Targeted Assessment of G0S2 Methylation Identifies a Rapidly Recurrent, Routinely Fatal Molecular Subtype of Adrenocortical Carcinoma Free

Adrenocortical carcinoma (ACC) is a rare cancer of the adrenal cortex affecting 0.5 to 2 individuals/million/year globally (1, 2). Though rare, ACC is frequently aggressive with 35% 5-year survival (3). Therapies for metastatic ACC are primarily palliative, limited to administration of adrenolytic drug mitotane and/or cytotoxic chemotherapy (3). Patients with locoregional ACC routinely receive surgery and adjuvant mitotane, but 50% to 70% recur and develop metastases even after complete (R0) resection (4, 5). Retrospective studies suggest adjuvant mitotane prolongs recurrence-free survival (6, 7), but its efficacy is limited by its poor pharmacokinetic properties and dose-limiting toxicities. Obtaining therapeutic serum levels of mitotane may take several months to achieve if at all (8). Furthermore, there is a substantial proportion of ACC patients who experience rapid recurrence (<12 months; refs. 4, 5, 9), whose aggressive disease course may preclude response to mitotane. These patients may instead benefit from more rapidly acting therapies; however, prospectively identifying this subgroup remains challenging.

Histologic grade based on cellular proliferation is the strongest predictor of recurrence following R0 resection in ACC; high-grade disease is associated with higher risk of recurrence (10–12). Despite its clinical utility, significant caveats complicate interpretation of grade on an individual basis. Evaluation of grade is prone to high interrater variability (13), and outcomes of patients with low- and high-grade disease remain heterogeneous, with rapidly recurrent patients in both strata (10, 14). Although some studies indicate clinical factors may be predictive of recurrence (9, 15), molecular profiling studies suggest biomarkers may better resolve this heterogeneity by identifying patients with homogeneously dismal outcomes.

We and others have shown that transcriptomes of aggressive ACC are characterized by pronounced cell-cycle activation (16), and a score based on mRNA levels of mitotic regulator BUB1B (BUB1 Mitotic Checkpoint Serine/Threonine Kinase B) and mitochondrial kinase PINK1 (PTEN Induced Putative Kinase 1) discriminates uniformly favorable from poor clinical outcomes (17, 18). Recent studies have implicated aberrant epigenetic patterning in ACC pathogenesis, highlighting that aggressive carcinomas bear widespread promoter CpG island hypermethylation (19, 20). Notably, the most comprehensive molecular study on ACC to date, The Cancer Genome Atlas project on ACC (ACC-TCGA), similarly identified that rapidly recurrent ACC is distinguished by a CpG island hypermethylation phenotype, “CIMP-high” (21).

Although these studies have illuminated molecular programs core to aggressive ACC biology, clinical translation of “big data”-derived biomarkers remains challenging. Thresholds for continuous data, for example mRNA-based biomarkers, vary across patient cohorts (17, 18), compromising biomarker utility for prospective clinical management of a rare malignancy. Furthermore, while targeted assessment of DNA methylation appears promising for prognosticating ACC (20, 22), measurement strategies frequently rely on several genomic loci, complicated data normalization procedures, and reference benign lesions (22). Finally, it remains unclear if validated biomarkers identify uniform ACC molecular subtypes amenable to clinical assessment of subtype-specific therapeutic approaches. It is therefore not surprising that mRNA and DNA methylation-based biomarkers have yet to be successfully translated clinically to prognosticate ACC, and highlights a strong need for identifying novel biomarkers with simplified, binary readouts and therapeutic import.

Here, we present a new analysis of ACC-TCGA data in which we demonstrate that CIMP-high ACC is a unique, rapidly recurrent ACC molecular subtype, bearing upregulation of cell cycle- and DNA damage-associated cellular programs. We identify that uniform hypermethylation and silencing of the gene G0S2 (G₀–G₁ Switch 2) is largely exclusive to CIMP-high ACC. We show in an independent cohort that targeted assessment of G0S2 methylation using an overnight assay independently identifies a subgroup of patients with rapidly progressive or fatal disease course. Our data demonstrates that G0S2 methylation status is essentially binary, and thereby has high potential to enable clinicians to prospectively identify ACC patients unlikely to exhibit durable response to standard adjuvant therapy. We also propose that rapidly acting adjuvant cytotoxic agents may benefit patients with this ACC subtype. Finally, our study demonstrates the utility of comprehensive databases like TCGA, and illustrates a pipeline to identify and test clinically relevant biomarkers for ACC that may be extended to other cancers.

