DNA methylation profiling predicts postsurgical regrowth in SF1-lineage nonfunctioning pituitary neuroendocrine tumors

Abstract

Background

Nonfunctioning pituitary neuroendocrine tumors (NFPitNETs) account for ∼30-35% of PitNETs; ∼75% arise from the SF1 lineage. Recurrence remains common despite resection (∼30% in 10 years), and routine histopathology/IHC has limited value in predicting recurrence risk. This study evaluated whether DNA methylation profiling improves recurrence risk stratification.

Materials and Methods

Genome-wide tissue methylation (Illumina EPIC v1, 850K) was analyzed in 117 retrospective NFPitNETs with clinical and imaging follow-up. Unsupervised consensus clustering defined methylation-based subgroups, followed by supervised differential methylation analysis to identify cluster-specific differentially methylated probes (DMPs). A classifier was trained using these signatures, with predicted subgroup memberships correlated with regrowth and progression-free survival (PFS). To ensure reliable estimations, longitudinal mixed-effects models were restricted to the interval of model stability (∼9 years), reflecting cohort follow-up. External validation was performed in 3 independent cohorts.

Results

Five clusters (k1-k5) emerged: 4 SF1-positive-predominant (k1, k2, k3, and k5) and 1 TPIT/PIT1-enriched NFPitNETs (k4). Among the 562 DMPs, many mapped to genes regulating cell-cycle and immune pathways. Compared with k1-k2, k3, k4, and k5 possessed significantly higher recurrence risk. Within SF1-lineage tumors, k3 exhibited postoperative tumor-volume expansion beginning at ∼6 years. The methylation-based classifier achieved ∼97% accuracy in assigning clusters and maintained prognostic separation across independent cohorts.

Conclusions

DNA methylation profiling identifies biologically and clinically distinct NFPitNET subgroups, particularly within the SF1 lineage, and may enhance prediction of recurrence risk. Prospective validation and demonstration of clinical utility are warranted to support integration into precision management workflows.

Key Points.

• DNA methylation profiling predicts regrowth in SF1-positive NFPitNET after surgery.

• A methylation-based machine-learning classifier achieves 97.5% accuracy across external cohort to stratify NFPitNET clusters.

• Methylation-based classification shows potential clinical utility for recurrence risk stratification.

Importance of the Study.

Nonfunctioning pituitary neuroendocrine tumors (NFPitNETs) lack reliable biomarkers for predicting postoperative regrowth, posing major challenges for individualized patient management. This study presents the largest genome-wide DNA methylation analysis of NFPitNETs with a long imaging follow-up to date, particularly of SF1-lineage, identifying 5 distinct methylation-defined clusters or subgroups with significant differences in progression-free survival. Among the SF1-positive NFPitNET, 1 subgroup demonstrated a markedly increased risk of long-term tumor regrowth. A machine learning classifier trained on subgroup-specific methylation signatures accurately predicted subgroup membership and maintained high performance in independent external cohorts and across tissue preservation methods, underscoring its robustness and translational potential. Building on prior studies that advanced prognostic classification using PitNET histopathological and clinical features, our findings suggest that epigenomic profiling can further refine prognostic assessment by revealing additional layers of biological and molecular heterogeneity not evident through conventional methods; potentially serving as a valuable complement to established clinical and pathological criteria. Together, these results highlight the clinical relevance of tumor methylation profiling and its potential to guide postoperative surveillance and therapeutic decision-making in NFPitNETs. Future efforts should focus on prospective validation and integration of DNA methylation classifiers into clinical workflows to refine risk assessment and improve personalized management strategies.

Pituitary neuroendocrine tumors (PitNETs) account for 17% of all intracranial tumors.¹ While 40% of PitNETs are prolactinomas, which are primarily treated nonsurgically, surgical removal is the mainstay treatment for nonfunctioning (NFPitNET) and functioning PitNETs.² PitNETs can additionally be divided into groups based on immunohistochemical (IHC) staining for transcription factor subtypes: SF1 (Steroidogenic factor 1; NR5A1) is expressed in gonadotroph PitNET, which accounts for 70%-75% of NFPitNETs,³ TPIT (T-box transcription factor; TBX19) in corticotroph PitNET, and PIT1 (Pituitary-Specific Positive Transcription Factor 1; POUF1) in PRL-, GH- and TSH-producing PitNETs.⁴

