A per- and polyfluoroalkyl substances-based gene signature links prognosis to immune landscapes in thyroid cancer

Abstract

Background

Thyroid cancer (THCA) is the most common endocrine malignancy with a rising global incidence and significant heterogeneity. Although per- and polyfluoroalkyl substances (PFAS) exposure is linked to thyroid dysfunction, the prognostic value of per- and polyfluoroalkyl substances-related genes (PFASRGs) and their role in the tumor immune microenvironment (TME) remain poorly understood. This study aims to systematically screen key PFASRGs and evaluate their prognostic value as biomarkers for THCA.

Methods

Utilizing The Cancer Genome Atlas (TCGA)-THCA transcriptomic data and PFASRGs, we constructed a prognostic model through differential expression analysis, univariate and multivariate Cox regression analyses, and the least absolute shrinkage and selection operator (LASSO). The model’s robustness was validated using receiver operating characteristic (ROC) curves, Kaplan-Meier analysis, and clinical nomograms. Furthermore, the TME, immunotherapy response, and drug sensitivities were systematically evaluated. Distinct molecular landscapes were characterized by stratifying the cohort via unsupervised consensus clustering analysis.

Results

The eight-gene prognostic model demonstrated robust performance, with area under the curve (AUC) values exceeding 0.85 across all validation cohorts. High-risk patients exhibited significantly shorter overall survival and an “inflamed” TME characterized by high immune scores and checkpoint expression. In contrast, the therapeutic efficacy of anti-programmed death-ligand 1 (PD-L1) agents was more pronounced in the low-risk category, as evidenced by a superior objective response. Furthermore, distinct molecular subtypes and risk-specific sensitivities to targeted agents, such as sorafenib and sunitinib, were identified, highlighting the model’s clinical utility for personalized treatment.

Conclusions

We established a novel THCA prognostic framework based on eight PFASRGs. This model exhibits superior performance in risk stratification, effectively distinguishing cohorts with divergent clinical trajectories, unique immune microenvironment features, and varied therapeutic responses. Our findings provide a powerful predictive tool for refining prognostic evaluation and facilitating the implementation of personalized management strategies for THCA patients.

Highlight box.

Key findings

• From transcriptomic data of thyroid carcinoma (THCA), eight differentially expressed genes related to per- and polyfluoroalkyl substances (PFAS) were identified, and a robust prognostic risk model was subsequently established. The high- and low- risk groups exhibited distinct immune infiltration profiles and biological functional activities.

What is known and what is new?

• Although PFAS exposure has been potentially linked to the rising incidence of THCA, the systemic prognostic and immunomodulatory functions of these substances in THCA remain poorly understood.

• This study systematically identified key PFAS-related genes (PFASRGs), established a prognostic evaluation model, and further elucidated the potential biological functions of these genes as well as their associated immune microenvironment features in THCA.

What is the implication, and what should change now?

• PFASRGs-derived biomarkers are applicable to both survival stratification and prognostic evaluation for individualized immunotherapy in THCA, thereby offering clinicians a valuable reference for refining prognostic judgment and tailoring precision therapeutic approaches.

Introduction

Thyroid cancer (THCA) represents the most prevalent malignancy of the endocrine system. While its incidence has demonstrated a rising trend in recent years, the overall prognosis remains favorable, with a 5-year relative survival rate of approximately 98% (1). The vast majority of THCA cases are classified as differentiated, with papillary thyroid carcinoma (PTC) being the most common histological subtype; its distinct clinical and molecular characteristics have been extensively characterized to facilitate subtype classification and prognostic evaluation (2). Diverse environmental and dietary factors contribute to its pathogenesis, in which exposure to ionizing radiation during childhood is a well-established risk factor (3), and variations in iodine intake levels correlate with different pathological subtypes (4,5). At the molecular level, mutations or gene rearrangements involving BRAF V600E, RAS, and RET/PTC constitute the core pathogenic foundation for THCA development and molecular stratification (6,7). For localized and well-differentiated THCA, surgical resection remains the primary therapeutic approach, often supplemented by postoperative radioactive iodine (RAI) therapy to eliminate residual lesions and reduce recurrence risk (8). For patients with RAI-refractory metastatic differentiated disease, multi-target tyrosine kinase inhibitors, such as sorafenib and lenvatinib, have emerged as critical therapeutic options by significantly prolonging progression-free survival (8,9). Despite significant advancements in clinical intervention, evidence supporting the impact of targeted therapies on extending overall survival remains limited (10). Furthermore, drug-related adverse events and acquired resistance frequently result in transient efficacy or necessitate treatment discontinuation (11), while the inherent heterogeneity of the tumor immune microenvironment (TME) and tumor latency pose persistent challenges for assessing therapeutic response and recurrence risk (12,13). Importantly, the immune system plays a distinct role in the development and progression of different THCA subtypes. For instance, PTC typically exhibits a relatively “cold” immune phenotype with low tumor-infiltrating lymphocytes (TILs) (9,14), whereas aggressive subtypes such as anaplastic thyroid carcinoma (ATC) often present a more inflamed microenvironment characterized by abundant TILs and elevated programmed death-ligand 1 (PD-L1) expression (15,16). Follicular thyroid carcinoma (FTC) and medullary thyroid carcinoma (MTC) also display subtype-specific immune profiles that correlate with their metastatic potential and clinical outcomes (17,18). Understanding these subtype-specific immune landscapes is valuable. Consequently, identifying prognostic biomarkers closely associated with immune regulation in THCA is essential. Such endeavors not only help elucidate key molecular pathways and therapeutic targets but also hold significant clinical value for enhancing the accuracy of prognostic stratification, optimizing immunotherapy strategies, and ultimately improving patient survival outcomes.