We downloaded the ACC-TCGA RNA-seq count table and raw data (IDAT files) from the Infinium HumanMethylation450 BeadChip (“450k”) platform from the GDC legacy archive (https://portal.gdc.cancer.gov/legacy-archive). We used R (23)/Bioconductor packages limma (24) and minfi (25) to obtain log2-normalized counts per million (CPM) values for gene expression and β and M values for methylation arrays. We used limma to nominate differentially expressed genes (Benjamini-Hochberg FDR-corrected P-value <0.05) between CIMP-high and non-CIMP-high ACC. We used goana (24, 26) to identify gene ontology terms enriched among differentially expressed genes in CIMP-high versus non-CIMP-high ACC. REVIGO (27) is an online tool that enables nonredundant visualization of large sets of GO terms based on semantic similarity. We used REVIGO with SimRel semantic similarity algorithm to plot the 200 most significant biological processes up (ranked by increasing P.Up, P.UP <0.05) or down (ranked by increasing P.Down, P.Down <0.05) in Fig. 1C. We used DMRcate (28) to interrogate differentially methylated regions (Stouffer-corrected P-value <0.05) across groups. We used logistic regression on the RNA-seq data to identify transcripts predictive of CIMP-high status. We used pheatmap (29) to perform unsupervised complete hierarchical clustering, and caret (30) to perform k-fold cross-validation.

Our study includes 114 adrenocortical tumors evaluated from 1989 to 2017. A total of 42 treatment-naive primary ACC, one primary ACC from a patient who received neoadjuvant etoposide/doxorubicin/cisplatin+mitotane, three non-primary ACC, and 14 cortisol-secreting adrenocortical adenomas (ACA) are from Faculdade de Medicina da Universidade de São Paulo (FMUSP), São Paulo, Brazil; 38 primary ACC, eight non-primary ACC, four aldosterone-secreting ACA, and four cortisol-secreting ACA are from the University of Michigan (UM), Ann Arbor, MI. Diagnosis of ACA/ACC was established by expert pathologic assessment (M.C.N.Z., T.J.G.) of surgical specimen using Weiss criteria (12). Diagnosis of ACA or ACC was assigned to samples with Weiss score <3 or ≥3, respectively. Informed consent was obtained from all participants, and studies were conducted in accordance with the Declaration of Helsinki with study protocols approved by FMUSP and UM Institutional Review Boards. Clinical, hormonal, and demographic data were collected retrospectively.

Immediately following surgical resection, samples were collected by an institutional pathologist and snap frozen in liquid nitrogen. Frozen tumor tissue was cryotome sectioned (6 μm) under RNase-free conditions to acquire ≥30 mg tissue per sample. ≥3 random noncontiguous sections from each tumor were prepared for rapid hematoxylin and eosin (H&E) staining to evaluate sample quality and tumor purity. Samples with >50% acellular material in >2/3 sections were excluded from downstream processing. Slides from samples included in downstream processing were assessed by T.J.G. to confirm typical ACC histology. Genomic DNA (gDNA) and total RNA were simultaneously extracted with AllPrep DNA/RNA Mini Kit (Qiagen; 80204) and optional on-column RNase A (Qiagen; 19101) and DNase I (RNase-free DNase Set, Qiagen; 79254) digests, respectively.

Samples collected immediately following surgical resection were snap frozen in liquid nitrogen, embedded in OCT freezing media (Miles Scientific), cryotome sectioned (5 μm), and evaluated by routine H&E by surgical pathologists. When possible, corresponding H&E sections from paraffin blocks were also evaluated. Areas of pure tumor (≥70% tumor cells) were selected for nucleic acid extraction. gDNA and RNA were extracted using one of the following methods: Trizol (Invitrogen/Thermo Fisher Scientific; 15596026) with acid–phenol:chloroform cleanup, RNeasy Mini Kit (Qiagen; 74104) or DNeasy Blood and Tissue Kit (Qiagen; 69504), or AllPrep DNA/RNA/Protein Mini Kit (Qiagen; 80004).

For all samples, RNA integrity was evaluated by agarose gel electrophoresis; purity (260/280, 260/230 ratios) and quantity of gDNA and RNA were measured by spectrophotometry (NanoDrop 2000 Spectrophotometer; Thermo Fisher Scientific; Catalog No. ND-2000). cDNA was synthesized (High-Capacity cDNA Reverse Transcription Kit with RNase Inhibitor; Applied Biosciences/Thermo Fisher Scientific; 4374966) from high integrity and high-quality RNA (visual 28S:18S rRNA ratio 2:1 and 260/280 ratio ≥2.00). qPCR was performed in the QuantStudio 3 Real-Time PCR System (Applied Biosciences/Thermo Fisher Scientific; A28136), using TaqMan Fast Advanced Master Mix (Applied Biosciences/Thermo Fisher Scientific; 4444557) and FAM-MGB-labeled TaqMan Gene Expression Assays (Applied Biosciences/Thermo Fisher Scientific) to evaluate expression of G0S2 (Hs00274783_s1), BUB1B (Hs01084828_m1), PINK1 (Hs00260868_m1), and housekeeping gene GUSB (Hs00939627_m1). TaqMan Gene Expression Assays were performed in triplicate. Gene expression levels were calculated using the ΔC_t method where ΔC_t(X) = C_t(X) – C_t(GUSB), and BUB1B-PINK1 score calculated as ΔC_t(BUB1B) – ΔC_t(PINK1).