Following surgical intervention, the ongoing challenge lies in identifying cases at higher risk of recurrence, especially for NFPitNETs.⁵ Several classifications have been proposed to better predict NFPitNET recurrence, including IHC staining for transcription factors (TFs),⁶ Hardy and Knosp imaging-based grading depicting invasive growth,^7 ,8 and preoperative tumor volume.⁹ More recently, a 9-tiered prognostic classification, combining evidence-based risk factors for recurrence (preoperative tumor size, tumor phenotype, invasion, and mass effect among other risk factors), was proposed by the Pituitary Society during the Pituitary Adenoma Nomenclature 3 (PANOMEN 3) clinical classification workshop.¹⁰ Retrospective data supported the prognostic utility of PANOMEN-3 to predict recurrence and the need for additional therapy; however, prospective multicenter validation is in progress. However, this classification does not differentiate recurrence risk specifically for the SF1 subtype.¹⁰

Histopathological analysis, including staining for hormonal markers and TFs, is crucial for clinical diagnosis, but it does not always capture the full complexity of the underlying tumor biology, particularly in heterogeneous tumors like PitNETs.^11,12

Molecular studies, specifically involving DNA methylation profiling, offer a more sensitive and comprehensive method for identifying distinct tumor subgroups that may not be apparent in histopathological analysis alone.^13–18 Some studies discovered a positive correlation between DNA methylation and aggressive behavior,¹² invasive growth¹⁹ and progression-free survival (PFS).^20,21 Suggesting the potential of DNA methylation profiling to personalize postoperative follow-up. However, the clinical application of DNA methylation patterns must be refined and further validated. We previously published a pilot study,²² which incorporated prognostically relevant methylation markers as potential predictors for recurrence. We identified methylation subclusters within the SF1-lineage subtype that exhibited variability in recurrence risk; however, survival differences remained statistically insignificant, likely due to limitations in sample size.

This retrospective study primarily aimed to validate our previously reported molecular subclassification of NFPitNETs by profiling DNA methylation in the largest SF1-lineage cohort to date, with detailed follow-up, assessing associations with PFS and regrowth, and evaluating the classifier across additional cohorts. Secondly, we explored the underlying biology of the methylation-defined subtypes through canonical pathway analyses.

Methods

Inclusion Criteria—Internal Cohort

Using a previously established database at the Department of Neurosurgery, Odense University Hospital, we included all patients older than 18 years who were operated on for NFPitNET from 2007 to 2017 and had available tumor tissue in the biobank at the Department of Pathology, along with complete clinical data and pre- and postoperative magnetic resonance imaging (MRI). The current study included patients and samples used in the pilot study (Batch #1)²² and new patients from the biobank (Batch #2), distinguished only by the date of DNA methylation profiling.

Patients with NFPitNET were followed in a multidisciplinary endocrinological and neurosurgical outpatient clinic and had annual clinical, radiological, and biochemical follow-up. Patients with PitNET surgery before the study period were classified as secondary surgeries for their first operation within the study period; any subsequent operations or radiotherapy were considered reinterventions.

Ethical clearance was granted by the Danish Health Research Ethics Committee system (internal grant ID: Acadre 17/46106, Project-ID: S-20170216). Furthermore, approval for data extraction and proper handling was granted by the Danish Data Protection Agency (journal nr. 16/25477).

Recurrence, Progression, and Volumetric Analysis

As previously described, 3D volumetric analysis of residual tumor was conducted longitudinally on annual follow-up MRIs in Horos (v4.0.1).²³ Briefly, a 3D volumetric analysis was performed on all available MRIs based on axial, coronal, or sagittal sections on T1-weighted images with gadolinium contrast. Tumor was delineated and volume was calculated through software in Horos. Preoperative invasive growth was evaluated by Hardy and Knosp classifications.^7 ,8 Invasive growth was defined as Hardy’s classification of 3 or 4 and/or D or E, and Knosp’s classification of 3 or 4. Progression was defined as any remnant tumor growth or recurrence following gross total resection.

Immunohistochemical Staining

All tissue samples were reevaluated according to the 2022 WHO Classification of Endocrine and Neuroendocrine Tumors,³ as detailed previously.²³ In short, formalin-fixed, paraffin-embedded (FFPE) blocks were sectioned (3 μm), deparaffinized, and subjected to heat-induced epitope retrieval in tris-based buffer (CC1). Endogenous peroxidase was blocked (1.5% H₂O₂). Sections were stained with primary antibodies against PIT1 (1:500), SF1 (1:500), and TPIT (1:1000), followed by species-specific peroxidase-conjugated secondary antibodies (OptiView). Visualization was performed using DAB (brown chromogen), and nuclei were counterstained with Mayer’s hematoxylin (blue). Positive controls included tissue microarrays and known PitNETs; negative controls were prepared through omitting the primary antibody.

