2.1 Setting and Overview

The incidence of TC, particularly DTC, is steadily increasing. Despite advances in diagnosis and treatment, the complexities of TC progression remain poorly understood. This study elucidates the role of lipid metabolism in TC by integrating transcriptomic and metabolomic data, identifying key LMGs, and developing a novel prognostic risk model. We examined the correlation between the risk model and immune cell infiltration and predicted the efficacy of potential therapeutic drugs in the TCGA_THCA database.

2.2 Data Collection

Four microarray datasets, namely GSE3678, GSE33630, GSE29265, and GSE3467, were retrieved from the GEO database (https://www.ncbi.nlm.nih.gov/geo/). Additionally, RNA-Seq data and clinical information for TCGA_THCA were obtained from the TCGA database (https://portal.gdc.cancer.gov/). A detailed description of each dataset is provided in Table S1 for reference. The raw data underwent rectification and normalization using the “limma” package in R software (x64 4.2.0). For a visual representation of the workflow, refer to Figure S1.

2.3 Serum Sample Preparation and Metabolomics Profiling

A total of 24 subjects were enrolled for LC/MS analysis, including 12 TC patients and 12 healthy controls (HC). Serum samples were collected from both groups. Clinical data, including age, gender, TNM stage, and so on, were documented (Table S2). TC diagnosis was confirmed through histopathological examination post-surgery. The analysis was performed utilizing an ultra high performance liquid chromatography (UHPLC) system (Vanquish UHPLC, Thermo) coupled with an Orbitrap mass spectrometer (Q Exactive HF) at Shanghai Applied Protein Technology Co. Ltd. Both electrospray ionization (ESI) positive and negative modes were utilized to ensure comprehensive coverage of the metabolome. HILIC separation was performed using an ACQUITY UPLC BEH Amide column (2.1 mm × 100 mm, 1.7 μm; Waters, Ireland). The raw MS data underwent preprocessing using ProteoWizard MSConvert before being imported into the XCMS software. Quality control (QC) samples, generated from combined aliquots of each processed sample, were injected every 10 samples to monitor data acquisition consistency and enable subsequent abundance normalization [20]. We utilized CAMERA (Collection of Algorithms of MEtabolite pRofile Annotation) for annotating isotopes and adducts. Only variables with more than 50% of nonzero measurement values in at least one group were retained in the extracted ion features. Metabolite compound identification relied on comparing accurate m/z values (< 10 ppm) and MS/MS spectra with an in-house database established using available authentic standards [21]. This study received approval from the Ethics Committee of Chongqing General Hospital, and all participating patients did so voluntarily.

2.4 Metabolomics Data Analysis

Following sum-normalization, the processed data underwent analysis using the R package (ropls), including Pareto-scaled principal component analysis (PCA) and orthogonal partial least squares discriminant analysis (OPLS-DA). Both PCA and OPLS-DA models were employed to investigate differentially abundant metabolites (DMs) between TC and HC. In the OPLS-DA model, the variable importance in the projection (VIP) value for each variable was calculated to indicate its contribution to classification. Further selection of DMs relied on the p value or fold change obtained from univariate analysis. After abundance matrix correction, Student's t-test was used to determine the significance of differences between two independent sample groups. To identify significant changes, a threshold of VIP > 1 and a p value < 0.05 were applied. Enrichment analysis of DMs was carried out using the OECloud tools available at https://cloud.oebiotech.com, with the cut-off criteria set at a p value < 0.05.