Assessment of G0S2 methylation by targeted bisulfite sequencing in physiologic tissues, ACA, and ACC was performed by Zymo Research Corporation. Zymo Research Co. designed/validated primers to amplify the G0S2 locus, chr1:209,848,443-chr1:209,848,900 (hg19), using a proprietary pipeline. Submitted gDNA with 260/280 ≥1.7, intact genomic band (≥5 kb) by gel electrophoresis, and sufficient quantity (≥100 ng) was subject to bisulfite conversion, targeted amplification, next-generation sequencing library indexing, and sequencing on Illumina MiSeq. Sequence data were demultiplexed and assessed for bisulfite conversion rate, read coverage, mapping efficiency, and CpG coverage. Bisulfite conversion rate was ≥99% for all samples. Average CpG coverage ranged from 5,000 to 50,000×. Methylation at each CpG was calculated from the ratio of methylated to total CpG count.

Available gDNA from ACC and ACA was subject to methylation-sensitive restriction digestion using EpiTect II DNA Methylation Enzyme Kit (Qiagen; Catalog No. 335452). This kit contains two enzymes: Methylation Sensitive Enzyme A (cannot cleave gDNA in the presence of CpG methylation in the proprietary restriction site) and Methylation Dependent Enzyme B (can cleave gDNA only in the presence of CpG methylation in the proprietary restriction site). Per manufacturer protocol, gDNA from each tumor was subject to four digests: “mock” digest (M_o, containing no restriction enzymes), methylation-sensitive digest (M_s, containing only Enzyme A), methylation-dependent digest (M_d, containing only Enzyme B), and double digest (M_sd, containing both enzymes). To measure intact gDNA following overnight restriction digestion, gDNA was amplified by qPCR using the EpiTect Methyl II PCR Primer Assay for Human G0S2 (Qiagen; Catalog No. EPHS101235-1A) and RT² SYBR Green ROX qPCR Mastermix (Qiagen; 330521). Percent G0S2 methylation was calculated arithmetically from M_o, M_s, M_d, and M_sd C_t values according to manufacturer instructions, using a Microsoft Excel spreadsheet provided by Qiagen.

We used Chi-square test to evaluate associations between categorical variables, Mann–Whitney test or Pearson correlation to compare continuous data from two groups, and Kruskal–Wallis with Dunn's multiple comparisons test to compare continuous data from >2 groups. We used pheatmap (29) to perform unsupervised complete hierarchical clustering. We used caret (30) to perform k-fold cross validation. We used receiver operating characteristic (ROC) curve analysis to estimate a cutoff of G0S2 expression that predicts methylation. We used Kaplan–Meier analysis with pairwise log-rank test to compare overall survival (OS) and disease-free survival (DFS), and Cox proportional hazards regression models to estimate hazard ratios for clinical/molecular variables. P-value <0.05 was significant for all analyses. Statistical analyses were performed in GraphPad Prism, MedCalc, and R (23).

In ACC-TCGA, comprehensive DNA methylome profiling of 79 treatment-naive primary ACC using the 450k platform clustered ACC into three DNA-methylation-based subtypes: “CIMP-low,” “CIMP-intermediate,” and “CIMP-high” (21). Although patients with CIMP-low and CIMP-intermediate carcinomas exhibited indistinguishable disease course (log-rank P = 0.22 for DFS of CIMP-low vs. CIMP-intermediate; Supplementary Fig. S1A), patients with CIMP-high carcinomas characteristically exhibited rapidly recurrent or deadly disease course with median DFS following R0/RX resection of 13.6 months (Fig. 1A) and median OS of 36 months (Supplementary Fig. S1B). Given the striking clinical phenotype associated with the CIMP-high signature, we sought to determine if other molecular classes and somatic alterations identified by ACC-TCGA were associated with this epigenetic program. We performed association tests between CIMP status and ACC-TCGA-defined transcriptome class (mRNA group), somatic copy number alteration profile (SCNA group), or somatic alterations. We observed that CIMP-high carcinomas were distinguished by a transcriptional signature featuring increased expression of steroidogenic and proliferative machinery (“Steroid-high + prolif.” transcriptional program), and a chromosomally “noisy” genomic landscape with numerous arm-level breaks and focal copy number gains and losses (Fig. 1B). CIMP-high ACC also frequently bore somatic alterations leading to activation of the cell cycle; however, CIMP-high status was not associated with an increased incidence of alterations leading to activation of Wnt signaling, present in ∼40% of ACC (Fig. 1B; ref. 21).