Per- and polyfluoroalkyl substances (PFAS) comprise a broad category of anthropogenic organofluorine chemicals extensively integrated into a myriad of manufacturing and household commodities—including non-stick coatings, food packaging, stain-resistant fabrics, and aqueous film-forming foams—owing to their exceptional hydrophobicity, thermal stability, and resistance to chemical degradation. Consequently, these substances exhibit remarkable environmental persistence and extensive bioaccumulation within biological systems (19). PFAS enter the human body through multiple routes, such as contaminated drinking water, dietary intake, inhalation, and dermal contact (20,21). Given their prolonged half-lives in human plasma and tissues, they present a significant public health concern characterized by chronic, low-dose, population-wide exposure (22,23). As quintessential endocrine-disrupting chemicals, PFAS have been demonstrated to perturb the synthesis, transport, and metabolism of thyroid hormones, thereby compromising the homeostasis of the hypothalamus-pituitary-thyroid axis (24). Mechanistic investigations have revealed that specific PFAS congeners can trigger mitochondrial dysfunction and induce the release of mitochondrial DNA, which subsequently activates the AIM2 inflammasome and promotes the secretion of pro-inflammatory cytokines, such as IL-1β. This cascade leads to chronic inflammation and tissue damage, providing direct molecular evidence for PFAS-driven remodeling of the immune microenvironment (25). Furthermore, emerging epidemiological evidence suggests a potential correlation between PFAS exposure and the rising incidence of THCA (26). Despite the growing body of literature implicating PFAS in THCA pathogenesis, a comprehensive and unified understanding of how these substances modulate tumor evolution and therapeutic sensitivity at the molecular and cellular levels remains to be established.

Building upon the molecular characteristics of THCA, we constructed a novel prognostic model by incorporating PFAS-related genes (PFASRGs). By performing systematic bioinformatics evaluations within The Cancer Genome Atlas (TCGA)-THCA cohort, we established and verified a robust risk stratification framework that effectively discriminates patient survival and delineates immune microenvironment features. This study offers fresh insights into THCA pathogenesis and lays a critical foundation for personalized clinical management. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0694/rc).

Methods

Data acquisition

Transcriptomic data for the THCA cohort (TCGA-THCA), comprising 59 normal and 513 tumor samples, were retrieved from the UCSC Xena database (https://xena.ucsc.edu/). This cohort consisted entirely of pathologically confirmed PTC. Corresponding copy number variation (CNV) and clinical metadata were sourced from TCGA database (https://portal.gdc.cancer.gov/). Individuals lacking survival information (i.e., incomplete follow-up records or missing vital status data) were excluded from the study cohort. Thereafter, we executed a random 60/40 split of the patient cohort, yielding a training cohort (302 cases) and an internal validation cohort (208 cases). Additionally, a total of 855 PFASRGs were curated from existing literature (27). The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Differentially expressed analysis and univariate Cox regression analysis

Transcriptional variations between THCA specimens and adjacent normal controls were characterized utilizing the “limma” R library (version 3.62.2), where genes meeting the criteria of |log fold change (FC)| >1 and adjusted P value <0.05 were defined as differentially expressed genes (DEGs). The intersection of these DEGs and the previously curated PFASRGs was designated as differentially expressed PFASRGs (DEPFASRGs). Prior to performing survival analysis, a 30-day survival threshold was established, whereby subjects with a shorter follow-up duration were discarded to minimize non-cancer-specific interference. To pinpoint candidates with significant survival relevance, univariable Cox proportional hazards modeling (P<0.05) was implemented leveraging the “survival” R package (version 3.5-8). Finally, the expression profiles of these prognostic-related genes were compared across groups, and their inter-gene correlations were evaluated.

Identification of prognostic factors and development of a prognostic model

Least absolute shrinkage and selection operator (LASSO) regression was implemented to refine the preliminary gene set and minimize the risk of overfitting, a process facilitated by the “glmnet” library (version 4.1-9) to ensure the robustness of the resulting model. To enhance model parsimony and mitigate the influence of multicollinear variables, the most suitable tuning parameter (λ) was identified through a decuple cross-validation procedure, facilitating effective coefficient shrinkage. Subsequently, multivariable Cox proportional hazards regression was conducted using the “survival” package (version 3.5-8) on the LASSO-selected genes to construct the final prognostic model. The risk score calculation method is as follows:

$$\text{Risk\hspace{0.17em}score}={\displaystyle \sum \left(\text{Gene\hspace{0.17em}}{\text{expression}}_i\times {\text{Coefficient}}_i\right)}$$ [1]

The median risk score was utilized as the optimal cutoff to bifurcate the patient cohort into high- and low-risk categories. This stratification allowed for the comparative analysis of survival outcomes through Kaplan-Meier curves. Validation of the model’s forecasting capability was facilitated by the “timeROC” package (version 0.4), which was employed to determine the 3-, 5-, and 10-year time-dependent area under the curve (AUC) scores and the associated receiver operating characteristic (ROC) manifolds. Finally, the distribution of risk scores and survival statuses was visualized. To establish its generalizability, the predictive performance of the framework was rigorously evaluated within both the internal validation set and the comprehensive TCGA-THCA cohort.

