HSPA5 Expression and Ferroptosis Related Transcriptomic Programs in Follicular Thyroid Carcinoma Across Three GEO Cohorts

Yinghan Liu¹,  Yongfu Zhao^1*

¹ Department of Thyroid Surgery, The Second Hospital of Dalian Medical University, Dalian, Liaoning Province, China

*Correspondence: Yongfu Zhao, zhaoyongfuwork@163.com

 

Abstract

Background: Follicular thyroid carcinoma (FTC) and follicular thyroid adenoma (FTA) share morphological and molecular features. HSPA5/GRP78 can stabilize GPX4 protein and suppress ferroptosis in pancreatic cancer, but the corresponding transcriptomic relationship in FTC is unknown.

Objective: To quantify HSPA5 expression in FTC versus FTA, evaluate HSPA5-GPX4 co-expression and KEGG ferroptosis pathway enrichment, and assess the reproducibility of HSPA5-based discrimination across independent public cohorts.

Methods: Depositor-processed expression matrices and official platform annotations were obtained from NCBI GEO for GSE82208 (27 FTC, 25 FTA; GPL570), GSE27155 (13 FTC, 10 FTA; GPL96), and GSE29315 (9 FTC, 17 FTA; GPL8300). Analyses were performed within cohort. Unambiguous probes were collapsed by highest median signal. HSPA5 differences were tested using two-sided Welch tests with Benjamini-Hochberg correction, summarized as Hedges g, and synthesized by random-effects meta-analysis. HSPA5-only ROC analyses used a direction fixed in GSE82208. FTC-only HSPA5-GPX4 Spearman correlations and preranked enrichment of KEGG hsa04216 were evaluated.

Results: HSPA5 expression was lower in FTC in GSE82208 (Hedges g = -0.855, 95% CI -1.416 to -0.293; P = 0.00269; FDR = 0.0811) and GSE27155 (g = -0.977, 95% CI -1.825 to -0.129; P = 0.0170; FDR = 0.275), but not significantly different in GSE29315 (g = -0.208, 95% CI -0.992 to 0.577; P = 0.607; FDR = 0.938). The pooled standardized effect was -0.708 (95% CI -1.135 to -0.281; P = 0.00117; I2 = 9.6%). Fixed-direction AUCs were 0.741, 0.769, and 0.569, respectively. HSPA5-GPX4 correlations were negative in all cohorts (rho = -0.380, -0.604, and -0.350), but none remained significant after within-cohort multiplicity correction. KEGG ferroptosis enrichment was not significant after correction in any cohort (FDR = 0.508, 0.117, and 0.355).

Conclusion: The available transcriptomic data do not support the proposed model of increased HSPA5 expression coupled to GPX4 and ferroptosis suppression in FTC. HSPA5 tended to be lower in FTC, HSPA5-GPX4 co-expression was inverse rather than positive, pathway enrichment was not reproducible, and single-gene discrimination was inconsistent. These findings should be validated at the protein and functional levels before mechanistic or clinical interpretation.

Keywords: follicular thyroid carcinoma; follicular thyroid adenoma; HSPA5; GRP78; GPX4; ferroptosis; transcriptomics; GEO

1. Introduction

Follicular thyroid carcinoma (FTC) is a malignant neoplasm of thyroid follicular epithelial origin and a recognized subtype of differentiated thyroid cancer. Because follicular tumors may share architectural and cytological features, differentiating FTC from follicular thyroid adenoma (FTA) remains challenging. The 2022 World Health Organization classification emphasizes invasive growth and the distinction between malignant, low-risk, and benign follicular-derived tumors [1]. Definitive diagnosis of conventional FTC requires histological demonstration of capsular and/or vascular invasion; fine-needle aspiration cannot reliably establish this distinction because invasion cannot be assessed in aspirated cells [2]. The reclassification of certain encapsulated follicular-patterned tumors illustrates why precise histological terminology is essential when assembling molecular cohorts [3].

Invasive FTC can cause hematogenous metastasis and recurrence, despite the generally favorable outcomes of differentiated thyroid cancer. Gene-expression studies have explored signatures that distinguish FTC from FTA, although reported diagnostic performance remains imperfect [4]. The landmark TCGA-THCA genomic analysis predominantly characterized papillary thyroid carcinoma (PTC) rather than providing a representative FTC discovery cohort [5]. More recent reviews of the WHO framework emphasize biologically and diagnostically meaningful distinctions among follicular-derived thyroid neoplasms [6]. Accordingly, FTC-specific evidence should not be inferred automatically from studies dominated by PTC.