We next analyzed RNA-seq data (n = 78) from ACC-TCGA to identify differentially expressed genes in CIMP-high compared with non-CIMP-high (CIMP-low + CIMP-intermediate) carcinomas (Supplementary Table S1). We performed gene ontology analysis on differentially expressed genes and identified that CIMP-high ACC exhibited transcriptional upregulation of numerous cell cycle- and DNA damage-associated biological processes, consistent with the enrichment of cell cycle-activating somatic alterations and chromosomal “noisiness” in this subgroup (Fig. 1C, left). Intriguingly, CIMP-high carcinomas exhibited transcriptional downregulation of a wide array of immunological processes (Fig. 1C, right), suggesting that CIMP-high ACC are relatively immune poor. The convergence of this unique transcriptional program, somatic alterations targeting the cell cycle, and “noisy” chromosomal landscape in CIMP-high carcinomas demonstrates that CIMP-high status defines a distinct molecular subtype of ACC characterized by rapidly recurrent or fatal disease course. Therefore, prospectively identifying CIMP-high carcinomas using targeted molecular markers may have strong clinical utility.

We sought to identify a single biomarker with strong discriminatory power between CIMP-high and non-CIMP-high ACC, straightforward to measure and interpret without reference samples or extensive data manipulation. We were therefore interested in genomic loci that are methylated and silenced exclusively in CIMP-high ACC. We analyzed DNA methylation data from ACC-TCGA to identify regions hypermethylated in CIMP-high compared with non-CIMP-high carcinomas (Supplementary Table S2). Among the top 10 most hypermethylated regions in our analysis was a 2kb region on chromosome 1 (chr1:209847618-209849445, hg19; Supplementary Fig. S2), encompassing 13 contiguous 450k probes and spanning the G0S2 gene locus (non-CIMP-high vs. CIMP-high: max β fold-change −0.709, mean β fold-change −0.508, Stouffer-corrected P-value 4.32 × 10⁻¹³⁴). Our analysis of differentially expressed genes in CIMP-high compared with non-CIMP-high ACC also revealed G0S2 was among the top five downregulated genes, nearly silenced in CIMP-high carcinomas (CIMP-high vs. non-CIMP-high: log2 fold change −5.21, Benjamini–Hochberg FDR-corrected P-value 2.31 × 10⁻¹⁰), and highly predictive of CIMP-high status (logistic regression coefficient −0.925, P-value 2.10 × 10⁻⁵; Supplementary Table S1). These results suggested G0S2 is silenced by hypermethylation in a subgroup of ACC as reported in a smaller ACC cohort (20), and that low G0S2 expression and hypermethylation predict CIMP-high status. This observation was particularly intriguing as analysis of GTEx RNA-seq data (31) revealed G0S2 is highly expressed in the physiologic adrenal gland (Supplementary Fig. S3).

We then plotted all 450k probes spanning the G0S2 locus in each tumor sample from ACC-TCGA, ranked by decreasing G0S2 expression. Strikingly, tumors exhibited an “all or none,” binary pattern of methylation, with uniform hypermethylation (probe β value >0.5) across the gene locus nearly restricted to CIMP-high carcinomas, and associated with reduced G0S2 expression (Fig. 2A). Indeed, average methylation level of probes residing in the G0S2 CpG island is significantly higher in CIMP-high compared with non-CIMP-high ACC (P < 0.0001, Kruskal–Wallis with Dunn's multiple comparisons test; Fig. 2B), expression of G0S2 is significantly lower in CIMP-high compared with non-CIMP-high ACC (P < 0.0001, Kruskal–Wallis with Dunn's multiple comparisons test; Fig. 2C), and both metrics are strongly inversely correlated (P < 2.2 × 10⁻¹⁶, r = −0.82, R² = 0.68, Pearson correlation; Fig. 2D). The inverse correlation between G0S2 methylation and expression in ACC-TCGA suggested that measurement of G0S2 methylation (or expression in the absence of genomic DNA) can enable identification of CIMP-high ACC.