PANOMEN-3 Grading

All samples were classified with the PANOMEN-3 system proposed by Ho et al.¹⁰ This grading (grades 0-3), developed by the Pituitary Society, integrates evidence-based risk factors from routine care to predict recurrence or need for therapy as suggested by retrospective study.¹⁰ Nine risk domains were scored for each patient on a 0/1 (some 2) scale: function status, secretory status, hypopituitarism, tumor size, mass effect, invasion, residual tumor (post-operative), histopathology, and syndromic genetics. The corrected score (sum divided by applicable domains) was calculated and mapped to Grade 0-3 (higher more severe) per PANOMEN-3 specification; pre- or postoperative versions were used as appropriate.

Methylation Analysis

Archived FFPE tissue samples for each patient were retrospectively collected from the Department of Pathology, Odense University Hospital. DNA extraction and methylation analysis were performed as previously described²² using the Illumina MethylationEPICv1 array (Illumina Inc.) according to the manufacturer’s protocol. Batch #1 (n = 42) and Batch #2 (n = 76) were profiled using EPICv1 (850K) in September 2022 and June 2024, respectively. Raw signal intensities were extracted using the function “readmetharray.exp” with appropriate hg38 annotations and corrected for background fluorescence intensities and red-green dye-bias using the function “preprocessIllumina.” Beta-values were calculated as (M/(M + U)), in which M and U refer to the (preprocessed) mean methylated and unmethylated probe signal intensities, respectively, using the function “getBeta.” Measurements in which the fluorescent intensity was not statistically significant above background signal (detection P-value > 10E-16) were removed from the dataset. Before analysis, we removed probes designed for sequences with known polymorphisms (SNP), probes with poor mapping quality and the X and Y chromosomes (a complete list of masked probes is provided by Zhou et al.²⁴ (Supplementary Figure S1a). Correction for batch effect was conducted through application of ComBat (v0.0.4),²⁵ with visualization provided in Supplementary Figure S1b and c. The top 1000 most variant probes were isolated across the cohort according to calculated beta-value row variance.

Unsupervised Methylation Clustering Analysis

Consensus clustering was performed using the ConsensusClusterPlus package (v1.68.0) on the top 1000 most variable probes across samples. Agglomerative hierarchical clustering with Euclidean distance and Ward.D2 linkage was applied over 5000 resampling iterations, allowing up to 12 clusters, as previously described.²⁶ The optimal number of clusters was determined based on inspection of the cumulative distribution function (CDF) plots and delta area plot to assess the relative increase in consensus stability, and visual inspection of consensus heatmaps to determine well-separated block structures. This was done on the full cohort, as well as on SF1-positive samples only.

Supervised Methylation Clustering Analysis

Following cluster generation, a genome-wide supervised analysis was performed to identify cluster-specific differentially methylated probes CpG (DMPs). CpG probe-level comparisons were conducted between samples of a single cluster and the remaining samples in the cohort using a nonparametric 2-sided Wilcoxon rank-sum test; P-values were adjusted for false discovery rate (FDR). To reduce the potential for background noise capture, we utilized volcano plots to visually guide selection of the DMPs, defined as CpG sites, which retained significant FDR-adjusted P-values and mean methylation differences (diff.mean) between groups (diff.mean ≥20%; Wilcoxon rank sum tests P-value_FDR <.05).

Predictive Modeling and Validation

To investigate the potential for prognostic application of DNA methylation signatures, we used random forest (RF) machine learning to construct a robust multiclass classifier that assigns samples to specific methylation clusters. Prior to the initiation of any model construction, cohorts were randomized through machine-driven processes into discovery (70%) and independent validation (30%) sets. The discovery set was further randomized into a training set (80%), used for identifying relevant signature sets and classifier construction, and a testing set (20%), used for selecting the most robust classifier across “unseen” data.