Ferroptosis was initially described as an iron-dependent, nonapoptotic mode of regulated cell death [7]. Subsequent work established that peroxidation-prone membrane phospholipids, iron availability, and antioxidant defenses jointly determine susceptibility to this process [8]. A later review placed these mechanisms in a broader biological and disease context [9]. Glutathione peroxidase 4 (GPX4) protects cells against ferroptosis by reducing lipid hydroperoxides [10]. ACSL4 contributes to sensitivity by shaping membrane fatty-acid composition [11], while oxidation of arachidonic- and adrenic-acid-containing phosphatidylethanolamines contributes to ferroptotic injury [12]. Ferroptosis suppressor protein 1 (FSP1) can provide a coenzyme Q-dependent protective route parallel to GPX4 [13]; an independent study likewise characterized its glutathione-independent activity [14].

In cancer settings, GPX4 dependence has been reported in therapy-resistant mesenchymal states [15] and drug-tolerant persister populations [16]. Cysteine depletion induced ferroptotic tumor injury in a pancreatic cancer model [17]. In separate experimental systems, activated CD8+ T cells were linked to tumor ferroptosis during immunotherapy [18], and radiotherapy combined with immunotherapy enhanced ferroptosis-associated lipid oxidation [19]. The tumor suppressor p53 can also promote ferroptotic susceptibility through repression of the cystine transporter subunit SLC7A11 [20]. These context-dependent findings motivate examination of ferroptosis-associated programs in FTC but do not establish that the same effects occur in this histological subtype.

The endoplasmic reticulum (ER) serves as a central organelle for protein folding and adaptive stress responses. Heat shock protein family A member 5 (HSPA5), also called glucose-regulated protein 78 (GRP78) or binding immunoglobulin protein (BiP), is a major ER-resident chaperone and unfolded protein response (UPR) regulator [21]. In cancer, its biological actions extend beyond protein quality control to include stress adaptation, apoptosis modulation, tumor cell survival, and drug response [22]. Importantly, experimental work in pancreatic ductal adenocarcinoma demonstrated that ATF4-dependent HSPA5 expression protects GPX4 from protein degradation, limits lipid peroxidation, and increases resistance to ferroptotic death [23]. These findings supply a mechanistic rationale for investigating HSPA5 in other epithelial malignancies, although they cannot establish an identical mechanism in FTC.

Evidence linking ferroptosis to thyroid tumor biology is emerging. In one published thyroid cancer study, GPX4 was associated with adverse clinical features and functional inhibition of GPX4 increased ferroptosis-related phenotypes in cultured thyroid cancer cells [24]. However, studies pooling histologically distinct thyroid cancers cannot resolve whether FTC exhibits the same molecular configuration. Furthermore, a finding that HSPA5 and GPX4 mRNA levels covary would not by itself demonstrate that HSPA5 stabilizes GPX4 protein; the originally reported interaction was post-translational [23]. An FTC-focused study must therefore distinguish direct observations in FTC tissue from mechanism imported from other tumor types.

Public gene-expression datasets offer an opportunity to investigate this question without obtaining new patient specimens. The GEO series GSE82208 includes profiles from 27 FTC and 25 FTA samples; GSE27155 provides 13 FTC and 10 FTA samples among several thyroid histotypes; and GSE29315 includes 9 FTC and 17 FTA samples. GSE82208 was linked to prior work on distinguishing malignant from benign follicular thyroid tumors [4]. These repositories represent starting points for an FTC-versus-FTA comparison, provided that sample identities, platform coverage, pathological classification, and potential cohort overlap are checked before analysis. Because the series were assembled for different research questions, they should not be presented as a harmonized prospective cohort.

Accordingly, we conducted an FTC-specific multi-cohort analysis with HSPA5 as a prespecified candidate. The primary objective was to quantify and replicate the HSPA5 expression difference between FTC and FTA. Secondary objectives were to evaluate FTC-only HSPA5-GPX4 co-expression, test enrichment of the KEGG human ferroptosis pathway, and estimate whether HSPA5 alone discriminated FTC from FTA. The working hypothesis was that HSPA5 might be linked to an anti-ferroptotic transcriptional context in FTC; the analysis was designed to report negative or contradictory findings without changing the prespecified direction after observing validation data.

2. Materials and Methods

2.1. Study design and reporting framework

This retrospective secondary analysis used de-identified public microarray data. The primary unit of analysis was the patient-derived tissue expression profile. Each series was analyzed separately because the platforms, probe sets, normalization procedures, and expression scales differed. Results were reported following the principle that observational transcriptomic associations cannot establish ferroptotic flux, GPX4 protein stabilization, or causality.

2.2. Data retrieval, cohort selection, and histopathological definitions

The NCBI Gene Expression Omnibus (GEO) accession records and files were accessed on 13 October 2026 [26]. GSE82208 comprised 27 FTC and 25 FTA samples on GPL570; GSE27155 contributed 13 FTC and 10 FTA samples on GPL96 after exclusion of other thyroid histologies; and GSE29315 contributed 9 FTC and 17 FTA samples on GPL8300 after exclusion of other histologies. The analyzed set therefore contained 49 FTC and 52 FTA expression profiles. Sample groups were reconstructed from GEO titles, source names, descriptions, and characteristics and were required to match the deposited counts before analysis.