Functional annotation and pathway profiling of gene sets

To pinpoint DEGs across the risk strata, the “limma” R package was employed with filtering thresholds established at |log2FC| >1 and false discovery rate (FDR) <0.05. To elucidate the functional roles of the identified DEGs, we performed Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) annotation analysis leveraging the “clusterProfiler” package (v4.14.6) to uncover the underlying biological pathways associated with the prognostic model.

Independent prognostic analysis

To determine whether the risk score serves as an autonomous predictor alongside clinicopathological factors, we implemented both univariable and multivariable Cox proportional hazards models, with the outcomes depicted via forest plots. Following this, we developed a clinical nomogram utilizing the “rms” package (version 8.0-0) to forecast survival likelihoods at 3-, 5-, and 10-year intervals. To evaluate the alignment between forecasted probabilities and actual outcomes, calibration analysis was implemented to validate the nomogram. Furthermore, we employed the “ggDCA” package (version 1.2) to facilitate decision curve analysis (DCA), thereby validating the clinical applicability of the established framework.

Risk model analysis based on clinicopathological features

The associations of the risk score with the phenotypic attributes of the TCGA-THCA cohort were evaluated through correlation tests. To evaluate the prognostic significance of key clinical parameters, patients were stratified into various subgroups based on their clinical features and risk categories. Specifically, patients were further categorized according to age (≤65 vs. >65 years), gender (female vs. male), clinical stage (stage I + II vs. III + IV), tumor (T) stage (T1+2 vs. T3+4), node (N) stage (N0 vs. N1), and metastasis (M) stage (M0 vs. M1). Subsequently, survival curves were generated via the Kaplan-Meier method to compare the prognostic outcomes across these pre-defined clinical stratifications.

Immune infiltration analysis

To comprehensively characterize the TME, we employed the single-sample gene set enrichment analysis (ssGSEA) algorithm via the “GSVA” package (version 2.0.7) and the estimation of stromal and immune cells in malignant tumor tissues using expression data (ESTIMATE) algorithm via the “estimate” package (version 1.0.13) to evaluate immune infiltration patterns and scores across different risk strata. In parallel, to corroborate the immune infiltration patterns, the cell-type identification and tracking by estimating relative subsets of RNA transcripts (CIBERSORT) method was utilized to estimate the proportions of infiltrating immune cells within the TME. For a deeper understanding of the model’s biological characteristics, bivariate Pearson analyses were conducted to correlate the signature’s genetic components with the cellular abundance of the immune microenvironment as estimated by CIBERSORT. Moreover, to investigate potential mechanisms of immune evasion, the tumor immune dysfunction and exclusion (TIDE) computational framework (http://tide.dfci.harvard.edu/) was implemented to contrast immunotherapeutic response scores across the two risk strata. Finally, the IMvigor210 cohort was employed to validate the translational value of the prognostic model in predicting the efficacy of immune checkpoint blockade. Individuals undergoing anti-PD-L1 intervention were bifurcated into polarized risk subsets leveraging our established framework, followed by a rigorous analysis of the correlation between risk profiles and clinical efficacy.

Somatic mutation burden profiling and therapeutic response estimation

To facilitate drug discovery, a multi-dimensional computational framework was utilized. First, we evaluated tumor mutational burden (TMB) and depicted the mutation profiles of the top 20 genes using the “maftools” package (version 2.22.0) to identify genomic disparities between risk groups. To uncover translational opportunities, we cross-referenced our signature genes with the CellMiner pharmacogenomic data, pinpointing agents whose clinical efficacy aligns with the expression of the identified feature genes. Finally, the “pRRophetic” package (version 0.5) was used to estimate and compare the half-maximal inhibitory concentration (IC50) values of specific compounds across the high- and low-risk cohorts, providing insights into individualized treatment strategies.

Subtype identification of PFASRGs

To delineate distinct molecular subtypes within the TCGA-THCA population, we implemented an unsupervised consensus clustering approach utilizing the “ConsensusClusterPlus” R library (version 1.70.0), driven by the expression profiles of the identified prognostic markers. Subsequently, survival analysis was conducted to evaluate the clinical outcomes across the identified molecular subtypes. To characterize the immunological heterogeneity across the identified clusters, the ssGSEA methodology was implemented to quantify variations in immune-associated pathways and the abundance of infiltrating immune populations. To further quantify the immune landscape, we utilized the MCP-counter algorithm (implemented via the “IOBR” package, version 0.99.0) and the CIBERSORT algorithm to analyze and compare immune infiltration patterns across the distinct subtypes.

Real-time quantitative polymerase chain reaction (RT-qPCR)

The normal thyroid cell line Nthy ori-3-1 and the THCA cell line BCPAP were obtained from the American Type Culture Collection (ATCC, Manassas, VA, USA). Cells were cultured in RPMI-1640 medium (Gibco, Grand Island, New York, USA) supplemented with 10% fetal bovine serum (Gibco) and maintained at 37℃ in a humidified incubator containing 5% CO₂. Total RNA was extracted using TRIzol reagent according to the manufacturer’s instructions. Complementary DNA was synthesized from the isolated RNA using the Hifair III One-Step RT-qPCR SYBR Green Kit (Yeasen, Shanghai, China). RT-qPCR was then performed with Hieff qPCR SYBR Green Master Mix (Yeasen) on the basis of the manufacturer’s protocol. Relative gene expression levels were calculated using the 2^−ΔΔCt method, with β-actin serving as the internal reference gene. The sequences of the primers used are listed in Table 1.

