Combining single cell and bulk transcriptomics with Mendelian randomization identifies gut microbiota related prognostic genes and mechanisms in thyroid cancer

Abstract

Background

Imbalance in gut microbiota (GM) may play a role in the development of thyroid cancer (TC), but the specific mechanisms remain unclear. This study aimed to identify prognostic genes associated with both TC and GM and investigate the underlying molecular mechanisms.

Methods

Public datasets were leveraged to perform differential expression, Mendelian randomization (MR), and machine learning analyses to pinpoint prognostic genes. A nomogram model was developed for survival prediction. Functional roles of candidate genes were explored through gene set enrichment analysis (GSEA), immune infiltration profiling, drug sensitivity prediction, and regulatory network construction. Single-cell RNA sequencing was employed to investigate cell-specific mechanisms in TC.

Results

MR analysis identified 27 GM traits causally linked to TC, corresponding to 124 genes. Three key prognostic genes (LRP1B, MCM6, and PPARG) were selected to construct a robust prognostic model, which demonstrated high predictive accuracy. GSEA revealed the involvement of lysosomal and oxidative phosphorylation pathways in TC pathogenesis. Immune profiling showed significant variations in five immune cell types, including monocytes, between risk groups, correlating with prognostic gene expression. Drug prediction suggested 138 potentially effective compounds, alongside regulatory elements such as hsa-miR-20a-5p. Single-cell analysis highlighted the pivotal role of fibroblasts in TC progression.

Conclusion

This study identified three prognostic genes (LRP1B, MCM6, PPARG) linked to TC and GM, offering new insights into the molecular mechanisms and presenting potential biomarkers and therapeutic targets for TC management.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-05132-8.

Introduction

Thyroid cancer (TC) ranks as one of the most prevalent endocrine malignancies, with the ninth highest global cancer incidence rate. It includes several histological subtypes, such as differentiated TC (DTC), anaplastic thyroid carcinoma (ATC), and medullary thyroid carcinoma (MTC) [1]. Among these, ATC is recognized as one of the most aggressive and lethal forms of cancer [2]. While DTC generally follows a slow progression and carries a favorable prognosis, a subset of patients may still develop distant metastases and experience disease-related mortality [3]. Current treatment strategies for TC encompass surgical resection, radioactive iodine (RAI) therapy, hormone suppression, and targeted molecular therapies [4, 5]. Despite the variety of available treatments, only a fraction of patients achieve long-term clinical benefits. Notably, around 30% of DTC cases eventually progress to RAI-refractory DTC (RAIR-DTC), a condition associated with significantly reduced survival and limited therapeutic options. The molecular mechanisms driving the development of RAIR-DTC remain insufficiently understood [6]. Consequently, elucidating the molecular mechanisms underlying TC and identifying reliable prognostic biomarkers are crucial for enhancing diagnostic accuracy, informing therapeutic decisions, and improving patient outcomes.

The gut microbiota (GM), a key component of the host’s micro-ecosystem, plays a pivotal role in the development and progression of various diseases. Growing evidence indicates that GM and its metabolites influence pathological processes by modulating host immunity, metabolism, and inflammation, thereby contributing to conditions such as cardiovascular disorders, osteoporosis, immune-related diseases, and various cancers [7]. Recent studies suggest that GM dysbiosis may promote certain thyroid diseases via mechanisms involving disruption of the hypothalamic–pituitary axis, immune dysregulation, and chronic inflammation [8, 9]. Furthermore, emerging evidence points to a potential link between GM and TC [10–12]. However, direct studies investigating the relationship between GM and TC remain scarce, and the underlying mechanisms remain incompletely understood. Therefore, systematically exploring the role of GM in the initiation, progression, and treatment of TC is of significant importance.

Mendelian randomization (MR) has become a powerful analytical tool for inferring causal relationships between exposures and disease outcomes by using genetic variants as instrumental variables (IVs). This approach effectively reduces confounding and reverse causality, offering robust insights into disease mechanisms [13]. Meanwhile, single-cell RNA sequencing (scRNA-seq) provides transcriptomic profiling at single-cell resolution, enabling the precise characterization of cellular heterogeneity and dynamic functional states within tissues. This technology offers unprecedented insights into the initiation and progression of diseases at the cellular level [14]. Although both MR and scRNA-seq have significantly advanced our understanding of various cancers and immune-related diseases, systematic investigations into the relationship between GM and TC using these methods remain limited.

This study systematically explored the potential causal relationship and underlying molecular mechanisms linking GM and TC. MR analysis was performed using genome-wide association study (GWAS) datasets to identify prioritized candidate genes implicated in the onset and progression of TC. Prognosis-related differentially expressed genes (DEGs) were then identified through publicly available TC transcriptomic datasets, and a risk scoring model along with a nomogram was developed to assess their clinical predictive value. Additionally, single-cell transcriptomic data were utilized to examine the distribution and functional alterations of key cell types, emphasizing cellular heterogeneity within TC. By integrating multi-level data, our study provides novel theoretical insights and identifies potential therapeutic targets to enhance the diagnosis and personalized treatment of TC.

Method

Data collection

Transcriptome data for TC were obtained from The Cancer Genome Atlas (TCGA-THCA) database (https://portal.gdc.cancer.gov/), comprising 510 tumor and 58 normal thyroid tissue samples, along with corresponding survival data. The dataset was randomly divided into training (358 tumor/42 normal) and validation (152 tumor/17 normal) sets at a 7:3 ratio using the rms package (version 6.8.1) [15].

Single-cell transcriptomic data from seven papillary thyroid carcinoma cases (GSE184362, platform GPL24676) were retrieved from the Gene Expression Omnibus (GEO) database.

Genetic data for the GM were sourced from the MiBioGen Consortium [16] (18,340 individuals from 24 cohorts, primarily European), comprising 211 taxa and 14,587 SNPs. MR data for TC (EBI-A-GCST90018929) [17] were obtained from the IEU GWAS database (1,054 cases, 490,920 controls, ~ 24 million SNPs). All data were accessed on February 8, 2025.

MR analysis

To investigate the causal relationship between GM and TC, MR analysis was performed with GM as the exposure and TC as the outcome, aiming to elucidate the genetic causal link. MR analysis was performed under three key assumptions: relevance, independence, and exclusion restriction of IVs.

IVs were extracted using the extract_instruments function from the TwoSampleMR package (version 0.6.4) [18]. The screening criteria included: (1) IVs with significant associations with the exposure (p < 5 × 10^⁻6) [19]; (2) exclusion of IVs in linkage disequilibrium (R² < 0.001, kb = 10); (3) exclude IVs with horizontal multicollinearity; (4) exclusion of IVs with an F-statistic < 10; (5) exclusion of SNPs with palindromic sequences, and exposures with fewer than three SNPs.