All FTC and FTA samples that passed metadata verification were retained; no expression sample was excluded. GEO sample accession identifiers were unique across the three series. Cross-study patient overlap could not be excluded conclusively from the deposited metadata, so pooled analyses were treated as exploratory. Other thyroid histotypes, normals, thyroiditis, and hyperplasia samples were excluded from the primary contrast. No survival analysis was performed because the series did not provide adequate FTC-specific follow-up.

2.3. Platform annotation, preprocessing, and quality assurance

The official NCBI platform annotation files GPL570.annot.gz, GPL96.annot.gz, and GPL8300.annot.gz were read between the platform-table delimiters. Probes mapping to multiple symbols were excluded. For genes represented by more than one unambiguous probe, the probe with the highest median signal across all samples in that cohort was selected independently of phenotype. GSE82208 used GCRMA log2-normalized intensities, GSE27155 used a custom log10-transformed quantile-normalized signal, and GSE29315 used RMA-normalized intensities [27]; native mean differences were therefore not pooled.

The processed matrices contained no missing values among the selected FTC and FTA samples after gene-level mapping. The primary HSPA5 probes were 211936_at for GPL570 and GPL96 and 36614_at for GPL8300. The selected GPX4 probes were 201106_at for GPL570 and GPL96 and 33931_at for GPL8300. A sensitivity analysis evaluated every unambiguous HSPA5 probe rather than selecting a probe by differential-expression significance.

2.4. HSPA5 differential expression and cross-cohort effect estimation

Within each cohort, HSPA5 expression in FTC and FTA was compared by a two-sided Welch t test. Benjamini-Hochberg FDR was calculated across all mapped genes [28], while the nominal P value was retained because HSPA5 was prespecified. Raw mean differences and 95% Welch intervals were reported on each depositor-specific scale. Standardized differences were expressed as Hedges g with an approximate 95% CI.

Cohort-specific Hedges g estimates were synthesized using a DerSimonian-Laird random-effects model. Heterogeneity was described using Cochran Q, tau2, and I2. Because only three modest-sized cohorts were available and patient overlap could not be excluded conclusively, the pooled estimate was interpreted as an exploratory cross-platform summary rather than definitive replication.

2.5. Ferroptosis gene-set definition and biological interpretation

The human KEGG Ferroptosis pathway hsa04216 was downloaded from the official KEGG REST record on 13 October 2026 and contained 42 unique gene symbols [30]. FerrDb V3 provides broader regulated-cell-death annotations [25], but it was not substituted for the version-locked KEGG set in this analysis. Depending on platform coverage, 40, 36, and 32 pathway genes were detected in GSE82208, GSE27155, and GSE29315, respectively.

Genes were ranked within cohort by the Welch t statistic for FTC versus FTA. Following the general preranked gene-set enrichment framework [29], a weighted enrichment score was calculated for hsa04216. Significance and normalized enrichment scores were estimated from 10,000 random gene sets matched to the number of detected pathway genes; P values were adjusted across the three cohorts by the Benjamini-Hochberg method. This competitive transcriptomic test does not measure ferroptosis activity or distinguish pro-ferroptotic from anti-ferroptotic pathway components.

2.6. HSPA5-centered co-expression and network exploration

Within FTC samples, Spearman correlations were calculated between HSPA5 and each detected KEGG ferroptosis gene. Benjamini-Hochberg correction was applied across the tested pathway genes separately in each cohort. GPX4 was the prespecified correlation of interest. An exploratory random-effects synthesis of GPX4 correlations used Fisher z transformation and inverse-variance weighting.

Correlation analysis was restricted to FTC to avoid inducing an association through the FTC-versus-FTA group difference. Correlation coefficients describe mRNA co-expression only and cannot establish HSPA5-GPX4 protein binding, GPX4 stabilization, or shared cellular origin.

2.7. Exploratory discrimination of FTC from FTA

HSPA5 expression was evaluated as a single-marker classifier. The score direction was fixed from the discovery cohort GSE82208 and applied unchanged to GSE27155 and GSE29315; because HSPA5 was lower in FTC in GSE82208, lower expression was assigned the positive FTC direction. AUCs and 95% percentile intervals were calculated using 5,000 stratified bootstrap replicates with random seed 20261013.

2.8. Sensitivity analyses and reproducibility

Probe sensitivity analysis included all unambiguous HSPA5 probes, with complete results reported in Table 5. Input file sizes and SHA-256 checksums, sample-level group tables, selected probes, complete gene-level statistics, correlation results, leading-edge genes, software versions, and analysis parameters were preserved in the accompanying analysis bundle. The workflow used Python 3 with pandas, NumPy, SciPy, scikit-learn, and Matplotlib.

