Long-term Survivors of Anaplastic Thyroid Cancer: A Genomic Predictive Model

Abstract

Context:

Longer-term survival is possible for some patients with anaplastic thyroid cancer (ATC). However, genomic factors associated with improved survival are poorly characterized.

Objective:

To develop a mathematical model to predict mutation-based survival risk in ATC.

Design:

Retrospective cohort study of 204 ATC samples from the cBioPortal database, divided into 80% training and 20% validation cohorts. Multivariate analysis identified prognostic genes, used to construct a point-based risk model. KEGG pathway enrichment and BRAF subanalyses were performed.

Setting:

Multi-institutional, international genomic database.

Patients or Other Participants:

Samples were included if sequencing and survival data were available (n = 204).

Intervention(s):

Not applicable.

Main Outcome Measure(s):

The prespecified primary outcome was overall survival.

Results:

Fourteen genes were associated with increased risk: TET1, MAPK12, ATP10A, PIK3CA, MUC4, PNPLA2, PLD4, EGLN2, BSN, FLNC, RADIL, ZMYND8, FRAS1, RECQL4. More aggressive (n = 37) and less aggressive cohorts (n = 128) were determined using the maximally selected rank statistic, yielding a point threshold of 0.27. The predictive performance of the risk model demonstrated a C-index of 0.74. On Kaplan-Meier analysis, 1-year survival differed for more aggressive patients (0%) compared to less aggressive patients (32%). For the validation cohort, survival remained significantly different between risk cohorts and on BRAF subanalysis. Each risk cohort subsequently underwent KEGG pathway enrichment analysis, which showed significantly increased enrichment across several pathways for more aggressive tumors.

Conclusion:

This model identifies mutated genes that are associated with the most aggressive ATCs and thus may aid in preoperative risk assessment when evaluating patients for surgery for curative intent.

Anaplastic thyroid cancers (ATC) remain a rapidly progressive and lethal malignancy with a median survival of 4 to 6 months. Treatment consists of surgery, radiation, and traditional chemotherapy agents such as doxorubicin and cisplatin (1). Targeted therapies have emerged such as dabrafenib and sorafenib; however, varied clinical benefit has been reported from relatively small cohorts (2). Despite an overall poor prognosis, a subset of patients experience long-term survival after receiving standard of care treatment—representing a cohort that has not been previously well described.

Most ATCs develop in the context of differentiated thyroid cancer, often arising from or co-occurring with papillary thyroid carcinoma. Recent reports have characterized the most frequent primary driver mutations as TP53, BRAF, NRAS, PIK3CA, and TERT (1). However, these aggressive malignancies exhibit broad genetic diversity, which may be key to elucidating differences in survival. Surgery represents the only possible option for cure but is most frequently performed palliatively to reduce the risk of local complications.

Routine molecular testing of all incisional biopsies presents an opportunity for preoperative risk stratification to discern between palliative and potentially curative surgery and to tailor counseling to individual patients. Mathematical modeling has been used to aid scoring systems for risk stratification across several malignancies. Perhaps most commonly, using a multivariable regression based on 21 and 12 genes, Oncotype DX generates a risk score for cancer recurrence for breast and colo-rectal cancers, respectively, to guide adjuvant chemotherapy (3, 4). A similar tool that has the ability to distinguish the most aggressive ATCs from those that have the best potential for longer-term survival has the potential to guide treatment paradigms, as well as possibly identify new targets for treatment. Herein, we propose a predictive model to risk-stratify patients with ATC using genomic profiles.

This study met exclusion criteria by the Weill Cornell Medical College Institutional Review Board.

Materials and Methods

ATC tumor samples from multiple studies compiled in cBioPortal with available whole-exome sequencing and targeted sequencing were included (1, 5). The publicly available cBioPortal website is a software developed by a multi-institutional team hosted by the Center for Molecular Oncology at Memorial Sloan Kettering Cancer Center. It is a portal for analyzing The Cancer Genome Atlas and other studies made available by authors that combines nonsynonymous mutations, DNA copy-number data, mRNA and microRNA expression data, protein-level data, epigenetic (methylation) data, and deidentified clinical data when available (6-8).

