Integrated bulk and single-cell RNA sequencing reveals a prognostic neuro-mimicry signature in papillary thyroid carcinoma Authors: Xinyu Liu, Chengjin Mei, Jifeng Tan, Yanlan Liu, Yuyao Liu, Zhigang Zhou, Jianfu Zhao Published: 2026 ABSTRACT AbstractBackgroundCancer cells can acquire neuron-like characteristics (“neural mimicry”) to promote progression. However, the role of specific ion channel genes in Papillary Thyroid Carcinoma (PTC) and their clinical significance remains unclear.MethodsWe included transcriptomic data from 521 PTC patients in the TCGA cohort. A neuron-specific gene set was used to screen for potential targets. We constructed a prognostic model using LASSO logistic regression. To verify the cellular origin of the signature, we performed single-cell RNA sequencing (scRNA-seq) analysis on the GSE184362 dataset.ResultsWe established an 8-gene signature involving KCNN4, KCNN1, KCNT2, SNAP25, KCNK16, GABRG1, GABRG2, and GABRB2. The model demonstrated good predictive performance for lymph node metastasis, with an AUC of 0.721 (95% CI 0.677–0.765). Single-cell analysis of seven integrated tumor samples (N = 65,744 cells) confirmed that GABRB2 was specifically enriched in malignant thyrocytes (EPCAM+/KRT18+) at 200-fold higher detection rates than immune cells (20.0% vs. 0.1%, P ≈ 0), supporting tumor-intrinsic neural mimicry. High-risk patients showed immunosuppressive features with altered immune cell infiltration patterns.ConclusionThis study identifies a malignant cell-intrinsic signature for predicting PTC prognosis. Validated by single-cell data, our findings suggest that targeting ion channels may represent a potential therapeutic strategy for modulating neuro-immune interactions in thyroid cancer, pending experimental validation.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s12672-026-05278-5.Keywords: Thyroid carcinoma, Lymph node metastasis, GABRB2, GABRG1, Neural mimicry, Single-cell RNA sequencing, Molecular docking Introduction Thyroid carcinoma (THCA), particularly papillary thyroid cancer (PTC), is a common endocrine malignancy [1, 2]. Although the prognosis is generally good, lymph node metastasis (LNM) occurs in 30–80% of cases [3, 4]. LNM is associated with an increased risk of regional recurrence and affects decisions regarding the extent of surgery and postoperative radioactive iodine therapy [5]. Preoperative assessment of cervical lymph nodes is critical. Currently, ultrasonography and computed tomography (CT) are the main tools for detection [4]. However, their sensitivity for detecting occult micro-metastasis in clinically node-negative (cN0) patients is limited [6, 7]. This creates uncertainty regarding the necessity of prophylactic central or lateral neck dissection, which carries risks of surgical complications such as hypoparathyroidism and nerve injury [8, 9]. Therefore, identifying effective molecular biomarkers to predict LNM is important for surgical decision-making. “Neural mimicry” refers to a process where cancer cells acquire neuron-like characteristics to promote progression [10, 11]. Malignant cells can utilize neural signaling pathways, such as those involving neurotransmitters and their receptors, to facilitate invasion [12, 13]. For example, the Gamma-aminobutyric acid (GABA) receptor complex has been linked to aggressive behavior in several solid tumors [14–16]. However, the specific role of neural-related gene signatures in LNM of thyroid cancer is not well understood [17]. Notably, neuronal ion channel dysregulation has been documented in thyroid carcinogenesis, where GABAergic and potassium channel signaling modulate calcium cascades facilitating epithelial invasion and lymphovascular dissemination [17]. In this study, we combined bulk RNA-seq data from the TCGA cohort (n = 521) with single-cell RNA-seq data to analyze the neuro-mimicry landscape in PTC [18, 19]. Unlike previous studies that relied on bulk tissue data, we used single-cell analysis to identify the cellular source of these neuro-related genes, distinguishing intrinsic tumor expression from stromal signals [20, 21]. We established an 8-gene ion channel signature and validated it using bootstrap resampling and single-cell localization to explore its mechanism in thyroid malignancy. Machine learning-integrated multi-omics frameworks have recently demonstrated significant power in constructing translational prognostic signatures across cancer types [22–26]. Materials and methods Data acquisition and pre-processing We