No post hoc threshold selection, direction reversal, imputation, cross-platform pooling of raw intensities, or exclusion based on HSPA5 values was performed. All three cohorts and all planned outcomes were reported irrespective of statistical significance or agreement with the original hypothesis.

2.9. Statistical considerations and research ethics

All tests were two-sided except the directional tail used to evaluate the signed enrichment score against matched random gene sets. Statistical significance was defined as FDR < 0.05 for genome-wide differential expression, pathway enrichment, and within-cohort correlation panels. Effect estimates, confidence intervals, nominal P values, and adjusted P values were reported together. The analysis of public de-identified data involved no intervention or contact with participants.

3. Results

3.1. Cohort selection and assay compatibility

Three GEO series satisfied the prespecified requirement for both FTC and FTA samples (Table 1; Figure 1). GSE82208 contributed 52 samples, GSE27155 contributed 23 of 99 series samples after histology restriction, and GSE29315 contributed 26 of 71 series samples. The final analysis included 49 FTC and 52 FTA profiles. Exact class counts matched the public series descriptions, and no selected sample contained missing expression values after mapping.

The platforms yielded 20,844, 12,501, and 8,631 unambiguously mapped representative genes, respectively. Their distinct dynamic ranges confirmed that raw expression values should not be pooled. All downstream comparisons were therefore performed within cohort, with standardized effects used only for cross-platform synthesis.

Figure 1. Analyzed cohort composition and prespecified workflow. Panel A shows the exact FTC and FTA samples retained from each GEO series. Panel B summarizes platform-specific mapping, HSPA5 testing, fixed-direction ROC analysis, FTC-only GPX4 correlation, and KEGG ferroptosis enrichment.

3.2. HSPA5 expression in FTC and FTA

HSPA5 was lower in FTC in all three cohorts, although precision and statistical support differed (Table 2; Figure 2). In GSE82208, mean expression was 12.814 in FTC and 13.721 in FTA (difference -0.907, 95% CI -1.482 to -0.332; P = 0.00269; FDR = 0.0811; Hedges g = -0.855, 95% CI -1.416 to -0.293). In GSE27155, the corresponding means were 4.157 and 4.321 (difference -0.164, 95% CI -0.295 to -0.033; P = 0.0170; FDR = 0.275; g = -0.977, 95% CI -1.825 to -0.129). GSE29315 showed a smaller, imprecise difference (9.473 versus 9.690; difference -0.216, 95% CI -1.087 to 0.655; P = 0.607; FDR = 0.938; g = -0.208, 95% CI -0.992 to 0.577).

Figure 2. HSPA5 expression in FTA and FTC. Points represent individual samples; horizontal and vertical black marks show the median and interquartile range. Values remain on each depositor-normalized platform scale. P values are from two-sided Welch tests, and FDR values are Benjamini-Hochberg adjusted across mapped genes.

3.3. Validation and possible diagnostic discrimination

The random-effects pooled standardized effect was -0.708 (95% CI -1.135 to -0.281; P = 0.00117), with low estimated heterogeneity (Q = 2.21, I2 = 9.6%, tau2 = 0.014; Figure 3). This summary was directionally consistent but was driven mainly by GSE82208 and GSE27155; the GSE29315 interval included no effect. The HSPA5-only AUC was 0.741 (95% bootstrap CI 0.599-0.868) in GSE82208, 0.769 (0.554-0.938) in GSE27155, and 0.569 (0.294-0.824) in GSE29315 when the lower-expression direction learned in GSE82208 was retained (Figure 4). The wide and overlapping intervals, particularly in validation cohorts, did not support a robust single-gene diagnostic test.

Figure 3. Standardized HSPA5 effects across cohorts. Negative Hedges g indicates lower HSPA5 expression in FTC than FTA. Horizontal lines denote 95% confidence intervals; the diamond is the DerSimonian-Laird random-effects estimate.

Figure 4. HSPA5-only discrimination of FTC from FTA. The classification direction was learned in GSE82208, where lower HSPA5 indicated FTC, and applied unchanged to both validation cohorts. Parentheses show 95% stratified-bootstrap intervals from 5,000 resamples.

3.4. HSPA5-associated ferroptosis transcriptional features

Within FTC samples, HSPA5 and GPX4 mRNA levels were inversely correlated in all three cohorts (Table 3; Figure 5): GSE82208 rho = -0.380 (P = 0.0507; within-cohort FDR = 0.185), GSE27155 rho = -0.604 (P = 0.0287; FDR = 0.275), and GSE29315 rho = -0.350 (P = 0.356; FDR = 0.670). None remained significant after correction across detected KEGG ferroptosis genes. The exploratory random-effects estimate was rho = -0.438 (95% CI -0.652 to -0.158; P = 0.00297; I2 = 0%), which was opposite to the anticipated positive co-expression pattern and should not be interpreted as a direct regulatory interaction.

