Identification of novel cholesterol metabolism-related biomarkers for thyroid cancer to predict the prognosis, immune infiltration, and drug sensitivity

Abstract

Background

With the sharp increase in the incidence of papillary thyroid carcinoma (PTC), the disease-specific survival rate has not improved significantly. Cholesterol metabolism plays a crucial role in tumor proliferation, regulation of tumor immune escape, and tumor drug resistance. However, there are few studies on the role of cholesterol metabolism in the occurrence and development of thyroid cancer (THCA). This study aimed to investigate the predictive value of cholesterol metabolism-related genes (CMRGs) in THCA and the relationship between immune invasion and drug sensitivity.

Methods

Cholesterol metabolism-related genes we identified from the molecular signatures database, and univariate Cox regression and least absolute shrinkage and selection operator(LASSO) were used to construct a predictive model of cholesterol metabolism-related genes based on the TCGA-THCA dataset. The TCGA dataset was randomly divided into a training group and a validation group to verify the model’s predictive value and independent prognostic effect. We then constructed a nomogram and performed enrichment analysis, immune cell infiltration, and drug sensitivity analysis. Finally, TCGA-THCA and GSE33630 datasets were used to detect the expression of signature genes, which was further verified by the HPA database.

Result

Six CMRGs (FADS1, NPC2, HSD17B7, ACSL4, APOE, HMGCS2) we identified by univariate Cox and LASSO regression to construct a prognostic model for 155 genes related to cholesterol metabolism. Their prognostic value was confirmed in the validation set, and a highly accurate nomogram was constructed combined with clinical features. We found that the mortality rate of high-risk patients increased by 11.41 times, and the infiltration of immune cells in the high-risk group was significantly reduced. Moreover, through drug sensitivity analysis, we obtained sensitive drugs for different risk groups. The GSE33630 dataset verified the expression of six CMRGs, and the HPA database verified the protein expression of the NPC2 gene.

Conclusion

Cholesterol metabolism-related features are a promising biomarker for predicting THCA prognosis and can potentially guide personalized immunization and targeted therapy.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-03483-2.

Introduction

The American Cancer Association predicts that by 2023, there will be approximately 43,720 newly diagnosed thyroid cancer (THCA) in the United States, with approximately 2,120 patients dying [1]. Cohort studies from U.S. cancer registries from 2000 to 2016 showed that the incidence of invasive papillary thyroid carcinoma (PTC) surged faster than that of highly differentiated PTC and anaplastic thyroid carcinoma (ATC) and that the aggressive PTC subgroup also exhibited heterogeneous clinical behavior and a broad risk of death [2]. In addition, an analysis of SEER data from 1992 to 2018 showed no improvement in disease-specific survival in patients with differentiated thyroid cancer (DTC) diagnosed with distant metastasis over the past three decades [3]. With the accumulation of THCA molecular expression level data, survival data, and clinical outcome data, in this study, we attempted to identify high-risk THCA patients more accurately through the combination of THCA key molecular level expression and clinical characteristics and to provide effective targeted therapy in combination with their disease characteristics.

As a vital lipid in mammals, cholesterol is an essential component of cell membranes and regulates the stability and fluidity of cell membranes. The disturbance of cholesterol metabolism plays an important role in the progression of tumors. Tumor cells are usually accompanied by increased cholesterol synthesis and uptake with decreased excretion. Some cancers increase cellular cholesterol concentration through up-regulation of SREBP2 to maintain the need for rapid proliferation of tumor cells [4, 5]. Studies have shown that the expression of low-density lipoprotein receptor (LDLR) is considered to be a risk factor and driver of tumor progression and is involved in tumor deterioration [6, 7]. At the same time, inhibition of intestinal cholesterol absorption or overexpression of specific cholesterol effector transporters can significantly inhibit the tumor proliferation of various malignant tumors [8]. In addition, cholesterol metabolism also plays an important role in the regulation of tumor immune escape, tumor drug resistance, autophagy, and ferroptosis. Studies have found that T cells in the tumor microenvironment (TME) lack cholesterol, while bone marrow cells and tumor cells that play an immunosuppressive role are rich in cholesterol [9]. This is because oxysterols such as 27-HC secreted by tumor cells significantly inhibit SREBP2 and activate LXR, thereby leading to T cell cholesterol deficiency and driving T cell depletion and dysfunction, especially for cytotoxic T cells [9]. Other studies have shown that SULT2B1b derivatives promote the activation and differentiation of monocyte-derived cells by interfering with LXR signals, thus delaying tumor growth [10]. Moreover, cholesterol-rich TME can also increase the expression of immune checkpoints [11]. Studies have shown that inhibition of PCSK9 combined with anti-PD1 immunotherapy can enhance the anti-tumor effect of some malignant tumors [12]. Given previous studies, we realize that the disturbance of cholesterol metabolism plays a vital role in tumor progression and treatment, but there are few studies on the role of cholesterol metabolism in the occurrence and development of THCA. Therefore, we hypothesized that changes in the expression levels of genes related to cholesterol metabolism might predict the risk of progression of THCA, which may be conducive to the identification of high-risk THCA, the discovery of new molecular targets, and the reference for subsequent treatment options for high-risk THCA.

In this study, we obtained information on cholesterol metabolism-related genes (CMRGs) from the molecular signatures database (Msigdb) using the THCA cohort in TCGA. CMRGs associated with prognosis were identified to construct a prognostic risk model, and the feasibility of the model was validated and evaluated. The risk model was verified as an independent prognostic factor for THCA, and the prognostic value of the nomogram was constructed and evaluated to increase its clinical application. We also performed a series of enrichment analyses, immune infiltration, and drug sensitivity analyses. In addition, we validated the expression of CMRGs using independent datasets and HPA databases. These results are promising to inform risk assessment and targeted therapies for THCA patients.