The effect alleles and effect sizes were harmonized, and exposure factors, IVs, and outcomes were aligned using the harmonise_data function. Five MR algorithms were applied via the mr function: Inverse Variance Weighted (IVW), MR Egger, Weighted mode, Simple mode, and Weighted median. The IVW method effectively integrates information from multiple instrumental variables, reduces random error, and exhibits high statistical power whilst satisfying the three core MR assumptions [20]. Consequently, we selected the IVW method as the primary MR approach. Exposure factors with p < 0.05 were considered potentially causally associated with TC. Odds ratios (OR) > 1 indicated risk factors, while OR < 1 indicated protective factors (95% Confidence Interval [CI] ≠ 1).

For ease of interpretation, several visualization functions were employed. The mr_scatter_plot function was used to visualize the relationship between exposure factors and outcomes, incorporating SNP-exposure and SNP-outcome effects. The mr_forest_plot function was utilized to assess the diagnostic effectiveness of estimated exposure parameters at each SNP locus, alongside the risk effects of exposure factors on outcomes at each SNP. The mr_funnel_plot function assessed whether MR analysis adhered to Mendel’s second law of random assortment. Sensitivity analysis was conducted to assess the robustness of findings, including heterogeneity testing (mr_heterogeneity function, p > 0.05), horizontal pleiotropy (mr_pleiotropy_test function, p > 0.05), and leave-one-out (LOO) analysis (mr_leaveoneout function). To confirm that forward analysis results were not influenced by reverse causal effects, Steiger directionality testing (directionality_test function, steiger pval < 0.05, correct_causal_direction = TRUE) was employed to validate causality direction.

GM taxa passing all analyses were considered causally linked to TC. Associated SNPs were mapped to genes using SNPnexus [21] to identify prioritized candidate genes.

Screening of candidate genes related to GM in TC and related functional analysis

To identify candidate genes linked to GM in the pathogenesis of TC, DEGs between TC and normal samples in the TCGA-THCA training set were analyzed using the DESeq2 package (version 1.42.0) [22] (|log₂FoldChange| > 0.5, p < 0.05). A volcano plot and heatmap were generated to visualize the top 10 upregulated and downregulated DEGs. These DEGs were then intersected with GM-associated prioritized candidate genes using a Venn diagram, and the overlapping genes were defined as candidate genes.

Functional enrichment analysis of the candidate genes was performed using Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analyses via the ClusterProfiler package (version 4.7.1.003) [23]. Specifically, the GO analysis focused on biological processes (BP), cellular components (CC), and molecular functions (MF). The top 5 GO terms for each of the three components (ordered by adjusted p-values) and the top 8 KEGG pathways (p < 0.05) were visualized.

To investigate the roles of the candidate genes at the protein level, a protein-protein interaction (PPI) network was constructed using the STRING database (http://string-db.org) with a confidence threshold > 0.15 and visualized using Cytoscape software (version 3.10.2) [24].

Identification of prognostic genes and development and evaluation of the prognostic model

For survival analysis, univariate Cox proportional hazards (PH) regression was performed using the survival package (version 3.7.0, https://CRAN.R-project.org/package=survival) on candidate genes with survival data from 358 tumor samples in the TCGA-THCA training set. Genes significantly associated with survival (hazard ratio [HR] ≠ 1, p < 0.05) were visualized using forest plots from the forestplot package (version 3.1.1). The PH assumption was tested using the cox.zph function, and genes with p > 0.05 were retained for further analysis.

Least absolute shrinkage and selection operator (LASSO) regression, applied using the glmnet package (version 4.1.8) [25], was used to select prognostic genes with non-zero coefficients at the lambda.min value. A prognostic model was developed using the predict.glmnet function, and risk scores were computed for both the TCGA-THCA training and validation sets. The formula is as follows:

x represents the gene expression level, and Coef represents the coefficient corresponding to the gene.

Based on the optimal cutoff, samples were classified into high-risk (HG) and low-risk (LG) groups. Risk distribution and survival status were visualized using the ggplot2 package (version 3.4.4), and a heatmap of prognostic gene expression was generated using the pheatmap package (version 1.0.12). Kaplan–Meier (K-M) survival analysis was conducted using the survminer package (version 0.4.9), and receiver operating characteristic (ROC) curves at 3, 5, and 7 years were plotted using the timeROC package (version 1.0.3.1) [26] to assess predictive performance, with an area under the curve (AUC) > 0.6 indicating good model accuracy.

Correlation of clinical characteristics and development of nomogram

To assess the relationship between risk score and clinical features, the Wilcoxon rank-sum test was applied to compare risk scores across clinical subgroups (age ≤ 46 vs. > 46, gender, stage I/II vs. III/IV, T1/T2 vs. T3/T4, N0 vs. N1, M0 vs. M1) in the TCGA-THCA training set (p < 0.05). Survival differences between HG and LG within each clinical subgroup were analyzed using K-M curves and log-rank tests.

To evaluate the prognostic utility of the risk score, univariate Cox PH regression was performed for the risk score and clinical variables (age, gender, stage, T, N, and M). Significant factors (HR ≠ 1, p < 0.05) were visualized using forest plots. PH assumption testing was conducted with the cox.zph function, and valid variables (p > 0.05) were included in multivariate Cox PH regression to identify independent prognostic factors. A nomogram based on these factors was constructed using the cph function from the rms package. Calibration curves were generated to evaluate the model’s predictive accuracy at 3-, 5-, and 7-year time points. The pROC package (version 1.18.5) [27] was used to plot ROC curves to assess the nomogram’s performance, with an AUC > 0.7 indicating good performance.

Gene set enrichment analysis (GSEA)

To explore pathways associated with prognostic genes, the “c2.cp.kegg.v7.5.1.symbols.gmt” set from MSigDB [28] was used as the reference. In the TCGA-THCA training set, Spearman correlations between each prognostic gene and all other genes were calculated using the stats package (version 2.4.3). Genes were ranked by correlation coefficients, and Gene Set Enrichment Analysis (GSEA) was performed using the ClusterProfiler package with thresholds |Normalized Enrichment Score (NES)| > 1 and p.adj < 0.05. The top 5 enriched pathways for each gene were selected for visualization.

Analysis of the immune microenvironment

Given the critical role of the immune system in TC, the CIBERSORT algorithm was used to estimate the infiltration proportions of 22 immune cell types in HG and LG samples from the TCGA-THCA training set (samples with p > 0.05 were excluded). Differences between groups were assessed using the Wilcoxon test (adjusted p-value < 0.05, Benjamini-Hochberg). Additionally, Spearman correlation analysis was performed using the psych package (version 2.4.6.26, https://CRAN.R-project.org/package=psych) to explore correlations between differentially infiltrated immune cells and prognostic genes, as well as among the prognostic genes themselves (|cor| > 0.3, p < 0.05). The corrplot package (version 0.92, https://github.com/taiyun/corrplot) was used to construct a heatmap to present the findings.