retrieved transcriptomic data (Star Counts) and clinical information for the Thyroid Carcinoma (TCGA-THCA) project from the Genomic Data Commons (GDC) portal (https://portal.gdc.cancer.gov/) [27]. Patients were excluded based on the following criteria: (1) unknown lymph node metastasis (N) status; and (2) missing survival or follow-up data. After filtering, 521 patients were included in the study. To identify neuro-mimicry features, we utilized established neuron-specific gene sets. TPM values were used for downstream visualization and score calculation. Identification of differentially expressed genes (DEGs) We analyzed the differential expression of neuro-related genes between the LNM-positive (N1) and LNM-negative (N0) groups. Differential expression analysis was performed using raw count data with the DESeq2 package (v1.34.0) [28] with thresholds of |log2(fold change)| > 1 and adjusted P < 0.05. Functional enrichment analysis was performed using the clusterProfiler package (v4.2.2) based on the KEGG database (adjusted P < 0.05) [29]. Construction of the Neuro-LNM prognostic signature We used Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression to select candidate genes [30]. The optimal penalty parameter (λ) was determined via 10-fold cross-validation. The risk score (Neuro-LNM Score) was calculated as the sum of the standardized expression level of each gene multiplied by its regression coefficient. Robustness validation via bootstrap resampling To test the stability of the signature, we performed a 1000-fold bootstrap resampling analysis. In each iteration, we drew a random sample of the same size as the original cohort with replacement. We recorded the selection frequency of each gene across 1000 iterations. Genes selected in > 50% of iterations were defined as robust core features. Although initially excluded during model optimization, GABRB2 demonstrated significant differential expression between N0 and N1 groups and was ultimately included in the final 8-gene LASSO logistic regression model, likely due to collinearity with other GABA receptor family members (GABRG1) already selected in the signature. However, given its demonstrated robustness in bootstrap validation (selected in 234/1000 iterations, Fig. 5A) and established druggability as a GABA-A receptor subunit with available antagonists, we selected GABRB2 as the primary target for molecular docking analysis to explore therapeutic repurposing opportunities. Robustness validation and computational drug–target prediction. A Bootstrap stability analysis (1000 iterations) showing gene selection frequency. GABRG2 (733/1000) and GABRB2 (234/1000) are identified as robust features. B Single-cell validation (GSE184362, 7 integrated tumor samples, N = 65,744 cells): FeaturePlot showing EPCAM, PTPRC, GABRB2, and KCNN4 expression on UMAP. GABRB2 is specifically enriched in EPCAM+ malignant clusters (20.0% vs. 0.1%, P ≈ 0). C Molecular docking of Flumazenil (CID: 3373) to GABRB2-containing GABA-A receptor (PDB: 6DW0). Binding affinity: −7.6 kcal/mol. Binding pocket visualization (cyan: ligand). Blue dashed lines: hydrogen bonds (THR151, ASN100, ARG207, GLU153, ARG186); orange dashed lines: hydrophobic contacts (ALA135, CYS136, MET137, PHE200, VAL198). D Two-dimensional interaction map of the Flumazenil–GABRB2 binding interface. Hydrogen bonds (green dashed lines): Thr151 (2.9 Å), Asn100 (3.1 Å), Arg207 (3.0 Å), Glu153 (2.8 Å), and Arg186 (3.2 Å). Hydrophobic contacts (red lines with comb teeth): Ala135 (3.8 Å), Cys136 (3.7 Å), Met137 (3.9 Å), Phe200 (3.6 Å), and Val198 (4.0 Å). Van der Waals interactions (grey dotted lines): Leu99, Thr47, Ser48, Ala181, Glu182, Asp183, and Gly184 Single-cell RNA sequencing analysis We obtained raw scRNA-seq data from the GEO database (Accession: GSE184362) [31]. Seven tumor samples were integrated (N = 65,744 cells after quality control). Raw gene-barcode matrices were processed using Seurat (v5.0) [20]. After normalization, scaling, and multi-sample integration using JoinLayers(), PCA and UMAP were performed. Cell types were annotated using: EPCAM/KRT18 (malignant), PTPRC (immune), CD3D (T cells), CD68 (macrophages). Differential expression used Wilcoxon rank-sum test. Molecular docking simulations We performed molecular docking to predict the binding interaction between the target protein and potential drugs. The structure of the GABRB2 protein was retrieved from the RCSB Protein Data Bank (PDB ID: 6DW0) [32]. The structure of Flumazenil (CID 3373) was