Methods

Data sources and pre-processing

Our comprehensive database of GEO (https://www.ncbi.nlm.nih.gov/geo/) to downloaded from the gene expression of thyroid cancer gene expression profile datasets. The GSE33630 dataset (n = 105), using the data platform GPL570, contained 11 samples of undifferentiated thyroid cancer, 49 samples of papillary thyroid cancer, and 45 samples of normal thyroid. We used the GEO platform annotation to match the probe ID with the gene symbol. The average expression value was taken for probes located in the same gene, and R software (version 4.2.1) was used for batch normalization. Regarding the expression of critical genes in the single-cell transcriptome of thyroid cancer, We use the Cancer Single-cell Expression Map online database (https://ngdc.cncb.ac.cn/cancerscem/index) to analyze the GSE148673 data set, The GSE148673 dataset included five samples of undifferentiated thyroid cancer. We use the cancer genome atlas (TCGA) database (https://portal.gdc.cancer.gov) to download THCA gene expression in patients with the spectrum and corresponding clinical information. The data type counts and fragments per kilobase exon model per million mapped fragments are selected and converted into transcript per kilobase exon model per million mapped reads (TPM) format. Excluding patients with incomplete data, the data included 56 normal samples and 497 tumor samples, of which 492 were definitively diagnosed with papillary thyroid cancer. When constructing the prognostic model and validating the model, we used the caret package in R software (version 4.2.1) to randomly divide the patients into a training set (n = 348) and a test set (n = 149) at a ratio of 7:3, without stratified data splitting. The clinical characteristics of all THCA patients were shown in Table 1. There was no statistically significant difference in clinical data between the two groups. We from MSigdb (http://www.gsea-msigdb.org/gsea/msigdb/collections.jsp) downloaded the genes associated with cholesterol metabolism, filtering the repetition and non-coding genes after receiving 93 [13]. A part of the IHC data was downloaded from the HPA database (https://www.proteinatlas.org/).

Table 1.

Baseline characteristics

	level	Overall	Training set	Test set	p
n		497	348	149
Age		46	46	48	0.713
(median [IQR])		[35.00, 58.00]	[36.00, 57.25]	[34.00, 59.00]
gender (%)	female	363 (73.0)	248 (71.3)	115 (77.2)	0.211
	male	134 (27.0)	100 (28.7)	34 (22.8)
T (%)	T1	142 (28.6)	103 (29.6)	39 (26.2)	0.746
	T2	162 (32.6)	109 (31.3)	53 (35.6)
	T3	169 (34.0)	118 (33.9)	51 (34.2)
	T4	22 ( 4.4)	16 ( 4.6)	6 ( 4.0)
	TX	2 ( 0.4)	2 ( 0.6)	0 ( 0.0)
N (%)	N0	227 (45.7)	166 (47.7)	61 (40.9)	0.222
	N1	221 (44.5)	152 (43.7)	69 (46.3)
	NX	49 ( 9.9)	30 ( 8.6)	19 (12.8)
M (%)	M0	282 (56.7)	197 (56.6)	85 (57.0)	0.566
	M1	8 ( 1.6)	4 ( 1.1)	4 ( 2.7)
	MX	206 (41.4)	146 (42.0)	60 (40.3)
	Unknow	1 ( 0.2)	1 ( 0.3)	0 ( 0.0)
stage (%)	I	279 (56.1)	199 (57.2)	80 (53.7)	0.766
	II	52 (10.5)	38 (10.9)	14 ( 9.4)
	III	111 (22.3)	76 (21.8)	35 (23.5)
	IV	53 (10.7)	34 ( 9.8)	19 (12.8)
	Unknow	2 ( 0.4)	1 ( 0.3)	1 ( 0.7)
PFI (%)	0	446 (89.7)	313 (89.9)	133 (89.3)	0.946
	1	51 (10.3)	35 (10.1)	16 (10.7)

Differential expression analysis

From TCGA, we divided the expression data into a normal group (01 A) and a tumor group (11 A). The R software Deseq2 [14] package was used to analyze the difference in prognostic-related cholesterol metabolism genes in different groups. The threshold value of DEGs was | Log2FC|>1 and FDR < 0.05. After building a predictive model, patients in the tumor group were divided into high-risk and low-risk groups based on risk scores. The Deseq2 package of R software was also used to analyze the difference in gene expression data of different groups. The threshold of DEGs was FDR < 0.05, where log2FC > 1 and log2FC < -1 indicated up-regulation and down-regulation of genes in high-risk groups, respectively. The final differences are shown in volcanic maps.

Risk model construction

The THCA training set in the TCGA cohort was used to construct the prognostic model. Univariate Cox regression analysis was then performed to calculate the difference in prognosis, with HR < 1 or > 1 as the protective or risk gene (p < 0.1), and the results were shown in the forest map. To avoid overfitting, we performed minimum absolute contraction and selection operator (LASSO) regression analysis using the “glmnet” software package [15] to identify signature genes. The optimal value of model development and the penalty coefficient λ of the gene were determined by 10x cross-validation of 1000 iterations. The penalty coefficient λ was identified based on a minimum criteria + 1SE rule. The risk score is calculated by the following formula: Risk score = ∑βi×Expi, βi is the Cox hazard ratio coefficient of mRNA, and Expi is the expression level of a gene. THCA patients were classified into low-risk and high-risk subgroups according to the optimal cut-off point. We used R software to create dot, line, and heat maps to show the relationship between the prognostic model and gene expression in the model. K-M survival analysis was used to evaluate the prognosis of low - and high-risk THCA patients. Receiver operating characteristic (ROC) curves over 1, 3, and 5 years were plotted against OS, and AUC values were calculated to evaluate the accuracy of the prognostic model. At the same time, the validity and robustness of the prognostic risk model were verified on the test data set and the whole TCGA data set. In addition, we performed a multifactor regression analysis of the entire TCGA dataset (p < 0.05) to assess whether the risk score obtained from the CMAGs prognostic model could be used as an independent predictor.

Nomogram

To further improve the clinical value, we constructed a nomogram based on age, clinical stage, and related characteristics of cholesterol metabolism and predicted the survival rate of patients at 1, 3, and 5 years through the R package “rms” [16]. ROC analysis was used to evaluate the differential performance of each factor in THCA prognosis. Calibration curves are used to compare expected and actual survival rates.