Table 1. Primers and their sequences for RT-PCR analysis.

Gene	Forward primer	Reverse primer
THRSP	GCAGCGAAGAGAATGGAACC	CTTCTATCATGTGAAGGGATCTTCC
CIDEC	AGGGCATCATGGCTTACAGT	GCTTCAGGGTTCCTAGTCTTGA
ALPL	AGTGCTCTGCGCAGGATTG	CGCCAGTACTTGGGGTCTTT
HGF	TGGCATCAAATGTCAGCCCT	GCTCGAAGGCAAAAAGCTGTG
AQP8	CCTGATGTCTGGAGAGATAGCC	AACCGTTCGTACCAGGACAC
APOE	GCCTCTAGAAAGAGCTGGGAC	TGATTGGCCAGTCTGGAGG
TF	CCAGAGTTTCCGCGACCATA	CGCTTCGTTTGCCGCAATG
MYH7	ACTTGAGTAGCCCAGGCACA	TAGCCGCTCCTTCTCTGACT

RT-PCR, real-time polymerase chain reaction.

Statistical analysis

All data processing and integrative analyses were executed within the R statistical framework (version 4.4.0), utilizing a diverse suite of project-specific libraries to ensure the reproducibility of the computational workflow. Statistical disparities across cohorts were appraised through the Wilcoxon rank-sum or Kruskal-Wallis H-tests, whereas linear associations between continuous variables were quantified using Pearson’s correlation method. Figures were produced mainly with the “ggplot2” package (version 3.5.2). The P values were summarized as follows: ***, P<0.001; **, 0.001<P<0.01; *, 0.01<P<0.05.

Results

Analysis of DEPFASRGs in THCA patients

A total of 2,770 DEGs (1,419 up-regulated and 1,351 down-regulated) were identified between the tumor and normal groups of the TCGA-THCA cohort (Figure 1A). A total of 146 overlapping genes, designated as DEPFASRGs, were identified by cross-referencing the THCA-associated DEGs with the PFASRG set (Figure 1B). The DEPFASRGs were subjected to univariable Cox regression analysis, yielding a subset of 23 candidate prognostic genes (Figure 1C, Table S1). Evaluation of the correlative dependencies among these genes revealed a tightly coupled regulatory network, with numerous pairs showing strong and significant associations (Figure 1D). Furthermore, the box plot was generated to illustrate the distinct expression profiles of these candidate genes in tumor versus normal tissues (Figure 1E).

Figure 1.

DEPFASRGs identification and functional analysis. (A) Volcano plot of differential expression analysis in TCGA-THCA. (B) Intersection analysis of DEGs and PFASRGs. (C) Univariate forest plot of prognostic genes. (D) Correlation heatmap of prognostic genes. (E) Box plot of prognostic gene expression. ***, P<0.001. CI, confidence interval; DEPFASRGs, differentially expressed PFAS-related genes; DEGs, differentially expressed genes; PFASRGs, PFAS-related genes; TCGA-THCA, The Cancer Genome Atlas Thyroid Cancer.

Construction and performance assessment of the predictive framework

To derive a robust predictive signature, we implemented a sequential regression-based pipeline designed to optimize feature selection and model stability. Initially, LASSO regression was performed on the 23 candidate genes, resulting in the selection of 13 key features (Figure 2A,2B, Table S2). Subsequently, stepwise multivariate Cox regression was utilized for model optimization, ultimately identifying eight core prognostic genes for THCA (Figure 2C, Table S3). Risk profiles for the patient cohort were quantified through our established model, followed by the construction of time-dependent ROC curves to validate its forecasting efficacy across 3-, 5-, and 10-year milestones. Within the training cohort, AUC scores across every temporal interval surpassed 0.9 (Figure 2D), reflecting remarkable predictive accuracy. Similarly, the AUC values remained above 0.85 in both the validation cohort and the entire TCGA-THCA dataset (Figure 2E,2F). Kaplan-Meier survival analysis further revealed that high-risk patients exhibited significantly lower survival rates compared to the low-risk group across the training, validation, and total cohorts (Figure 2G-2I). To provide a visual representation of these disparities, risk score and survival status distribution maps were plotted (Figure 2J-2L). Notably, divergent expression levels were observed for all eight prognostic genes when comparing the high-risk and low-risk populations (Figure S1A), with AQP8 and CIDEC standing out as independent indicators of survival (Figure S1B). Gene expression analysis indicated that THRSP, ALPL, AQP8, and APOE were upregulated in tumor tissues compared to normal tissues, while the remaining four genes were significantly downregulated (Figure S1C). The results of the RT-qPCR experiment were consistent with those of this analysis (Figure S1D). To decode the underlying molecular heterogeneity between the identified risk strata, we performed comprehensive functional annotation and pathway enrichment profiling. GO analysis highlighted that the DEGs between risk groups were primarily associated with the humoral immune response, serine-type endopeptidase activity, and glycosaminoglycan binding (Figure 2M). KEGG enrichment results indicated that these genes were significantly linked to critical pathways, including neuroactive ligand-receptor interaction, calcium signaling pathway, complement and coagulation cascades, and the renin-angiotensin system (Figure 2N).