2.5 Bioinformatic Analysis

The identification of DEGs utilized the R package “limma” (version 3.50.3) [22], with statistically significant values defined by a threshold of |Log2FC| > 1 and adjusted p < 0.05. The VennDiagram package (version 1.7.3) was employed to analyze the overlap among the four GEO datasets. Result diagrams, such as volcano plots, heatmaps, and forest plots, were generated using the online platform for data analysis and visualization at https://www.bioinformatics.com.cn [23]. Functional enrichment analysis, focusing on genes related to lipid metabolism pathways, was conducted using the Sangerbox platform to access Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) databases [24]. A significance threshold of p < 0.05 and false discovery rate (FDR) < 0.25 was applied. Prognostic analyses were conducted using the R packages “survival” (version 3.4.0) and “survminer” (version 0.4.9) for both univariate and multivariate assessments of independent prognostic values of specific gene signatures. A least absolute shrinkage and selection operator (LASSO) regression analysis was conducted using the R package “glmnet” (version 4.1–7). Subsequently, a Cox proportional hazards regression prognostic model was constructed by incorporating predictive prognostic genes into the regression equation to obtain risk scores, expressed as: risk score = SUM [gene expression × coefficient]. Risk assessment was conducted using the R packages “ggrisk” (version 1.3) and “rms” (version 6.3–0). Samples in the TCGA_THCA cohort were categorized into high- or low-risk groups according to their median risk scores. Receiver operating characteristic (ROC) and Kaplan–Meier (K–M) analyses were performed between the two groups. The R package “survival” was employed to compare overall survival (O.S.). The prognostic performance of the model was evaluated at 1-, 3-, and 5-year using the “pROC” package to calculate the ROC and area under the curve (AUC).

2.6 Development of Nomogram and Gene Set Enrichment Analysis (GSEA)

Age, stage, and risk score were utilized to construct a nomogram employing R packages such as “survival”, “rms” (version 6.3–0), “nomogramFormula,” and “ggplot2.” Calibration curves for the nomogram were generated to evaluate the concordance between predicted and actual survival. To investigate the enrichment of signaling pathways and biological processes related to LMGs between high- and low-risk groups, GSEA was employed (http://software.broadinstitute.org/gsea/index.jsp). The algorithm's parameters included a minimum gene set size of 5 and a maximum gene set size of 5000.

2.7 Immune Infiltration Analysis and Gene Set Cancer Analysis (GSCA)

The proportion of 22 immune cell types in each THCA sample from the TCGA_THCA database was computed using CIBERSORT [25], a deconvolution algorithm. To ensure result reliability, 1000 permutations were performed, and samples with p < 0.05 were retained. Estimation of TIICs and lipid metabolism biomarkers was based on the risk signature. The Wilcoxon rank-sum test was used to examine the difference in immune cell proportions between high- and low-risk groups.

Correlation analysis between LMGs and immune cells was conducted using the “psych” R package (version 2.3.9) to visualize their relationship. Additionally, the GSCA platform was employed for immunogenomic gene set cancer analysis [26], focusing on the correlation between variations in LMGs and the abundance of immune cells.

2.8 Prediction the Sensitivity of Potential Therapeutic Drugs Based on Risk Stratification

The half-maximal inhibitory concentration (IC₅₀) of each THCA patient from the TCGA_THCA database was predicted utilizing the “pRRophetic” R package, leveraging data from the Genomics of Drug Sensitivity in Cancer (GDSC) dataset. This approach was employed to evaluate therapeutic drug sensitivity [27]. With the emerging concept of personalized treatment for TC, predicting treatment response is crucial. Chemotherapy response for 251 drugs was forecasted across different risk groups. While most are not FDA-approved for certain types of TC, they are recommended by the National Comprehensive Cancer Network (NCCN) [28]. We selected 12 drugs relevant to TC treatment, including sorafenib [29], dabrafenib [30], trametinib [31], Lenvatinib [32], axitinib [33], doxorubicin [34], gemcitabine [35], imatinib [36], paclitaxel [37], pazopanib [38], sunitinib [39], and bexarotene [40].

2.9 Cell Lines and Cell Culture

The human thyroid follicular epithelial cell line (Nthy-ori3-1) was obtained from iCellbioscience (Shanghai, China), and the human TC cell lines (BCPAP, KTC-1, and TPC-1) were purchased from Procell (Wuhan, China). All cells were cultured in RPMI-1640 (Gibco) supplemented with 10% fetal bovine serum (FBS) and 1% penicillin/streptomycin (P/S), and maintained in a humidified incubator with 5% CO₂ at 37°C. Authentication of all cell lines was performed using a Short Tandem Repeat (STR) assay.

