tumors from the latter two cases, which revealed positive PD-L1 expression on tumor cells, accompanied by abundant infiltration of CD8⁺ T cells and macrophages (CD68⁺) in the tumor microenvironment (TME) (Figures 6I and 6J). These clinical findings fully support the in silico results above.

Subtype-based precision treatment paradigm and real-world data validation

In light of the intrinsic biological features of the three molecular subtypes, we here propose a subtype-based precision treatment paradigm for advanced DTC. First, given that the CC1 subtype can achieve good and durable therapeutic effects with RAI treatment, RAI monotherapy is generally sufficient for patients with CC1 tumors. Second, the CC2 subtype has a moderate response to RAI therapy but exhibits higher expression of the anti-angiogenic mTKI target PDGFR. Thus, employing RAI as the first-line treatment for patients with CC2 tumors and considering anti-angiogenic targeted therapy (i.e., mTKI) as the second-line treatment following RAI resistance might be a reasonable approach. Third, the CC3 subtype has a relatively poor response to RAI treatment and exhibits the strongest immunogenicity and moderate expression of anti-angiogenic mTKI targets. Therefore, combination immunotherapy and targeted therapy may be a potential treatment for the CC3 subtype after RAI therapy failure (Figure 1).

To evaluate the potential clinical relevance of our proposed therapeutic paradigm, we established a large cohort of 110 advanced DTC patients receiving systemic therapies, which integrates real-world data from four approved therapies and seven clinical trials (External Cohort 2, Figure 7A; Table S4). For this cohort, pre-systemic therapy core needle biopsies or surgical samples were obtained from all patients, with their CC subtypes determined by the G-P framework. As anticipated, this cohort had a markedly lower proportion of CC1 tumors than the Discovery Cohort (7.3% vs. 38.1%) but a higher proportion of CC3 tumors (50.0% vs. 26.5%) (Figure 7B). This is explained by the fact that systemic therapy (targeted and/or immunotherapy) serves as second-line or later treatment for advanced DTC, following RAI. RAI-avid CC1 tumors seldom progress to the stage requiring such treatment, in contrast to CC3 tumors, whereas CC3 tumors frequently require second-line or exploratory therapies. Consequently, this difference in cohort composition further confirms the biological validity of our subtyping system.

Figure 7.

Establishment of the real-world treatment cohort undergoing systemic therapy (External Cohort 2) and validation of our proposed subtype-based precision treatment paradigm

(A) Selection process of the External Cohort 2. Finally, 110 patients with advanced DTC were enrolled in this cohort.

(B) Pie chart showing the distribution of CC subtypes in External Cohort 2 (n = 110).

(C) Bar plots comparing therapeutic efficacy of patients using mTKI alone or using mTKI + PD-1 inhibitor combination across CC subtypes in External Cohort 2 (n = 110), evaluated by RECIST v.1.1 criteria.

(D) Chest CT of a representative CC2 tumor in External Cohort 2 showing complete remission (CR) of the pulmonary metastases after 13 months of donafenib administration (right), compared with baseline (left, red arrows). Scale bar, 50 μm.

(E) Neck magnetic resonance imaging of a representative CC3 tumor in External Cohort 2 showing partial remission (PR) of the neck mass after 10 months of lenvatinib combined with camrelizumab (right, red arrow), compared with baseline (left, red arrow).

(F) Chest CT of a representative CC3 tumor in External Cohort 2 showing PR of the pulmonary metastases after 5 months of lenvatinib combined with camrelizumab (right, red arrow), compared with baseline (left, red arrows).

(G) Chest enhanced CT of a representative CC3 tumor in External Cohort 2 showing PR of the mediastinal mass after 10 months of lenvatinib combined with camrelizumab (right, red arrow), compared with baseline (left, red arrow). Scale bar, 50 μm. Also see Table S4.

We found that real-world treatment data from this cohort validate our subtype-based precision treatment paradigm. Among patients receiving anti-angiogenic mTKItargeted therapy alone, those with CC2 tumors had the best response (objective response rate [ORR]: 50%; disease control rate [DCR]: 86.1%), followed by CC3 (ORR: 37.5%; DCR: 79.2%), while CC1 had the worst efficacy (ORR: 25%, DCR: 50%). Conversely, in patients receiving combination mTKI targeted therapy and anti-PD-1 immunotherapy, the CC3 subtype demonstrated superior efficacy, with an ORR of 56.5% and DCR of 95.6%, followed by CC2 (ORR: 42.1%, DCR: 78.9%), with CC1 again having the worst outcomes (ORR: 25%, DCR: 50%) (Figure 7C, with representative cases shown in Figures 7D–7G).

Our preliminary findings are derived from a retrospective, non-randomized cohort and may be influenced by confounding factors such as prior treatment. Nevertheless, they reveal a treatment response pattern across subtypes that aligns with, and provides initial support for, our proposed paradigm. While these data are exploratory and require validation in prospective studies, they offer a clinically informed framework for tailoring therapy in advanced DTC.

Discussion

While some population-based retrospective studies demonstrate that RAI therapy improves survival outcomes in advanced DTC,^21,22,23 approximately 50% of patients with locally advanced or metastatic disease eventually develop RAIR, characterized by diminished iodine avidity and poor prognosis.^24,25 This therapeutic challenge persists despite recent advancements; with median survival drastically dropping from >10 years to 3–5 years following RAI resistance.^26,27,28 Concurrently, the treatment landscape is rapidly evolving with anti-angiogenic mTKI-based targeted therapies and emerging checkpoint blockade immunotherapies. Optimizing precision medicine for advanced DTC requires a deeper characterization of its molecular heterogeneity.

Previous studies of advanced DTC have focused on the relationship between genetic alterations and aggressive tumor behavior and therapeutic resistance. For instance, BRAF^V600E mutations correlate with reduced radioiodine avidity and increased recurrence risk due to impaired iodine metabolism pathways.²⁹ Similarly, TERT promoter mutations synergize with BRAF^V600E to drive tumor progression and higher mortality, serving as independent predictors of poor prognosis.^14,15 Although TP53 mutations are less common in DTC compared to anaplastic subtypes, their presence in advanced cases, especially co-presence with TERT promoter mutations, may indicate dedifferentiation and resistance to RAI therapy.^19,30,31 Beyond their role in RAI therapy, genomic alterations serve as critical biomarkers for targeted therapies (e.g., inhibitors of BRAF, RET, and NTRK) and immunotherapies (e.g., microsatellite instability [MSI] and tumor mutational burden [TMB]). Despite this, the role of proteomics in advanced thyroid cancer remains poorly investigated, even though proteins are the primary executors of biological activities.