For each cluster, the most significant cluster-specific DMPs (∼100) were selected through filtration of significance (P  _FDR < .05) and sorting of significant CpGs by descending absolute mean methylation difference (diff.mean). These signatures were then used to train the classifier through an automated process. Using the identified signatures, we generated a classifier using the function “train” (caret, v6.0.94) with 500 decision trees; 10-fold cross-validation was conducted across the training set. Following construction, the classifier was tested across the remaining testing set; receiver operating characteristic (ROC) curves were generated. Steps taken from splitting of the discovery set to classifier testing and ROC generation were repeated for 50 total iterations. The optimal classifier was selected using traditional ROC parameters.

Independent External Cohorts

Following model selection, we conducted a 2-step validation process, first applying the model to the aforementioned independent validation set, in addition to obtained DNA methylation profiles from 3 publicly available data sets (Silva-Junior et al.³ [GSE207937], Neou et al.⁴ [E-MTAB-7762], and Belakhoua et al.⁵ [GSE283928]).

To evaluate whether the model was a reliable prognostic tool regardless of tissue sample preservation method, we also assessed its performance using paired tumor tissues preserved as FFPE or snap-frozen at −80°C immediately after resection. DNA methylation analysis was conducted on these samples following the same procedures used for the primary cohort.

Integrative Analysis Between Publicly Available Paired DNA Methylation and RNA-Sequencing

To identify differentially expressed genes (DEGs) across clusters, we performed supervised analyses on the 2 publicly available paired DNA methylome and transcriptome datasets (GSE207937 and E-MTAB-7762), namely Silva-Junior (n = 77) and Neou cohorts (n = 46). First, the methylation dataset was classified according to k-cluster memberships using our RF-based classifier. Then, using matching expression data of genes targeted by methylated probes and DESeq2 (version 1.44.0), we identified DEGs that were cluster-specific, primarily focusing on the methylation group k3 NFPitNET, which showed more regrowth relative to the other clusters. Subsequently, we mapped the DEGs to CpGs located in their regulatory elements (promoters and enhancers), selecting those CpGs that simultaneously presented differential cluster-specific DNA methylation levels and negative correlation with the expression of putative target genes, termed probe-gene pairs (PGPs). To infer a potential epigenetic regulatory action of the DMPs identified through the supervised comparisons between clusters, DMPs were mapped to their putative target gene using the EPIC manifest (hg38). Enhancer elements were defined using the GeneHancer database (hg38) provided by the UCSC Genome Browser. Promoter elements were defined using GENCODE v.31 annotations,²⁷ with consideration of CpGs located 200 bp up/downstream from the target gene.

Gene Set Enrichment Analysis

We performed gene set enrichment analysis (GSEA) using the DEGs to identify significantly enriched pathways and cellular processes across each cluster-specific set of PGPs and potentially relevant intracellular signaling processes through the package clusterProfiler (v4.14.4).²⁸

Copy Number Variation Inference Across NFPitNETs (Emphasis on SF1-Lineage)

We inferred copy-number variation (CNV) from the methylation arrays through application of Conumee2 v2.1.2, using 5 nontumoral pituitary controls as reference²⁹ Gene lists used as input for CNV calling were compiled through comprehensive literature searches and inclusion of differentially methylated probes (DMPs) located in gene regulatory regions, derived from inter-cluster comparisons (Figure 1a). After removal of noninformative and poorly annotated loci, we obtained a final set of 716 unique genes: 221 from DMP analyses (Supplementary File S4) and 495 derived from literature (Supplementary File S5). Gene-level amplifications and deletions were defined through thresholds of |alteration|>0.25. To investigate differences across SF1-NFPitNETs clusters, we evaluated CNV burden using 2 designs applied to both gene lists: (1) one-vs-rest comparisons for each SF1-lineage cluster (eg, k1 vs k2+k3+k5; k2 vs k1+k3+k5) and (2) k1+k2 versus k3+k5, based on the PFS differences observed between these groups on Kaplan-Meier analysis (Figure 2b). Proportional differences between clusters were tested using Fisher’s Exact or Chi-square tests with Yates correction.

Figure 1.

(a) DNA methylation heatmap displaying 562 of the most differentially methylated probes (DMPs; β-values) across PitNET unsupervised k-means clusters (n = 117). Samples are sorted into methylation-based clusters and are annotated with relevant clinicopathological/molecular features. Vertical tracks (right): genomic annotations. (b) Volcano plots illustrating the selected cluster-specific DMPs derived from cluster-oriented supervised comparisons (nonparametric Wilcoxon rank-sum test).

Figure 2.