To further characterize the immune microenvironment, stromal score, immune score, ESTIMATE score, and tumor purity were calculated using the estimateScore function, with intergroup differences assessed by the Wilcoxon test (p < 0.05). Additionally, expression differences for 38 immune checkpoints identified from a publication [29] were compared between HG and LG using the Wilcoxon test (adjusted p-value < 0.05, Benjamini-Hochberg).

Drug sensitivity analysis and construction of molecular regulatory network

Tumor-specific drug sensitivity data were obtained from the GDSC database [30]. In the TCGA-THCA training set, the pRRophetic package (version 0.5) [31] was used to estimate the half-maximal inhibitory concentration (IC₅₀) values for 138 commonly used chemotherapy drugs in each sample. Spearman correlation analysis was performed to assess associations between drug IC₅₀ values and risk scores (adjusted p-value < 0.05, |cor| > 0.3). Wilcoxon tests compared IC₅₀ values between high-risk (HG) and low-risk (LG) groups (adjusted p-value < 0.05), and the top 6 most significantly different drugs were visualized.

To investigate potential regulatory mechanisms of prognostic genes, potential regulatory microRNAs (miRNAs) for each prognostic gene were predicted using miRTarBase [32] and TarBase [33], with overlapping results identified. Using the ENCORI database [34] (CLIP-Data ≥ 1, degradation-data ≥ 0, pan-Cancer ≥ 0) and the MirNet database [35], long non-coding RNAs (lncRNAs) targeting these miRNAs were predicted, and the intersection of these lncRNAs was taken. The regulatory network of prognostic genes, miRNAs, and lncRNAs was visualized using Cytoscape.

Processing scRNA-seq data and screening for key cells

In the GSE184362 dataset, quality control (QC) was performed using the CreateSeuratObject function in the Seurat package (version 5.1.0) [36]. QC criteria included: 200 < nFeature_RNA < 3,000, percent.mt < 10%, and nCount_RNA < 13,000. Genes expressed in fewer than 3 cells were excluded. Data normalization was performed using the NormalizeData function (LogNormalize, scale factor = 10,000), and the top 2,000 highly variable genes (HVGs) were identified using the FindVariableFeatures function based on variance analysis. The top 10 HVGs were highlighted using the LabelPoints function.

The data were normalized using the ScaleData function, followed by dimensionality reduction via Principal Component Analysis (PCA). The significance of principal components (PCs) was assessed using the JackStrawPlot function, and their variance contributions were visualized through a scree plot generated by the ElbowPlot function (p < 0.05). Based on the PCA results, the FindNeighbors and FindClusters functions were used for unsupervised clustering of cells. Clustering was refined using the RunUMAP function (resolution = 0.5), which divided the cells into distinct subpopulations. DEGs across clusters were identified using the FindAllMarkers function. Cell-type annotation was performed by referencing established marker genes from the literature [37] and the CellMarker database [38]. A bubble plot was generated to visualize marker gene expression across annotated clusters. Finally, the RunUMAP function was used again to visualize the annotated clusters, and average expression levels of prognostic genes were calculated using Seurat and displayed via another bubble plot to identify key cell populations involved in TC progression.

Pseudotime trajectory analysis and cell communication analysis

To investigate the differentiation trajectory of key cells in TC, normalization was performed using the ScaleData function in the Seurat package, followed by PCA with the RunPCA function and clustering via FindNeighbors and FindClusters functions. Cells were divided into subclusters (p < 0.05), and the RunTSNE function (resolution = 0.06) was employed for visualization. Pseudotime analysis was conducted using the Monocle package (version 2.26.0) [39], with differentiation trajectories inferred using the DDRTree package (version 0.1.5, https://github.com/cole-trapnell-lab/DDRTree). Expression trends of prognostic genes along the trajectory were also assessed.

Intercellular communication was analyzed using the CellChat package (version 1.6.1) [40] on annotated cells from the GSE184362 dataset. First, based on the annotated Seurat objects, CellChat objects were constructed separately for each group, with cell type serving as the identity label. The analysis utilises the built-in human ligand-receptor database, CellChatDB.human. By screening for genes associated with signalling pathways, ligands and receptors that are significantly highly expressed between different cell populations are identified (screening criteria: p < 0.05 and logFC > 0.1), thereby inferring potential intercellular interactions. The probability of cell-cell communication was calculated using CellChat’s `computeCommunProb` function, which simulates communication strength by integrating ligand-receptor expression levels with cell population abundance, thereby inherently accounting for the impact of differences in cell population size within the analysis. Finally, the results were statistically evaluated using CellChat’s built-in 100-replacement test, with a significance threshold set at p < 0.05.

Statistical analysis

Statistical analyses were performed in R (version 4.3.2), with significant group differences evaluated using the Wilcoxon test. A p-value of less than 0.05 was considered statistically significant.

Results

GM with causal relationships with TC

Using the IVW method for MR analysis, 27 GM components were identified with significant causal relationships to TC (p < 0.05), comprising 11 risk factors (OR > 1, 95% CI ≠ 1) and 16 protective factors (OR < 1, 95% CI ≠ 1) (Table 1). In scatter plots, a positive slope indicated a risk factor, while a negative slope represented a protective factor. Intercepts near zero suggested no confounding (Supplementary Fig. S1). SNP effect sizes were < 0 for protective factors and > 0 for risk factors (Supplementary Fig. S2). SNP distribution was generally symmetrical and uniform, consistent with Mendel’s second law (Supplementary Fig. S3). These 27 components were selected for further analysis.

Table 1.