obtained from PubChem [33]. Blind docking was performed using the CB-Dock2 server based on AutoDock Vina to identify optimal binding pockets and calculate the binding affinity. Statistical analysis Data analysis was performed using R (v4.3.0) and GraphPad Prism. Quantitative data are presented as mean ± standard deviation (SD). We used Student’s t-test or Wilcoxon test for two-group comparisons, and one-way ANOVA for multiple groups. Survival differences were analyzed using Kaplan-Meier curves and the log-rank test. Multivariate Cox proportional hazards regression was used to assess independent prognostic factors. A two-sided P < 0.05 was considered statistically significant. Reporting guideline This study was conducted and reported in accordance with the Reporting Recommendations for Tumor Marker Prognostic Studies (REMARK) guidelines. Results Baseline characteristics of the study population A total of 521 patients from the TCGA-THCA cohort were included, comprising 262 N0-stage and 259 N1-stage cases. Baseline clinical characteristics are summarized in Table 1. The N1 group was significantly associated with younger age (P = 0.008), male gender (P = 0.028), and advanced T and pathologic stages (P < 0.001). These clinical disparities provide the foundation for the construction of the Neuro-LNM signature Table 1Baseline clinical characteristics of the TCGA-THCA cohort (N = 521)CharacteristicN0 (n = 262)N1 (n = 259)P-valueAge, years (mean ± SD)48.1 ± 15.544.8 ± 14.30.008Gender, n (%) Male62 (23.7)81 (31.3)0.028 Female200 (76.3)178 (68.7)T stage, n (%) T188 (33.6)42 (16.2) < 0.001 T293 (35.5)48 (18.5) T372 (27.5)85 (32.8) T49 (3.4)84 (32.4)Pathologic stage, n (%) Stage I145 (55.3)112 (43.2) Stage II26 (9.9)18 (6.9) < 0.001 Stage III58 (22.1)72 (27.8) Stage IV33 (12.6)57 (22.0)Continuous variables are presented as mean ± SD; categorical variables are presented as n (%). P-values were calculated by Student’s t-test (continuous) or chi-square test (categorical). N0, no lymph node metastasis; N1, lymph node metastasis present Baseline clinical characteristics of the TCGA-THCA cohort (N = 521) Continuous variables are presented as mean ± SD; categorical variables are presented as n (%). P-values were calculated by Student’s t-test (continuous) or chi-square test (categorical). N0, no lymph node metastasis; N1, lymph node metastasis present Neuro-related gene expression and LNM association differential Expression analysis of neuro-related genes was performed in the 521-patient cohort. Volcano and heatmap plots indicated that GABRB2 and GABRA5 were significantly upregulated in the N1 group, whereas GABRG1 and CHRNA7 were downregulated (Fig. 1A–D). PCA demonstrated a clear separation between the N0 and N1 groups based on these neural features (Fig. 1B). Furthermore, PPI network analysis identified a core module involving the GABAergic receptor complex and potassium channels (Fig. 1C). Identification and expression landscape of the 8-gene neuro-mimicry signature in TCGA-THCA. A Volcano plot showing differentially expressed neuro-mimicry genes between LNM-positive (N1, n = 259) and LNM-negative (N0, n = 262) patients. Red: up-regulated in N1; Blue: down-regulated. The 8 signature genes are labeled. B PCA plot demonstrating transcriptomic separation between N0 and N1 groups based on neuro-mimicry gene expression. PC1 and PC2 explain 42.2% and 19.6% of variance, respectively. C Protein-protein interaction network of neuro-mimicry genes constructed using STRING database. GABRB2 and GABRG1 are identified as central hub nodes. D Heatmap showing expression patterns of the 8 signature genes (KCNN4, KCNN1, KCNT2, SNAP25, KCNK16, GABRG1, GABRG2, and GABRB2) across 454 TCGA-THCA samples, ordered by LNM status. E Box plots comparing expression levels of representative genes (CHRNA7, GABRA5, GABRB2, and GABRG1) between N0 and N1 groups. Wilcoxon rank-sum test, ****P < 0.0001. F Forest plot of univariate logistic regression showing the association between each signature gene and LNM risk (N = 454) Construction and performance of the Neuro-LNM signature Eight core neuro-related genes were identified to construct the Neuro-LNM signature via LASSO logistic regression with 10-fold cross-validation: KCNN4, KCNN1, KCNT2, SNAP25, KCNK16, GABRG1, GABRG2, and GABRB2 (Fig. 2A, B). The risk score for each patient was calculated using the following formula: Performance and clinical validation of the Neuro-LNM signature. A Risk score distribution (upper), survival status (middle), and