Figure 2.

Prognostic model development and validation. (A) LASSO coefficient distributions for the 23 identified markers, illustrating the shrinkage process during feature selection. (B) Partial likelihood deviance for the LASSO regression via 10-fold cross-validation to select the optimal penalty parameter (λ). (C) Multivariate Cox proportional hazards results for the 8 key predictors, with the forest plot delineating hazard ratios and confidence intervals. (D-F) Time-dependent ROC plots assessing the predictive accuracy for survival at the 3-, 5-, and 10-year marks within the training, validation, and entire cohorts, respectively. (G-I) Kaplan-Meier survival curves comparing the overall survival between high- and low-risk groups in the training, validation, and entire cohorts. (J-L) Distributions of risk scores and survival statuses for patients in the training, validation, and entire cohorts. (M) Functional annotation of the transcriptional signatures distinguishing high-risk from low-risk patient subsets via GO analysis. (N) Functional annotation of the transcriptional signatures distinguishing high-risk from low-risk patient subsets via KEGG analysis. *, P<0.05; **, P<0.01; ***, P<0.001. AUC, area under the curve; CI, confidence interval; FC, fold change; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic.

Analysis of independent prognostic factors

The prognostic merit of the established model was appraised via univariable Cox regression, confirming that the riskScore acted as a prominent determinant of patient outcomes (Figure 3A). Adjusting for potential confounding factors, the multivariate Cox proportional hazards model verified that the risk score remains a statistically significant and autonomous prognostic indicator (Figure 3B). Building upon these findings, a clinically practical nomogram integrating these independent factors was constructed to quantify the prognosis for each patient in the TCGA-THCA cohort (Figure 3C). DCA demonstrated that the nomogram yielded significant clinical net benefits in predicting 3-, 5-, and 10-year survival rates (Figure 3D). Furthermore, calibration curves indicated that the nomogram-predicted probabilities were highly congruent with the ideal outcomes (Figure 3E).

Figure 3.

Development and validation of the nomogram. (A,B) Systematic screening (A) and independent validation (B) of prognostic predictors via univariate and multivariate Cox analyses. (C) A clinical nomogram developed by integrating the risk score with clinical characteristics to predict 3-, 5-, and 10-year overall survival. (D) DCA quantifying the net clinical gain and practical value of the nomogram across varying threshold probabilities. (E) Calibration curves assessing the alignment between nomogram-predicted probabilities and observed outcomes. CI, confidence interval; DCA, decision curve analysis; M, metastasis; N, node; T, tumor.

Stratified assessment of the risk signature across clinicopathological features

The distribution of risk scores across key clinical parameters was visualized using violin plots. Statistical analysis revealed significant disparities in risk scores associated with age and stage (Figure 4A). The clinical robustness of our framework was subsequently challenged by segregating the patient cohort into discrete categories according to individual clinical attributes. Consistently, Kaplan-Meier analysis validated that high-risk status was associated with an inferior prognosis within the M0, N1, and female subgroups compared to the low-risk group (Figure 4B). Table 2 presents a comparative overview of the initial clinicopathological landscape for patients categorized into different risk strata.

Figure 4.

Stratified assessment of the risk index across distinct clinicopathological categories. (A) Variation patterns of the risk scores among major clinicopathological features. (B) Kaplan-Meier curves contrasting high- versus low-risk subsets across various clinical strata. M, metastasis; N, node; T, tumor.

Table 2. Comparison of baseline clinical features between distinct risk categories.

Name	High (n=151)	Low (n=151)	P
Fustat			0.002
0	140 (92.7)	151 (100.0)
1	11 (7.3)	0 (0.0)
Age (years)	49.0±16.5	44.4±15.0	0.01
Gender			0.79
Female	111 (73.5)	108 (71.5)
Male	40 (26.5)	43 (28.5)
Stage			0.81
Stage I	83 (55.0)	89 (58.9)
Stage II	15 (9.9)	15 (9.9)
Stage III	37 (24.5)	32 (21.2)
Stage IV	15 (9.9)	15 (9.9)
Unknown	1 (0.7)	0 (0.0)
T			0.23
T1	48 (31.8)	32 (21.2)
T2	49 (32.5)	59 (39.1)
T3	47 (31.1)	53 (35.1)
T4	6 (4.0)	7 (4.6)
Unknown	1 (0.7)	0 (0.0)
N			0.39
N0	73 (48.3)	66 (43.7)
N1	63 (41.7)	74 (49.0)
Unknown	15 (9.9)	11 (7.3)
M			0.67
M0	85 (56.3)	81 (53.6)
M1	2 (1.3)	4 (2.6)
Unknown	64 (42.4)	66 (43.7)

Data are presented as n (%) or mean ± standard deviation. M, metastasis; N, node; T, tumor.