Causal associations between GM and TC by IVW MR analysis

Exposure	id.outcome	nSNP	b	pval	qvalue	OR	OR_lci95	OR_uci95
family.Streptococcaceae.id.1850	Thyroid_cancer	41	-0.47594	1.14E-04	0.011615	0.621301	0.487849	0.791259
family.Porphyromonadaceae.id.943	Thyroid_cancer	9	-1.05563	4.22E-05	0.008562	0.347974	0.209955	0.576722
order.Bifidobacteriales.id.432	Thyroid_cancer	79	-0.17847	0.0031	0.104886	0.836552	0.743241	0.941578
genus.Bifidobacterium.id.436	Thyroid_cancer	81	-0.15961	0.006784	0.146824	0.852477	0.759449	0.9569
class.Actinobacteria.id.419	Thyroid_cancer	75	-0.17848	0.00522	0.146824	0.836541	0.738065	0.948157
family.Bifidobacteriaceae.id.433	Thyroid_cancer	79	-0.17847	0.0031	0.104886	0.836552	0.743241	0.941578
phylum.Actinobacteria.id.400	Thyroid_cancer	62	-0.22303	0.009193	0.146824	0.800093	0.676485	0.946288
phylum.Firmicutes.id.1672	Thyroid_cancer	11	-0.47227	0.041594	0.312721	0.623584	0.395911	0.982183
class.Methanobacteria.id.119	Thyroid_cancer	8	0.360327	0.010126	0.146824	1.433798	1.089464	1.886961
family.Methanobacteriaceae.id.121	Thyroid_cancer	8	0.360327	0.010126	0.146824	1.433798	1.089464	1.886961
order.Methanobacteriales.id.120	Thyroid_cancer	8	0.360327	0.010126	0.146824	1.433798	1.089464	1.886961
genus.unknowngenus.id.1868	Thyroid_cancer	13	0.359358	0.026933	0.245989	1.432409	1.041862	1.969355
order.Rhodospirillales.id.2667	Thyroid_cancer	10	0.351689	0.032234	0.260799	1.421467	1.030255	1.96123
genus.Streptococcus.id.1853	Thyroid_cancer	35	-0.45605	3.98E-04	0.020987	0.63378	0.492411	0.815736
family.Oxalobacteraceae.id.2966	Thyroid_cancer	14	-0.23472	0.033403	0.260799	0.790796	0.637003	0.981719
genus.RuminococcaceaeUCG003.id.11,361	Thyroid_cancer	14	0.397464	0.022688	0.230286	1.488046	1.057155	2.094567
genus.LachnospiraceaeUCG001.id.11,321	Thyroid_cancer	11	-0.45316	0.008528	0.146824	0.635617	0.453472	0.890924
genus.Ruminococcustorquesgroup.id.14,377	Thyroid_cancer	8	-0.47524	0.043459	0.315079	0.621735	0.391984	0.986148
genus.Terrisporobacter.id.11,348	Thyroid_cancer	3	0.606033	0.00656	0.146824	1.833144	1.184212	2.837684
genus.Blautia.id.1992	Thyroid_cancer	9	-0.60664	0.016525	0.189109	0.545181	0.331988	0.895281
genus.Anaerostipes.id.1991	Thyroid_cancer	12	0.808902	4.14E-04	0.020987	2.24544	1.433229	3.517932
family.Alcaligenaceae.id.2875	Thyroid_cancer	24	-0.37528	0.013802	0.186789	0.687095	0.509666	0.926293
genus.Eggerthella.id.819	Thyroid_cancer	6	0.377096	0.029082	0.245989	1.458044	1.039171	2.045758
genus.FamilyXIIIAD3011group.id.11,293	Thyroid_cancer	10	0.512048	0.0177	0.189109	1.668705	1.092982	2.547688
order.Burkholderiales.id.2874	Thyroid_cancer	11	-0.55123	0.01658	0.189109	0.576241	0.367082	0.904577
class.Betaproteobacteria.id.2867	Thyroid_cancer	10	-0.54366	0.026505	0.245989	0.580618	0.359181	0.93857
genus.Barnesiella.id.944	Thyroid_cancer	10	0.542286	0.017531	0.189109	1.719934	1.099472	2.690541

Subsequent MR analysis using a fixed-effect IVW method confirmed the absence of horizontal pleiotropy across all 27 components (p > 0.05) (Supplementary Table S1). LOO analysis indicated that no single SNP significantly influenced the outcomes, supporting the stability of the results (Supplementary Fig. S4). The Steiger test excluded reverse causality (p < 0.01, output = TRUE) (Table 2). Together, these results support the significant causal associations between the 27 GM components and TC. SNPs from these components were mapped to 124 prioritized candidates associated with both GM and TC.

Table 2.

Steiger test results for directionality of GM–TC associations

Exposure	id.exposure	snp_r2.exposure	snp_r2.outcome	correct_causal_direction	steiger_pval
family.Streptococcaceae.id.1850	WbrAWN	0.0660412011284667	1.23059411783262e-4	TRUE	1.04533126097413e-245
family.Porphyromonadaceae.id.943	NLOI5r	0.0148711221802351	5.22610314789029e-5	TRUE	4.54071377907409e-53
order.Bifidobacteriales.id.432	78NVnr	0.204932707052485	1.02723917659645e-4	TRUE	0
genus.Bifidobacterium.id.436	uI3cUx	0.207149110602244	1.08129437642212e-4	TRUE	0
class.Actinobacteria.id.419	NOog7C	0.200472976587323	8.88724244796754e-5	TRUE	0
family.Bifidobacteriaceae.id.433	5eOvD7	0.204932707052485	1.02723917659645e-4	TRUE	0
phylum.Actinobacteria.id.400	9USikH	0.123394659212622	6.88464688549622e-5	TRUE	0
phylum.Firmicutes.id.1672	bqVoQp	0.0178398904739177	2.79218659784313e-5	TRUE	5.02665736935706e-66
class.Methanobacteria.id.119	7rLga6	0.0102112787978053	2.01815418457107e-5	TRUE	5.5153467023338e-38
family.Methanobacteriaceae.id.121	96Entt	0.0102112787978053	2.01815418457107e-5	TRUE	5.5153467023338e-38
order.Methanobacteriales.id.120	nmOE6A	0.0102112787978053	2.01815418457107e-5	TRUE	5.5153467023338e-38
genus.unknowngenus.id.1868	bljYDn	0.0213691314172709	2.41306091708428e-5	TRUE	7.33182652190296e-80
order.Rhodospirillales.id.2667	awKY0v	0.0132448907152603	2.68479718846155e-5	TRUE	8.53943052648822e-49
genus.Streptococcus.id.1853	gmA4Xq	0.0611362270986087	8.86144294153182e-5	TRUE	3.77137484706205e-229
family.Oxalobacteraceae.id.2966	OrtrLt	0.0196599001532464	4.09078450353477e-5	TRUE	8.86142536903793e-72
genus.RuminococcaceaeUCG003.id.11,361	v2GceQ	0.0174466289159408	3.03275379592204e-5	TRUE	2.57047975749143e-64
genus.LachnospiraceaeUCG001.id.11,321	nx05h2	0.0154887482373532	3.21003790745729e-5	TRUE	8.69425605251561e-57
genus.Ruminococcustorquesgroup.id.14,377	rNahyn	0.012945279852931	2.16293741886142e-5	TRUE	4.04260388573127e-48
genus.Terrisporobacter.id.11,348	IUojC7	0.00491287888815759	2.29874755552054e-5	TRUE	3.40236480857763e-18
genus.Blautia.id.1992	YmE6gN	0.0112045456372634	2.16490289313433e-5	TRUE	1.3992077819655e-41
genus.Anaerostipes.id.1991	7DOb53	0.0156690517700517	3.91648788555371e-5	TRUE	6.44383887036322e-57
family.Alcaligenaceae.id.2875	SdBd6B	0.0434441932556281	6.04677893399337e-5	TRUE	1.25022744798837e-161
genus.Eggerthella.id.819	aQSBgt	0.00833994757272742	1.69836228296059e-5	TRUE	2.96148356207416e-31
genus.FamilyXIIIAD3011group.id.11,293	ESia9v	0.0136656741218356	3.64358972873035e-5	TRUE	1.22993486358433e-49
order.Burkholderiales.id.2874	aWSi1n	0.0185568226949131	2.93463379027717e-5	TRUE	1.30876581039103e-68
class.Betaproteobacteria.id.2867	ZE0xnu	0.0161047370829759	2.58877148892425e-5	TRUE	1.19995136538947e-59
genus.Barnesiella.id.944	RTw8Z3	0.0146265183323488	3.7755349797331e-5	TRUE	3.99479384294003e-53