Data Acquisition and Processing

Patient-level mutation data and clinical outcomes for ATC were extracted from cBioPortal. Genetic mutation status for 745 genes was recorded in the dataset, along with survival time and survival status. Mutation status was categorized as wild-type, mutant, or not available. Genes with missing values exceeding 75% were removed. Missing values in the remaining genes were imputed using the column median.

Training and Validation Cohort Creation

Mutation burden was calculated as the sum of mutated genes per patient. The dataset was stratified into a training cohort (80%) and a validation cohort (20%) using stratified random sampling based on mutation burden.

Univariate Cox Proportional Hazards Analysis

Univariate Cox proportional hazards regression was performed in the training cohort to identify genes associated with overall survival. Genes with a P-value < .05 were retained. Genes mutated fewer than 5 times in the cohort were excluded (Fig. 1) (9).

Figure 1.

Gene inclusion flowchart.

LASSO Cox Regression

To reduce overfitting by limiting the number of genes included in the panel based on the cohort size, a LASSO Cox proportional hazards model was fitted using the glmnet package to select a parsimonious set of prognostic genes. The optimal penalty parameter (λ) was determined by 10-fold cross-validation using the minimum lambda criterion (λmin). Selected genes were assigned nonzero coefficients representing their contribution to the risk score (10).

Risk Score Calculation and Risk Stratification

Risk scores were calculated for each patient in both the training and validation cohorts as a weighted sum of the selected gene mutation statuses, using the LASSO coefficients. Given the overall poor survival of the population, groups were described as more aggressive or less aggressive. Patients were stratified into more aggressive and less aggressive groups based on the optimal cutoff determined using the maxstat package with the maximally selected rank statistic (10).

Survival Analysis and Validation

Kaplan-Meier survival curves were generated and compared using the log-rank test. Twelve-month survival probabilities were estimated for the more aggressive and less aggressive cohorts. Concordance indices (C-indices) were calculated to evaluate the discriminatory ability of the risk score in both the training and validation cohorts. Coefficient plots and hazard ratio (HR) plots were created to visualize the contribution of each gene to the risk score and its association with survival.

KEGG Pathway Analysis

Patients from the combined training and validation cohorts were stratified into more aggressive and less aggressive groups based on the risk classification from the LASSO Cox model. Mutation profiles of genes (745 genes) were compared between the more aggressive and less aggressive risk groups. KEGG cancer-related pathway enrichment analysis was conducted separately for more aggressive and less aggressive gene sets using the clusterProfiler package. Significantly enriched pathways were defined using an adjusted P-value < .05. For each enriched pathway, the proportion of patients with at least 1 mutation in any of the pathway-associated genes was calculated, and rates were compared between the more aggressive and less aggressive risk cohorts using Fisher’s exact test. Alteration rates were visualized with a scatter plot, labeling pathways altered more 1.5 times higher in either group (11).

Software

All analyses were conducted using R (version 4.3.1).

Results

Two hundred four tumor samples met inclusion criteria with available survival data. Samples originated from 2 multi-institutional international studies (1, 5). Median survival for the whole cohort was 4.8 months. Similar to existing literature, the most frequently identified mutations were TP53 (34.3%), BRAF (25.9%), TERT (12.2%), NRAS (11.2%), ANKRD11 (10.7%), and PIK3CA (9.8%).

Following filtering for statistically significantly associated genes with survival on univariate analysis (n = 85) and with nonzero coefficients on LASSO (n = 14), multivariable Cox regression was performed on selected genes (Fig. 2A). This analysis yielded 14 mutated genes associated with increased risk (HR > 1): TET1, MAPK12, ATP10A, PIK3CA, MUC4, PNPLA2, PLD4, EGLN2, BSN, FLNC, RADIL, ZMYND8, FRAS1, and RECQL4. Here, RECQL4 mutations contributed the greatest risk demonstrated on an HR-based nomogram (Fig. 2B). A maximally selected rank statistic determined an optimal risk cutoff point threshold of 0.27 cumulative points to differentiate more and less aggressive cohorts (Fig. 2B). The predictive performance of the nomogram at this threshold was found to stratify groups between more and less aggressive risk cohorts at a C-index of 0.74 for the training cohort.

Figure 2.