To date, no proteogenomic study has specifically focused on advanced DTC, despite the publication of several proteogenomic investigations on thyroid cancer. Among these, two have centered on medullary thyroid carcinoma and poorly differentiated/anaplastic thyroid carcinoma,^32,33 which are biologically distinct from DTC. Of the two existing proteogenomic studies on PTC, one was specifically devoted to pediatric cases,³⁴ while the other encompassed tumors across all disease stages,³⁵ revealing a clear distinction from the present study. Since the vast majority of PTC cases are biologically indolent with high cure rates in early and mid stages, incorporating patients from all stages into a single classification model dilutes the proportion of advanced cases. Consequently, the critical biological heterogeneity of advanced-stage patients, who are in greater need of refined treatment strategies, may not be adequately captured. In contrast, our study exclusively focuses on advanced DTC. This specific focus enabled us to further expand the predictive spectrum for treatment efficacy at the proteomic level, encompassing a wider spectrum of therapeutic modalities beyond conventional RAI approaches.

Our study revealed marked molecular heterogeneity of advanced DTC. Using a discovery cohort and two external validation cohorts, this study identified and validated three proteomic subtypes of advanced DTC and characterized their intrinsic biological features: the canonical subtype (CC1), characterized by a higher degree of differentiation, more frequent RAS mutations, and strong avidity for RAI therapy; the stromal subtype (CC2), characterized by a stroma-rich microenvironment, elevated expression of certain anti-angiogenic mTKI targets, and response to RAI therapy; the immunogenic subtype (CC3), characterized by enrichment of immune cells, a higher frequency of the BRAF^V600E and TERT promoter mutation duet, and greater RAI refractoriness. Based on these features, we proposed subtype-based precision treatment strategies, which were validated using real-world therapeutic data (Figure 1).

A high proportion of advanced DTCs in our cohort exhibited co-occurring mutations. The frequencies of tumors harboring ≥2 mutations in the CC1, CC2, and CC3 subtypes were 30.2%, 35.0%, and 46.7%, respectively. Similarly, the proportions of tumors with ≥3 mutations across these subtypes were 14.0%, 5.0%, and 13.3%, respectively. This observation suggests that the prognosis and response to radioiodine therapy in advanced DTC are not determined solely by the mutation count but are profoundly influenced by the specific driver mutations and the TME they help shape. Importantly, not all accumulated mutations carry the same prognostic weight. For instance, the co-occurrence of BRAF and TERT promoter mutations is a well-established and extensively studied marker of high-risk disease in DTC.^14,36 Our results align with this, showing a statistically significant increasing gradient in the proportion of the BRAF/TERT promoter genetic duet from CC1 to CC3. Thus, specific mutation combinations are likely more critical determinants of aggressive behavior than the absolute number of hits. Consequently, the favorable prognosis of CC1, despite a higher prevalence of ≥3 mutations, may be attributed to the relative lack of such particularly aggressive co-mutations, thereby allowing the tumors to maintain a more differentiated state. In contrast, the poor prognosis of CC3 is likely driven by a combination of high-risk molecular events and its immune microenvironment, rather than by mutation burden alone. This observation supports the concept that tumor aggressiveness is a product of complex genotype-phenotype interactions. Therefore, our CC classification system, by integrating these multifaceted dimensions of biology, more effectively captures the underlying disease drivers than a simple tally of mutations.

As described above, genomic profiling plays an indispensable role in predicting prognosis and treatment responses in advanced DTC, while pathomics—a cutting-edge field in precision oncology—complements genomics by computationally analyzing tumor morphology, spatial architecture, and cellular heterogeneity from histopathologic images. In this study, the key innovation lies in developing a clinically applicable integrative framework that synergizes pathomics features with genomic data to effectively infer the three molecular subtypes, which enables seamless projection of proteomic heterogeneity in advanced DTC onto single-cell/spatial transcriptomic layers and subsequent validation through real-world clinical data integration. The G-P framework-inferred subtyping demonstrated high consistency in biological characteristics across multiple independent cohorts, indicating robust stability and generalizability of the established deep-learning computational model. Of particular interest is that our proposed molecular subtyping demonstrates robust predictive value for the efficacy of different therapeutic paradigms in advanced DTC. In our Discovery Cohort, the efficacy of RAI therapy in CC1 tumors was significantly higher than that in CC2 and CC3 tumors. In the systemic therapy cohort, which predominantly comprised very advanced iodine-refractory patients, the CC2 subtype showed superior response rates to targeted therapies and the CC3 subtype exhibited better efficacy with targeted plus immunotherapy compared to other subtypes. This demonstrates that our proposed subtyping can effectively guide precision treatment strategies, and our clinically accessible G-P framework provides a practical tool for future prospective subtype-based umbrella clinical trials.

Particular attention should be paid to the fact that our real-world cohort has a clinically critical subgroup of patients across all proteomic subtypes who demonstrate resistance to RAI, targeted therapy, and immunotherapy. The mechanisms underlying this multi-therapy resistance are multifactorial. RAIR is frequently associated with diminished sodium-iodide symporter expression and MAPK pathway-mediated tumor dedifferentiation.³⁷ Resistance to anti-angiogenic mTKIs commonly arises through activation of alternative pro-angiogenic signaling pathways (such as FGF) and adaptive remodeling of the TME.^38,39 Meanwhile, immunotherapy failure is often attributable to a profoundly immunosuppressive microenvironment characterized by T cell exhaustion and compromised antigen presentation.^40,41 To address this therapeutic challenge, future efforts should prioritize the approved genotype-directed targeted therapies (e.g., BRAF, RET, or NTRK inhibitors) in genetically eligible patients. Notably, beyond inhibiting their primary genetic drivers, some of these agents have demonstrated a potential redifferentiation effect, thereby potentially resensitizing tumors to RAI therapy.^42,43 Novel therapeutic avenues, such as CAR-T cell therapy and radionuclide therapy,^26,44,45,46 represent promising strategies currently under preclinical investigation. For this very high-risk population, achieving improved outcomes will ultimately require the implementation of deeply personalized combination therapies, guided by longitudinal molecular profiling and dynamic biomarker monitoring.

In conclusion, our study reveals the biological heterogeneity of advanced DTC and proposes a molecular subtyping system with potential clinical utility. The G-P integrative model provides a practical tool for subtyping, and the associated treatment hypotheses merit further investigation in prospective, biomarker-directed clinical trials.