Finally, we sought to evaluate the ability of G0S2 methylation alone to classify ACC-TCGA samples by CIMP status. We performed unsupervised hierarchical clustering analysis using the logit-transformed β values of 450k probes lying within the G0S2 CpG island (Supplementary Fig. S4A). This analysis identified two distinct clusters of samples: one cluster with samples bearing either no or low levels of G0S2 methylation (“G0S2 unmethylated”) corresponding to 2/3 of ACC-TCGA, and one with samples bearing high levels of uniform or heterogeneous G0S2 methylation (“G0S2 methylated”) corresponding to 1/3 of ACC-TCGA. The G0S2 methylated cluster was strongly enriched for CIMP-high ACC (P < 0.0001, Fisher exact test), capturing 18/19 CIMP-high samples. To evaluate the performance of a logistic regression model utilizing G0S2 methylation to discriminate CIMP-high from non-CIMP-high ACC, we performed an internal k-fold cross validation (k = 5, 20 repeats) on the average of the logit-transformed β values of probes residing in the G0S2 CpG island. Our fitted logistic regression model is described in Supplementary Table S3, and the ROC curve (ROC AUC = 0.928; 95% CI, 0.8235–1) is depicted in Supplementary Fig. S4B. At average G0S2 methylation >0.5200819 (measured by 450k array), we can predict assignment to CIMP-high using G0S2 methylation alone at 94.87% accuracy, with 94.74% sensitivity, 94.92% specificity, 85.71% positive predictive value, and 98.25% negative predictive value. This analysis demonstrates that G0S2 hypermethylation has high discriminatory power to distinguish CIMP-high from non-CIMP-high ACC, and shows that unsupervised clustering of G0S2 CpG island methylation enables reliable identification of CIMP-high samples. Taken together, our analysis of ACC-TCGA suggests that assessment of G0S2 methylation and/or expression can reliably identify CIMP-high ACC without comprehensive DNA methylome data.

We sought to evaluate G0S2 methylation in an independent ACC cohort, and determine if physiologic tissues and ACA exhibit G0S2 methylation. We collected gDNA and mRNA from a retrospective cohort of 80 treatment-naive primary ACC, 22 ACA, and 12 non-naive/non-primary ACC, summarized in Supplementary Table S4. We also collected gDNA from extra-adrenal tissues, microdissected adult adrenal cortex, and total adult adrenal cortex. We performed targeted bisulfite sequencing of G0S2 and determined that uniform hypermethylation throughout the locus is pathologic, exclusive to a subset of primary ACC and nonprimary/recurrent ACC (Fig. 3A; Supplementary Table S5). These findings are supported by unsupervised hierarchical clustering analysis on logit-transformed targeted bisulfite sequencing data (Supplementary Fig. S5A), in which we recapitulate G0S2 unmethylated and G0S2 methylated clusters we identified in ACC-TCGA. We also demonstrate that physiologic tissue and benign adrenocortical tumors cluster with G0S2 unmethylated ACC, whereas only ACC with high levels of uniform or heterogeneous G0S2 methylation reside in the G0S2 methylated cluster. The association of physiologic adrenal cortex samples with G0S2 unmethylated ACC is consistent with the high expression of G0S2 in the physiologic adrenal gland (Supplementary Fig. S3).

The uniform pattern of G0S2 methylation in ACC-TCGA and our cohort indicated that locus methylation may be accurately measured by methylation-sensitive restriction digestion/qPCR-based methods instead of bisulfite-based approaches. We evaluated G0S2 methylation using one such approach, EpiTect (Qiagen). EpiTect and targeted bisulfite sequencing were highly concordant (Fig. 3B; Supplementary Fig. S5C), demonstrating that ACA have no measurable G0S2 methylation, whereas ACC have a bimodal distribution (Fig. 3C; 40% of ACC in FMUSP+UM Primary ACC Cohort have G0S2 hypermethylation). We then sought to evaluate the concordance between EpiTect and binary G0S2 methylation status defined by unsupervised hierarchical clustering analysis (Supplementary Fig. S5A). For all samples with paired EpiTect and targeted bisulfite sequencing data (n = 74; 60 ACC, 14 ACA), we performed an internal k-fold cross validation (k = 5, 20 repeats) to evaluate a logistic regression model utilizing EpiTect measurements to discriminate these two classes. Our fitted logistic regression model is described in Supplementary Table S6 and ROC curve (ROC AUC = 1) depicted in Supplementary Fig. S5B, and enables us to obtain a perfect classification with an EpiTect cutoff of 4.696%. These analyses demonstrate that EpiTect enables accurate assessment of binary G0S2 methylation status defined by gold-standard targeted bisulfite sequencing, reinforcing its potential clinical utility.

As in ACC-TCGA, tumors with G0S2 hypermethylation have minimal transcript expression compared with ACA or ACC without G0S2 methylation (Fig. 3D). Interestingly, nonprimary/non-naive ACC also exhibited the G0S2 methylation/expression inverse relationship (Supplementary Table S7). Finally, we used ROC curve analysis to identify a threshold of G0S2 expression that reliably predicts G0S2 hypermethylation (ROC AUC = 0.8557, P < 0.0001; Supplementary Fig. S6). At ΔC_t(G0S2) >3.944, we could predict G0S2 hypermethylation with 92.31% specificity (95% CI, 79.13%–98.38%) and 48.15% sensitivity (95% CI, 28.67%–68.05%); we used this cutoff to infer G0S2 methylation status of 10 primary ACC for which gDNA was unavailable.