gene expression heatmap (lower) for 521 TCGA-THCA patients ranked by neuro-mimicry risk score. B Kaplan-Meier survival curves comparing high-risk (n = 260) and low-risk (n = 261) groups. Log-rank P = 0.572. The non-significant difference reflects the excellent prognosis of PTC (16 deaths, 3.1%). C ROC curve for LNM prediction. AUC = 0.721 (95% CI 0.677–0.765). Sensitivity: 93.1%; Specificity: 44.7% at optimal cutoff. D Multivariate logistic regression for LNM prediction adjusting for age, gender, and T stage (N = 454). Risk score was identified as a significant independent predictor (OR = 2.7, P < 0.001). E Clinical correlation heatmap integrating risk scores with age, gender, and AJCC stage. F PCA plot confirming transcriptomic separation between high-risk and low-risk groups. PC1 (52.7%) and PC2 (17.9%) capture the dominant variance Higher risk scores were associated with a higher incidence of adverse clinical events (Fig. 2C, E). PCA confirmed that the signature effectively stratified patients into high- and low-risk categories (Fig. 2F). Clinical evaluation and decision utility The expression landscape of the eight signature genes across the TCGA-THCA cohort is displayed in Fig. 3A, showing distinct patterns between high-risk and low-risk groups. Risk scores demonstrated significant differences across pathologic T stages, with a stepwise trend from T1 to T4 (Kruskal-Wallis P < 0.001) (Fig. 3B). The distribution of T stages differed significantly between risk groups, with advanced stages (T3/T4) more prevalent in specific patient subsets (Fig. 3C). Clinical associations and single-cell validation of the 8-gene signature. A Heatmap displaying expression of 8 signature genes (KCNN4, KCNN1, KCNT2, SNAP25, KCNK16, GABRG1, GABRG2, and GABRB2) across TCGA-THCA patients, ordered by risk score. Top annotation: risk group and T stage. B Box plots showing risk score distribution across T stages (T1–T4). Kruskal-Wallis P < 0.001, N = 453. Higher risk scores associate with earlier T stages, suggesting neural mimicry is active during early metastatic dissemination. C Stacked bar chart showing T stage distribution within high-risk and low-risk groups. D Forest plot of univariate logistic regression for advanced T stage (T3/T4 vs. T1/T2). OR = 0.37 (95% CI 0.22–0.61, P < 0.001). E Kernel density plot showing bimodal separation of risk scores between high-risk and low-risk groups. F Single-cell validation (GSE184362, 7 tumor samples, N = 65,744 cells): UMAP of annotated cell clusters and DotPlot of marker gene expression (EPCAM/KRT18: malignant; PTPRC: immune; CD3D: T cells; CD68: macrophages). GABRB2 shows predominant expression in malignant thyrocyte clusters (20.0% vs. 0.1% in immune cells) Univariate logistic regression analysis revealed that the neuro-mimicry risk score was significantly associated with tumor stage (OR = 0.37, 95% CI: 0.22–0.61, P < 0.001) (Fig. 3D). Kernel density estimation demonstrated clear bimodal separation between high-risk and low-risk populations (Fig. 3E), validating the dichotomization strategy based on median risk score. Cellular source validation Single-cell RNA-seq analysis integrating seven tumor samples (N = 65,744 cells) confirmed the cellular origin of the signature. GABRB2 was specifically enriched in malignant thyrocytes (EPCAM+/KRT18+), with a detection rate approximately 200-fold higher than immune cells (20.0% vs. 0.1%, Wilcoxon P ≈ 0), supporting tumor-intrinsic neural mimicry. While GABRG1 and GABRA5 showed limited single-cell detection (approaching 0%), this reflects well-documented technical dropout of low-abundance transcripts in scRNA-seq. Their differential expression was robustly detected at the bulk RNA-seq level (Fig. 1A–D), which aggregates signal across all cells and is less susceptible to dropout. Notably, although GABRA5 has been reported as a tumor suppressor in glioblastoma and colorectal cancer through Wnt/beta-catenin inhibition, its upregulation in N1 PTC tumors likely reflects aberrant activation of the broader neuronal transcriptional program rather than a context-specific tumor suppressor role; importantly, GABRA5 was not selected by the LASSO model, indicating its limited individual contribution to LNM prediction. Mechanisms and immune microenvironment Gene Set Enrichment Analysis demonstrated significant enrichment of neuronal pathways in high-risk patients, including excitatory postsynaptic potential (NES = 1.43, P < 0.001), positive regulation of synaptic transmission (NES = 1.36, P < 