Signature gene expression, mutation analysis, and protein-protein interaction networks

We used the GSE33630 dataset and the whole dataset of the TCGA-THCA cohort to conduct the Wilcoxon.test by R software to analyze the expression differences of signature genes in thyroid cancer and normal thyroid samples and expressed them by boxplot (* <0.05, ** <0.01, *** <0.001, **** <0.0001). GSE33630 for ATC samples containing, so we will split the dataset to analyze the difference between PTC and normal thyroid samples and the analysis of the difference between ATC and normal thyroid samples, the results also using boxplot said (* <0.05, ** <0.01, *** <0.001, **** <0.0001). To define the signature gene expression level in thyroid carcinoma and the relationship between the gene mutation, we sign by analyzing the cBioportal database (https://www.cbioportal.org/) the level of mutations in the gene in thyroid carcinoma. To clarify the Expression of some signature genes in the thyroid cancer single-cell transcriptome, we used the Cancer Single-cell Expression Map database to analyze ATC samples from the GSE148673 dataset. In addition, the GeneMANIA database (https://genemania.org/) was used to predict protein function and identify proteins that interact with signature genes.

Enrichment analysis and PPI

Through the “clusterProfler” package [17], we conducted Kyoto Encyclopedia Pathway enrichment analysis (KEGG) [18] and gene ontological functional annotation (GO) for differential genes in high-risk and low-risk groups. p < 0.05 was considered statistically significant. Obtained from our MsigDB [13]”c2.cp.kegg_legacy.v2023.2.Hs.symbols.gmt” and “h.all.v2023.2.Hs.symbols.gmt” subset, Gene collection enrichment analysis (GSEA) was performed using the clusterProfler package [19] to evaluate the enrichment differences of pathways and feature functions between the two groups, and P < 0.05 was considered statistically significant. “enrichplot” package are used to create GSEA maps that show the significantly enriched pathways and biosignatures of high-risk groups obtained through GSEA analysis. In addition, we performed a single sample gene collection enrichment analysis (GSVA) for gene expression in the high-risk and low-risk groups using the GSVA package [20]. In “c2.cp.kegg_legacy.v2023.2.Hs.symbols.gmt” GSVA analysis Settings in the subset of P < 0.05, | log2FC | < 0.25, In the “h.all.v2023.2.Hs.symbols.gmt” GSVA analysis Settings in the subset of P < 0.01, | log2FC | < 0.15, as a result, with the heatmap displays. We also analyzed protein-protein interactions (PPI) of the top 200 DEGs significantly upregulated in the high-risk group using the String database (https://string-db.org/, accessed March 10, 2024) [21]. The interaction score was set to 0.4. According to the analysis results, the PPI map was drawn by Cytoscape software (v3.10.1) [22], and Hub genes were searched by the MCC algorithm of the cytoHubba plug-in.

Immunotherapy analysis

Using R package “IOBR” [23], through single-sample gene set enrichment analysis(ssGSEA) methods to analyze related gene activation of tumor microenvironment (https://github.com/IOBR/IOBR) [24]. cibersort [25], epic [26], mcpcounter [27], and estimate [28] were used to estimate the relationship between risk score and immune cells based on the expression profiles of samples in the TCGA-THCA cohort.

Drug sensitivity analysis

The Cancer Therapeutics Response Portal (CTRP, https://portals.broadinstitute.org/ctrp/) is about The Cancer drug sensitivity and drug reaction of molecular marker information public database. Cancer drug response data and genomic sensitivity markers we identified from the CTRP database. We used the pRRophetic algorithm [29] to predict the sensitivity of patients in different clinical variable groups to common anticancer drugs or small molecular compounds. Wilcoxon signed-rank test was used to compare the semi-inhibitory concentration values of low and high-risk THCA, and the statistically significant cut was set as P < 0.05.

Statistical analyses

All data calculation and statistical analysis were performed using R software (https://www.r-project.org/, version 4.2.1). The variance between the non-normally distributed variables was analyzed using the Wilcoxon rank sum test to compare two sets of continuous variables. Univariate and multifactorial Cox analyses were used to identify independent prognostic factors. The “Survival” package in R was used for survival analysis, the Kaplan-Meier survival curve was used to represent survival differences, and the log-rank test was used to evaluate the significance of survival time differences between the two groups. The ROC curve was plotted using the “pROC” package of R software, and the AUC was calculated to assess the accuracy of the ROC curve in distinguishing between cancer and adjacent tissue. The Pearson correlation analysis method was used to calculate the correlation coefficient between genes and risk scores. All statistical P-values were bilateral, and P < 0.05 was considered a statistically significant difference.

Results

Technology roadmap, CMRGs in THCA

The flow chart of this study is shown in Fig. 1. We obtained the thyroid expression data of 553 cases from TCGA. We randomly split the cohort into a training set and a validation set at 7:3. Statistical analysis was conducted on the subtypes of the training set, validation set, and the whole data set. There was no statistical difference in the characteristics of each subtype in each group’s baseline (Table 1). In the Msigdb database, we obtained 155 CMRGs involved in cholesterol metabolism and finally obtained 93 CMRGs by removing the replication and filtering out the non-coding RNA. DEseq2 package was used to analyze the differentially expressed CMRGs in tumor samples compared with the normal group. FDR < 0.05, |log2FC| >1 was set as the screening criteria. A total of 11 up-regulated genes and six down-regulated genes we identified. A volcano map was drawn, showing the five genes with the highest and lowest expression levels (Fig. 2A). We performed univariate COX regression analysis on CMRGs genes in the entire dataset (p < 0.1), and obtained ten genes that were significantly associated with prognosis (Fig. 2B), of which 6 were prognostic risk factors (HR > 1) and the remaining 4 were prognostic protective factors (HR < 1). Subsequently, prognostic CMRGs were used to build prognostic models through LASSO regression analysis in the training set, and the models were validated both in the validation set and the entire dataset. Subsequently, multivariate COX regression showed that risk score played an independent prognostic role among clinical factors. The nomogram shows the model more intuitively and improves its clinical application. We analyzed the expression of the CMRGs gene in the model compared with the normal group in the GSE33630 dataset. Also, we analyzed the mutation level of the CMRGs gene, PPI, and its expression level in single cells of thyroid cancer. Further, we divided tumor samples into high-risk and low-risk groups by risk score. We analyzed differential genes, functional enrichment, immune infiltration, tumor metabolic characteristics, and drug sensitivity between the two groups. Finally, immunohistochemical data from the HPA database were used to verify the protein expression levels of NPC2 in thyroid cancer.

Fig. 1.

Flow diagram of the study

Fig. 2.