Identification of 39 candidate genes and analysis of related functions

Differential expression analysis of the TCGA-THCA training set identified 6,498 DEGs, with 3,510 upregulated and 2,988 downregulated genes in the TC group (Fig. 1a). A heatmap of 20 representative DEGs demonstrated distinct expression differences between groups based on log₂FoldChange values (Fig. 1b). By intersecting the 6,498 differentially expressed genes with the 124 prioritized candidates associated with GM and TC, 39 candidate genes were identified (Fig. 1c).

Fig. 1.

Identification of candidate genes related to GM and TC, and analysis of their functional roles (a) Volcano plot of DEGs between TC and normal samples in the TCGA-THCA training set (b) Heatmap of the top 20 representative DEGs, including 10 upregulated and 10 downregulated genes (c) Venn diagram showing the intersection of 6,498 DEGs and 124 prioritized candidate identified via MR analysis (d) GO enrichment analysis of the 39 candidate genes (e) KEGG pathway enrichment analysis of the candidate genes (f) PPI network of the 39 candidate genes

GO and KEGG enrichment analyses were performed on these 39 genes. GO analysis identified 248 significantly enriched terms, including 187 BPs, 22 CCs, and 39 MFs. Key BPs included cell recognition and negative regulation of sodium ion transmembrane transport; CCs involved cargo receptor activity and lipoprotein receptor activity; MFs included nuclear chromosome, Golgi cisterna membrane, and microvillus (Fig. 1d, Supplementary Table S2). KEGG analysis revealed 18 significantly enriched pathways, including efferocytosis, amphetamine addiction, insulin secretion, longevity regulation, and aldosterone synthesis and secretion (Fig. 1e, Supplementary Table S3).

PPI analysis of the 39 genes (excluding 4 isolated proteins) identified a network of 35 proteins with 67 interactions (Fig. 1f). Notably, WWOX, ANKS1B, PTPRG, OPCML, and LRP1B exhibited the most frequent interactions.

LRP1B, MCM6, and PPARG as prognostic genes and construction of a prognostic model

To identify prognostic genes and construct a prognostic model, univariate Cox regression analysis of 39 candidate genes revealed that LRP1B, MCM6, and PPARG were significantly associated with overall survival (OS) (Fig. 2a). Among these, MCM6 acted as a protective factor (HR < 1), while LRP1B and PPARG were risk factors (HR > 1). All three genes met the PH assumption (p > 0.05) (Fig. 2b). LASSO regression identified the optimal model at λ = 0.00068296, retaining these three genes (Fig. 2c–d).

Fig. 2.

Construction and validation of a prognostic model based on LRP1B, MCM6, and PPARG (a) Univariate Cox regression identified three prognosis-related genes (HR ≠ 1, p < 0.05) (b) Proportional hazards assumption tests (p > 0.05) (c, d) LASSO regression determined the optimal gene set at λ = 0.00068296 (e, f) Risk scores, survival time, and status in HG and LG in training and validation sets (g) Heatmap showing gene expression patterns across risk groups (h) Kaplan–Meier curves showing significantly reduced OS in HG (i) Time-dependent ROC curves assessing 3-, 5-, and 7-year survival prediction performance

A prognostic model was then constructed using the expression levels and LASSO coefficients:

1

Risk scores were calculated for all tumor samples in both the TCGA-THCA training and validation sets. Based on the optimal cutoffs (-0.1080 and 0.4278), samples were divided into HG and LG. The training set comprised 166 HG and 192 LG samples, while the validation set included 23 HG and 129 LG samples.

In both datasets, higher risk scores correlated with shorter survival and increased mortality (Fig. 2e–f). LRP1B and PPARG exhibited higher expression in HG, whereas MCM6 was more highly expressed in LG (Fig. 2g). K-M curves demonstrated significantly lower survival in HG for both the training set (p < 0.01) and validation set (p = 0.044) (Fig. 2h).

ROC analysis in the training set showed strong predictive performance, with AUCs of 0.825 at 3 years, 0.819 at 5 years, and 0.846 at 7 years. The validation set yielded AUCs of 0.640, 0.657, and 0.613, respectively (Fig. 2i), demonstrating good model reliability across datasets.

Correlation between risk score and clinical characteristics and development of a nomogram

In the TCGA-THCA training dataset, significant differences in risk scores were observed only within the N stage subgroups (Fig. 3a). Survival differences between HG and LG were significant in the T3/T4 stage (p < 0.0001), N0 (p = 0.042), N1 (p = 0.00062), M0 (p = 0.029), male (p = 0.034), female (p = 0.0038), and age > 46 (p = 0.00015) (Fig. 3b).

Fig. 3.

Correlation between risk score and clinical features and construction of a prognostic nomogram (a) Violin plot showing the differences in risk scores among clinical subgroups (b) Kaplan–Meier curves demonstrating survival differences between HG and LG in clinical subgroups (c) Univariate Cox regression showing risk score and stage as significant prognostic factors (d) PH assumption test results (p > 0.05) for risk score and stage (e) Multivariate Cox regression confirming risk score and stage as independent prognostic factors (f) Nomogram integrating risk score and stage to predict 3-, 5-, and 7-year OS (g) Calibration curves showing strong agreement between predicted and actual survival (h) ROC curves of the nomogram indicating excellent predictive performance (AUC > 0.87)

Univariate Cox analysis indicated that only risk scores and stage were significantly associated with survival (Fig. 3c), and both passed the PH assumption test (Fig. 3d). Multivariate Cox analysis further confirmed that risk scores and stage were independent prognostic factors (Fig. 3e).