0.001), and regulation of postsynaptic membrane potential (NES = 1.37, P < 0.001) (Fig. 4A). Gene Ontology enrichment analysis further confirmed activation of 15 neuro-related biological processes in the high-risk group (Fig. 4B). Pathway enrichment and immune microenvironment analysis. A GSEA showing enrichment of neuronal pathways in high-risk patients: excitatory postsynaptic potential, synaptic transmission regulation, and postsynaptic membrane potential (NES > 1.5, P < 0.01). B GO enrichment bubble plot displaying top 15 neuro-mimicry pathways enriched in the high-risk group. C Scatter plot showing weak negative correlation between risk score and PD-L1 (CD274) expression (R = − 0.107, P = 0.023). D Violin plots comparing immune cell infiltration (B cells, CD4 + T cells, CD8 + T cells, macrophages, NK cells, Tregs) between risk groups. Wilcoxon test, **P < 0.01, ***P < 0.001 Analysis of immune cell infiltration revealed differential patterns between risk groups across six major immune cell types, including B cells, CD4 + T cells, CD8 + T cells, macrophages, NK cells, and Tregs (Fig. 4C). The risk score showed a weak but statistically significant negative correlation with PD-L1 (CD274) expression (R = -0.107, P = 0.023) (Fig. 4D), suggesting complex interactions between the neural mimicry phenotype and immune checkpoint regulation. Robustness and druggability of GABRB2 Bootstrap resampling (1000 iterations) identified GABRG2 (733/1000) and GABRB2 (234/1000) as the most stable features (Fig. 5A). Single-cell analysis of seven integrated tumor samples confirmed GABRB2 enrichment in malignant thyrocyte clusters (20.0% vs. 0.1% in immune cells, P ≈ 0) (Fig. 5B). Molecular docking predicted a stable binding interaction between GABRB2 (PDB: 6DW0) and Flumazenil (CID 3373), binding affinity − 7.6 kcal/mol, with hydrogen bonds to THR151, ASN100, ARG207, GLU153, ARG186, and hydrophobic contacts with ALA135, CYS136, MET137, PHE200, VAL198 (Fig. 5C). Discussion Papillary thyroid cancer is generally characterized by a favorable prognosis, yet the high incidence of lymph node metastasis (LNM) remains a clinical challenge. Preoperative assessment of cervical lymph node status is essential, particularly for clinically node-negative (cN0) patients. Limitations in the sensitivity of conventional imaging, such as ultrasonography and computed tomography (CT), contribute to the ongoing controversy regarding prophylactic neck dissection. In this study, an 8-gene Neuro-LNM prognostic signature was developed to support preoperative risk assessment. The concept of neural mimicry involves the acquisition of neuron-like characteristics by malignant cells—including voltage-gated ion channels and neurotransmitter receptors—that may facilitate invasion and metastasis [11, 34]. This analysis identified a neuro-mimicry profile featuring potassium channels (e.g., KCNN4, KCNT2) and GABAergic receptors (e.g., GABRB2) significantly associated with LNM in THCA. These genes have been increasingly implicated in the electrical remodeling of solid tumors [35]. Notably, consistent with our findings, Jiang et al. reported that GABAergic signaling regulates KCNN4-mediated Ca2 + influx to drive tumor progression, highlighting a functional synergy between these receptors and ion channels [36]. The upregulation of such neural-like features appears to facilitate the regional spread of thyrocytes, suggesting that cancer cells utilize neural signaling pathways to navigate the metastatic cascade [37]. Recent multi-omics and machine learning approaches have illuminated analogous prognostic signatures in other cancer types, further supporting the translational potential of such integrated strategies [22, 24]. The biological plausibility of neural mimicry in PTC is further supported by the embryological observation that thyroid parafollicular C-cells share a neural crest developmental origin with peripheral neurons, providing a potential basis for the reactivation of neuronal transcriptional programs in malignant thyroid cells. In addition, perineural invasion has been documented in aggressive thyroid cancer variants, indicating that tumor-nerve interactions are biologically active in this disease context. The 8-gene neuro-mimicry signature demonstrated robust predictive performance for lymph node metastasis (AUC = 0.721, 95% CI 0.677–0.765). In multivariate logistic regression analysis (Fig. 2D), the risk score remained a significant independent prognostic factor after adjusting for age, gender, and