Identification of genes related to cholesterol metabolism in THCA patients. A Volcano maps showed differential expression of 93 genes related to cholesterol metabolism in normal and THCA tumor tissues. Red dots represent genes up-regulated in tumor tissue, and blue dots represent down-regulated genes in tumor tissue. B Univariate Cox regression analysis showed that ten genes related to cholesterol metabolism were significantly correlated with THCA prognosis. C and D Multivariate Cox regression and Lasso analysis of 93 cholesterol metabolism-related genes. E Signature genes and risk coefficients of prognostic models obtained by Lasso analysis

Prognostic model construction and validation

Based on the expression data of the training group, 10 CMRGs associated with prognosis were analyzed by LASSO regression using univariate COX regression. Finally, six CMRGs (FADS1, NPC2, HSD17B7, ACSL4, APOE, and HMGCS2) were selected to construct the risk model related to cholesterol metabolism (Fig. 2C and D). The list of 6 risk genes and their calculated coefficients are shown in Fig. 2E, and the C index is 0.849. Based on the ranking of risk scores from high to low and survival analysis, we selected the top 30% risk score to divide THCA patients in the training set into a high-risk group (n = 104) and a low-risk group (n = 243) (Fig. 3A). According to survival analysis, the mortality rate of the high-risk group was 11.41 times that of the low-risk group (HR 11.41, 95%CI 2.46–52.90, p < 0.001, Fig. 3B). Risk models predicting 1 -, 3 -, and 5-year OS had AUC values of 0.987, 0.811, and 0.817, respectively (Fig. 3C). Based on the cholesterol metabolism-related risk model, risk scores for each patient in the validation set and the entire TCGA-THCA dataset were calculated to verify the prognostic value of the model. In the validation set and the whole data set, patients in the top 30% of the risk score were also taken as the high-risk group (n = 45, n = 149), and the rest were taken as the low-risk group (n = 104, n = 347) (Fig. 3D and G). According to the survival curve, the survival probability of patients in the high-risk group of the validation set and the whole data set was significantly reduced (HR 9.41, 95%CI 1.05–84.24, p < 0.014; HR 10.81, 95%CI 3.08–37.93, p < 0.001; Fig. 3E and H), and the results were consistent with the training set. The validation set predicted the AUC values of the 1-year, 3-year, and 5-year OS risk models at 0.763, 0.815, and 0.844, respectively (Fig. 3F). The entire data set predicted the AUC values of the 1-year, 3-year, and 5-year OS risk models at 0.854, 0.813, and 0.803, respectively (Fig. 3I). We ranked the risk scores of patients in the three groups and analyzed their distribution; the number of deaths increased with the increase in the risk score, and the heat map described the expression patterns of the six risk genes involved in the risk model in the different risk groups (Fig. 3A, D and G). Our validation results demonstrate the applicability and high stability of the cholesterol metabolism-related prognostic model in predicting the prognosis of patients with THCA. To assess whether risk scores and clinical features are independent prognostic factors for patients with PTC, we used a multifactor Cox analysis to analyze the entire TCGA dataset. The results showed that after adjusting for other covariates, risk score remained an independent predictor of prognosis (P < 0.009, HR = 2.2,95% CI = 1.2–3.9, Fig. 4A).

Fig. 3.

Evaluation of prognostic model in the training set, validation set, and whole group. A, B and C Presentation of the prognostic model in the training set (scatter plot, heat map), Kaplan-Meier survival curve, and ROC curve. D, E and F The prognosis model was presented in the validation set model (scatter plot, heat map), Kaplan-Meier survival curve, and ROC curve. G, H and I prognosis model in the whole TCGA-THCA model display (scatter plot, heat map), Kaplan-Meier survival curve, and ROC curve

Fig. 4.

Prognostic value of a predictive risk model for six metabolism-related genes in THCA. A Multivariate Cox regression analysis of risk model score, age, and tumor stage for prognostic value. B A clinical prognostic nomogram can predict 1, 3, and 5-year survival. C Time-dependent ROC curve analysis for predicting 1, 3, and 5-year OS based on risk and clinical features. Clinical features: age and tumor stage

Nomogram

Using clinical data from the TCGA dataset, we created a nomogram based on risk score, clinical stage, and age with an overall C-index of 0.953 that can be used to predict 1 -, 3 -, and 5-year survival (Fig. 4B). Calibration curves show that the nomogram accurately predicts survival (S1 Fig). In addition, ROC analysis showed that the AUC (0.974, 0.953, 0.940) of the 1, 3, and 5-year survival rate predicted by the nomogram was higher than that of other factors (risk score 0.854, 0.813, 0.803; clinical stage 0.841, 0.706, 0.775; Age 0.946, 0.926, 0.928; Fig. 4C). These results show that the nomogram model performs well regarding robustness and validity.

Expression, mutation level, and protein interaction in CMRGs

To further clarify the differential expression of 6 CMRGs in thyroid cancer, two independent datasets, TCGA-THCA and GSE33630, were used for expression analysis. The results showed that APOE, HSD17B7, and NPC2 in the TCGA cohort were significantly up-regulated, HMGCS2 was down-regulated considerably, and there was no statistically significant difference in ACSL4 and FADS1 (Fig. 5A). The GSE33630 dataset (n = 105) contained 49 papillary thyroid carcinoma samples, 11 undifferentiated thyroid carcinoma samples, and 45 normal thyroid samples. Expression analysis was performed using the entire GSE33630 dataset, papillary carcinoma and normal thyroid samples, and undifferentiated carcinoma and normal samples, respectively (Fig. 5B and D). In the overall dataset, we observed that APOE, HSD17B7, NPC2, ACSL4, and FADS1 were significantly up-regulated, and HMGCS2 was down-regulated (Fig. 5B). In the papillary carcinoma and normal thyroid samples, the differential expression was consistent with the whole dataset, but the difference changes in HSD17B7 and ACSL4 were reduced compared with the entire dataset (Fig. 5C). In the samples of undifferentiated cancer and normal thyroid, APOE, ACSL4, and FADS1 were significantly up-regulated, the difference changes of APOE and HSD17B7 were decreased compared with the difference changes in the whole dataset, the difference changes of ACSL4 were increased compared with the difference changes in the entire dataset, and HSD17B7 was slightly up-regulated. The differences between NPC2 and HMGCS2 compared with the normal group were not statistically significant (Fig. 5D). Combined with the prognostic effects of 6 CMRGs in thyroid cancer, we found that the changing trend of their expression levels in papillary thyroid cancer and undifferentiated thyroid cancer was consistent with their prognostic effects.