Figure 5. HSPA5-GPX4 co-expression within FTC samples. Each point is one FTC sample; lines are visual linear trends, whereas the annotated coefficients and P values are Spearman statistics. None of the correlations remained significant after within-cohort correction across detected KEGG ferroptosis genes.

3.5. Pathway enrichment and network context

The KEGG ferroptosis pathway was not significantly enriched after correction in any cohort (Table 4; Figure 6). GSE82208 showed a negative but null enrichment score (NES = -0.834; permutation P = 0.508; FDR = 0.508). GSE27155 showed nominal positive enrichment (NES = 1.381; P = 0.0389), but this did not remain significant after correction (FDR = 0.117) and was not reproduced. GSE29315 also showed nonsignificant positive enrichment (NES = 0.974; P = 0.237; FDR = 0.355). Thus, the pathway analysis provided no reproducible evidence of a ferroptosis-associated transcriptional shift in FTC.

Figure 6. Preranked enrichment of KEGG hsa04216 Ferroptosis. Genes were ranked by the FTC-versus-FTA Welch t statistic; positive positions favor higher expression in FTC. Red ticks mark detected pathway genes. P values were estimated from 10,000 matched random gene sets, and FDR values were adjusted across cohorts. No cohort met FDR < 0.05.

4. Discussion

This multi-cohort analysis produced a result that contradicted the initial mechanistic expectation. HSPA5 was not increased in FTC; its expression was lower than in FTA in all three cohorts, with statistically informative effects in GSE82208 and GSE27155 but an imprecise near-null effect in GSE29315. The pooled standardized estimate supported an overall negative association, yet the absence of genome-wide FDR significance for HSPA5 within any individual cohort and the small number of studies require restraint. The most defensible conclusion is that lower HSPA5 transcript abundance is associated with FTC in two datasets, not that HSPA5 is a universal FTC biomarker.

The direction of the HSPA5-GPX4 relationship also failed to support the proposed anti-ferroptotic transcript model. Zhu and colleagues showed that HSPA5 can protect GPX4 from protein degradation in pancreatic cancer cells [23], a post-translational mechanism that does not require positively correlated transcripts. In the present FTC samples, correlations were negative in every cohort and did not survive within-cohort multiplicity correction. The pooled inverse estimate may indicate shared differentiation or stress-state variation, cell-composition effects, or chance correlation in small cohorts; it does not demonstrate that HSPA5 represses GPX4.

HSPA5 is a central endoplasmic-reticulum chaperone and unfolded-protein-response regulator [21]. Its cancer-related stress-adaptation functions have also been reviewed [22]. Reduced HSPA5 mRNA in FTC relative to FTA could reflect altered follicular differentiation, secretory activity, tumor-cell composition, or platform-specific probe behavior rather than reduced protein function. Conversely, HSPA5 protein abundance or localization may diverge from transcript measurements through translational control, stress-dependent stabilization, and cellular redistribution. Protein-level assays are therefore necessary before the transcript difference can be interpreted biologically.

The pathway analysis similarly did not establish a coordinated ferroptosis program. No cohort met the FDR threshold for enrichment of KEGG hsa04216. The nominal signal in GSE27155 was not observed in the discovery cohort and was not significant after correction. Because KEGG ferroptosis contains genes with different mechanistic roles, a positive enrichment score would indicate coordinated ranking rather than ferroptotic activity. The present null and heterogeneous results are compatible with limited power, incomplete legacy-array coverage, or predominantly post-transcriptional regulation.

The ROC analysis provided only limited diagnostic evidence. HSPA5 achieved moderate AUCs in GSE82208 and GSE27155 when lower expression was used to identify FTC, but performance approached chance in GSE29315 and all validation intervals were wide. These findings do not justify a clinical threshold or a claim that HSPA5 alone can distinguish carcinoma from adenoma. Multigene approaches have been explored in follicular thyroid tumors [4], but developing a new classifier would require a separate prespecified design with nested validation to limit optimism.

Probe sensitivity supported the direction of the primary result (Table 5). GPL570 contained two unambiguous HSPA5 probes, and both were significantly lower in FTC by nominal Welch testing; GPL96 and GPL8300 each contributed one unambiguous HSPA5 probe. The primary rule selected the highest-median probe without using phenotype labels, reducing the risk of choosing the most favorable probe. Nevertheless, microarray cross-hybridization and annotation changes remain possible sources of bias.

A strength of the study is the use of three public cohorts with exact sample-level phenotype verification and platform-specific analysis. Fixed ROC direction, outcome-independent probe selection, complete reporting of negative results, input checksums, and release of gene-level outputs improve reproducibility. The convergence of the first two datasets is notable because they use different processed-data scales, although statistical pooling cannot remove biological and pathological heterogeneity.