2.10 Real-Time Quantitative PCR (RT-qPCR) and Western Blot

Total RNA from cell samples was extracted using the RNAiso Plus reagent (#9108, TaKaRa), and cDNA was synthesized using the Prime Script RT Reagent Kit (#RR047A, TaKaRa). For RT-qPCR analysis of selected gene expression, the TB Green Premix Ex Taq II (#RR820A, TaRaKa) was employed, with GAPDH serving as the endogenous control. The QuantStudio5 real-time PCR system (Applied Biosystems) was used to measure relative mRNA expression levels, with data analysis performed using the 2^{(−ΔΔC(T))} method. Each sample was examined in triplicate, and the experiment was repeated independently three times. Primer sequences are provided in Table S3.

Cell proteins were extracted using RIPA lysis buffer (#P0013B, Beyotime), and the protein concentration was determined using a BCA protein assay kit (#WB6501, NCM Biotech). Equal amounts of proteins were isolated using SDS-PAGE and blotted onto a polyvinylidene fluoride membrane (Millipore, USA), followed by blocking with 5% nonfat milk for 1.5 h at room temperature. Membranes were then incubated overnight with the designated primary antibodies at 4°C, followed by incubation with the secondary antibody for 1 h. Signal detection was performed using the enhanced chemiluminescence (ECL) detection kit (#P2200, NCM Biotech) with an automatic chemiluminescence/fluorescence image analysis system (Bio-Rad, USA). The antibodies used are listed in Table S4.

2.11 Statistical Analysis

Statistical analyses were conducted using R 4.2.0 (University of Auckland, New Zealand), SPSS 25.0 (IBM, USA), and GraphPad Prism 8.0 (San Diego, CA, USA). The data were presented as the mean ± standard error of the mean. Significant differences between the means of two groups were determined using Student's t-test. Spearman's correlation coefficient was employed for correlation analysis, and the Wilcoxon test was used to compare data from different groups. All analyses were two-tailed, and the p value < 0.05 was considered indicative of statistically significant results.

3.1 Screening and Analysis of 62 Differential LMGs

The four GEO datasets (GSE3678, GSE33630, GSE29265, and GSE3467) include 90 normal and 85 PTC samples, identifying a total of 304 intersecting DEGs, comprising 125 upregulated and 179 downregulated genes (Figure S2a,b). Furthermore, transcriptomic data from TCGA_THCA were obtained, featuring 59 normal and 504 DTC samples. A total of 1603 DEGs were identified, with 895 genes upregulated and 708 genes downregulated (Table S1). Genes associated with lipid metabolism processes were manually screened based on GO and KEGG enrichment analyses (Figure S2c,d), resulting in the selection of 62 differentially expressed LMGs. Among these, 34 were downregulated, and 28 were upregulated (Figure 1a–c). Further enrichment analysis revealed that these genes were linked to the adipocytokine signaling pathway, glycerophospholipid metabolism, PPAR signaling pathway, sphingolipid signaling pathway, as well as fatty acid biosynthesis (Figure 1d). The expression pattern of LMGs in TCGA_THCA was depicted using a heatmap (Figure 1e). Then, the survival analysis was performed using the TCGA_THCA cohort. Univariate Cox regression analysis was employed to explore the potential predictive value of 62 LMGs in TC, identifying 17 LMGs significantly associated with O.S. based on TCGA_THCA (p < 0.05; Figure 2a). LASSO regression with tenfold cross-validation was subsequently executed, resulting in the identification of an optimal lambda value and association with 12 of the 17 prognostic LMGs significantly linked to O.S. (Figure 2b,c).