Fig. 5.

Expression, mutation level, and protein interaction of CMRGs. A and B Based on the TCGA-THCA and GSE33630 data sets, respectively, the differential expression of 6 genes related to cholesterol metabolism in normal and THCA tumor tissues. C: Differential expression of six cholesterol metabolism-related genes in normal and PTC tumor tissues based on the GSE33630 dataset. D Differential expression of six genes related to cholesterol metabolism in normal and ATC tumor tissues based on the GSE33630 dataset. E TCGA-THCA mutation levels of 6 genes related to cholesterol metabolism. F PPI analysis of 6 genes related to cholesterol metabolism. G, H and I indicated the APOE, NPC2, and FADS1 expression levels in ATC single cells

We analyzed and visualized the genomic changes of 6 CMRGs in the TCGA-THCA cohort through the cBioportal database, as well as the corresponding TMB, BRAF mutations, and survival status information (Fig. 5E), and found that their genomic changes were not noticeable. The change in transcriptome expression level may be related to the epigenetic modification of genes or the transcription process. Since the differential expression of APOE and NPC2 was the most significant, and FADS1 had the most significant relationship with prognosis, APOE, NPC2, and FADS1 were selected for further analysis. We analyzed the GSE148673 dataset using the Cancer Single-cell Expression Map online database. NPC2 was significantly expressed in almost all annotated immune cells, especially in dendritic cells, monocytes, macrophages, CD8 + T cells, and fibroblasts (Fig. 5G). APOE is mainly expressed in macrophages, dendritic cells, and fibroblasts (Fig. 5H). The association between APOE and NPC2 and immune activity in thyroid cancer was partially verified. FADS1 was significantly expressed mainly in fibroblasts (Fig. 5I).

In addition, we conducted protein-protein interaction (PPI) analysis using the GeneMANIA website and found that the most co-expressed and co-localized proteins with FADS1 were found, followed by HSD17B7 (Fig. 5F). NPC2 has a gene regulatory relationship with ELOVL6 and CYP51A1, and APOE has a regulatory relationship with HMGCS2 (Fig. 5F). NPC2 is enriched in the isoprenoid metabolic and fatty acid derivative biosynthetic processes. APOE is enriched in the cholesterol metabolic process, isoprenoid metabolic process, and fatty acid derivative biosynthetic process. FADS1 is enriched in oxidoreductase activity, fatty acid metabolic process, and fatty acid derivative biosynthetic process.

Enrichment analysis

According to the Lasso-Cox regression method, all THCA patients were divided into a high-risk group and a low-risk group according to 3:7. We analyzed differentially expressed genes (FDR < 0.05, |log2FC| < 1) in the two groups by DEseq2, and showed them by volcano map, in which 1258 genes were up-regulated, and 682 genes were down-regulated (Fig. 6A). The five genes with the most significant down-regulated and up-regulated were labeled in the figure, respectively. GO and KEGG analyses were performed to understand which biological functions and signaling pathways were enriched in the high-risk group compared to the low-risk group (P < 0.05, Fig. 6B and C). The results showed that GO analysis was mainly concentrated in potassium ion transport and organic hydroxy compound transport at the BP level. CC layer is enriched in the synaptic membrane, ion channel complex, collagen-containing extracellular matrix, neuronal cell body, and immunoglobulin complex. The MF layer is mainly enriched in passive transmembrane transporter activity, ion channel activity, gated channel activity, and metal ion transmembrane transporter activity (Fig. 6B). KEGG is mainly enriched in neuroactive ligand-receptor interaction and calcium signaling pathways (Fig. 6C). In addition, we performed GSEA enrichment analysis on the KEGG pathway set and the hallmark gene set (P < 0.05), respectively, as shown in the ridge map (Fig. 6D and E), and showed the significant enrichment pathway and hallmark in the high-risk group in the GSEA enrichment results as shown in the gseaplot plot (Fig. 6F and G). We observed that the signal pathway GSEA analysis in the high-risk group was mainly concentrated on focal adhesion, valine leucine and isoleucine degradation, lysine degradation, and propanoate metabolism (Fig. 6D). Hallmark GSEA analysis in the high-risk group was mainly concentrated on kras signaling, pancreas beta cells, myogenesis, oxidative phosphorylation, and hypoxia (Fig. 6E). Next, we performed GSVA enrichment analysis for a single sample pathway (P < 0.01, |log2FC| < 0.15) and hallmark set (P < 0.05, |log2FC| < 0.25), and the results were shown by heat map (S2 and S3 Fig). Regarding the signaling pathway, GSVA enrichment analysis verified that valine leucine and isoleucine degradation, lysine degradation, and propanoate metabolism are significantly enriched at high risk. Regarding the hallmark set, GSVA enrichment analysis verified that oxidative phosphorylation is significantly enriched at high risk (S2 and S3 Fig). Finally, to clarify which genes play an important role in the high-risk risk group, the top 200 up-regulated genes in the differentially expressed genes between the high-risk and low-risk groups were selected for PPI analysis. The top 10 hub genes were found through Cytoscape software. PPARG, CD36, FABP4, BCL2, ACACB, PPARGC1A, PRKCA, KIT, SLC5A5, IYD (Fig. 6H and I).

Fig. 6.

Differentially expressed genes, enrichment analysis of GO, KEGG, GSEA, and PPI in high and low-risk groups. A Volcano maps showed differentially expressed genes in high and low-risk groups. Red dots represent genes up-regulated in tumor tissue, and blue dots represent down-regulated genes in tumor tissue. B and C GO function and KEGG enrichment analysis of differentially expressed genes in high and low-risk groups. D and E GSEA enrichment analysis was performed for differentially expressed genes in high and low-risk groups based on KEGG pathway sets and hallmark gene sets, respectively. F and G represent significantly up-regulated pathways and hallmarks in the high-risk group, respectively. H and I PPI and Hub gene analysis of differentially expressed genes in high and low-risk groups

Immune-infltration analysis

We used the ssGSEA algorithm to analyze the immune microenvironment characteristics of the high-risk and low-risk groups. We found that other immune cells, except B cells and APC, significantly decreased in a high-risk group (Fig. 7A). Genes associated with tumor-associated macrophages and immune checkpoint inhibitors differed most significantly between the two groups (Fig. 7B and C). We can see that the expression of genes related to immune checkpoint inhibitors is reduced considerably in the high-risk group, suggesting that THCA patients in the high-risk group have a worse immunotherapy response and poor prognosis than those in the low-risk group (Fig. 7C). We further selected the five genes most differentially expressed in tumor-associated macrophages between the two groups and analyzed the correlation between their expression and risk score. Results It was observed that FABP4, PCED1B, and RAMP1 were significantly negatively correlated with risk scores, while IPCEF1 and PARM1 were significantly positively correlated with risk scores (Fig. 7D and H). These results suggest that risk score is closely related to the degree of immune infiltration.