Together with ACC-TCGA, these data illustrate that uniform G0S2 hypermethylation and silencing is exclusive to a subset of ACC, and that G0S2 methylation can be accurately measured using restriction digest/qPCR-based methods or inferred from G0S2 expression when gDNA is unavailable.

High histologic grade is an established predictor of dismal outcomes in ACC (10–12). In the FMUSP+UM Primary ACC Cohort, patients with high-grade tumors accordingly exhibited rapidly recurrent disease following R0/RX resection (median DFS of 7.8 months). However, 3/10 of patients with high grade tumors remain disease free after >48 months follow-up and 11/32 patients with low-grade disease exhibited recurrence, demonstrating that proliferation-based grade alone stratifies patients into heterogeneous groups (Fig. 4A). In striking contrast, stratification by G0S2 methylation (measured by EpiTect or inferred from G0S2 expression when gDNA unavailable) demonstrates that patients with tumors bearing G0S2 hypermethylation homogeneously exhibited rapidly recurrent or fatal disease course (median DFS following R0/RX resection of 14 months and median OS of 17 months; Fig. 4B and C). Remarkably, G0S2 hypermethylation was identified at comparable frequency in low- and high-grade tumors (P = 0.076, Fisher exact test), with G0S2 hypermethylation in 13/44 low-grade tumors (Fig. 4D), suggesting that G0S2 hypermethylation identifies aggressive disease in tumors inadequately stratified by tumor grade. Finally, carcinomas with G0S2 hypermethylation were identified at comparable frequency in patients with localized ACC (ENSAT I-II), localized ACC with locoregional invasion or lymph node involvement (ENSAT III), and ACC with distal metastases (ENSAT IV) at diagnosis (P = 0.31, Chi-square test; Fig. 4E). Notably, among 17 ENSAT I-III patients with R0/RX resection and G0S2 hypermethylation, only one patient remains disease free at >24 months.

We performed Cox proportional hazards regression analysis to evaluate the significance of G0S2 hypermethylation at predicting recurrence and death compared with other clinical metrics in the FMUSP+UM Primary ACC Cohort (Table 1). High-grade and G0S2 hypermethylation were the only variables that significantly predicted recurrence as univariates (high-grade vs. low-grade HR = 3.15, G0S2 methylated vs. unmethylated HR = 6.91). In contrast, cortisol secretion, ENSAT IV, tumor size, tumor weight, high grade, and G0S2 hypermethylation all significantly predicted death as univariates [cortisol-secreting vs. non-cortisol-secreting HR = 2.86, ENSAT IV vs. II and I HR = 5.26, tumor size (cm) HR = 1.16, tumor weight (g) HR = 1.0007, high-grade vs. low-grade HR = 3.42, G0S2 methylated vs. unmethylated HR = 2.65]. G0S2 hypermethylation remained significant in all multivariate models (Table 1). These observations demonstrate that G0S2 hypermethylation independently predicts rapidly recurrent disease course prior to detection of macroscopic disease spread, and routinely fatal disease course in the setting of disseminated disease.

Though G0S2 hypermethylation independently predicts uniformly dismal disease course, patients without G0S2 methylation exhibited heterogeneous outcomes (Fig. 4B and C). We sought to determine if alternative molecular predictors could resolve this heterogeneity by separating patients with certain favorable prognosis from those with intermediate recurrence risk. We and others have shown that a score derived from expression of BUB1B and PINK1 (BUB1B-PINK1) can stratify ACC into “good prognosis” and “bad prognosis” groups (17, 18). The disease course of “good prognosis” ACC has been likened to that of patients with ACA, as patients were primarily cured by surgery. Interestingly, “good prognosis” ACC had BUB1B-PINK1 indistinguishable from ACA (17).

We evaluated BUB1B-PINK1 in FMUSP+UM Primary ACC and ACA Cohorts. We then performed an internal k-fold cross validation (k = 5, 20 repeats) on BUB1B-PINK1 score to evaluate the performance of a logistic regression model predicting any history of metastasis in G0S2 unmethylated ACC (Supplementary Fig. S7 depicts fitted logistic regression model ROC curve with ROC AUC 0.840; 95% CI, 0.7177–0.9619; model is described in Supplementary Table S8). At BUB1B-PINK1 <5.200, we predicted metastasis in patients with G0S2 unmethylated carcinomas with 100% sensitivity and 31.58% specificity. We assigned carcinomas from the FMUSP+UM Primary ACC Cohort to three groups: ACC I (G0S2 unmethylated, BUB1B-PINK1 >5.200), ACC II (G0S2 unmethylated, BUB1B-PINK1 <5.200), and ACC III (G0S2 methylated).