Leveraging the prognostic model to delineate immune infiltration patterns and forecast potential immunotherapy benefit

To delineate the unique immunological landscapes of the identified risk strata, we performed an extensive assessment of the tumor-infiltrating immune microenvironment. A comparative analysis of ssGSEA enrichment scores revealed a markedly attenuated immune profile in the low-risk category, characterized by significant reductions in both functional immune pathways and the abundance of infiltrating subsets relative to the high-risk cohort (Figure 5A). ESTIMATE algorithm analysis further revealed that patients in the low-risk group exhibited lower ESTIMATE, Immune, and Stromal Scores, but significantly higher tumor purity (Figure 5B). CIBERSORT analysis showed that the low-risk group was characterized by significantly elevated infiltration of Tregs, M0 macrophages, M2 macrophages, and resting mast cells, whereas the high-risk group was predominated by naive B cells, plasma cells, and M1 macrophages (Figure 5C). The high-risk subgroup was characterized by a higher transcript abundance of most immune checkpoint markers compared to the low-risk group (Figure 5D). Interestingly, individuals categorized into the low-risk stratum manifested an enhanced sensitivity to PD-L1 blockade, as evidenced by their elevated objective response rates (Figure 5E). Compared to the low-risk group, the high-risk subset exhibited considerably greater TIDE scores (Figure 5F), implying a stronger tendency toward immune evasion. Finally, Pearson correlation analysis elucidated the intricate associations between the feature genes of the prognostic model and the landscape of immune cell infiltration (Figure 5G).

Figure 5.

Characterization of the tumor immune microenvironment. (A) Boxplots showing the ssGSEA scores of immune cell subsets and immune-related functions between the high- and low-risk groups. (B) TME metrics—encompassing stromal and immune infiltration, alongside estimated tumor purity—between the two risk strata as determined via the ESTIMATE framework. (C) Bar-and-whisker plots depicting the landscape of immune infiltration across different cohorts based on CIBERSORT-derived estimations. (D) Boxplot showing the expression levels of immune checkpoint molecules in the high- and low-risk groups. (E) Comparison of clinical responsiveness between high- and low-risk categories undergoing anti-PD-L1 therapy in the IMvigor210 population. (F) Divergence in TIDE-calculated immune evasion potential across the polarized risk strata. (G) Correlation heatmap between the feature genes of the prognostic model and the levels of immune cell infiltration. *, P<0.05; **, P<0.01; ***, P<0.001. CIBERSORT, cell-type identification and tracking by estimating relative subsets of RNA transcripts; ESTIMATE, estimation of stromal and immune cells in malignant tumor tissues using expression data; PD-L1, programmed death-ligand 1; ssGSEA, single-sample gene set enrichment analysis; TIDE, tumor immune dysfunction and exclusion; TME, tumor immune microenvironment.

Tumor mutational landscape and drug sensitivity

We initially explored the link between somatic mutational burden and our prognostic model; notably, both risk-stratified cohorts consistently exhibited high mutational frequencies in the BRAF and NRAS genes (Figure 6A). Leveraging the “pRRophetic” package, we estimated the IC50 of prevalent medications to investigate the potential for personalized clinical interventions. The results revealed distinct sensitivities between the two risk strata: compared to the high-risk group, patients in the low-risk group were more sensitive to doxorubicin and trametinib, whereas they exhibited greater resistance to sorafenib and sunitinib (Figure 6B). To explore potential therapeutic avenues, we quantified the association between our prognostic features and anticancer drug sensitivity using the CellMiner repository, identifying agents whose efficacy significantly correlated with the model’s gene expression. The results demonstrated that ALPL expression was significantly negatively correlated with trametinib (r=−0.330, P=0.01) and positively correlated with lenvatinib (r=0.270, P=0.03). Additionally, APOE expression showed a significant positive correlation with dabrafenib (r=0.298, P=0.02), while CIDEC expression was significantly negatively correlated with vandetanib (r=−0.455, P<0.001) (Figure 6C, Table S4).

Figure 6.

Tumor mutation and drug sensitivity analysis. (A) Visualization of mutational profiles for the 20 genes with the highest aberration frequencies in the high- (upper panel) and low-risk (lower panel) cohorts, annotated with TMB status. (B) Violin plots comparing the estimated IC50 values of four candidate drugs between the high- and low-risk groups. (C) Visualization of gene-drug interactions, depicting the interrelationship between the expression profiles of the prognostic candidates and drug potency data via the CellMiner platform. IC50, half-maximal inhibitory concentration; TMB, tumor mutational burden.

Subtype identification

Based on the expression of feature genes within the prognostic model, three distinct THCA-associated subtypes were identified through unsupervised consensus clustering (Figure 7A-7C). Principal component analysis (PCA) further confirmed the robust discrimination and effective separation among these three molecular subtypes (Figure 7D). Kaplan-Meier survival analysis indicated significant prognostic disparities, with group 2 patients exhibiting the worst overall survival, whereas group 3 patients demonstrated the most favorable outcomes (Figure 7E). Subsequently, differential expression analysis across these subtypes yielded 37 overlapping genes (Figure 7F). Functional profiling through GO enrichment indicated that these markers are mainly associated with biological processes such as monocarboxylic acid transport, lipoprotein particle organization, and lipid transporter activity. Furthermore, KEGG analysis indicated that these genes were mainly associated with cytokine-cytokine receptor interaction, linoleic acid metabolism, and the NF-kappa B signaling pathway (Figure 7G). Additionally, gene expression analysis showed significantly heterogeneous expression levels of these markers across the three subtypes (Figure 7H). In contrast to groups 1 and 3, group 2 displayed a more prominent TME signature, as evidenced by significantly higher ESTIMATE, immune, and stromal indices, which inversely correlated with its lower tumor purity levels (Figure 7I). Consistent with these findings, ssGSEA analysis indicated that both immune cell infiltration and immune function scores were significantly elevated in group 2, characterized by a higher abundance of B cells (Figure 7J). Further dissection via CIBERSORT revealed an increased infiltration of naive B cells in group 2 (Figure 7K). Finally, MCP-counter analysis corroborated these results, showing significant differences in B lineage infiltration among the three subtypes (Figure 7L).