Fig. 7.

The immune microenvironment characteristics of high-risk and low-risk groups were analyzed by ssGSEA algorithm. A Differences in the immune microenvironment between high and low-risk groups. B Differential expression of genes associated with immune checkpoint inhibitors in high-low risk groups. C Differential expression of tumor-associated macrophage gene sets in high-low risk groups. D-H Correlation between the five genes most significantly differentially expressed in tumor-associated macrophages between the two groups and risk scores

To further verify the difference in immune infiltration between the two groups, the Estimate, EPIC, CIBERSORT, and MCPcounter algorithms were used for analysis (Fig. 8A and D). Estimations showed significant differences between the two groups, with patients in the high-risk group having fewer stroma components, lower immune infiltration, and higher tumor purity, suggesting a significantly poorer prognosis and immunotherapy response in the high-risk group than in the low-risk group (Fig. 8A). EPIC analysis showed that immune cell infiltration decreased significantly in high-risk groups, with macrophages, CD8 + T cells, and NK cells being the most significant (Fig. 8B). CIBERSORT also verified that there was less infiltration of immune cells in the high-risk group, among which T cells, NK cells, monocytes, dendritic cells, mast cells, and eosinophils showed the most apparent differences between the two groups (Fig. 8C). MCPcounter analysis showed that the expression of T cells, CD8 + T cells, cytotoxic T cells, myeloid dendritic cells, B lienage, and Myeloid dendritic cells at high risk decreased. In contrast, the expression of endothelial cells increased. Among them, T cells, CD8 + T cells, and cytotoxic T cells showed the most significant differences in infiltration between the two groups (Fig. 8D). The above analysis further verified that the risk score constructed according to cholesterol metabolism was closely related to the degree of immune cell infiltration.

Fig. 8.

A-D Indicates that the Estimate, EPIC, CIBERSORT, and MCPcounter algorithms analyze the difference in immune cell infiltration between high-risk and low-risk groups, respectively

Drug sensitivity verification

Targeted drug therapy is another essential tumor therapy in addition to immunotherapy. To explore the relationship between drug sensitivity and high and low-risk groups, we analyzed the therapeutic sensitivity of various anticancer drugs in the CTRP database for THCA patients. The results showed that the sensitivity of the high-risk group to BRD-K09587429, BRD-K16147474, BRD-K52037352, IPR-456, PD318088, and pevonedistat(MLN4924) was significantly higher than that of the low-risk group (p < 0.05, Fig. 9A and F). Conversely, patients in the low-risk group, including sirolimus and BRD-K45681478, were more sensitive than those in the high-risk group (p < 0.05, Fig. 9G and H). These findings can potentially provide insights into future targeted therapies and resistance prevention for THCA patients.

Fig. 9.

Drug sensitivity analysis and immunohistochemistry. A-H Drug sensitivity analysis in high-risk and low-risk groups. I and J The protein expression level of NPC2 in normal thyroid and PTC tissue was detected by CAB032888 antibody, respectively. K and L The protein expression level of NPC2 in normal thyroid and PTC tissue was detected by HPA000835 antibody, respectively

Validation of key genes by IHC from HPA

In addition, we obtained the IHC protein expression results of NPC2 in normal thyroid tissue and papillary thyroid carcinoma tissue from the HPA database, and the staining results of antibody CAB032888 and HPA000835 were shown in the figure (Fig. 9I and L). Moderately positive expression of NPC2 in normal thyroid tissue and strong positive expression in papillary carcinoma further validated our previous analysis.

Discussion

In the last 30 years, the incidence of thyroid cancer has increased dramatically [30]. Studies have shown that the incidence of aggressive PTC has increased significantly, and the surge rate exceeds that of highly differentiated PTC and ATC [2]. Compared with the mortality rate of most other cancers, the mortality rate of thyroid cancer has not been significantly improved in recent years [3]. Therefore, early identification of high-risk patients and appropriate intervention is essential. Thyroid cancer is closely related to obesity and metabolic disorders [31, 32]. In recent years, more and more studies have been conducted on the role of cholesterol metabolism in tumorigenesis and progression. Since cholesterol and its metabolites play an essential role in constructing cell membranes and maintaining their stability and fluidity, and the cell membrane is a critical pathway in signal transmission, many studies have confirmed that the disturbance of cholesterol metabolism promotes tumor proliferation [4, 5]. In addition, studies have found that the cholesterol content of tumor cells in the tumor microenvironment increases. Still, the cholesterol content of anti-tumor immune cells decreases, thus promoting the occurrence of tumor immune escape and significantly reducing the sensitivity of immunotherapy [9]. In addition, cholesterol metabolism also plays a vital role in tumor-related iron death and autophagy. Therefore, it is significant to study how CMRGs affect the clinical prognosis, immune infiltration, and drug sensitivity of THCA. To our knowledge, we are the first to examine the relationship between CMRGs and THCA prognosis. In this study, we conducted univariate COX regression analysis on 155 CMRGs and obtained 10 CMRGs associated with prognosis. Further, through LASSO regression analysis, We constructed a prognostic risk model based on six genes (FADS1, NPC2, HSD17B7, ACSL4, APOE, and HMGCS2) and verified the model’s reliability. Multivariate COX regression shows that our risk score has independent predictive value. We also constructed a nomogram to improve clinical application. The results showed that the AUC of 1 -, 3 - and 5-year survival rates predicted by the nomogram were 0.974, 0.953, and 0.940, respectively, suggesting that it had excellent prognostic value. We also compared the enrichment pathway, immune infiltration, and drug sensitivity between the high-risk and low-risk groups based on the risk model. The expression of CMRGs in the risk model was validated in the TCGA-THCA cohort and GEO dataset, and the expression of CMRGs with the most significant difference and the most prognostic correlation in single cells of thyroid cancer was also evaluated. Finally, we obtained IHC results from the HPA database to verify the protein expression of NPC2 in thyroid cancer tissues. In conclusion, we constructed a prognostic risk model based on CMRGs. We identified potential molecular targets associated with prognosis, which is accurate in identifying high-risk thyroid cancer patients and provides a reference for subsequent drug treatment selection. Moreover, this prognostic risk model has certain potential and value for clinical application.