Limitations of the study

Our study had some limitations. First, the current conclusions are derived from retrospectively analyzed datasets, particularly evident in the real-world systemic treatment cohort that had limited patients in each treatment arm. These findings and subtype-based treatment suggestions are still preliminary at this stage, and prospective validation through rigorously designed clinical trials remains imperative to establish higher-level evidence supporting these observations. Second, given that proteogenomic analyses are often based on limited tissue sampling, variability within tumors or biopsy bias may influence subtype classification and biomarker interpretation. Third, due to the potential strong intratumoral heterogeneity of tumors, pathomics-related studies often face limitations in tissue sampling, which may fail to fully represent the entire tumor’s characteristics. Particularly in our real-world clinical data validation cohort that included numerous core needle biopsy samples, this inevitably introduces bias into the pathomics prediction results. Fourth, the use of a targeted gene panel precluded the assessment of broader immunotherapy biomarkers, including TMB and MSI. Fifth, it is well established that numerous other driver genes, such as TP53 mutations, components of the PI3K-AKT signaling pathway, and epigenetic regulators, play significant roles in the biology, aggressiveness, and treatment response of thyroid cancer. However, in our cohort, these alterations individually occur at a relatively low frequency. Given the current limited sample size of our study, integrating these low-frequency alterations, despite their biological importance, could compromise the stability and generalizability of the G-P diagnostic model. Therefore, focusing on the most prevalent and impactful alterations (BRAF, RAS, and BRAF+TERT) was a pragmatic choice to ensure model robustness. Sixth, further mechanistic studies are warranted to understand the specific mechanisms of response within each CC cluster.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Xiao Shi (xshi11@fudan.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •Data statement: Sequencing data for seven newly profiled patients (P1, P2, P3, P4, P6, P8, and P9) have been deposited in the Genome Sequence Archive (GSA) for Human under accession number GSA: HRA011340. Data for three advanced DTC patients (P5, P7, and P10) were derived from our prior study²⁰ and have been deposited under accession number GSA: HRA001107 (corresponding to samples PTC5-T, PTC10-T, and PTC1-T). Proteomics data: mass spectrometry-based quantitative proteomics raw data have been deposited in the integrated Proteome resources (iProX) database under accession number iProX: IPX0011848000 at https://www.iprox.cn.
• •Code statement: Computational resources: custom CNN models, analysis scripts, and consensus clustering algorithms developed for this study are available at the following GitHub repository: https://github.com/xshi11/Consensus-Clustering-of-ADTC. All computational analyses were performed using publicly available software packages and tools as detailed in the respective subsections below and referenced appropriately throughout the manuscript. Software and algorithms: detailed information regarding all software packages, algorithms, and computational tools employed in this study, including version numbers and specific parameters, are provided in relevant method subsections. All analysis pipelines utilize open-source tools to ensure reproducibility and accessibility for the research community.
• •General statement: Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request.

Acknowledgments

The study was supported by the National Natural Science Foundation of China (82373008, 82573036, and 82002830 to X.S.; 82473361 and 82072951 to Y.W.; and 82002827 to T.Z.), the Science and Technology Commission of Shanghai Municipality (22Y21900100/23DZ2305600 to Y.W. and 23ZR1412000 to X.S.), the Shanghai Anticancer Association Foundation (SACA-AX202213 to Y.W.), and Shanghai Municipal Health Commission and Shanghai Medicine and Health Development Foundation (WJWRC202302 to X.S.). We thank Ms. Yuqing Yang from Shanghai OE Biotech Co., Ltd and Ms. Lingling Tan from Westlake Omics (Hangzhou) Biotechnology Co., Ltd. for their assistance in the analyses of proteomics data.

Author contributions

Conceptualization, X. Shi, Yu Wang, T.Z., and W.C.; methodology, T.Z., W.C., H.T., Cenkai Shen, Y. Zhang, Y.S., Y. Zhou, W.D., F.Z., and T.X.; validation, T.Z. and W.C.; formal analysis, X. Shi, T.Z., W.C., H.T.; investigation, Chentian Shen, H.X., C.W., C.L., Y.D., Yulong Wang, N.H., and X.M.; resources, X. Shi, Yu Wang, Q.J., Z.L., and D.J.; data curation, X. Shi, T.Z., W.C., H.T., and Cenkai Shen; writing – original draft, X. Shi, T.Z., W.C., and H.T.; writing – review and editing, Yu Wang, W.W., X. Sun, and D.J.; visualization, T.Z., W.C., H.T., Cenkai Shen, Y.S., Y. Zhou, F.Z., and T.X.; supervision, X. Shi, Yu Wang, Q.J., W.W., and X. Sun; funding acquisition, X. Shi, Yu Wang, and T.Z.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
anti-human CD4	Abcam	Cat# ab133616; RRID:AB_2750883
anti-human CD8	Abcam	Cat# ab245118; RRID:AB_3068617
anti-human CD68	Abcam	Cat# ab201340; RRID:AB_2920880
Anti-human PD-L1	Abcam	Cat# ab237726; RRID:AB_2884992
HRP-conjugated secondary antibodies	Agilent	Cat# K5007; RRID:AB_2888627
Anti- human CD134	Abcam	Cat# ab270727
Anti-pan-CK	Abcam	Cat# ab215838; RRID:AB_2922672
Anti- human CD56	Abcam	Cat# ab204446
Goat anti-rabbit IgG H&L	Abcam	Cat# ab205718; RRID:AB_2819160
Chemicals, peptides, and recombinant proteins
Optimal Cutting Temperature (OCT) compound	SAKURA	Cat# 4583
AllPrep DNA/RNA Kit	Qiagen	80204
RPMI 1640	Corning	Cat# 10-040-CVR
0.04% BSA	MACS	Cat# 1000076
0.2% collagenase II	Gibco	Cat# 17101015
Human CD45 MicroBeads	Miltenyi	Cat# 130-045-801
the 10x Genomics Chromium Next GEM Single Cell 3′ Reagent Kits v3.1	10x Genomics	Cat# PN-1000268
BD 40-μm cell strainer	Falcon	Cat# 352340
Critical commercial assays
Barocycler system	PressureBioSciences	NEP2320-45k
SOLAμ solid-phase extraction cartridges	Thermo Fisher Scientific™	N/A
nanoElute® nanoflow LC system	Bruker Daltonics	N/A
NGS-based panel	RigenBio	N/A
Deposited data
Raw and analyzed proteomics data of proteogenomic advanced DTC Cohort	This paper data	https://www.iprox.cn/page/home.html (iProX: IPX0011848000)
Raw and analyzed data of Single-Cell and Spatial Transcriptomics advanced DTC Cohort	This paper data	https://ngdc.cncb.ac.cn/gsa-human/ (GSA: HRA011340)
Kyoto Encyclopedia of Genes and Genomes (KEGG) database	Kanehisa	https://www.kegg.jp/
Software and algorithms
SPSS software (version 25.0)	IBM Corp	https://www.ibm.com/products/spss-statistics
GraphPad Prism (version 9.0.0)	GraphPad Software	https://www.graphpad.com/
R (version 4.2.3)	R Core Team	RRID: SCR_001905
DIA-NN software (version 1.8.1)	Github	https://github.com/ComplexData/DiaNN
Space Ranger software (version 2.0.1)	10x Genomics	https://www.10xgenomics.com/
ImageJ software (v1.54g)	Github	https://imagej.nih.gov/ij/
Cell Ranger software (version 9.0.0)	10X Genomics	https://www.10xgenomics.com/
Other
Custom convolutional neural network (CNN) models	This paper	https://github.com/xshi11/Consensus-Clustering-of-ADTC

Experimental model and study participant details

Discovery cohort: Patient selection and characterization of the proteogenomic cohort

A total of 196 patients with advanced DTC who underwent surgical treatment in the Department of Head and Neck Surgery at Fudan University Shanghai Cancer Center (FUSCC) between January 2008 and December 2022, followed by postoperative RAI therapy, were retrospectively identified. This study was approved by the Institutional Review Board of FUSCC (IRB# 2311286-19), and informed consent was obtained from all patients included in the study.