A nomogram incorporating risk score and stage was constructed (Fig. 3f). Calibration curves for 3-, 5-, and 7-year survival predictions closely aligned with the ideal 45° line (Fig. 3g), and the nomogram demonstrated robust predictive performance with AUCs of 0.992 at 3 years, 0.879 at 5 years, and 0.916 at 7 years (Fig. 3h), supporting its clinical applicability for TC prognosis.

Functional enrichment analysis of LRP1B, MCM6, and PPARG

During disease progression, significant differences were observed in multiple gene-related signaling pathways. LRP1B was enriched in 50 pathways, notably including ribosome, cell adhesion molecules, valine, leucine, and isoleucine degradation, and cytokine–cytokine receptor interaction (Fig. 4a, Supplementary Table S4). MCM6 was involved in 76 pathways, such as oxidative phosphorylation, allograft rejection, type I diabetes mellitus, Leishmania infection, and Parkinson’s disease (Fig. 4b, Supplementary Table S5). PPARG was enriched in 19 pathways, including Parkinson’s disease, valine, leucine, and isoleucine degradation, oxidative phosphorylation, Huntington’s disease, and Alzheimer’s disease (Fig. 4c, Supplementary Table S6). All three genes were commonly enriched in seven pathways: valine, leucine, and isoleucine degradation, butanoate metabolism, propanoate metabolism, DNA replication, lysosome, Parkinson’s disease, and oxidative phosphorylation.

Fig. 4.

GSEA-based functional enrichment of three prognostic genes (a Top 5 pathways significantly enriched for LRP1B (b) Top 5 pathways significantly enriched for MCM6 (c) Top 5 pathways significantly enriched for PPARG

The occurrence of TC was regulated by multiple immune cells and genes

In the TCGA-THCA dataset, immune cell infiltration differed between HG and LG. CD4 + T cells, M2 macrophages, and M0 macrophages showed the highest infiltration levels (Fig. 5a). Four immune cell types exhibited significant differences between groups (adjusted p-value < 0.05): Naive B cells, Regulatory T cells (Tregs), and resting dendritic cells were upregulated in the LG, while Monocytes were upregulated in HG (Fig. 5b). A significant negative correlation was observed between LRP1B and MCM6 (cor = -0.34, p < 0.01), and between LRP1B and Tregs (cor = -0.30, p < 0.05) (Fig. 5c, Supplementary Table S7). Regarding immunotherapy response, four immune scores showed significant differences (p < 0.05), with immune, stromal, and ESTIMATE scores reduced in HG (Fig. 5d). Among 30 immune checkpoints, most were downregulated in HG (adjusted p-value < 0.05), except for IDO2, CD27, and NRP1, which were upregulated (Fig. 5e).

Fig. 5.

Immune infiltration landscape and immunological differences between high- and low-risk groups (a) Infiltration proportions of 22 immune cells across TCGA-THCA samples (b) Five immune cells showed significant differences between HG and LG (p < 0.05) (c) Correlation heatmap among prognostic genes and differentially infiltrated immune cells (d) Comparison of stromal score, immune score, ESTIMATE score, and tumor purity between HG and LG (e) Expression patterns of 30 differentially expressed immune checkpoints in HG and LG. (*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001)

Correlation between risk score and drug responsiveness, and the molecular regulatory network of prognostic genes

Risk scores were significantly correlated with the IC₅₀ values of 30 out of 138 drugs: 11 drugs showed a significant negative correlation (adjusted p-value < 0.05, cor < -0.3), while 19 drugs showed a positive correlation (adjusted p-value < 0.05, cor > 0.3) (Fig. 6a, Supplementary Table S8). Drug sensitivity also varied by risk group, with significant differences in IC₅₀ values for 94 drugs (adjusted p-value < 0.05). The six most significantly different drugs were A.443,654, AICAR, Axitinib, AZD8055, CGP.60,474, and CMK (adjusted p-value < 0.05) (Fig. 6b, Supplementary Table S9).

Fig. 6.

Correlation of risk score with drug sensitivity and construction of a regulatory network (a) Spearman correlation between risk scores and IC50 values of 138 drugs; 30 drugs showed significant correlation (b) Top six drugs with significantly different IC50 values between HG and LG (c) ceRNA network involving 3 prognostic mRNAs, 14 miRNAs, and 34 lncRNAs regulating LRP1B, MCM6, and PPARG. (****p < 0.0001)

Additionally, 14 miRNAs were predicted to regulate LRP1B (3 miRNAs), MCM6 (4 miRNAs), and PPARG (9 miRNAs). hsa-miR-20a-5p was shared by all three genes, while hsa-miR-1-3p was shared by MCM6 and PPARG (Supplementary Table S10). These miRNAs predicted 34 associated lncRNAs, among which KCNQ1OT1, NEAT1, MALAT1, OIP5-AS1, and XIST showed frequent interactions. The mRNA–miRNA–lncRNA network included 181 interactions among 3 mRNAs, 14 miRNAs, and 34 lncRNAs (Fig. 6c, Supplementary Table S11).

A total of 8 types of cells were identified in single-cell analysis

Raw data from the GSE184362 single-cell dataset consisted of 65,907 cells and 23,050 genes. After QC processing, 10,661 cells and 27,522 genes were retained (Fig. 7a–b). The top 2,000 HVGs were identified, with the top 10 including IGLC2, IGLC3, IGKC, IGHA2, S100A2, IGHG1, CCL17, IGHA1, CCL21, and IGHG4 (Fig. 7c). Based on the scree plot, 21 PCs were selected for further analysis (Fig. 7d–e).

Fig. 7.

Single-cell RNA sequencing reveals 8 cell types and identifies key PPARG-expressing fibroblasts (a, b) Violin plots showing QC metrics before and after filtering of cells (c) Top HVGs among 2,000 identified (d, e) PCA scree plots identifying 21 principal components (f) UMAP plot showing 28 cell clusters (g) Annotation of clusters into 8 cell types (h) Marker gene expression supporting cell type annotations (i) Bubble plot showing prognostic gene expression; PPARG is highly expressed in fibroblasts

After dimensionality reduction and clustering, 28 cell clusters were identified (Fig. 7f), which were subsequently annotated into 8 cell types: myeloid cells, T cells, NK cells, thyroid cells, fibroblasts, endothelial cells, B cells, and plasma cells (Fig. 7g), supported by marker gene expression (Fig. 7h). Among these, PPARG exhibited high expression in fibroblasts, identifying them as the key cell type for this study (Fig. 7i).