NPC2 is a protein-coding gene associated with the innate immune system and plasma lipoproteins’ assembly, remodeling, and clearance. Regarding the participation of NPC2 in cholesterol metabolism, studies have found that SREBP-1 is activated during cholesterol consumption, which can upregulate critical autophagy and NPC2, promote the release of cholesterol from lysosomes, and thus promote tumor growth [33, 34]. On the contrary, studies have shown that independent of cholesterol metabolism, downregulation of NPC2 can activate MAPK/ERK and is associated with poor clinicopathological features. NPC2 is also involved in the induction and maintenance of autophagy, thus playing a protective factor in tumor growth [35]. In addition, NPCs are significantly correlated with immune T cells such as Th1, Th2, CD4_T, and CD8_T, and NPC2 secreted by tumors can inhibit the recruitment of immature macrophages to the tumor microenvironment and play a role in inhibiting tumor growth [36]. Our research also observed that the reduction of NPC2 is associated with poor prognosis, and the immune cells in the high-risk group are significantly downregulated. Single-cell analysis also reveals that NPC2 is expressed at higher levels in immune cells, including dendritic cells, macrophages, and monocytes. This suggests that NPC2 may play an anti-tumor growth role in the tumor growth process by participating in immune responses [37–39]. In the future, more detailed basic studies are needed to clarify the mechanism of action of NPC2 in thyroid cancer diseases. This may provide information for more accurate targeted therapy to facilitate better prognosis for thyroid cancer patients.

APOE is an apolipoprotein involved in lipoproteins’ production, transformation, and clearance [40]. Our study found that APOE is highly expressed in thyroid cancer. In thyroid cancer, qPCR and WB have been used to verify that APOE is significantly highly expressed in thyroid cancer tissues [41], and other studies have suggested that FTO inhibits glycolysis and PTC growth by reducing the stability of APOE mRNA in a N6-methyladenosine-dependent manner [42]. In other cancers, there are also many studies on the mechanism of APOE up-regulation promoting cancer progression [43–46]. For example, studies have found that in gastric cancer cells, exosomes derived from M2 macrophages mediate the transfer of functional APOE protein from TAMs to tumor cells and then activate the PI3K-Akt signaling pathway to promote the migration of gastric cancer cells [43]. Other studies have reported that APOE is also significantly increased in peripheral blood mononuclear cells of patients with pancreatic ductal adenocarcinoma, and APOE can induce the expression of immunosuppressive factors CXCL1 and CXCL5 through LDLR and NF-KB signaling pathway, promoting tumor progression [44]. In lung cancer cells, down-regulating APOE expression can inhibit the growth and metastasis of cancer cells by increasing the antitumor activity of TREM-1-dependent NK cells [45]. APOE promotes glioma progression by mediating the secretion of interleukin IL-1b by the NLRP1 inflammasome [46]. On the other hand, APOE has also been shown to play a protective factor in tumor progression, which may be related to inflammation, immunity, different genotypes, tissue-specific mechanisms, etc. For example, APOE-deficient mice are more responsive to inflammatory stimuli and are associated with increased sensitivity to developing inflammation-related colorectal cancer [47]. Immunologically, APOE activation induces a robust antitumor response and increases T-cell activation [48]. The function of APOE is also related to genotype. Studies have found that compared with APOE2 mice, APOE4 mice showed enhanced antitumor immune activation, and mice expressing the human APOE4 allele showed reduced melanoma progression and metastasis and improved survival rate [49]. Another study found that APOE ε2/ε4 and ε3/ε4 play a protective role in the development of laryngeal squamous cell carcinoma, while APOE ε3/ε3 May promote the growth of laryngeal squamous cell carcinoma [50]. Our study found that immune cell infiltration was significantly reduced in the high-risk group, and the expression of APOE was related to immune cell infiltration. In addition, other studies have confirmed that APOE levels are significantly associated with B cells, CD8 + T cells, neutrophils, and dendritic cells in PTC, as well as immune biomarkers. In the future, further discoveries are needed to clarify the mechanism of action of APOE in THCA, including whether it exhibits tissue-specific mechanisms, whether it can play a protective role by regulating immune cells, and whether different APOE genotypes in THCA affect tumor growth and prognosis.

The protein encoded by FADS1, a member of the fatty acid desaturase gene family, is a key rate-limiting enzyme of polyunsaturated fatty acids and can catalyze dihomogamma-linolenic acid to arachidonic acid (AA) [51]. Our study found that FADS1 was highly expressed in THCA and positively correlated with poor prognosis. Currently, there are no basic studies on FASD1 in thyroid cancer. Studies have found that the expression of FADS1 is up-regulated in laryngeal squamous cell carcinoma. The level of prostaglandin E2 (PGE2), the downstream metabolite of AA, is also increased in laryngeal cancer tissues. The high expression of FADS1 is closely related to the advanced clinical features and poor prognosis of patients with recurrent laryngeal cancer after chemotherapy [51]. FADS1 can promote the proliferation, migration, and invasion of laryngeal cancer cells by activating the AKT/mTOR signaling pathway [51]. In colon cancer, FADS1 induces enrichment of Gram-negative bacteria through a high AA microenvironment [52]. By activating the TLR4/MYD88 pathway in colon cancer cells, Gram-negative microorganisms metabolize the FADS1-AA axis to PGE2 and promote the growth of colorectal cancer [52]. Elimination of Gram-negative bacteria can eliminate the FADS1 effect [52]. In breast cancer, studies suggest that high levels of AA induce a pro-inflammatory environment, and these pro-inflammatory effects may be related to breast cancer progression [53]. In addition, studies have found that when TP53 is mutated, the expression level of FADS1 is higher, and FADS1 may be the downstream target of TP53 mutation [54, 55]. In most cancer types, tumors with increased FADS1 expression also showed increased features of fibroblast and macrophage infiltration [55]. Our research also found that FADS1 has the highest expression in fibroblasts. Combined with the high expression of FADS1 in THCA and the close correlation between TP53 mutations and thyroid cancer, we speculate that FADS1 may play a role in the progression of thyroid cancer through the TP53-FADS1-AA/PGE2 axis. In addition, studies have found that high expression of LINC00460 negatively regulates the proliferation and metastasis of osteosarcoma by insertion into the miR-1224-5p/FADS1 axis and is positively correlated with distant metastasis and poor overall survival of patients [56]. LINC01569 helps tumor cells escape inherent immune surveillance through the miR-193a-5p/FADS1 signaling axis and promotes the occurrence and development of laryngeal cancer [57]. In the future, we need further basic research to confirm the mechanism of action of FADS1 in THCA.