Patients were screened according to the following inclusion and exclusion criteria. Inclusion Criteria: (1) Pathologically confirmed DTC, including papillary thyroid carcinoma (PTC) or follicular thyroid carcinoma (FTC) or their pathological variants; (2) Locally advanced disease or distant metastasis confirmed by pathological or imaging examinations. Specifically, in this study, locally advanced DTC refers to a disease state of DTC where the primary tumor or metastatic lymph nodes have extensively invaded surrounding critical structures such as the trachea, esophagus, larynx, common carotid artery, or skin, such that complete R0 resection is either unachievable or would entail profound surgical risks or functional deficits; (3) RAI therapy administered for therapeutic purposes (rather than prophylactic ablation) with objectively evaluable lesions present for assessing RAI avidity/sensitivity.

Exclusion Criteria: (1) Unavailability of formalin-fixed paraffin-embedded (FFPE) pathological blocks; (2) Presence of poorly differentiated components or high-grade DTC (HGDTC) features identified on pathological review; (3) Concurrent diagnosis of other malignancies; (4) Loss to follow-up. All pathological sections underwent independent review and diagnosis by two experienced pathologists.

Ultimately, 113 patients met the criteria and were included in the final analysis (13–76 years of age, 69 female and 44 male, with all patients’ sex aligning with their gender, Table S1). Demographic information, clinicopathological characteristics, survival follow-up data, and other relevant clinical data were collected. Digital pathology images are obtained by digitizing Hematoxylin and Eosin (H&E)-stained sections of patient tumor tissue. Genetic data are acquired via next-generation sequencing (NGS) performed on fresh frozen tissue (when available) or sections from FFPE blocks.Overall survival (OS) of patients is defined as the period from the first day of RAI therapy until the patient’s death from the tumor or the last follow-up visit before losing follow-up for any other reason. Progression-free survival (PFS) is similarly defined as the period from the day of the patient’s first RAI therapy following surgery until the patient experiences disease recurrence and/or lymph node or distant metastasis, or until the last follow-up visit before loss to follow-up for any reason.

External Cohort 1: Single-cell and spatial transcriptomics cohort

Single-cell transcriptomics (scRNA-seq) using fresh samples from 7 advanced DTC samples (P1, P2, P3, P4, P6, P8 and P9), four of which (P1, P2, P3 and P9) were synchronized with spatial transcriptomics (spRNA-seq) to provide comprehensive molecular characterization. These patients with advanced DTC underwent initial or repeated surgical interventions at the Department of Head and Neck Surgery of FUSCC between January 2022 and December 2024. The scRNA-seq data for three additional samples (P5, P7 and P10) were obtained from our previously published dataset and integrated into the current analysis.²⁰ These two parts of data constituted the final External Cohort 1 (10 patients, aged 15–79 years, 4 female and 6 male, with all patients’ sex aligning with their gender, Table S3). All cases had pathologically confirmed DTC (PTC, FTC, or their variants), with locally advanced disease or distant metastasis confirmed through pathology or imaging. Comprehensive NGS genetic testing results and digital pathology slides are available for all patients. All histological sections underwent independent review and diagnosis by two experienced pathologists to ensure diagnostic accuracy and consistency. Collection and use of relevant data was approved by the IRB of FUSCC (IRB# 2311286-19), and informed consent was obtained from all patients included in the study.

External Cohort 2: Real-world systemic treatment cohort of advanced DTC

Patients with locally advanced or distant metastatic DTC who received immunotherapy or anti-angiogenic mTKI targeted therapy at FUSCC between January 2018 and December 2024 were retrospectively identified, including a subset of patients treated via participation in clinical trials. The therapeutic regimens included 7 clinical trials: Surufatinib (ClinicalTrials.gov identifier: NCT02614495), Toripalimab combined with Surufatinib (ClinicalTrials.gov identifier: NCT04524884), Famitinib combined with Camrelizumab (ClinicalTrials.gov identifier: NCT04521348), Anlotinib (ClinicalTrials.gov identifier: NCT04309136), Lenvatinib (ClinicalTrials.gov identifier: NCT02966093), Lenvatinib combined with Toripalimab/Camrelizumab/Pembrolizumab (ClinicalTrials.gov identifier: NCT06195228), Donafenib (ClinicalTrials.gov identifier: NCT03602495). It is worth noting that this cohort’s enrollment initiation date was set at 2018. Prior to 2018, only sorafenib (2017) was approved in China for treating advanced DTC. Subsequently, three additional drugs covered in our study—lenvatinib (2020), anlotinib (2022), and donafenib (2022)—were successively approved in China. After these approvals, prescriptions for these three drugs could be issued directly in clinical practice without being incorporated into clinical trials. In summary, this cohort integrated data of 110 patients from four approved drugs and seven clinical trials (aged 29–79 years, 56 female and 54 male, with all patients’ sex aligning with their gender, Table S4). These study protocols were separately approved by the IRB of FUSCC prior to the initiation of each clinical trial, and were performed in full accordance with the Guideline for Good Clinical Practice and the Declaration of Helsinki. The acquisition of clinical trial patient data was approved by both the Principal Investigators of all mentioned clinical trials and the respective enterprises. The use of data from this cohort was approved by the IRB of FUSCC (IRB# 2311286-19), and informed consent was obtained from all participating patients.

Patients were screened according to predefined criteria. Inclusion criteria included: (1) Pathologically and radiologically confirmed diagnosis of locally advanced or distant metastatic DTC; (2) Availability of histopathological specimens (either from surgery or core needle biopsy) prior to systemic therapy to prevent bias in histopathological analysis (e.g., post-systemic treatment tumors may exhibit extensive necrosis); (3) At least 3 months of post-treatment efficacy evaluation and over 6 months of clinical follow-up data. Exclusion criteria included: (1) Presence of poorly differentiated or anaplastic components in pathological specimens; (2) Concurrent diagnosis of other malignancies; (3) Incomplete clinical-pathological data or loss to follow-up. After applying these criteria, 110 patients were ultimately included in the analysis. Demographic information, clinical characteristics, oncological treatment history, therapeutic regimens, adverse events, molecular profiling results, survival outcomes, and other relevant data were systematically collected.

Method details

Specimen acquisition and preparation

Specimens from patients with advanced DTC were systematically collected from representative clinical sites by surgical resection or core needle biopsy. The acquisition of tumor specimens was performed through surgical resection for all patients in the Discovery Cohort, External Cohort 1, and the majority of patients (67/110, 60.9%) in External Cohort 2. A subset of specimens in External Cohort 2 (43/110, 39.1%) was obtained through core needle biopsy procedures to accommodate patients where surgical resection was not clinically indicated. All tissue collection procedures adhered to standardized protocols to ensure specimen quality and minimize pre-analytical variables that could impact molecular profiling outcomes.