Multivariable Cox regression and nomogram for target genes. (A) Fourteen-gene multivariable analysis. (B) Point-based nomogram from β coefficient.

Point-based assessment of cohorts yielded 128 less aggressive samples and 37 more aggressive samples. Median and 1-year (%) survival differed significantly for less aggressive patients (7.2 months, 32%) compared to more aggressive patients (2.4 months, 0%) (P < .001) (Fig. 3). Less aggressive patients demonstrated consistent long-term survival with 32, 6, and 3 patients surviving greater than 1, 3, and 5 years.

Figure 3.

Kaplan-Meier training cohort. Survival curves for training cohorts based on a 0.27-point mutation score cutoff.

Application of the same point threshold to the validation cohort yielded 29 less aggressive and 11 more aggressive samples (Fig. 4). Survival remained significantly different between cohorts with a median and 1-year survival of 5.9 months and 34% for less aggressive patients and 2.5 months and 10% for more aggressive patients (P = .043). Similarly, less aggressive patients had a high incidence of long-term survival with 9, 5, and 2 patients surviving greater than 1, 3, and 5 years, respectively. With this validation cohort, the model retained the ability to adequately discriminate between risk features with a C-index of 0.62.

Figure 4.

Kaplan-Meier validation cohort. Survival curves for validation cohort based on a 0.27-point mutation score cutoff.

Given the increasing utilization of targeted therapy for BRAF variants in recent years, the same model was applied to the BRAF altered cohort to determine if similar survival differences were observed in this subpopulation. Forty-three less aggressive and 10 more aggressive tumors demonstrated significantly different median and 1-year survivals of 7.2 months (26.8% at 1 year) and 3.1 months (0% at 1 year), respectively (P = .001). This corresponded to a C-index of 0.641.

Finally, to assess this cutoff among the whole cohort of 204 samples, the same threshold was applied, yielding a C-index of 0.713. This produced 157 less aggressive samples and 48 more aggressive samples with significantly different survival rates (P < .001).

Each risk-cohort subsequently underwent KEGG pathway enrichment analysis. Thirteen pathways exhibited enrichment across both cohorts with 4 exhibiting significantly enhanced enrichment among more aggressive tumors. These included PD-L1 expression and PD-1 checkpoint pathway in cancer, central carbon metabolism in cancer, mRNAs in cancer, and proteoglycans in cancer (Fig. 5).

Figure 5.

Pathway enrichment comparison. Pathways enriched more than 1.5 times in either cohort labeled.

Discussion

This study used a large, multi-institutional, international dataset of sequenced ATCs to create and validate a genomics-based predictive model of survival. Fourteen mutated genes were associated with altered survival, which were used to develop a nomogram for genomically predicting longer-term survivors for the first time in the literature with adequate accuracy. Patients in the more aggressive cohort were found to exhibit significantly greater oncologic pathway enrichment in the PD-L1 expression and PD-1 checkpoint pathway in cancer, central carbon metabolism in cancer, mRNAs in cancer, and proteoglycans in cancer pathways.

At present, clinical features and adjuvant therapies are pre-dominantly featured in ATC prognostic literature. Increasing patient age, nodal involvement, distant metastases, and leuko-cytosis are all associated with poor survival (1). Although they have not characterized a specific low- or high-risk population, previous studies have also found several prognostically pertinent genomic influences on overall survival. Specifically, in their study comparing poorly differentiated thyroid cancer and ATC genomics, Landa et al found that tumors with TERT promoter alterations exhibited worse overall survival, especially when combined with NRAS or BRAF mutations (5). Although TERT mutations were consequently well represented in this dataset (12% of all samples), TERT did not emerge as a prognostically significant gene in the context of our larger cohort—pointing to TERT’s role as a driver for ATC but not a predictor of survival (1). Thus, apart from PIK3CA, the novel set of prognostically pertinent variants described in this study differs from previously characterized driver mutations, warranting further evaluation at the bench level (1, 12).

Among genes conferring a poor prognosis in the model, most have established clinical and therapeutic oncologic roles studied in humans, with several already associated with thyroid malignancies. PIK3CA is a well-described oncogene across numerous malignancies, including ATC (1). Further, it exhibits a central role in oncogenesis as driver mutations in the canonical PI3 K pathway are common to 89% of cancers studied in The Cancer Genome Atlas (12). MAPK12 mutations are known to produce increasing invasiveness by way of promoting epithelial-mesenchymal transition. Further, MAPK12 has emerged as a possible immunotherapy target for thyroid cancer (13).

