Numerous studies have focused on finding one or more genes to build tumor risk assessment models in recent years. At the same time, the impact of ferroptosis on tumor biology is also a hot topic today[13, 14]. Those bring us a new idea for FRGs combinations to construct a reasonable and practical risk assessment model.
This study obtained complete RNA sequence data of 246 FRGs in thyroid cancer through TCGA and the FerrDb databases. Next, we analyzed the relationship between FRGs and the patient's OS and the difference in FRGs expression between cancerous and normal tissues. According to the association between DE-FRGs and the patient's OS, we obtained 14 PR-DE-FRGs. Motivated by this, a risk assessment model with superior performance was proposed. Since this study data only comprised samples from a single cohort of TCGA, we randomly divided the samples into the training set, test set, and whole set to achieve the effect of mutual multi-set verification of the results. We split each training/test/whole set of samples into high-risk and low-risk groups based on the median value of the risk score. We found that low risk is significantly associated with better OS. Surprisingly, when we drew the ROC curve to verify this result, the 1-year, 2-year, and 3-year AUC of the training/test/whole sets were more outstanding than 0.9. The AUC of the risk score is the largest among all prognostic-related factors. It proves that the risk assessment model has exceptional predictive value for the prognosis of patients.
Previous studies have demonstrated that ferroptosis was involved in cancer development and severely impacted the prognosis of cancer patients. For example, Liang et al. and Tang et al. found that the prognosis model constructed by ferroptosis-related genes in the high-risk group of Hepatocellular Carcinoma had significantly more inferior OS than the low-risk group [38, 39]. Mou et al. found that when genes' expression level that promotes ferroptosis is low, the OS of clear cell renal cell carcinoma is also inferior . Similar results were found in lung adenocarcinoma by Yao et al.  and Tian et al. . However, few studies have established a risk assessment model based on ferroptosis-related genes to predict the prognosis of THCA. This study used up to 14 genes to build models to minimize the risk of overfitting screening features, accurately predicting THCA patients' prognosis.
These 14 genes can be roughly divided into oxidative metabolism (PGD, DPP4, CDKN2A, MIOX, ARNTL, BID, CAPG, BID), lipid metabolism (ANGPTL7, TXNRD1, GPX4, SRXN1, DRD4), iron metabolism (TFRC, ISCU, DRD4) three categories  . In the risk assessment model, eight genes (ANGPTL7, DRD4, SRXN1, TXNRD1, CDKN2A, MIOX, PGD, and TFRC) were identified as prognostic risk factors (HR > 1), while the remaining six genes were opposite (CAPG, GPX4, ARNTL, ISCU, BID and DPP4) (HR < 1). PGD reduced NADP to NADPH, preventing erastin-induced ferroptosis in Calu-1 cells . Xie et al. found that high expression of DPP4 promotes plasma-membrane-associated DPP4-dependent lipid peroxidation, which ultimately leads to ferroptosis . P53, through the sequestration of DPP4, achieves anti-ferroptotic function . ARF was initially identified as an alternative transcript of the Ink4a tumor suppressor locus that encodes p16INK4a, an inhibitor of CDKN2A . Combination of ARF induction and ROS treatment-induced ferroptotic cell death and Knockdown of endogenous ARF protected cells from ROS-induced cell death . MIOX is a proximal tubular enzyme that promotes ROC production when overexpressed, thereby promoting ferroptosis . ARNTL in the blockade of ferroptotic cancer cell death through control of the EGLN2-HIF1A pathway . CAPG may inhibit ferroptosis by regulating GSH . For lipid metabolism, SRXN1 and TXNRD1's expression were up-regulated during ferroptosis induced by erastin or RSL3 . In contrast, ANGPTL7's expression was downregulated during ferroptosis caused by erastin or RSL3. Those results suggested that the expression of these three genes may be related to cancer ferroptosis, but the specific mechanism of action is still unclear. According to reports, TXNRD1 can enhance the cytotoxicity of Lysine oxidase, which can activate ferroptosis ; this suggests that TXNRD1 may promote ferroptosis. The transcriptional activation of the SRXN1 gene is strongly responsive to oxidative stress , and SRXN1 plays a vital role in removing excess ROS . The study of Kim et al. found that SRNX1 may be related to tumor progression , consistent with our research that SRNX1 is a prognostic risk factor for thyroid cancer. Song et al. also observed a higher expression level of ANGPTL7 in the esophageal squamous cell carcinoma study, and it was used as a prognostic risk gene . Our research results are consistent with these studies. GPX4 is the phospholipid hydroperoxidase that inhibits ferroptosis because it can convert lipid hydroperoxides into non-toxic lipid alcohols. BID is the mediator of mitochondrial dysfunction downstream of glutathione (GSH) depletion in the model of oxytosis . For iron metabolism, dopamine reduced erastin-induced ferrous iron accumulation, glutathione depletion, and malondialdehyde production . This is why DRD4 may inhibit ferroptosis. Transferrin is an iron carrier protein in serum that can be transported into the cell via receptor-mediated endocytosis, RNAi of transferrin receptor (TFRC) inhibited ferroptosis . Transferrin can only interact with TFRC and be transported into the cell when loaded with iron . Overexpression of ISCU (Iron-sulfur cluster assembly enzyme, a mitochondrial protein) significantly attenuated Dihydroartemisinin induced ferroptosis by regulating iron metabolism, rescuing the mitochondrial function, and increasing the level of GSH. Oncogenic mutations in RAS family members (included KRAS, HRAS, and NRAS) are common in 30% of human tumors. Many studies have shown that NRSA gene mutations may be associated with the occurrence of ferroptosis [13, 25]. Dietrich et al. proved that NRAS is overexpressed in hepatocellular carcinoma and is a biomarker of poor patient prognosis and found that it contributes to Sorafenib resistance in hepatocellular carcinoma . Many studies have proven that NRAS mutations in the thyroid are important for cancer, but the relationship between NRAS mutations and patient prognosis has not yet been determined [61, 62]. We confirmed a significant association between NRAS and risk score, but no significant association was found between NRAS and OS in patients with thyroid cancer.