Several limitations constrain inference. The cohorts are small, retrospective, and based on legacy microarrays. Original diagnoses may not have been rereviewed under the 2022 WHO framework [1], and detailed invasion, metastatic, treatment, and survival variables were unavailable. Cross-study patient overlap could not be excluded conclusively from public metadata. Bulk tissue expression is also sensitive to tumor purity and stromal or immune admixture. Finally, the DerSimonian-Laird variance estimate is unstable with only three cohorts.

The study addresses FTC versus FTA status rather than disease progression. A difference between malignant and benign follicular tumors cannot establish that HSPA5 controls proliferation, invasion, metastasis, recurrence, or survival. The available data therefore do not support prognostic or therapeutic claims. Prospective diagnostic samples and clinically annotated FTC cohorts would be required to assess generalizability and clinical utility.

Most importantly, transcriptomics cannot test the defining biochemical events of ferroptosis. The matrices contain no measurements of GPX4 protein turnover, glutathione, labile iron, lipid peroxidation, oxidized phospholipids, or rescue by ferroptosis inhibitors. HSPA5 perturbation in authenticated FTC models, combined with GPX4 protein and lipid-ROS assays and ferroptosis-specific rescue experiments, would be needed to evaluate causality.

The negative and inconsistent findings refine the original hypothesis. Rather than supporting an HSPA5-high, GPX4-linked anti-ferroptotic state, the data indicate lower HSPA5 transcript abundance in two cohorts, inverse HSPA5-GPX4 co-expression, and no reproducible KEGG ferroptosis enrichment. Future work should investigate whether the transcript decrease reflects thyroid differentiation or cellular composition before testing a direct ferroptosis mechanism.

These results also illustrate why candidate mechanisms derived from other tumor types should be evaluated without assuming conservation across histologies. The HSPA5-GPX4 protein-stability mechanism remains biologically credible in pancreatic cancer [23], but the present FTC tissue data do not reproduce its predicted transcriptional pattern. A mechanistic model should therefore remain conditional until supported by FTC-specific perturbation experiments.

An appropriate experimental extension would quantify HSPA5 and GPX4 proteins in independently reviewed FTC and FTA tissues and examine their cellular localization. In FTC cell or organoid models, HSPA5 depletion or overexpression should be paired with GPX4 turnover, lipid reactive oxygen species, oxidized phospholipid profiling, and viability under ferroptosis-inducing conditions. Rescue by ferrostatin-1 or liproxstatin-1 and orthogonal apoptosis controls would be required to establish specificity.

Overall, the analysis provides evidence against overinterpreting HSPA5 as an anti-ferroptotic FTC driver on the basis of public mRNA data. The observed lower expression may still have biological or diagnostic relevance, but its inconsistency in GSE29315 and limited discrimination require external validation. Reporting this direction honestly offers a more reliable foundation for subsequent protein-level and functional work than preserving the original positive hypothesis.

5. Conclusion

Across three GEO cohorts comprising 49 FTC and 52 FTA samples, HSPA5 transcript expression was consistently lower in FTC, but individual-cohort FDR thresholds were not met and one validation cohort was imprecise. HSPA5-GPX4 co-expression was inverse, and KEGG ferroptosis enrichment was not reproducible or significant after correction. HSPA5 showed moderate discrimination in two cohorts but near-chance performance in the third. These data do not support an HSPA5-high anti-ferroptotic mechanism or clinical single-gene biomarker in FTC; independent protein-level, functional, and prospective validation is required.

Tables

Table 1. NCBI GEO cohorts included in the FTC-versus-FTA analysis.

Accession	Analysis role	FTC	FTA	Platform
GSE82208	Discovery	27	25	GPL570
GSE27155	Validation 1	13	10	GPL96
GSE29315	Validation 2	9	17	GPL8300
Total	Cross-platform summary	49	52	3 platforms

Abbreviations: FTC, follicular thyroid carcinoma; FTA, follicular thyroid adenoma. Other histologies in GSE27155 and GSE29315 were excluded.

Table 2. Cohort-specific HSPA5 differential-expression and fixed-direction ROC results.

Cohort	n FTC / FTA	HSPA5 probe	Hedges g (95% CI)	P / FDR	AUC (95% CI)
GSE82208	27 / 25	211936_at	-0.855 (-1.416, -0.293)	0.00269 / 0.0811	0.741 (0.599, 0.868)
GSE27155	13 / 10	211936_at	-0.977 (-1.825, -0.129)	0.017 / 0.275	0.769 (0.554, 0.938)
GSE29315	9 / 17	36614_at	-0.208 (-0.992, 0.577)	0.607 / 0.938	0.569 (0.294, 0.824)

Hedges g is standardized as FTC minus FTA. AUC direction was fixed in GSE82208 as lower HSPA5 expression indicating FTC and was applied unchanged to both validation cohorts.

Table 3. HSPA5-GPX4 Spearman correlations within FTC samples.