ACA and ACC I tumors had no difference in BUB1B-PINK1 (P > 0.05, Kruskal–Wallis with Dunn's multiple comparisons test), whereas ACC II and ACC III had different BUB1B-PINK1 from ACA (P < 0.0001) and ACC I (II vs. I: P < 0.005, III vs. I: P < 0.0001). ACC II and ACC III had indistinguishable BUB1B-PINK1 (P > 0.05), suggesting BUB1B-PINK1 cannot further stratify G0S2 methylated carcinomas (Fig. 5A). Using this combination of BUB1B-PINK1 and G0S2 methylation status, we stratified the FMUSP+UM Primary ACC Cohort into three groups with variable risk of recurrence (Fig. 5B) and death (Fig. 5C). In patients with G0S2 unmethylated carcinomas, we could now distinguish those who remain disease free and alive (ACC I) from those with history of recurrence and death (ACC II). All clinical and molecular data are summarized in Supplementary Table S9.

These results demonstrate the combined utility of G0S2 methylation and BUB1B-PINK1 score in stratifying patients into three groups, two of which have uniformly favorable or dismal outcomes. These data illustrate a strategy for implementing molecular biomarkers in series to precisely define risk categories in ACC, with high potential to impact clinical management.

ACC is a rare cancer with variable outcomes inadequately stratified by clinical and histologic metrics. ACC-TCGA identified three molecular subtypes of ACC and posited that clinical heterogeneity arises from unique transcriptional and epigenetic programs driving each class (21). We noted that the genomes of rapidly recurrent carcinomas are characterized by aberrant methylation directed to promoter CpG islands, “CIMP-high.” In this study, we also identified that CIMP-high carcinomas comprise a distinct molecular subtype of ACC, bearing upregulation of cell cycle- and DNA damage-associated cellular programs. However, prospective assessment of this complex signature is infeasible for routine molecular diagnostics.

Here, we identified that hypermethylation and silencing of G0S2 is a hallmark of ACC-TCGA CIMP-high carcinomas. In an independent cohort, we determined that G0S2 hypermethylation is restricted to 40% of ACC, absent from ACA and physiologic tissues. We then demonstrated that measurement of G0S2 methylation using a straightforward, overnight assay independently identifies a homogeneous subgroup of ACC patients with rapidly recurrent and fatal disease course. G0S2 methylation is essentially binary (carcinomas are either G0S2 methylated or G0S2 unmethylated), subverting a requirement for complicated analytical strategies and reference samples. G0S2 hypermethylation almost invariably predicts rapidly recurrent and fatal disease in patients with localized, locoregional, and disseminated ACC, including one third of patients with low-grade disease. Interestingly, we observed only one patient with tumor G0S2 hypermethylation who remains disease free >24 months following R0/RX resection. Given that adjuvant mitotane therapy is the standard of care at FMUSP and UM, our data suggest that G0S2 hypermethylation predicts short-lived remission regardless, reinforcing the need to develop improved adjuvant therapies for high-risk patients.

Expert opinion proposes that adjuvant cytotoxic chemotherapy should be considered as alternative to mitotane in high-risk patients (32, 33). However, a precise definition of “high risk” is lacking, relying on histologic grade and subjective clinical assessment. Our study suggests that prospective assessment of G0S2 methylation would objectively identify uniformly high-risk patients. Additionally, we illustrated that G0S2 methylation can be combined in series with validated biomarkers (BUB1B-PINK1) to stratify ACC into three groups, with uniformly favorable (recurrence free), intermediate, and uniformly dismal (inevitable recurrence) clinical outcomes. Such a strategy could dramatically improve clinical management and direct future trials on adjuvant therapies (Fig. 5D). The major ongoing clinical trial evaluating the efficacy of adjuvant mitotane in low-intermediate risk ACC (“ADIUVO,” NCT00777244) defines risk using grade; our study suggests this criterion is inadequate, as up to one third of these patients will have tumor G0S2 hypermethylation and likely recur on adjuvant mitotane. As new clinical trials are designed to evaluate adjuvant therapies in high-risk patients, we propose assessment of G0S2 methylation to determine risk as in Fig. 5D.