Proteomics

Protein extraction and digestion

The proteomic analysis was conducted on 145 samples obtained from 125 patients, which included technical replicates and pooled quality control samples for evaluating experimental workflow and instrument stability. Among these 125 patients, 113 patients constituted the final Discovery Cohort, while the remaining 12 were excluded from the Discovery Cohort due to complete loss to follow-up. Depending on tumor content, one or two 5 μm-thick FFPE tissue slides per sample were used for protein extraction. The slides were processed using the pressure cycling technology (PCT)-assisted sample preparation pipeline as described in our previous publications.^47,48,49 Prior to peptide extraction, we performed microdissection on the FFPE sections, which had been cross-referenced with corresponding HE-stained sections, to ensure thorough removal of surrounding non-tumorous tissue areas. The peptide extraction protocol involved a Barocycler system (NEP2320-45k, PressureBioSciences Inc, South Easton, USA) for cycling processing. The denaturation, reduction and alkylation steps were carried out at 30°C for 90 cycles. Each cycle consisted of a high-pressure phase (45,000 psi) maintained for 30 s followed by an ambient pressure phase lasting 10 s. Denaturation was achieved by incubating samples in urea/thiourea buffer (6 M urea, 2 M thiourea). Subsequent reduction and alkylation steps were processed with 10 mM tris (2-carboxyethyl) phosphine (TCEP) and 40 mM iodoacetamide (IAA), respectively. The digest step was subsequently performed at 30°C for 120 cycles. Each cycle consisted of a 50-s high-pressure phase (20,000 psi) followed by a 10-s ambient pressure phase. The digest process is treated with Lys-C and trypsin at a 1:50 (w/w) enzyme-to-protein ratio. Upon reaction completion, desalting was carried out using SOLAμ solid-phase extraction cartridges (Thermo Fisher Scientific, San Jose, USA).⁴⁷

LC-MS/MS analysis

Peptide samples were separated using a nanoElute nanoflow LC system (Bruker Daltonics, Germany) equipped with a precolumn (3 μm, 100 Å, 20 mm × 75 μm i.d.) and an analytical column (custom-packed with C18 material, 1.9 μm, 120 Å, 150 mm × 75 μm i.d.). The separation was performed at a flow rate of 300 nL/min with a 60-min linear gradient: buffer B (0.1% formic acid in 100% acetonitrile) increased from 5% to 27% over 50 min, then to 40% over the next 10 min, with buffer A consisting of 0.1% formic acid in water.

Eluted peptides were analyzed using a timsTOF Pro mass spectrometer (Bruker Daltonics, Germany) coupled with the liquid chromatography system, equipped with a CaptiveSpray nano-electrospray ion source and operated in Data-Independent Acquisition (DIA) mode with Parallel Accumulation–Serial Fragmentation (PASEF). The TIMS unit was configured with an accumulation and ramp time of 100 ms each, resulting in a total cycle time of 1.17 s. Each cycle comprised 14 PASEF scans covering four ion mobility–m/z isolation windows per scan, with the ion mobility range set from 0.6 to 1.6 Vs/cm². Full MS1 and MS2 scans were acquired across the m/z range of 100–1700 Th. Precursors with a single charge state were excluded from fragmentation.

Database search

MS data processing followed an established protocol.⁵⁰ All data-independent acquisition (DIA) raw files underwent computational analysis using DIA-NN software (version 1.8.1) operating in library mode configuration.⁵¹ The analysis employed a comprehensive spectral library containing over 12,000 proteins and 215,000 precursors to ensure robust protein identification and quantification.⁵² Proteolytic digestion parameters specified trypsin and Lys-C as the enzymatic cleavage agents for peptide generation.

Post-translational modification parameters included carbamidomethylation of cysteine residues (+57.021464 Da) as a fixed modification to account for alkylation during sample preparation, while methionine oxidation (+15.994915 Da) was incorporated as a variable modification to capture naturally occurring oxidative changes. Peptide selection criteria encompassed sequences ranging from 7 to 30 amino acids in length, with precursor ion mass-to-charge ratios constrained to 300–1800 m/z and fragment ion detection limited to 200–1800 m/z to optimize spectral quality and identification confidence. Quality control measures implemented a stringent 1% false discovery rate (FDR) threshold at the precursor level using target-decoy database search strategy to minimize false positive identifications. All additional DIA-NN computational parameters remained at software default settings to maintain analytical consistency and reproducibility across sample processing workflows.

Quantification of global proteome data

Proteins with abnormal quantities of identification were removed from the dataset. The remaining 8,216 proteins with less than 80% missing values were used in the subsequent step of analysis. Column-median centering normalization technique was used to normalize the Extracted Ion Chromatogram (XIC) value of every protein to remove variations in equal loading from the different samples, followed by log2-transforming the data. The SeqKnn algorithm (v1.0.1) was then used to impute missing value.⁵³ For every batch of pool samples, as well as the technical replications, the coefficient of variation (CV) was calculated.

Consensus clustering and subtype characterization

Of the 8216 quantifiable proteins, those with a CV in the top 50% (n = 4108) were included in downstream clustering analyses. Consensus clustering (CC) was performed on the remaining proteomic profiles using the ConsensusClusterPlus R package (v1.50.0),⁵⁴ with the following parameters: maxK = 5, reps = 500, pItem = 0.8, clusterAlg = “kmdist”, distance = “pearson”, and corUse = “everything”. Optimal cluster number was determined based on the consensus matrix, cumulative distribution function (CDF), delta area plot, tracking plot, and quantitative approaches including Proportion of Ambiguously Clustered pairs (PAC) and Monte Carlo Reference-based Consensus Clustering (M3C).

To investigate pathway-level differences among subtypes, single-sample Gene Set Enrichment Analysis (ssGSEA) was conducted using the Kyoto Encyclopedia of Genes and Genomes (KEGG) database, generating Enrichment Scores (ES) for signaling pathways in each sample. Subgroups exhibiting distinct molecular or clinicopathological characteristics were prioritized for further investigation. Ultimately, the 113 samples in the Discovery Cohort were stratified into three subtypes comprising 43, 40, and 30 cases, respectively.

Differential protein identification

Differentially expressed proteins were identified using Welch’s t test (unequal variances) with Benjamini-Hochberg false discovery rate (FDR) correction. Proteins meeting significance thresholds (FDR <0.05; fold change >2) were classified as differentially expressed.

Functional enrichment analysis

GO term and KEGG pathway enrichment analyses were performed using clusterProfiler (v4.7.1.002).⁵⁵ Significance was determined by Fisher’s exact test with Bonferroni correction. Pathways with FDR <0.05 were considered significantly enriched.

Gene Set Enrichment Analysis

Gene Set Enrichment Analysis (GSEA) was implemented via clusterProfiler (v4.7.1.002) using the MSigDB Hallmark gene sets.⁵⁶ Enrichment significance was evaluated through 1,000 phenotype permutations.