Pseudotime analysis results of fibroblasts

Before pseudotime analysis, dimensionality reduction and clustering of fibroblasts identified 5 clusters using 13 PCs (Fig. 8a). Pseudotime trajectory analysis revealed that fibroblasts branched from an initial developmental state (Fig. 8b). Cluster 1 represented the early stage, clusters 0 and 2 were at the terminal stages, and clusters 3 and 4 corresponded to transitional phases. During differentiation, the expression of LRP1B and MCM6 remained stable, while PPARG expression initially remained constant and then gradually increased along pseudotime (Fig. 8c).

Fig. 8.

Pseudotime trajectory reveals fibroblast subclusters and PPARG dynamics (a) t-SNE plot showing 5 fibroblast clusters after PCA-based dimensionality reduction (13 PCs) (b) Pseudotime trajectory reveals differentiation from cluster 1 (early stage) toward clusters 0 and 2 (terminal), with clusters 3 and 4 as transitional states (c) Expression dynamics of LRP1B, MCM6, and PPARG across pseudotime; PPARG increases gradually during fibroblast differentiation

Cell communication analysis of fibroblasts

In the intercellular communication network, fibroblasts exhibited strong and frequent interactions with other cells, ranking highly in both communication frequency and intensity (Fig. 9a–b). The strongest communication occurred between fibroblasts and thyroid cells, followed by interactions between fibroblasts and NK cells (Fig. 9c). NK cells also showed frequent interactions with other cell types. Fibroblasts received numerous signals from thyroid and NK cells, with the top ligand–receptor pairs being CD99–CD99, COL1A2–SDC4, COL1A2–CD44, and FN1–CD44 (Fig. 9d–e).

Fig. 9.

Cell–cell communication analysis highlights fibroblast interactions in the tumor microenvironment (a, b) Interaction network highlights strong communication between fibroblasts and thyroid cells, followed by NK cells (c) Heatmaps showing overall number and strength of interactions; fibroblasts exhibit high communication activity (d, e) Top ligand–receptor pairs involving fibroblasts include CD99–CD99, COL1A2–SDC4, COL1A2–CD44, and FN1–CD44

Discussion

TC, the most prevalent endocrine malignancy, has shown a rising incidence worldwide [1]. Recent insights into host-microbe interactions suggest a potential role of the GM in cancer development through immune regulation, metabolic modulation, and chronic inflammation [7]. The present study integrated MR, bulk and single-cell transcriptomics, and computational modeling to explore the relationship between GM and TC. Three key prognostic genes (LRP1B, MCM6, and PPARG) were identified, and a robust risk model was constructed. Multi-layered analyses of functional pathways, immune infiltration, drug sensitivity, regulatory networks, and cell-specific expression provided novel insights into TC pathogenesis and therapeutic vulnerabilities.

Among the identified genes, MCM6 and LRP1B exhibited expression patterns suggesting distinct functional roles in TC progression, potentially influenced by tumor stage. MCM6, a component of the minichromosome maintenance complex, is crucial for DNA replication initiation and cell cycle regulation. In this study, MCM6 was upregulated in the low-risk group and positively associated with overall survival, suggesting a context-dependent tumor-suppressive role. Pathway analysis linked MCM6 to butanoate metabolism, raising the possibility that it modulates butyrate synthesis or catabolism. Butyrate has been reported to have anticancer effects, including suppression of inflammation and tumor growth [41], and can regulate gene expression through epigenetic mechanisms such as KEAP1 methylation and NRF2-ARE signaling inhibition [42, 43], thereby enhancing its anticancer effects. We therefore speculate that MCM6 may influence TC progression by regulating key enzymes involved in butyrate synthesis or catabolism, ultimately affecting intracellular butyrate levels. By contrast, MCM6 is often overexpressed in hepatocellular, colorectal, and breast cancers, where it correlates with proliferation and poor prognosis [44]. This discrepancy may be attributable to differences in study design or tumor microenvironment, and may indicate dual roles of MCM6—metabolic regulation in TC versus canonical DNA replication in other cancers. In differentiated TC, where proliferation is relatively slow, MCM6 may primarily function to maintain genomic stability [45]. Overall, these findings underscore the context-specific functions of MCM6, which warrant validation through functional assays and multi-omics approaches.

LRP1B, a member of the low-density lipoprotein receptor family, is generally considered a tumor suppressor due to its role in inhibiting cell migration and modulating extracellular matrix interactions [46]. However, emerging evidence has highlighted its complex behavior, with studies [47, 48] suggesting that high LRP1B expression may be associated with poor prognosis in certain contexts. Consistently, in our study, LRP1B was identified as a risk factor (HR > 1, p < 0.05), supporting its potential pro-tumorigenic role in TC. Moreover, its expression varied significantly across T, N, and pathological stages, suggesting a stage-specific risk profile and further underscoring its clinical significance. PPARG, a nuclear hormone receptor, regulates lipid metabolism, glucose homeostasis, and adipocyte differentiation. Its role in cancer remains highly context-dependent. On one hand, PPARG has been reported to exert tumor-suppressive effects through anti-inflammatory and pro-differentiation signaling pathways [49]. Short-chain fatty acids (SCFAs) produced by the gut microbiota can enhance PPARG activity [50]. Among them, butyrate promotes the differentiation of naive CD4⁺ T cells into the Th9 subset by increasing transforming growth factor-β (TGF-β) levels and concurrently upregulating PPARG expression [51, 52]. In addition, Filippone et al. reported that sodium propionate suppresses tumor cell growth through activation of PPARG signaling [53]. On the other hand, PPARG activation has been linked to tumor-promoting effects [54]. In oral squamous cell carcinoma (OSCC), PPARG has been shown to promote disease progression by driving Th17 polarization and activating CEBPA/IL-17 C signaling [55]. In our study, high PPARG expression in TC appeared to be associated with a more aggressive phenotype. Given the unique microenvironment of thyroid tissue, the precise mechanisms by which SCFAs influence PPARG activity and their clinical relevance in this context warrant clarification using thyroid-specific experimental models. Collectively, integrating these three genes into our risk model provides prognostic value and mechanistic insights into the interplay between host metabolism, immune modulation, and microbial influences in thyroid carcinogenesis.

Functional enrichment analysis revealed that LRP1B, MCM6, and PPARG are involved in several metabolism- and immunity-related pathways, many of which are strongly associated with GM-derived metabolites. Notably, all three genes were enriched in pathways related to valine, leucine, and isoleucine degradation, oxidative phosphorylation, butanoate metabolism, and propanoate metabolism. These SCFAs, primarily produced by the GM, are known to regulate host energy metabolism, mitochondrial function, and immune responses. For instance, butyrate has been shown to promote anti-inflammatory signaling and influence histone acetylation [56], while propionate modulates T cell differentiation and systemic metabolic pathways [57]. These findings suggest that the prognostic genes may participate in the metabolic regulation of TC, although the precise mechanisms require further investigation. Additionally, LRP1B’s involvement in cytokine-cytokine receptor interaction pathways may reflect microbiota-induced immune modulation. Collectively, these findings imply that the identified genes could act as intermediaries, linking microbial metabolic activity to TC progression through immunometabolic mechanisms.