(a) Kaplan-Meier survival curves depicting progression-free survival (PFS) for samples stratified by their k-clustering (n = 111, vertical ticks: censorship). Median survival time across clusters was compared using log-rank tests. (b) Kaplan-Meier curves for PFS in SF1-positive PitNETs only. (c, d) Linear mixed-effect model of tumor volume, adjusted for preoperative tumor volume, invasive growth, and reintervention: (c) all samples and (d) SF1-positive PitNET only.

Statistical Analysis

All methylation analysis was performed in R (v. 4.4.1). Nonparametric 2-sided Wilcoxon rank-sum tests and testing adjustments (eg, FDR) were used to identify significant differences in binary group comparisons. Machine-learning classifiers were formulated using a RF algorithm. Receiver operating characteristic curves were used to estimate the predictive power for each iteration of the RF classifier (false positive rate (1−specificity [SP]), true positive rate (sensitivity [Se])). Kaplan-Meier survival analysis was used to estimate PFS by DNA methylation clusters. PFS was defined as any remnant tumor growth or recurrence following gross total resection. Hazard ratios were calculated using a Cox proportional hazards model with cluster 1 as reference.

To assess potential differences in tumor regrowth patterns among DNA methylation clusters, we used a linear mixed-effects model to account for the variability in tumor growth (volume) over time, conducted in STATA (v18.5). This model adjusted for preoperative tumor volume, invasive growth, and reinterventions, with a random intercept for each patient (Supplementary File S1). Analyses were confined to the first 10 postoperative years as many patients had irregular follow-ups beyond this period. A 95% confidence interval (CI) was reported for each annual cluster comparison, with statistical significance defined as P < .05. Fisher’s exact test was used to evaluate associations between invasive growth, primary surgery, and reintervention.

Data/Code Availability

The raw de-identified (SNP-removed) DNA methylation intensity data files (EPIC Array v1; .idat), as well as generated classifier, have been deposited to Mendeley Data under accession DOI: 10.17632/vr2wb38r6t.1. Source codes necessary for production of each of the main and supplemental findings may be found through GitHub [https://github.com/gherrgo/SF1-Post-Surgical-Regrowth]. Any additional necessary information may be found in the Supplementary Files.

Results

Clinical Characteristics

Genome-wide DNA methylation was performed on 118 NFPitNET samples. The results from 42 of these samples were published in a pilot study.²² One sample (Sample ID#142) did not meet the criteria for inclusion based on quality control assessment (Supplementary Figure S1a), leaving 117 samples for further preprocessing and methylation analysis. Clinical characteristics for the cohort are shown in Table 1. Most of the NFPitNETs (77.8%) were SF1-positive. Over half had invasive growth preoperatively (52.1%), and 25% needed reintervention after initial surgery. The mean preoperative tumor volume was 8.9 (7.5-10.3) cm³, and the mean postoperative tumor volume was 3.6 (2.4-4.9) cm³. Remnant tumor tissue was present in 94 cases (82.5%) equaling gross total resection in 20 cases (17.5%). Six patients were lost to follow-up. The median follow-up time was 7 years and 1 month (85 months).

Table 1.