Next-generation sequencing and bioinformatics analysis

Library preparation and sequencing

Targeted DNA and RNA libraries were prepared using an NGS-based panel (RigenBio, Shanghai, China), a multiplex PCR-based assay for point mutations and insertions/deletions of 22 thyroid cancer-related genes (AKT1, ALK, BRAF, CTNNB1, CHEK2, EIF1AX, EZH1, FGFR1, FLT3, GNAS, HRAS, KIT, KRAS, NRAS, ZNF148, PIK3CA, PTEN, RET, SPOP, TERT promoter, TP53 and TSHR) and thyroid cancer-related gene fusions including RET, NTRK1, NTRK3, and PPARG.¹² Briefly, genomic DNA and total RNA were isolated from fresh-frozen tumor specimens using AllPrep DNA/RNA Kit (Qiagen). RNA samples were reverse-transcribed into cDNA, and both DNA and cDNA templates were subjected to multiplex PCR to amplify target regions. Unique dual indices and Illumina sequencing adapters were incorporated during a secondary PCR step. Amplified libraries were purified, quantified using a Qubit fluorometer (Thermo Fisher), and sequenced on the Illumina NovaSeq 6000 platform (Illumina, San Diego, USA) to generate 150 bp paired-end reads.

Bioinformatic processing and variant calling

Raw sequencing data underwent initial quality control using FastQC (v0.11.9) and ReSeqTools (v0.25). Adapter trimming and low-quality base removal were performed with Trimmomatic (v0.39). Processed DNA reads were aligned to the human reference genome (hg19) using BWA (v0.7.17). Single nucleotide variants (SNVs) and small insertions/deletions (indels) were identified with VarScan2 (v2.4.4) and annotated using the Ensembl Variant Effect Predictor (VEP). For RNA-derived reads, alignment to a customized reference genome was performed using BWA, and gene fusion events were detected using in-house scripts designed for fusion transcript identification. All analyses were conducted according to quality-controlled, pre-validated pipelines.

Treatment response assessment and biochemical evaluation

Structural response to RAI therapy, targeted therapy and immunotherapy interventions was evaluated using Response Evaluation Criteria in Solid Tumors (RECIST v1.1) or immune-modified iRECIST guidelines.⁵⁷ Specifically, response categories were defined as complete response (CR) for disappearance of all target lesions, partial response (PR) for ≥30% decrease in sum of target lesion diameters compared to baseline, progressive disease (PD) for ≥20% increase in sum of target lesion diameters compared to the smallest recorded sum or the appearance of new lesions, and stable disease (SD) for changes insufficient to qualify as PR or PD. For patients exhibiting variable therapeutic responses during treatment with identical agents or different drugs within the same class, best overall response (BOR) sustained for minimum six weeks was selected for statistical analysis.

Biochemical response to RAI therapy was also assessed through serial serum thyroglobulin (Tg) measurements, with baseline Tg concentration defined at initial I¹³¹ administration. Patients were stratified into three response groups based on relative Tg changes: Group G1 demonstrated ≥50% Tg decrease relative to baseline, Group G2 showed <50% Tg decrease or <10% increase, and Group G3 exhibited ≥10% Tg increase following therapy.⁵⁸ This biochemical classification system enabled correlation of Tg dynamics with therapeutic efficacy across CC subtypes.

Assessment of thyroid differentiation score

Thyroid differentiation status was quantitatively assessed through computation of Thyroid Differentiation Score (TDS) adapted from the methodology established in the landmark TCGA-PTC study using 16 thyroid metabolism and function genes (TG, TPO, SLC5A5, SLC26A4, SLC5A8, DIO1, DIO2, PAX8, FOXE1, NKX2-1, GLIS3, DUOX1, DUOX2, THRA, THRB, TSHR, TFF3).¹³

For the scRNA/spRNA-seq cohort (External Cohort 1), TDS was calculated using the algorithm containing all 16 genes. Specifically, transcriptomic expression data underwent initial log₂ transformation of Counts Per Million (CPM) values for scRNA-seq or spRNA-seq data, followed by mean-centering normalization across the dataset. Individual TDS value was computed as the arithmetic mean of normalized expression levels for the 16 signature genes.

For the proteome-based Discovery Cohort, TDS was calculated per sample using expression values of proteins corresponding to the gene set established in mRNA-based scoring frameworks.¹³ Specifically, among the 16 genes mentioned above, 9 were covered by our global proteome data in this study including DIO1, DUOX1, DUOX2, FOXE1, NKX2-1, PAX8, TG, TPO, TSHR. Expression levels of these 9 proteins were log2-transformed and subsequently centered on the mean. TDS score for each sample was defined as the arithmetic mean of the normalized expression values of these proteins, calculated as: TDS = Mean of log₂(Fold Change) for the 9 proteins.

Histological and multiplex immunostaining procedures

Hematoxylin and eosin staining

FFPE and fresh-frozen tissue sections were prepared for H&E staining. Paraffin sections were deparaffinized with Environmentally Friendly Dewaxing solutions and rehydrated through a graded ethanol series, followed by tap water rinsing. Frozen sections were equilibrated from −20°C to room temperature, fixed in tissue fixative for 15 min, and washed under running water. Hematoxylin staining was performed for 3–5 min, followed by tap water rinsing, brief differentiation, and bluing before a final rinse. Eosin counterstaining followed gradient dehydration (85% and 95% ethanol, 5 min each), and slides were incubated in eosin solution for 5 min. Final dehydration was performed in anhydrous ethanol (three changes, 5 min each), cleared in xylene, and mounted using neutral resin. Bright-field images were acquired for downstream analysis.

Immunohistochemistry and multiplex immunofluorescence

For immunohistochemistry (IHC), FFPE slides were dewaxed using three Environmentally Friendly Dewaxing solutions, rehydrated through absolute ethanol, and washed in distilled water. Endogenous peroxidase was quenched with 3% hydrogen peroxide for 25 min, and nonspecific binding was blocked using 3% BSA (or rabbit serum for goat-derived antibodies). Primary antibodies (Anti-CD4: Cat# ab133616; Anti-CD8: Cat# ab245118; Anti-CD68: Cat# ab201340; Anti-PD-L1: Cat# ab237726; Abcam, Cambridge, UK) were applied and incubated overnight at 4°C. After PBS washes, HRP-conjugated secondary antibodies (Dako REAL EnVision Detection System, Cat# K5007, Agilent, Santa Clara, USA) were added and incubated for 50 min. Signal detection was achieved with 3,3′-diaminobenzidine (DAB), followed by hematoxylin counterstaining, ethanol dehydration, xylene clearing, and sealing.

Multiplex immunofluorescence (mIF) was performed using a TSA-based protocol. Slides were baked, dewaxed in xylene, and rehydrated. Antigen retrieval was performed by microwave heating in AR buffer, followed by washing in PBST. Tissue boundaries were encircled with a hydrophobic barrier, and blocking solution was applied in a humidified chamber. Slides were sequentially incubated with primary antibodies (Anti-CD4: Cat# ab133616; Anti-CD8: Cat# ab245118; Anti-CD68: Cat# ab201340; Anti-PD-L1: Cat# ab237726; Anti-CD134: Cat# ab270727; Anti-pan-CK: Cat# ab215838; Anti-CD56: Cat# ab204446; Abcam) and HRP-conjugated secondary antibodies (Goat anti-rabbit IgG H&L: Cat# ab205718; Abcam), followed by TSA fluorophore labeling. After each cycle, previous antibody complexes were removed with heat-induced stripping. This staining-amplification-stripping sequence was repeated for each target. Finally, slides were counterstained with DAPI and mounted using antifade medium.