3.2 Identification and Enrichment Analysis of DMs

There was no statistical difference in age, gender, BMI, diabetes, and smoking history between the TC patients and HC. Except for HDL-C (p = 0.010), there was no significant difference in other biochemical indexes between the two groups, including TSH, CEA, Creatinine, ALT, AST, T-CHOL, TRIG, HDL-C, LDL-C, and TRIG/HDL-C. All TC patients were pathologically diagnosed with PTC post-operation, with TNM stages I-II, and 83.33% had the BRAF^V600E gene mutation (Table S2). In the metabolomic cohort, both PCA and OPLS-DA results demonstrated a clear separation between TC and HC groups (Figure S3a–d). In both ESI positive and negative modes, a total of 692 metabolites were identified. Through the adjusted p values and p value thresholds, 59 metabolites of interest were selected, indicating differences in metabolite abundance between TC and HC. Examples include isocitric acid, malate, phosphorylcholine, phytosphingosine, and trans-dehydroandrosterone, with characteristics and statistics summarized in Table S5. Importantly, the analysis of major metabolic pathways associated with DMs in TC revealed involvement in pathways such as the citrate cycle (TCA cycle), steroid hormone biosynthesis, glycerophospholipid metabolism, sphingolipid metabolism, and other pathways closely related to lipid metabolism (Figure 2d). Notably, these pathways intersect with those enriched in differentially expressed LMGs. The heatmap analysis of DMs was presented in Figure 2e. Serum TCA cycle intermediates have been investigated in relation to various underlying pathological processes [41, 42]. Isocitric acid and malate as intermediates of the TCA cycle were shown to be more abundant in TC in our analysis. These data suggested a significant involvement of lipid metabolism in the development of TC.

3.3 Development and Evaluation of the Prognostic Risk Model

To identify key LMGs involved in TC processes, transcriptomes and metabolomes were integrated for analysis. The pathways of 62 LMGs were co-enriched with the top 20 pathways from the metabolome, resulting in the identification of 29 key LMGs, including 11 overlapping pathways (Figure S3e). Interestingly, 6 genes (FABP4, PPARGC1A, AGPAT4, ALDH1A1, TGFA, and GPAT3) were found to be intersection genes among the 29 key LMGs and the previously identified 12 prognostic genes. Multivariate Cox regression analysis was then performed to investigate the effects of the 6 LMGs, which were subsequently selected as signature genes to establish a prognostic risk model for TC patients (Figure S3f). The prognostic LMGs were represented by the risk score formula: Risk score = (0.414 * GPAT3 + 0.034 * FABP4 + 0.265 * PPARGC1A + 0.451 * AGPAT4–0.266 * TGFA—0.587 * ALDH1A1). THCA patients were stratified into high- and low-risk groups based on the median risk score (−2.067) in the TCGA_THCA cohort, revealing differential expression of the 6 LMGs between the two groups (Figure 2f). K–M analysis revealed that the high-risk group exhibited lower O.S. compared to the low-risk group (p = 0.0045; Figure 3a). The 1, 3, and 5-year O.S. rates for the high-risk group were 93.55%, 93.83%, and 94.42%, respectively, compared to 96.15%, 99.30%, and 98.45% for the low-risk group (p = 0.0144). Figure 3b illustrates the distribution of risk scores and gene expression, correlated with survival status, indicating shorter survival times among patients with high-risk scores. Moreover, ROC curve analysis was conducted to evaluate the predictive capability of the LMGs prognostic signature. The AUC for the 1-, 3-, and 5-year survival rates were 0.761 (95% CI 0.613–0.909), 0.853 (95% CI 0.687–0.999), and 0.878 (95% CI 0.757–0.999), respectively, suggesting that the 6 LMGs signature exhibited robust predictive ability in TCGA_THCA (Figure 3c).

3.4 Prognostic Risk Score Was Associated With Clinical Outcome

Univariable and multivariable Cox regression analyses were employed to examine the relationship between clinical parameters, risk scores, and O.S. The results indicated that the risk score, age, and stage were independent risk factors (p < 0.05; Figure 3d,e). Gender and N0/N1 (presence or absence of lymph node metastasis) showed no significance. Subsequently, a nomogram was constructed by integrating these independent prognostic factors to enhance the accuracy and reliability of the predictive model (Figure 3f). The nomogram for predicting 1-, 3-, and 5-year O.S. was calibrated, and the calibration curve revealed excellent concordance between actual observations and nomogram predictions (Figure 3g). The AUC values of the nomogram for predicting 1-, 3-, and 5-year O.S. were 0.940 (95% CI 0.905–0.974), 0.925 (95% CI 0.875–0.976), and 0.828 (95% CI 0.757–0.999), respectively, suggesting that the nomogram exhibited good predictive value for the clinical outcomes of TC patients (Figure 3h).