Immune infiltration analysis revealed distinct immune landscapes between HG and LG. In HG, decreased infiltration of Tregs, and naive B cells, alongside increased monocyte levels, suggests a shift toward a pro-inflammatory and less immune-regulated state. The observed negative correlation between LRP1B and Tregs, as well as between LRP1B and MCM6, may indicate indirect regulatory relationships, though current evidence does not support direct antagonism. High MCM6 expression may be associated with a more immune-permissive microenvironment and a lower risk, while LRP1B may contribute to poor prognosis, potentially by promoting immune suppression or influencing cell cycle regulation. Notably, immune, stromal, and ESTIMATE scores were significantly reduced in HG, reflecting a phenotype of impaired immune function and a weakened anti-tumor immune response. Moreover, widespread downregulation of immune checkpoints in HG except for IDO2, CD27, and NRP1 suggests potential immune escape mechanisms. These findings indicate that the three-gene signature not only stratifies prognosis but also reflects underlying immunological states that may influence tumor initiation and progression, potentially mediated in part by microbiota-related immune regulation.

The risk score was significantly associated with differential drug sensitivity, highlighting its potential utility in guiding personalized therapy. Notably, IC50 values for 30 drugs showed strong correlations with the risk score, with agents such as A.443,654, AICAR, and AZD8055 exhibiting significantly lower IC50 values in LG, indicating higher sensitivity. These drugs target pathways related to AKT signaling [58], AMPK activation [59], and mTOR inhibition [60], which may intersect with the metabolic and proliferative pathways enriched in our gene set. This suggests that tumors in LG may be more responsive to metabolic and kinase-targeted therapies. Additionally, the construction of an mRNA–miRNA–lncRNA regulatory network revealed potential upstream regulators of the prognostic genes. Hsa-miR-20a-5p, a known miRNA involved in immune and metabolic regulation [61], was predicted to target al.l three genes, suggesting it may function as a central modulator in the risk network. Hsa-miR-1-3p, shared by MCM6 and PPARG, has been implicated in tumor suppression and differentiation [62]. Several interacting lncRNAs, including KCNQ1OT1, MALAT1, NEAT1, and XIST, play roles in cancer cell proliferation, immune escape, and chromatin regulation. These non-coding RNAs may influence gene expression and treatment response through competing endogenous RNA (ceRNA) mechanisms. Together, these findings highlight a complex regulatory axis that may not only impact tumor biology but also influence therapeutic vulnerability, offering potential targets for RNA-based interventions in TC.

Pseudotime and cell communication analyses further revealed that fibroblasts play an active role in TC progression, both as signal senders and receivers. They particularly interact with thyroid and NK cells through ECM-related ligand-receptor pairs, such as COL1A2–CD44 and FN1–CD44. Among the three prognostic genes, PPARG exhibited a gradual increase during fibroblast differentiation, suggesting its involvement in changes to fibroblast functional states, such as activation or stromal remodeling, although further studies are needed to confirm its specific role.

In conclusion, our integrative approach provides a multifaceted view of TC progression shaped by genetic, metabolic, immune, and microenvironmental factors. LRP1B, MCM6, and PPARG not only serve as prognostic markers but also converge on distinct oncogenic processes.

Limitations

• The GM GWAS data were derived primarily from European populations, which may restrict the generalizability of our findings to other ethnic groups.

• The prognostic risk model based on LRP1B, MCM6, and PPARG showed lower performance in the internal validation set compared with the training set, suggesting possible overfitting.

• Although the nomogram demonstrated favorable performance, its generalizability requires further confirmation. Future studies should therefore include more diverse cohorts to assess the robustness of the identified biomarkers and models across populations, and validate model performance in larger, independent clinical datasets.

• Functional studies, such as gene knockout and in vitro/in vivo assays, are needed to elucidate the biological roles of these genes in TC.

Conclusion

This study integrated transcriptomic and scRNA-seq data to identify prognostic genes in TC potentially linked to GM. MR revealed 27 GM traits associated with TC, and intersecting SNP-mapped genes with DEGs yielded 39 candidate genes. LRP1B, MCM6, and PPARG emerged as key prognostic markers through Cox regression and machine learning analyses. A risk model based on these genes effectively stratified patient prognosis and was incorporated into a predictive nomogram. Enrichment and immune infiltration analyses revealed distinct molecular and immunological landscapes between risk groups. Drug sensitivity analysis identified compounds with differential efficacy, providing insights into personalized therapy. Single-cell data highlighted fibroblasts as key contributors to TC progression. These findings offer a multi-dimensional perspective on TC pathogenesis and may inform future prognostic and therapeutic strategies.

Electronic Supplementary Material

Below is the link to the electronic supplementary material.

Abbreviations

GM

Gut microbiota

TC

Thyroid cancer

MR

Mendelian randomization

IVs

Instrumental variables

OR

Odds ratio

CI

Confidence Interval

DEGs

Differentially expressed genes

HR

Hazard ratio

HG

High-risk group

LG

Low-risk group

ROC

Receiver operating characteristic

AUC

Area under the curve

Author contributions

X.W. conducted the data analysis and wrote the manuscript. B.W. participated in data preprocessing and figure preparation. S.D. contributed to literature review and result interpretation. T.W. assisted with statistical analysis and manuscript editing. W.Q. and L.Z. supervised the project and revised the manuscript. W.Q. and L.Z. are co-corresponding authors. All authors have read and approved the final manuscript.

Funding

The study was supported by the Key Research and Development Projects of Shaanxi Province (grant numbers:2015SF186).

Data availability

The datasets supporting the findings of this study are available in publicly accessible repositories. The thyroid cancer transcriptome and clinical data were obtained from The Cancer Genome Atlas (TCGA-THCA) database (https://portal.gdc.cancer.gov/). The single-cell RNA sequencing data of papillary thyroid carcinoma were retrieved from the Gene Expression Omnibus (GEO) under accession number GSE184362 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE184362). The gut microbiota genome-wide association study (GWAS) summary data were sourced from the MiBioGen Consortium. The thyroid cancer GWAS summary data were obtained from the IEU GWAS database (study ID: EBI-A-GCST90018929, https://gwas.mrcieu.ac.uk/). All datasets used in this study are publicly available and were accessed on February 8, 2025. No new sequencing data was generated in this study; all analyses were based on previously published datasets.

Declarations

Ethics approval and consent to participate

This study only involved publicly available data and did not involve any experiments on humans or animals. Ethical approval was not required.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.