Cohort	FTC n	Spearman rho	P value	Within-cohort FDR
GSE82208	27	-0.380	0.0507	0.185
GSE27155	13	-0.604	0.0287	0.275
GSE29315	9	-0.350	0.356	0.67
Random-effects	49	-0.438	0.00297	Not applicable

The exploratory pooled rho was -0.438 (95% CI -0.652 to -0.158; I2 = 0.0%). Within-cohort FDR values were calculated across all detected KEGG ferroptosis genes.

Table 4. Preranked enrichment of KEGG hsa04216 Ferroptosis.

Cohort	Detected genes	ES	NES	Permutation P	BH FDR
GSE82208	40	-0.228	-0.834	0.508	0.508
GSE27155	36	0.389	1.381	0.0389	0.117
GSE29315	32	0.266	0.974	0.237	0.355

ES, enrichment score; NES, normalized enrichment score. P values were estimated from 10,000 matched random gene sets and adjusted across cohorts.

Table 5. Sensitivity analysis of all unambiguous HSPA5 probes.

Cohort	Probe	Primary	FTC-FTA difference	Hedges g	P value
GSE82208	211936_at	Yes	-0.907	-0.855	0.00269
GSE82208	230031_at	No	-1.514	-1.006	0.000557
GSE27155	211936_at	Yes	-0.164	-0.977	0.017
GSE29315	36614_at	Yes	-0.216	-0.208	0.607

Differences are reported on each cohort’s depositor-specific expression scale.

Declarations

Ethics approval: Not applicable because this secondary analysis used only de-identified, publicly available GEO data and involved no direct interaction with human participants.

Consent to participate and publication: Not applicable; no new participants or identifiable individual data were included.

Data availability: The depositor-processed matrices and platform annotations are available from NCBI GEO under GSE82208, GSE27155, and GSE29315. The accompanying reproducibility bundle contains the executed analysis code, SHA-256 input manifest, sample-group audit, probe mappings, complete gene-level results, pathway and correlation tables, software versions, and figure files.

Funding: The authors received no external funding for this work.

Conflicts of interest: The authors declare no competing interests.

Author contributions: Y.L.: Conceptualization, Methodology, Software, Formal analysis, Investigation, Data curation, Visualization, Writing – original draft. Y.Z.: Conceptualization, Validation, Supervision, Project administration, Writing – review and editing. Both authors read and approved the final manuscript.

Artificial intelligence disclosure: Generative artificial intelligence was used to assist language editing and document preparation. All data analyses were executed on the cited public datasets using the accompanying reproducible code, and the authors remain responsible for verification, interpretation, and the final manuscript.