GPX4 is an essential target for studying the mechanism of ferroptosis in cancer and a promising strategy for exploring potential treatments for cancer. At the same time, more and more shreds of evidence showed that miR-1287-5p plays a substantial carcinogenic or anti-cancer effect in human tumors, such as Cervical Cancer, Breast cancer, THCA, and so on. Many previous studies have demonstrated that the miR-1287-5p/GPX4 axis plays a vital role in regulating the apoptosis and ferroptosis of cancer cells. We verified the AL928654.4/ miR-1287-5p/ GPX4 regulatory axis, and the survival curve revealed the high expression of GPX4, the low expression of miR-1287-5p, and the increased expression of AL928654.4, and the better prognosis is relevant. AL928654.4 is significantly negatively correlated with miR-1287-5p and similar to the miR-1287-5p with GPX4. The above results may explain the regulation mode of GPX4 in the prognosis of THCA. The study by Wang et al. showed similar results to ours. We have not found previous studies on AL928654.4, which led to our lack of sufficient evidence, but this also provides us with a new idea for studying the regulatory role of AL928654.4/ miR-1287-5p/ GPX4 in THCA.
Although ferroptosis has become a research focus, there are few studies on ferroptosis from thyroid cancer. Moreover, investigating the biological processes and pathways of thyroid cancer ferroptosis is even more limited. Therefore, based on 14 PR-DE-FRGs, GSEA was performed. Similar to expectations, Adipocytokine signaling pathway, Fatty acid metabolism, Hedgehog signaling pathway, mTOR signaling pathway, regulation of autophagy enriched, and Wnt signaling pathway in the high-risk group are all closely related to ferroptosis. Adipocytokine signaling pathway, Fatty acid metabolism can affect ferroptosis by regulating lipid metabolism . Research by Yang et al. found that the Hedgehog signaling pathway, regulation of autophagy enriched, and Wnt signaling pathway play a vital role in the cancer stem cell process . FZD7 is a classic Wnt receptor that can alter GSH metabolism and protect ovarian cancer from oxidative stress . Studies have shown that promoting the mTOR signaling pathway can induce ferroptosis in trophoblast cells .
Since we found in GO analysis that immune-related biological processes and pathways were enriched, we further analyzed the impact of ferroptosis on immunomodulatory. Surprisingly, except for B cells, CD8 + T cells, and Tfh, the scores of the other 13 immune cells and 13 immune functions in the high-risk group were significantly lower than those in the low-risk group. The study has speculated that ferroptotic cells will release distinct signals, including lipid mediators, which will attract APC and other immune cells to the site of ferroptotically dying cells. A study found that ferroptotic cancer cells are efficiently engulfed by macrophages in vitro. In addition, oxidized lipids and lipid droplets regulate the anti-tumor immune response . Lipid bodies containing oxidatively truncated lipids block antigen cross-presentation by dendritic cells in cancer, leading to anti-tumor immunodeficiency. NK cells play a central role in cancer immune insurance by killing cancer cells. . Many studies have shown that tumor-associated neutrophils and lymphocytic are associated with a better prognosis of tumors, but the specific mechanism is still controversial. In addition, 13 immune-related functions in the high-risk group were significantly impaired. Taken together, these results demonstrate that the poor prognosis of thyroid cancer patients in our high-risk group is probably related to low anti-tumor immunity. Stem cell characteristics reveal new drug targets for anti-cancer therapy. Cancer stem cells (CSCs) usually show high iron levels in the cells . The Ferroptosis pathway can selectively induce CSC death . We found that the low-risk and N0 stage groups had higher mDNAsi scores. This conclusion indicated that cancerous cells in our low-risk group are less aggressive and can invade surrounding tissues, which is also reflected in the study of glioblastoma. .
Another advantage of our research is combining the risk assessment model with clinical applications. For example, the risk score is significantly correlated with the expression of three IPS, five ICIs, and the IC50 of five chemotherapeutics. These results all implied that our risk scoring model could effectively predict the chemosensitivity of patients receiving these five chemotherapy agents and predict the effect of immunotherapy. Clinically, an operational decision support tool was needed. Therefore, we used factors that may affect the prognosis (age, stage, gender, and risk score) to build a Nomogram for predicting thyroid cancer's 1, 3, and 5-year overall survival.
There are several limitations to this study. First, we only obtained the differential expression results of the 14 PR-DE-FRGs used to construct the model by analyzing the sharing data. However, due to the difficulty of collecting clinical specimens and the limitation of experimental conditions, we, unfortunately, failed to verify these gene expression differences through experiments. To make up for this limitation as much as possible, we used images of PR-DE-FRGs protein expression between thyroid cancerous and normal tissues obtained by immunohistochemical staining based on clinical specimens to verify our results. Second, the conclusion of the relationship between the risk score and the tumor microenvironment may need to be verified by experiments with bulk samples and a substantial workload. Third, we have not obtained transcriptome data with survival information in other databases to validate our model. To make up for this shortcoming, we randomly divided the TCGA sample into three sets. Then we verified the model similar to the external set and finally achieved brilliant verification results.