High-risk CIMP-high/G0S2 methylated ACC is associated with a unique transcriptional, copy number, and mutational landscape in ACC-TCGA, suggesting a common biological program underlies this aggressive ACC subtype (21). We demonstrated that CIMP-high carcinomas are chromosomally noisy, frequently bear somatic alterations leading to activation of cell cycle, and exhibit a transcriptional program characterized by increased expression of steroidogenic enzymes, proliferation machinery, and genes coordinating DNA damage-associated processes. Cell cycle and DNA damage-associated genes upregulated in CIMP-high ACC include MELK, AURKB, CDK6, PLK1, and TOP2A which have been successfully targeted in preclinical and translational models of ACC (34–38), and may even predict clinical responsiveness to combination therapy with etoposide, doxorubicin, cisplatin, and mitotane (39). Although there is currently little data to support a clinical trial evaluating utility of demethylating agents alone in ACC (40, 41), studies in other solid tumors demonstrate that epigenetic priming with demethylating agents may increase efficacy of cytotoxic therapies and targeted DNA repair inhibitors (42–44). Together, these observations suggest that therapies targeting the cell cycle, DNA repair, and epigenetics may be efficacious in patients with CIMP-high/G0S2 methylated ACC and warrant future study.

The molecular mechanisms driving CpG island hypermethylation in IDH1/2-wild-type cancers including CIMP-high ACC are still poorly understood (45). Our data and studies identifying G0S2 hypermethylation in other cancer types (46, 47) suggest that methylation of this locus is driven by the same unknown molecular programs supporting hypermethylation in other regions of the CIMP-high cancer genome. However, the high expression of G0S2 in lipid-rich tissues including the adrenal gland (Supplementary Fig. S3) suggests that G0S2 may have tissue-specific tumor suppressor roles. Although G0S2 has been best characterized as a regulator of lipid metabolism (48), recent studies have demonstrated that methylation-dependent silencing of G0S2 in breast cancer augments oncogenic PI3K/mTOR signaling (49) and MYC transcriptional activity (50). These studies suggest that G0S2 may have important roles in adrenocortical biology, including a similar tumor suppressor function worthy of future investigation.

In conclusion, our study is the first to reduce the complex genome-wide CpG island hypermethylation signature from ACC-TCGA to a single, binary molecular marker, amenable to targeted assessment using routine molecular diagnostics. Assessing G0S2 methylation as we have here is inexpensive, straightforward, compatible with a timeline feasible for clinical decision-making, and will enable the direction of efficacious adjuvant therapies for patients with uniformly aggressive ACC. Future studies will be directed towards evaluating G0S2 methylation prospectively, in circulating tumor DNA, and in readily available clinical samples including formalin-fixed paraffin-embedded tissues.

D.R. Mohan, A.M. Lerario, and G.D. Hammer are listed as co-inventors on a provisional patent application on compositions and methods for characterizing cancer that is owned by The Regents of the University of Michigan. T. Else is a consultant/advisory board member for HRA Pharma. G.D. Hammer is a consultant/advisory board member for Millendo Therapeutics. No potential conflicts of interest were disclosed by the other authors.

Conception and design: D.R. Mohan, A.M. Lerario, G.D. Hammer

Development of methodology: D.R. Mohan, A.M. Lerario, B.M.P. Mariani, M.C.N. Zerbini, S.K.N. Marie

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): D.R. Mohan, A.M. Lerario, T. Else, M.Q. Almeida, M. Vinco, J. Rege, B.M.P. Mariani, M.C.N. Zerbini, B.B. Mendonca, A.C. Latronico, S.K.N. Marie, W.E. Rainey, M.C.B.V. Fragoso, G.D. Hammer

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): D.R. Mohan, A.M. Lerario, T. Else, B. Mukherjee, T.J. Giordano, G.D. Hammer

Writing, review, and/or revision of the manuscript: D.R. Mohan, A.M. Lerario, T. Else, B. Mukherjee, M.Q. Almeida, M.C.N. Zerbini, A.C. Latronico, S.K.N. Marie, W.E. Rainey, T.J. Giordano, G.D. Hammer

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): D.R. Mohan, A.M. Lerario, B.B. Mendonca

Study supervision: A.M. Lerario, G.D. Hammer

This work is supported by the University of Michigan Rogel Cancer Center (grant to G.D. Hammer), and supported (in part) by the NIH through the University of Michigan's Cancer Center Support Grant (5 P30 CA46592). D.R. Mohan was/is supported by the University of Michigan Medical Scientist Training Program (5 T32 GM7863), the University of Michigan Doctoral Program in Cancer Biology, the University of Michigan Rogel Cancer Center, and The Drew O'Donoghue Fund. The authors would like to express their deepest gratitude to all study participants for their generous contribution, without whom advances in ACC research and clinical care would be impossible. The authors would also like to thank Sueli Oba-Shinjo, PhD; Stella Goncalves, BS; Isabele F. Moretti, BS; and Thais F. Galatro, PhD; for their assistance in processing patient samples from Faculdade de Medicina da Universidade de São Paulo.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.