References

1. Baloch ZW, Asa SL, Barletta JA, et al. Overview of the 2022 WHO classification of thyroid neoplasms. Endocr Pathol. 2022;33:27–63. doi:10.1007/s12022-022-09707-3.
2. Haugen BR, Alexander EK, Bible KC, et al. 2015 American Thyroid Association Management Guidelines for Adult Patients with Thyroid Nodules and Differentiated Thyroid Cancer: The American Thyroid Association Guidelines Task Force on Thyroid Nodules and Differentiated Thyroid Cancer. Thyroid. 2016;26(1):1–133. doi:10.1089/thy.2015.0020.
3. Nikiforov YE, Seethala RR, Tallini G, et al. Nomenclature Revision for Encapsulated Follicular Variant of Papillary Thyroid Carcinoma: A Paradigm Shift to Reduce Overtreatment of Indolent Tumors. JAMA Oncol. 2016;2(8):1023–1029. doi:10.1001/jamaoncol.2016.0386.
4. Wojtas B, Pfeifer A, Oczko-Wojciechowska M, et al. Gene expression (mRNA) markers for differentiating between malignant and benign follicular thyroid tumours. Int J Mol Sci. 2017;18:1184. doi:10.3390/ijms18061184.
5. Cancer Genome Atlas Research Network. Integrated genomic characterization of papillary thyroid carcinoma. Cell. 2014;159:676–690. doi:10.1016/j.cell.2014.09.050.
6. Basolo F, Macerola E, Poma AM, Torregrossa L. The 5th edition of WHO classification of tumors of endocrine organs: changes in the diagnosis of follicular-derived thyroid carcinoma. Endocrine. 2023;80(3):470–476. doi:10.1007/s12020-023-03336-4.
7. Dixon SJ, Lemberg KM, Lamprecht MR, et al. Ferroptosis: an iron-dependent form of nonapoptotic cell death. Cell. 2012;149:1060–1072. doi:10.1016/j.cell.2012.03.042.
8. Stockwell BR, Friedmann Angeli JP, Bayir H, et al. Ferroptosis: a regulated cell death nexus linking metabolism, redox biology, and disease. Cell. 2017;171:273–285. doi:10.1016/j.cell.2017.09.021.
9. Jiang X, Stockwell BR, Conrad M. Ferroptosis: mechanisms, biology and role in disease. Nat Rev Mol Cell Biol. 2021;22:266–282. doi:10.1038/s41580-020-00324-8.
10. Yang WS, SriRamaratnam R, Welsch ME, et al. Regulation of ferroptotic cancer cell death by GPX4. Cell. 2014;156:317–331. doi:10.1016/j.cell.2013.12.010.
11. Doll S, Proneth B, Tyurina YY, et al. ACSL4 dictates ferroptosis sensitivity by shaping cellular lipid composition. Nat Chem Biol. 2017;13:91–98. doi:10.1038/nchembio.2239.
12. Kagan VE, Mao G, Qu F, et al. Oxidized arachidonic and adrenic phosphatidylethanolamines navigate cells to ferroptosis. Nat Chem Biol. 2017;13:81–90. doi:10.1038/nchembio.2238.
13. Bersuker K, Hendricks JM, Li Z, et al. The CoQ oxidoreductase FSP1 acts parallel to GPX4 to inhibit ferroptosis. Nature. 2019;575:688–692. doi:10.1038/s41586-019-1705-2.
14. Doll S, Freitas FP, Shah R, et al. FSP1 is a glutathione-independent ferroptosis suppressor. Nature. 2019;575:693–698. doi:10.1038/s41586-019-1707-0.
15. Viswanathan VS, Ryan MJ, Dhruv HD, et al. Dependency of a therapy-resistant state of cancer cells on a lipid peroxidase pathway. Nature. 2017;547:453–457. doi:10.1038/nature23007.
16. Hangauer MJ, Viswanathan VS, Ryan MJ, et al. Drug-tolerant persister cancer cells are vulnerable to GPX4 inhibition. Nature. 2017;551:247–250. doi:10.1038/nature24297.
17. Badgley MA, Kremer DM, Maurer HC, et al. Cysteine depletion induces pancreatic tumor ferroptosis in mice. Science. 2020;368:85–89. doi:10.1126/science.aaw9872.
18. Wang W, Green M, Choi JE, et al. CD8+ T cells regulate tumour ferroptosis during cancer immunotherapy. Nature. 2019;569:270–274. doi:10.1038/s41586-019-1170-y.
19. Lang X, Green MD, Wang W, et al. Radiotherapy and immunotherapy promote tumoral lipid oxidation and ferroptosis via synergistic repression of SLC7A11. Cancer Discov. 2019;9:1673–1685. doi:10.1158/2159-8290.CD-19-0338.
20. Jiang L, Kon N, Li T, et al. Ferroptosis as a p53-mediated activity during tumour suppression. Nature. 2015;520:57–62. doi:10.1038/nature14344.
21. Hetz C. The unfolded protein response: controlling cell fate decisions under ER stress and beyond. Nat Rev Mol Cell Biol. 2012;13:89–102. doi:10.1038/nrm3270.
22. Lee AS. Glucose-regulated proteins in cancer: molecular mechanisms and therapeutic potential. Nat Rev Cancer. 2014;14:263–276. doi:10.1038/nrc3701.
23. Zhu S, Zhang Q, Sun X, et al. HSPA5 regulates ferroptotic cell death in cancer cells. Cancer Res. 2017;77:2064–2077. doi:10.1158/0008-5472.CAN-16-1979.
24. Chen H, Peng F, Xu J, Wang G, Zhao Y. Increased expression of GPX4 promotes the tumorigenesis of thyroid cancer by inhibiting ferroptosis and predicts poor clinical outcomes. Aging (Albany NY). 2023;15:230–245. doi:10.18632/aging.204473.
25. Zhou N, Peng L, Luo Q, et al. FerrDb V3: expanding the manually curated resource for regulators and disease associations from ferroptosis to regulated cell death. Nucleic Acids Res. 2026;54:D572–D582. doi:10.1093/nar/gkaf1119.
26. Barrett T, Wilhite SE, Ledoux P, et al. NCBI GEO: archive for functional genomics data sets—update. Nucleic Acids Res. 2013;41:D991–D995. doi:10.1093/nar/gks1193.
27. Irizarry RA, Hobbs B, Collin F, et al. Exploration, normalization, and summaries of high density oligonucleotide array probe level data. Biostatistics. 2003;4:249–264. doi:10.1093/biostatistics/4.2.249.
28. Benjamini Y, Hochberg Y. Controlling the false discovery rate: a practical and powerful approach to multiple testing. J R Stat Soc B. 1995;57:289–300. doi:10.1111/j.2517-6161.1995.tb02031.x.
29. Subramanian A, Tamayo P, Mootha VK, et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc Natl Acad Sci USA. 2005;102:15545–15550. doi:10.1073/pnas.0506580102.
30. Kanehisa M, Goto S. KEGG: Kyoto Encyclopedia of Genes and Genomes. Nucleic Acids Res. 2000;28:27–30. doi:10.1093/nar/28.1.27.