Slide scanning and quantitative analysis

Stained slides were scanned using a Pannoramic MIDI digital pathology scanner (3DHISTECH, Budapest, Hungary) and evaluated in CaseViewer (v2.4). Two board-certified pathologists, blinded to clinical and prognostic data, independently reviewed all slides. Immunostaining was quantified using the average optical density (AOD), derived from five representative 20× fields per slide. The integrated optical density (IOD) and corresponding area were measured using ImageJ software (v1.54g),⁵⁹ and AOD was calculated as: AOD = IOD/Area. Final AOD scores were averaged from both pathologists' readings.

Integrated genomics-pathomics framework to infer CC subtypes

Deep learning-based histopathological classifier as input for subsequent integrated models

We established an integrated computational framework combining deep learning-based histopathological analysis with genomic profiling for molecular subtype prediction. Histopathological whole-slide images (WSIs) from tumor specimens underwent systematic preprocessing wherein images were partitioned into non-overlapping 256 × 256-pixel tiles. Quality control measures included automated selection of high-quality tissue regions and color normalization to mitigate inter-batch staining variations, generating tens of thousands of standardized tiles per WSI. The deep learning architecture employed a two-stage convolutional neural network (CNN) implementation using PyTorch framework. The primary CNN served as a tissue-type discriminator to identify histologically relevant regions, while the secondary CNN performed molecular subtype classification on the filtered high-quality tiles.

Genomic mutation as input for subsequent integrated models

Concurrently, based on classification accuracy and variable parsimony, the three most prevalent gene mutations (BRAF^V600E mutations, RAS mutations, and TERT promoter mutations) were ranked using Mean Decrease Accuracy and Mean Decrease Gini metrics to serve as genomic input results for subsequent integrated models.

Model integration and performance validation

To enhance predictive accuracy beyond either modality alone, a final integrative random forest (RF) model was developed by combining inputs from the histopathological CNN model with inputs from genomic features. Specifically, the probability scores generated by the CNN model and the mutational profiles of BRAF^V600E, RAS and TERT promoter were used as input features for a new ensemble model constructed using the RF algorithm (Genomics-Pathomics framework, G-P framework). The model was implemented in R using the “randomForest” package, with the number of trees (ntree) set to 500.⁶⁰

Model performance was quantitatively assessed using receiver operating characteristic (ROC) analysis, with area under the curve (AUC) serving as the primary classification metric for both tissue discrimination and molecular subtype prediction tasks. The integrated model demonstrated superior predictive performance compared to individual models built solely on pathomics or genomic features, justifying its selection as the final classification framework. External validation was performed on independent datasets including the scRNA/spRNA-seq cohort (External Cohort 1) and real-world systemic treatment cohort (External Cohort 2) to assess model generalizability and robustness across diverse patient populations and institutional protocols.

Single-cell transcriptomics (scRNA-seq) experimental and bioinformatic method

Single-cell suspension preparation

Under sterile conditions, freshly collected differentiated thyroid carcinoma tissues were washed twice with ice-cold RPMI 1640 + 0.04% BSA, minced into ∼0.5-mm³ fragments using surgical scissors, and digested in freshly prepared enzyme solution at 37°C for 30–60 min with gentle inversion every 5–10 min. The digestion mixture contained RPMI 1640 (Corning, Cat# 10-040-CVR), 0.04% BSA (MACS, Cat# 1000076), and 0.2% collagenase II (Gibco, Cat# 17101015). The digested suspension was filtered through a BD 40-μm cell strainer (Falcon, Cat# 352340) 1–2 times and centrifuged (300 × g, 5 min, 4°C). The cell pellet was resuspended in medium, mixed with an equal volume of red blood cell (RBC) lysis buffer (Miltenyi, Cat# 130-094-183), and incubated for 10 min at 4°C. After centrifugation (300 × g, 5 min), the supernatant was discarded. The pellet was washed once with medium, followed by another centrifugation again and the final supernatant was removed. CD45⁺ immune cells were enriched using Human CD45 MicroBeads (Miltenyi, Cat# 130-045-801) according to the manufacturer’s protocol (sorting in the ratio of positive: negative = 1:2). Finally, the cells were re-suspended with 100 μL RPMI 1640 medium (Corning, Cat# 10-040-CVR) + 0.04% BSA. Single-cell suspension concentration and cell viability were then evaluated using Luna-FL cell counter (Logos Biosystems, Korea) or Trypan Blue staining method.

scRNA-seq library construction

Freshly prepared single-cell suspensions were adjusted to a concentration of 700–1200 cells/μL and processed according to the manufacturer’s protocol for the 10x Genomics Chromium Next GEM Single Cell 3′ Reagent Kits v3.1 (Cat# PN-1000268, 10x Genomics, Pleasanton, USA) for loading onto the Chromium Controller and subsequent library preparation. The constructed libraries were subjected to high-throughput sequencing on the Illumina Nova 6000 PE150 platform.

scRNA-seq analysis method

The FASTQ files were processed and aligned to reference genome (human: GRCh38) using Cell Ranger software (version 9.0.0) from 10x Genomics, with unique molecular identifier (UMI) counts summarized for each barcode. The UMI count matrix was then analyzed using the Seurat (version 4.0.0) R package.⁶¹ To remove low-quality cells and likely multiplet captures, retaining high-quality cells meeting these criteria: Cells were filtered by (1) gene numbers <200, (2) UMI <1000, (3) log10GenesPerUMI <0.7, (4) proportion of UMIs mapped to mitochondrial genes >15% and (5) proportion of UMIs mapped to hemoglobin genes >5%. Potential doublets were removed using DoubletFinder (v2.0.3).⁶² To obtain the normalized gene expression data, library size normalization was processed using the NormalizeData function. Specifically, the global-scaling normalization method “LogNormalize” normalized the gene expression measurements for each cell by the total expression, multiplied by a scaling factor (10,000 by default), and log-transformed the results.

For dimensionality reduction and clustering, the top 2000 highly variable genes (HVGs) were identified using FindVariableFeatures (mean.function = FastExpMean, dispersion.function = FastLogVMR). Principal-component analysis (PCA) was performed to reduce the dimensionality with RunPCA function. Then, the RunHarmony function in harmony (version 1.0) R package was performed to remove the batch effects.⁶³ Cellular clustering based on gene expression profiles was conducted via graph-based partitioning using the FindClusters function. Results were visualized in two dimensions with Uniform Manifold Approximation and Projection (UMAP) algorithm with the RunUMAP function.

The FindAllMarkers function (test.use = presto) was used to identify marker genes of each cluster. Differentially expressed genes (DEGs) were selected using the function FindMarkers (test.use = presto). Identified markers were visualized via VlnPlot and FeaturePlot.