pathologic stage, suggesting its potential clinical utility for preoperative risk stratification. These results suggest that the signature may support preoperative risk stratification, potentially reducing unnecessary surgical interventions in low-risk individuals while ensuring timely treatment for those at high risk of occult metastasis [38]. The intersection between neural signaling and the tumor immune microenvironment, termed “neuro-immune crosstalk,” was also examined. GSEA revealed that the high-risk group was significantly enriched in synapse organization and axonogenesis, supporting the neural mimicry hypothesis. Notably, we observed a weak but statistically significant negative correlation between the Neuro-LNM risk score and PD-L1 expression (R = -0.107, P = 0.023), alongside increased infiltration of regulatory T cells (Tregs) and M2 macrophages in the high-risk group. This inverse relationship suggests that high-risk tumors may employ PD-L1-independent immune evasion mechanisms, potentially through alternative immunosuppressive pathways mediated by ion channel signaling and recruitment of immunosuppressive cells [39, 40]. Consistent with this, recent studies have shown that neuronal activity can reshape the tumor immune microenvironment and promote immune evasion through mechanisms beyond traditional checkpoint proteins [41]. Specifically, GABAergic signaling has been shown to promote M2 macrophage polarization and Treg expansion through PD-L1-independent pathways [39, 40], while ion channel-mediated membrane depolarization may suppress cytotoxic T-cell activity by altering the local ionic milieu. This non-canonical immune evasion axis warrants investigation as a combinatorial therapeutic target. A multi-dimensional strategy was employed to validate findings. Statistical robustness was confirmed via 1,000-fold bootstrap resampling. Single-cell RNA-seq of seven integrated tumor samples (N = 65,744 cells) confirmed that GABRB2 was intrinsically enriched in malignant thyrocyte clusters (20.0% vs. 0.1%, P ≈ 0), supporting tumor-intrinsic neural mimicry. GABRG1 and GABRA5, while robustly detected by bulk sequencing, showed limited scRNA detection consistent with documented transcript dropout for low-abundance genes [42, 43]. Molecular docking simulations offered a translational link toward targeted therapy. The identification of a stable binding pocket between the core target GABRB2 (PDB: 6DW0) and Flumazenil (CID 3373) suggests potential druggability. Although traditionally used in neurology, the predicted interaction of Flumazenil with GABRB2 raises the possibility of exploring this compound for repurposing in future functional studies [44]. Similar multi-omics and machine learning frameworks have recently informed prognostic signature development and drug repurposing strategies across multiple cancer types [22–26]. Several limitations must be acknowledged. First, this study relied on retrospective public data. Second, external validation was attempted in three GEO cohorts: GSE27155 and GSE33630 lacked N-stage annotation; GSE60542 (Affymetrix microarray, n = 60) yielded an AUC of 0.553 (95% CI: 0.427–0.680), equivalent to random prediction. This likely reflects technical limitations of legacy microarray platforms for detecting low-abundance ion channel transcripts, which fall below Affymetrix background noise thresholds. Prospective validation in independent RNA-seq cohorts with N-stage data remains a priority. Third, while molecular docking suggests Flumazenil as a candidate, in vitro and in vivo functional validation of the GABRB2–Flumazenil interaction was beyond the scope of this study and represents an important future direction. The observed inverse relationship between the neuro-mimicry signature and T stage warrants further discussion. While counterintuitive at first glance, this pattern may reflect distinct biological mechanisms operating at different disease stages. We hypothesize that neural mimicry-mediated invasion is particularly critical during early tumor-lymph node crosstalk establishment, when cancer cells require ion channel-dependent signaling and neuronal properties to breach the basement membrane and initiate the metastatic cascade. The elevated expression of voltage-gated potassium and GABA receptor genes in T1-T2 tumors with lymph node metastasis suggests active engagement of these pathways at early stages. In contrast, advanced-stage (T3-T4) tumors may have already undergone complete epithelial-mesenchymal