Among genes with oncologic significance studied outside of thyroid cancer, several present plausible mechanisms for the development of metastases or possess significant clinical associations. RECQL4 and ZMYND8 are associated with several malignancies due to their role in chromosomal instability. Further, ZMYND8 is a retinoic acid-inducible gene, suggesting that all-trans-retinoic acid, a drug commonly used in the treatment of leukemias, could be effectively used to augment its expression (14, 15). TET1 and BSN both act as WNT pathway regulators, with variants promoting epithelial-mesenchymal transition and chemotherapy resistance (16, 17). MUC4 mutants cause defective endothelial cell function, which may lead to metastatic progression (18). With regard to clinical associations, PNPLA2 deletion mutations are found in 38% of lung cancers, while ATP10A, EGLN2, and FLNC exhibit strong associations with tumorigenesis and progression in melanoma, breast cancer, and glioblastoma, respectively (Vegliante) (19-21).

While this study focused on developing a unique, predictive genomic model, we also aimed to further characterize longer-term survivors based on tumor pathway enrichment. The higher incidence of commonly observed variants within pathways associated with more aggressive tumors may suggest 2 unique ATC phenotypes based on the degree of pathway enrichment. Specifically, increased enrichment in pathways, such as PD-L1 expression, containing well-described driver mutations such as TP53, NRAS, CTNNB1, RET, PIK3CA, PTEN, EGF, MTOR, and ALK were detected within more aggressive tumors. Although many of these variants may not independently confer an impact on survival, the lower incidence of pathway enrichment containing these mutations seen in the less aggressive cohort may further characterize this population as a biologically unique subgroup of longer-term survivors.

Newer trials have investigated the role of targeted therapy for ATC that was not consistently available at the time of sample collection for the data presented in this study. Subbiah et al achieved a 1-year survival of 51.7% using dabrafenib plus trametinib in BRAF mutated unresectable or metastatic ATC (2). Therefore, the prognostic advantage of targetable mutations may not have been captured in our study. However, our sub-analysis of BRAF variants demonstrates that more aggressive and less aggressive phenotypes are significantly maintained in this cohort, which may point to different responses to targeted therapy in future studies.

This study presents several limitations inherent to using datasets with limited clinical information for an already rare malignancy. Although multivariable analysis controlled for all significant genes, the type of adjuvant therapy administered was not able to be controlled for. However, the primary sources report that the majority of patients underwent traditional chemotherapy, as samples were obtained before the regular use of targeted therapy was introduced (1). To this point, in addition to dabrafenib and the described agents under investigation, future targeted therapies may augment survival and change the prognostic value of this tool. Further, information on the sequence of treatment (upfront surgery, neoadjuvant vs adjuvant systemic therapy, or radiotherapy) was not available. However, ATC’s nearly ubiquitous presentation as a progressive, systemic disease with dismal median survival suggests that the genomic data presented here may offer important insights absent of clinical factors.

Longer-term survival among patients with ATC is rare but possible with standard of care treatment. Using a multi-institutional, international cohort, this study used a large series of sequenced ATCs to create a genomics-based predictive survival model. If prospectively validated, this model may better prognosticate outcomes, aiding in clinicians’ ability to effectively counsel patients. Further, when paired with individual patient characteristics, this model could help tailor treatment toward palliative therapies for local complications or more aggressive treatment strategies, such as formal resections for curative intent. Differences in pathway enrichment point to unique genotypes between more aggressive and less aggressive ATC cohorts. In the era of targeted BRAF inhibition, predictable differences in BRAF variant survival may point to variable drug effects by tumor phenotype. Further investigation into the possibility of novel targeted therapies for this gene panel may expand treatment options and survival for ATC. Finally, prospective validation of this risk score will be essential to accurately risk-stratify and counsel patients.

Data Availability

The data generated in this study are publicly available in cBioPortal at https://www.cbioportal.org/study/summary?id=thyroid_gatci_2024, https://www.cbioportal.org/study/summary?id=thyroid_mskcc_2016.