Figure 7.

Subtype identification and immune profiling. (A) Heatmap depicting the consensus clustering results of THCA specimens derived from the expression profiles of prognostic markers. (B) CDF curves reflecting the relative stability of the consensus clustering iterations for cluster numbers (k) spanning from 2 to 9. (C) Fluctuations in the area under the CDF curve as k ranges from 2 to 9. (D) PCA scores plot highlighting the distinct spatial segregation and non-overlapping distribution of the three molecular clusters. (E) Kaplan-Meier survival curves evaluating overall survival outcomes across the three clusters. (F) Set intersection analysis depicting the shared differentially expressed transcripts across the three identified molecular clusters. (G) Functional annotation, including GO and KEGG pathways, performed on the overlapping genes. (H) Boxplot illustrating the heterogeneous expression levels of prognostic genes across the three subtypes. (I) Comparison of Stromal, Immune, and ESTIMATE scores, alongside tumor purity, among the subtypes using the ESTIMATE algorithm. (J-L) Evaluation of the immune infiltration landscape and immune-related functions using (J) ssGSEA, (K) CIBERSORT, and (L) MCP-counter algorithms across the identified subtypes. ns, not significant; *, P<0.05; **, P<0.01; ***, P <0.001; ****, P<0.0001. BP, biological process; CC, cellular component; CDF, cumulative distribution function; CIBERSORT, cell-type identification and tracking by estimating relative subsets of RNA transcripts; ESTIMATE, estimation of stromal and immune cells in malignant tumor tissues using expression data; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; MF, molecular function; MCP, Monte Carlo permutation; PCA, principal component analysis; ssGSEA, single-sample gene set enrichment analysis; THCA, thyroid cancer.

Discussion

The rapid expansion of global industrialization has led to the widespread environmental distribution and long-term bioaccumulation of PFAS in the human body (19,28). PFAS have been confirmed to disturb thyroid hormone synthesis and metabolic homeostasis, and exposure risks are potentially linked to the rising incidence of THCA. Nevertheless, the precise functional roles of PFASRGs in driving THCA progression, alongside their modulation of the TME, have not been fully characterized. To address this issue, this study integrated multiple bioinformatics methods to successfully construct a prognostic model based on eight core genes. We not only achieved effective stratification of THCA patients but also evaluated the clinical application potential of these genes as novel biomarkers, thereby providing a new theoretical basis for the precision medicine and individualized intervention of THCA.

We identified a signature of eight genes (THRSP, CIDEC, ALPL, HGF, AQP8, APOE, TF, and MYH7) closely associated with the occurrence and progression of THCA. THRSP (thyroid hormone responsive) encodes a nuclear lipogenic protein sensitive to thyroid hormones, carbohydrates, and insulin, primarily participating in fatty acid synthesis and metabolic regulation (29). In THCA, THRSP expression is correlated with tumor differentiation and lymph node metastasis; its expression levels may reflect the metabolic status of the tumor and possess potential predictive value (30). Cell death inducing DFFA like effector C (CIDEC) is a key protein regulating lipid droplet fusion and lipid storage, playing a vital role in adipocyte differentiation (31). In thyroid tumor models, CIDEC is one of the adipogenic genes regulated by PPFP/PPARγ, and its induction is associated with the adipocyte-like differentiation of tumor cells (32). Alkaline phosphatase, biomineralization associated (ALPL) encodes tissue-nonspecific alkaline phosphatase and is linked to stemness phenotypes and immune infiltration in various malignancies (33). Prior reports indicate that ALPL overexpression correlates with unfavorable clinical outcomes (34), highlighting its viability as a prognostic indicator for THCA—an observation that aligns with our current results. Hepatocyte growth factor (HGF) is a significant mesenchymal-derived growth factor that promotes cell proliferation, migration, and invasion by binding to its receptor, c-met (35). In THCA, PTC is associated with the significant overexpression of the HGF/c-met axis (36). Aquaporin 8 (AQP8) is a membrane-bound water channel protein that, in addition to water permeability, facilitates the transmembrane transport of small molecules such as hydrogen peroxide, thereby modulating cellular redox homeostasis and signaling pathways (37). Research indicates that AQP8 is significantly upregulated in THCA. Furthermore, in gliomas, it has been shown to promote cell proliferation and invasion via the ROS/PTEN/AKT signaling axis, suggesting its potential functional role in tumor biology (38). Apolipoprotein E (APOE) serves as a crucial mediator in lipoprotein metabolism and cholesterol transport, while also modulating immune cell polarization and the TME (39). APOE has been implicated in the tumorigenesis of various cancers, including lung, gastric, and THCA, where it is associated with an elevated risk of metastasis (40-42). Furthermore, bioinformatics investigations across multiple THCA cohorts have demonstrated that APOE expression correlates with immune infiltration, supporting its potential as a prognostic and immune-related biomarker (43,44). TF (transferrin) encodes plasma transferrin, the primary protein responsible for iron transport in the body, which is involved in cellular iron acquisition and the maintenance of systemic iron metabolic homeostasis (45). Recently, within a ferroptosis-related gene model for THCA, TF was identified as a candidate gene associated with prognosis (46). Myosin heavy chain 7 (MYH7) encodes the β-cardiac myosin heavy chain, a central molecule in the contraction of cardiac and slow-twitch skeletal muscle fibers, with its expression being regulated by thyroid hormones (47). While MYH7 is primarily a structural muscle gene rather than a canonical oncogenic driver, its regulation by thyroid hormones—coupled with tumor-associated systemic metabolic and muscular alterations—suggests its potential utility as an indirect biological marker (48). Consequently, these eight feature genes collectively drive the initiation, malignant progression, and metastatic dissemination of THCA by orchestrating key molecular cascades.