Bonferroni correction was applied for multiple testing adjustments of p values. p value <0.05 and |log₂foldchange| > 0.58 was set as the threshold for significantly differential expression. Cell type annotation was performed with SingleR (v1.4.1) by correlating each cell’s expression profile with reference datasets via Spearman correlation, assigning the cell type from the reference dataset showing the highest correlation.⁶⁴

Distinguishing malignant and non-malignant thyrocytes

As described in the landmark TCGA-PTC study and our previous publication, the proportion of CNV at the single-cell level in thyroid cancer is less pronounced than in many other malignancies.^13,20 Consequently, methods commonly used in other cancer types – such as inferCNV and other CNV-based approaches – lack efficacy for distinguishing malignant cells in thyroid cancer. Therefore, in our prior work, we developed a machine learning classifier trained on TCGA bulk transcriptome profiles to discriminate between malignant and non-malignant thyrocytes within scRNA-seq data, achieving exceptional accuracy.²⁰ In the current study, we directly applied this classifier to distinguish malignant and non-malignant thyrocytes in our scRNA-seq datasets.

Due to the overall low level of CNV in PTC, it means there may be substantial overlap in the distribution of CNV levels between benign and malignant thyrocytes. For this reason, we believe that CNV is not a highly specific or reliable criterion for distinguishing benign from malignant thyrocytes. Although it is difficult to make a determination for individual cells, the average CNV level in a group of malignant thyrocytes is still likely to be significantly higher than that in a group of benign thyrocytes. Thus, although not suitable for primary classification, comparing CNV levels between pre-defined benign and malignant cells can serve to corroborate the accuracy of the classification. CNV profiles were inferred from scRNA-seq data using the inferCNV package (v1.0.4).⁶⁵ Briefly, gene expression levels were transformed by first sorting genes according to their genomic positions and then applying a moving-average smoothing with a 101-gene window. The smoothed values were then mean-centered. The analysis was performed with a cutoff value of 0.1, followed by a de-noising step to generate the final CNV profiles.

Pseudotime trajectory evolutionary analysis of scRNA-seq

The developmental pseudotime was determined with the Monocle2 package (version 2.9.0).⁶⁶ The raw count was first converted from Seurat object into CellDataSet object with the importCDS function in Monocle. The differentialGeneTest function of the Monocle2 package was used to select ordering genes (FDR <0.01) which were likely to be informative in the ordering of cells along the pseudotime trajectory. The dimensional reduction clustering analysis was performed with the reduceDimension function, followed by trajectory inference with the orderCells function using default parameters. Gene expression was plotted with the plot_genes_in_pseudotime function to track changes over pseudo-time. The Pearson correlation coefficient quantifies the linear relationship between pseudotime value and TDS scores by computing the ratio of their covariance to the product of their standard deviations, thereby standardizing the result to a comparable range of [-1, 1].

Spatial transcriptomics (spRNA-seq) experimental and bioinformatic method

spRNA-seq library construction

Freshly collected advanced DTC tumors were dissected into appropriately sized blocks, surface moisture absorbed with lint-free paper, embedded in Optimal Cutting Temperature (OCT) compound (Cat# 4583, SAKURA, Osaka, Japan), snap-frozen on dry ice, and stored at −80°C. OCT-embedded samples were sent to Shanghai OE Biotech Co., Ltd. (Shanghai, China) for spatial transcriptomic library preparation, sequencing and data analysis. Frozen sections (10 μm thickness) were prepared using a Leica CM1950 Microtome Cryostat (Leica Microsystems, Wetzlar, Germany) at −20°C. Tissue sections were subjected to methanol fixation, H&E staining, imaging and destaining following the 10x Genomics recommended experimental procedure (CG000614). Probe hybridization, probe release and transferred to the 10x Genomics Visium CytAssist slide, and the library construction was performed using the Visium CytAssist Spatial Gene Expression for FFPE kit (PN-1000520 for Human, 6.5mm), according to the 10x Genomics experimental flow (CG000495). Final libraries were sequenced (PE100) on the BGI DNBSEQ-T7 platform (BGI Genomics, Shenzhen, China).

spRNA-seq analysis process

The FASTQ files were processed using Space Ranger software (version 2.0.1) from 10x Genomics with unique molecular identifier (UMI) counts summarized for each barcode, align reads to the human reference genome (GRCh38), and quantify spot-level metrics (total spots, reads/spot, genes detected, UMIs) for quality assessment. The filtered UMI count matrix was then analyzed using Seurat (version 4.1.0) R package.⁶¹ Sctransform was used to normalize data and identify top 3000 highly variable genes (HVGs).⁶⁷ Principal component analysis (PCA) was conducted to reduce dimensionality on the log transformed gene-barcode matrices of top variable genes. The top 3000 highly variable genes (HVGs) were identified with Seurat’s FindVariableFeatures, followed by PCA dimensionality reduction on HVGs and batch effect correction via RunHarmony (v1.0).⁶³ Marker gene identification for each cluster utilized FindAllMarkers (bimod test), while DEG detection between conditions employed FindMarkers (presto test) with Bonferroni-adjusted p values. Statistical significance was defined by an adjusted p value <0.05 and an absolute log₂-fold-change >0.58.

AddModuleScore calculating TDS, stromal and immune scores in spRNA-seq

The AddModuleScore function in Seurat R package was used to define the score of TDS and immune and stromal gene sets (Table S5). Specifically, AddModuleScore computes spot-level gene program activity scores through mean target expression normalized by background control feature gene sets. All analyzed features were binned based on averaged expression, and the control features were randomly selected from each bin.

Quantification and statistical analysis

Statistical methodologies for omics-based analyses are detailed within their respective sections. For clinical data analysis, standard statistical tests were applied as appropriate to the data type and study design. For categorical variables, comparisons were performed using Pearson’s chi-square test with Bonferroni correction for multiple testing. For continuous variables, two-tailed Student’s t test or Mann-Whitney U test was applied to compare two groups for normally distributed data or non-normally data, respectively, while comparisons involving more than two groups were performed using ANOVA followed by post hoc Tukey’s honestly significant difference (HSD) test.

Survival analyses were conducted using Kaplan-Meier estimates and compared using the log rank test. Multivariable Cox regression models evaluated associations between somatic gene alterations and PFS. Forward stepwise variable selection was conducted using likelihood ratio tests at α = 0.05 entry threshold.

Clinical data analyses were primarily conducted using SPSS software (version 25.0; IBM Corp., Armonk, USA). Receiver operating characteristic (ROC) curves and area under the curve (AUC) values were calculated using the “pROC” packages and visualized using “ggplot2” packages in R (version 4.2.3).^68,69 Data visualization and graphing were performed using GraphPad Prism (version 9.0.0; GraphPad Software, San Diego, USA) and R (version 4.2.3). All statistical tests were two-sided, and p < 0.05 were considered statistically significant unless otherwise specified.

Published: March 6, 2026