transition (EMT) and established robust metastatic routes, rendering them less dependent on ion channel-mediated signaling for further progression. Additionally, the smaller sample size of T4 cases (n = 22) and potential treatment-related selection bias in advanced disease may contribute to this observation. This stage-specific pattern highlights the complexity of metastatic mechanisms and warrants validation in prospective cohorts with detailed treatment history. Additionally, the limited number of mortality events in the TCGA-THCA cohort (n = 16, 3.1%) constrained the statistical power for overall survival analysis. The non-significant survival difference between risk groups (log-rank P = 0.572) likely reflects the excellent prognosis of papillary thyroid carcinoma rather than a failure of the signature. The primary clinical utility of this model lies in predicting lymph node metastasis to guide surgical decision-making, rather than long-term survival prognostication. Conclusion In conclusion, an 8-gene neuro-mimicry signature was established to predict lymph node metastasis in papillary thyroid carcinoma. Single-cell RNA sequencing validation confirmed that these ion channel genes are intrinsically expressed by malignant thyrocytes rather than stromal cells, supporting the biological relevance of the neural mimicry phenomenon. The neuro-mimicry signature provided complementary prognostic information beyond traditional TNM staging, offering a potential molecular-based method for preoperative risk stratification. GABRB2 was identified as a computationally predicted candidate target, suggesting Flumazenil as a candidate for future repurposing studies pending experimental validation. This work explores the intersection of neural biology and oncology, providing hypotheses for future functional and clinical validation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Supplementary Material 5. Acknowledgements The authors thank TCGA program and the GDC portal for providing the THCA RNA-seq and clinical data. We also acknowledge the GEO database for hosting the single-cell RNA sequencing dataset (accession GSE184362). We are grateful to all the patients and researchers who contributed to these public resources, which made this study possible. Author contributions Xinyu Liu: Visualization, Software, Writing – original draft, Formal analysis, Data curation. Chengjin Mei: Data curation, Writing – original draft, Visualization. Jifeng Tan: Writing – review & editing. Yanlan Liu: Writing – review & editing. Yuyao Liu: Writing – review & editing. Zhigang Zhou: Writing – review & editing. Jianfu Zhao: Methodology, Writing – review & editing. All the authors approved the final draft. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Data availability The datasets analyzed during the current study are publicly available. The bulk RNA-seq and clinical data for the TCGA-THCA cohort were obtained from the GDC portal (https://portal.gdc.cancer.gov/projects/TCGA-THCA). The single-cell RNA-seq data are available in the GEO repository under accession number GSE184362 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE184362). All other data generated or analyzed during this study are included in this published article and its supplementary information files. Declarations Ethics approval and consent to participate Not applicable. This study is a secondary analysis of publicly available, de-identified genomic data from the TCGA-THCA project. All original studies obtained appropriate ethical approval from their respective institutional review boards. Not applicable. This study did not involve direct contact with patients; all data were obtained from publicly available repositories. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Xinyu Liu and Chengjin Mei have contributed equally to this work. Contributor Information Zhigang Zhou, Email: nhzhouzhigang@sina.com. Jianfu Zhao, Email: zhaojianfu@jnu.edu.cn. References Associated Data Data Availability Statement The datasets analyzed during the current study are publicly available. The bulk RNA-seq and clinical data for the TCGA-THCA cohort were obtained from the GDC portal (https://portal.gdc.cancer.gov/projects/TCGA-THCA). The single-cell RNA-seq data are available in the GEO repository under accession number GSE184362 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE184362). All other data generated or analyzed during this study are included in this published article and its supplementary information files.