Immune cell composition is a core element of the TME, profoundly impacting tumorigenesis, progression, and therapeutic response through complex intercellular communication networks (49). Based on the immune infiltration analysis in this study, we found that the infiltration levels of B cells and M1 macrophages in the low-risk group were significantly lower than those in the high-risk group, suggesting that the high-risk group maintains an immunological state characterized by chronic inflammation. Although B cells are generally associated with better survival outcomes (50) and are typically considered to exert anti-tumor effects by producing antibodies and promoting T cell responses (51), mounting evidence suggests that the enrichment of specific subpopulations (such as regulatory B cells) or B cell activation under chronic inflammatory conditions may inhibit anti-tumor immunity by secreting factors like IL-10, thereby promoting tumor growth (52). This indicates that a simple increase in B cell abundance does not necessarily equate to anti-tumor efficacy. Similarly, although M1 macrophages are traditionally regarded as possessing pro-inflammatory and anti-tumor activities (53), their abundance was paradoxically elevated in our high-risk group. This may reflect a non-specific, persistent inflammatory response induced by PFAS exposure rather than effective anti-tumor killing. Notably, PFAS exposure not only leads to an increase in macrophages (54) but also induces the release of pro-inflammatory cytokines by activating NF-κB and AIM2 inflammasome pathways, leading to tissue damage and chronic inflammation (25). Furthermore, PFAS may interfere with B cell development and antibody secretion, leading to impaired humoral immunity (55). These findings suggest that PFAS may drive the progression of THCA toward a more aggressive, high-risk phenotype by inducing a persistent inflammatory microenvironment and modulating B cells and macrophages. However, whether PFAS directly regulates these immune cells through specific molecular axes remains to be further validated by subsequent in vivo and in vitro experiments.

In the TMB analysis of this study, a compelling finding was that the BRAF mutation frequency in the low-risk group (44%) was slightly higher than that in the high-risk group (35%). This disparity primarily stems from the significant differences in tumor purity between the two strata. In the high-risk group, the presence of intense immune infiltration and chronic inflammation resulted in lower tumor purity, which likely compromised the sensitivity of mutation detection. Conversely, the higher tumor purity in the low-risk group significantly enhanced the signal-to-noise ratio of sequencing data, thereby facilitating the identification of a higher BRAF mutation rate. Furthermore, BRAF and RAS family members typically exhibit robust mutual exclusivity during the progression of THCA (56). In our high-risk group, we observed not only NRAS mutations (5%) but also an enrichment of HRAS (3%) driver variants. This more diverse genomic landscape biologically “diluted” the relative proportion of BRAF mutations within the high-risk cohort. Collectively, the eight-gene prognostic model developed in this study—grounded in PFAS-related transcriptional regulation—is capable of identifying a specific clinical subgroup that is not exclusively driven by the BRAF pathway but exhibits marked aggressiveness through environmental interference, such as PFAS exposure. Therefore, this inconsistency in mutational distribution further underscores the robustness of the PFAS-related risk score as an independent prognostic instrument.

Despite the construction and validation of a robust PFAS-related prognostic model in this study, several limitations warrant acknowledgment. First, our analysis was primarily reliant upon public databases such as TCGA and CellMiner. Although standardized analytical pipelines were implemented, batch effects across different platforms and the intrinsic heterogeneity of sample preparation may still exert potential interference on the precision of gene expression quantification. Second, while cross-validation and normalization strategies were utilized to mitigate the risk of overfitting, the sample size remains relatively modest. Consequently, the generalizability of the model needs further validation in broader, independent large-scale cohorts. Furthermore, although the eight identified genes and their associations with drug sensitivity achieved statistical significance, the underlying molecular mechanisms and pharmacological responsiveness lack direct confirmation through in vitro or in vivo experiments (e.g., cell-based functional assays or animal models). Therefore, the clinical translational potential of the model and the pathophysiological functions of the feature genes remain to be elucidated through future prospective multi-omics research and in-depth functional validation.

Conclusions

In conclusion, by integrating PFAS-related molecular features, this study successfully established and validated a robust prognostic framework for THCA. The model not only facilitates precise risk stratification—categorizing patients into subgroups with significantly divergent survival outcomes—but also systematically elucidates the heterogeneity across core biological pathways, immune landscapes, and pharmacological sensitivities. Overall, our prognostic model demonstrates substantial potential for clinical prognostic assessment and provides a solid theoretical foundation for future mechanistic exploration and targeted therapeutic strategies in THCA.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.