HSD17B7 encodes an enzyme involved in steroid and cholesterol biosynthesis, and its regulation has been found to determine the proliferation and carcinogenicity of primary keratinocytes and squamous cell cancer cells, as well as mitochondrial OXPHOS activity [58]. ACSL4 is a long-chain polyunsaturated acyl-CoA synthetase that drives the preferential binding of polyunsaturated fatty acids into phospholipids, resulting in changes in the composition and properties of membrane phospholipids [59]. ACSL4 is upregulated in metastatic triple-negative breast cancer, ACSL4-mediated phospholipid remodeling of the cell membrane induced lips-raft localization and activation of integrin β1 in a CD47-dependent manner, which led to downstream focal adhesion kinase phosphorylation that promoted metastasis [59]. HMGCS2 is the rate-limiting enzyme of ketone formation. It mainly synthesizes ketone bodies, β-hydroxybutyrate, and acetoacetate [60]. HMGCS2 down-regulation mediated ketone reduction promotes the clinicopathological progression of liver cancer by regulating c-Myc/cyclin D1 and Caspase-dependent signaling [60].

In this study, through functional enrichment analysis, we found that most immune cells were significantly down-regulated in the high-risk group, especially CD8 + T cells, Treg cells, macrophages, dendritic cells, and neutrophils. The expression of immune checkpoint-related genes and TAM-related genes was significantly different in the high and risk groups, and the immune checkpoint-related genes were significantly down-regulated in the high-risk group. These results suggest that our prognostic model is closely associated with immune cell infiltration, that the high-risk group has poorer immune cell infiltration, and that immunotherapy outcomes are worse than those in the low-risk group [61]. We used four algorithms, namely Estimate, EPIC, CIBERSORT, and MCPcounter, to analyze the immune cell infiltration in the low-risk and high-risk groups. Estimate quantifies the relative abundance of stromal and immune components in tumor samples, along with tumor purity. EPIC estimates the absolute proportions of immune cell subsets and fibroblasts in samples while simultaneously calculating tumor cell proportion. CIBERSORT estimates the relative proportions of 22 specific immune cell subtypes within samples [62]. MCP-counter quantifies the relative abundance scores of 8 major immune cell types and 2 stromal cell types in samples. Cross-validation using different algorithms can improve the reliability of the results. The analysis using the Estimate algorithm revealed that the stromal and immune scores in the low-risk group were significantly higher than those in the high-risk group, indicating that the tumor purity in the high-risk group was higher, which is usually associated with a poor prognosis. These findings are consistent with previous studies, in which cholesterol metabolism can affect the immune system, affect antitumor immunity, and reduce the efficacy of immunotherapy [63]. Excess cholesterol can lead to depletion of CD8 + T cells in some tumor models, and the targeted cholesterol pathway may be synergistic with PD-1 checkpoint inhibitors to treat some cancers [63]. In addition, we explored drug sensitivity to antitumor therapy in patients with different risk states. We found that patients in the high-risk group were more sensitive to BRD-K09587429, BRD-K16147474, BRD-K52037352, IPR-456, PD318088, and pevonedistat. Studies have reported that IPR-456 blocked the invasion of breast cancer cells by inhibiting the binding of uPA and uPAR [64]. At present, there are many studies on Pevonedistat, a small molecule inhibitor of NEDD8-activating enzyme, which can up-regulate TNF-α and IFN-γ in patient-derived CD8 + T cells and show enhanced cytotoxicity [65, 66]. In addition, pevonedistat can up-regulate PD-L1 in adult T-cell leukemia/lymphoma cells and increase the therapeutic efficacy of immune checkpoints [67]. It is of significant clinical significance to further investigate in the future whether pevonedistat can improve the prognosis of THCA patients by regulating the immune mechanism. Finally, we verified the significantly high expression of NPC2 in thyroid cancer tissues through the HPA database. In the future, further basic research is needed to explore the detailed biological functions and mechanisms of the six CMRGs involved in the prognosis of THCA [68].

At present, there is no research on cholesterol metabolism in the field of THCA. We established a prognostic model of THCA on genes related to cholesterol metabolism, which is of great significance for clinical application and treatment selection and lays a theoretical foundation for future mechanism research. However, this study still has certain limitations. All our prognostic analyses were conducted in the TCGA cohort. Due to the lack of follow-up survival information in other cohorts, our study only had internal validation and relied on retrospective data. Future prospective studies are needed to provide more high-quality data to further verify the accuracy of the model. Meanwhile, the cohort sources of the TCGA and GEO datasets are different, and the potential patient characteristics have certain differences. Moreover, the sample size of GSE33630 is relatively small. Therefore, some in vivo and in vitro experiments will be needed in the future to verify the prognostic risk of cholesterol-related genes and their relationship with immune cell infiltration. In the future, we will further improve the relevant basic research to better study the functional mechanism of cholesterol metabolism-related genes in THCA.

Conclusions

In conclusion, based on the RNA sequencing data of the TCGA-THCA cohort, we analyzed and obtained 6 CMAGs to construct a THCA prognostic model on cholesterol metabolism and a nomogram to perform immunoinfiltration analysis and drug sensitivity analysis. The six CMAGs may also provide new targets for the prognosis and treatment of THCA, providing information for stratifying THCA prognosis and personalized targeted therapy.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

XL conceived and designed the study had full access to all of the data in the study and took responsibility for the integrity of the data and the accuracy of the data analysis. FW conceived and designed the study. DL collected the data. XL performed the statistical analysis. XL and PS drafted this manuscript. All authors contributed to the manuscript revision and read and approved the submitted version.

Funding

None.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.