3.5 Association of the Risk Model and Immune Microenvironment

Considering the significant impact of lipid metabolic reprogramming on cancer progression and therapeutic response through immune microenvironment remodeling [41], we investigated the relationship between immune-cell characteristics and the risk model based on LMGs. Using the CIBERSORT algorithm to analyze the abundance of TIICs in the TCGA_THCA cohort, a heatmap was generated to illustrate the differences in immune characteristics (Figure 4a). Figure 4b depicted the abundance of various immune cell infiltrates in TC samples. Further analysis of the differences in the proportion of each of the 22 TIICs between the high- and low-risk groups uncovered significantly higher infiltration levels of B cell memory, monocytes, CD8+ T cells, and CD4+ T cells memory resting in the high-risk group compared to the low-risk group (Figure 4c). Notably, the relationship between risk groups and immune cell abundance requires further data validation. Subsequently, we assessed the correlation between the expression of the 6 LMGs and the levels of the 22 TIICs. The results showed a negative correlation between the infiltration of five immune cell types (B cells naïve, CD4+ T cells memory activated, T cells regulatory (Tregs), dendritic cells resting/activated, and macrophages M0) and the expression of five genes (FABP4, PPARGC1A, AGPAT4, ALDH1A1, and GPAT3), while it was positively correlated with B cells memory, CD8+ T cells, CD4+ T cells memory resting, mast cells activated, and eosinophils. The correlation of TGFA was just the opposite (Figure 4d). The results of immunogenomics analysis on the GSCA platform were consistent with our findings, essentially validating our analysis (Figure 4e). Interestingly, the methylation levels of PPARGC1A and AGPAT4 exhibited significant correlations with the infiltration levels of immune cells (Figure 4f).

3.6 Association of the Risk Model With Efficacy of Therapeutic Drugs

We explored the relationship between risk scores and the efficacy of TC treatment-related drugs by calculating IC₅₀. Although some drugs are not FDA approved, they can be considered if clinical trials or other systemic therapies are not available or appropriate [28, 43]. The benefits and risks of available treatment options should be weighed in clinical practice or clinical trials. We observed differences in drug responses between the high- and low-risk groups. The IC₅₀ values of sorafenib, dabrafenib, lenvatinib, axitinib, imatinib, and pazopanib were lower in the high-risk group, indicating that high-risk patients may benefit more from these drugs (p < 0.05; Figure 5a–l). Conversely, trametinib, doxorubicin, gemcitabine, paclitaxel, sunitinib, and bexarotene exhibited higher drug responses in the low-risk group. These findings imply that patients' responses to potential therapeutic drugs for TC can be predicted based on the risk score.

3.7 Potential Mechanisms of the Prognostic Risk Genes in Lipid Metabolism Pathway

To delve deeper into the potential mechanisms involving the six LMGs, we performed GSEA by comparing the high and low expression levels of these signature genes using the TCGA_THCA database. The analysis revealed significant enrichment of lipid metabolism-related processes in each prognostic risk gene. Specifically, FABP4, PPARGC1A, AGPAT4, ALDH1A1, and GPAT3 were associated with negative modulation of fatty acid metabolism, the PPAR signaling pathway, steroid biosynthesis, glycerophospholipid metabolism, and the adipocytokine signaling pathway. In contrast, TGFA exhibited the opposite regulatory ability (Figure 6a–f). These results suggested an association between the expression of prognostic risk genes and lipid metabolism signaling pathways in TC patients.

3.8 Validation of mRNA and Protein Expressions of Prognostic Signature Genes

We conducted RT-qPCR and Western blot analyses to assess the expression levels of the six LMGs in Nthy-ori 3–1, BCPAP, KTC-1, and TPC-1 cell lines. Compared to the Nthy-ori 3–1 cell line, TGFA exhibited a marked overexpression at both mRNA and protein levels in BCPAP, KTC-1, and TPC-1 cells (p < 0.05). Conversely, the expression of FABP4, PPARGC1A (PGC1A), AGPAT4, ALDH1A1, and GPAT3 was significantly lower in BCPAP, KTC-1, and TPC-1 cells (Figure 7a,b). These findings validated the results of our bioinformatic analysis, underscoring the reliability of our research.