Characteristics of the NFPitNET Clusters

	Total	Cluster 1	Cluster 2	Cluster 3	Cluster 4	Cluster 5
	n = 117	18 (15.3)	32 (27.4)	27 (23.1)	20 (17.1)	20 (17.1)
Median age (years) (min-max)	62.5 (26-87)	64 (41-80)	62 (32-81)	67 (26-87)	58 (35-72)	56 (43-74)
Gender
Males	77 (65.8)	11 (61.1)	26 (81.2)	20 (74.1)	5 (25)	14 (70)
Females	40 (34.2)	7 (38.9)	6 (18.8)	7 (25.9)	15 (75)	6 (30)
Lineage
SF1	91 (77.8)	16 (88.9)	30 (93.7)	24 (88.9)	1 (5)	20 (100)
PIT1	5 (4.3)	0	0	0	5 (25)	0
TPIT	13 (11.1)	0	0	0	13 (65)	0
Unknown	8 (6.8)	2 (11.1)	2 (6.3)	3 (11.1)	1 (5)	0
Invasive growth
No	56 (47.9)	8 (44.4)	18 (56.3)	11 (40.7)	8 (40)	10 (50)
Yes	61 (52.1)	10 (55.6)	14 (43.7)	16 (59.3)	12 (60)	10 (50)
PANOMEN 3
Grade 1	8 (6.8)	1 (5.6)	2 (6.2)	1 (3.7)	2 (10)	2 (10)
Grade 2	53 (45.3)	6 (33.3)	17 (53.1)	10 (37)	9 (45)	11 (55)
Grade 3	56 (47.9)	11 (61.1)	13 (40.6)	16 (59.3)	9 (45)	56 (47.9)
Mean preoperative tumor volume (cm3) (95% CI)	8.9 (7.5-10.3)	9.1 (6.7-11.5)	7.6 (5.6-9.7)	9.7 (7.6-11.8)	10.9 (4.1-17.7)	8.1 (6.4-9.8)
Remnant tumor
No	20 (17.5)	2 (11.1)	7 (21.9)	5 (18.5)	3 (15)	4 (20)
Yes	94 (82.5)	16 (88.9)	25 (78.1)	22 (81.5)	17 (85)	16 (80)
Mean postoperative tumor volume (cm3) (95% CI)	3.6 (2.4-4.9)	3.5 (1.5-5.5)	2.4 (1.4-3.5)	4.8 (2.9-6.7)	7.4 (0.6-14.2)	3.1 (2.0-4.2)
Patients lost to follow-up, n	6 (5.3)	0 (0)	2 (6.2)	1 (3.7)	1 (5)	2 (10)
Reintervention
No	83 (74.8)	14 (77.8)	26 (86.7)	18 (69.2)	12 (63.2)	13 (72.2)
Yes	28 (25.2)	4 (22.2)	4 (13.3)	8 (30.8)	7 (36.8)	5 (27.8)
Median follow-up time (months) (min-max)	85 (6-174)	97 (46-155)	82.5 (23-167)	86.5 (6-159)	85 (8-174)	82.5 (26-169)

Otherwise indicated, data are depicted as number of patients (%).

PANOMEN-3 Grading Distribution Does Not Differ Across Methylation Clusters

Based on the PANOMEN 3 classifier, 6.8% were grade 1, 45.3% were grade 2, and 47.9% were grade 3. The distribution of PANOMEN grades did not significantly differ across the methylation clusters (P = .72) (Table 1).

Methylation-based Clustering Identified Five Distinct NFPitNET Subgroups

After DNA methylation preprocessing, unmasking procedures, and corrections for batch differences, a total of 619 329 probes were available for final analysis. According to CDF and delta area plots, the greatest improvement in stability occurred between clusters k = 4 and k = 5, after which the gains plateaued. At k = 5, the consensus heatmap displayed clear, well-separated block structures, and the mean consensus score for each cluster exceeded 0.8, indicating high stability. Importantly, the k = 5 also aligned with biologically interpretable groups, that is, 4 predominantly SF1-lineage clusters (k1,2,3,5) and one enriched for TPIT/PIT1-lineage tumors (k4), supporting its selection (Figure 1a). The initial clustering observed within the pilot study correlated well with our new findings. Notably, invasive growth (P = .909), secondary surgery (P = .100), and reintervention (P = .646) were randomly distributed across clusters.

Supervised analysis revealed 562 cluster-specific probes that were significantly differentially methylated across clusters (Figure 1a). Volcano plots illustrate cluster-specific DMPs for each cluster compared to the remaining samples across the cluster-specific DMPs (Figure 1b).

When clustering was repeated only within SF1-lineage tumors, 2 samples (1.7%) changed cluster assignment compared with the original clustering done on the full dataset.

Differential Tumor Regrowth Patterns Across Methylation Clusters

K-clusters 3, 4, and 5 had a significantly higher risk of regrowth compared to reference k1 (hazard ratio [HR] = 2.57, 3.63, and 2.69, with P-values of .02, .002, and .018, respectively). No statistically significant difference was observed between k2 and reference k1 (HR = 1.1, P = .81) (Figure 2a). In SF1-lineage NFPitNETs only, increased hazards for regrowth were observed in k3 and k5 compared to reference k1 (HR = 2.62 and 2.73, P-values = .018 and 0.17, respectively), while no significant difference was observed for k2 (HR = 1.11, P = .81) (Figure 2b).

Linear mixed-effects analysis of tumor volume after surgery showed no significant differences across the entire NFPitNET cohort (Figure 2c). For SF1-lineage NFPitNETs, k3 showed a statistically significant increase in tumor volume after 6 years of follow-up, in contrast to other SF1 NFPitNET clusters (P = .004; Figure 2d). Because follow-up imaging data became sparse beyond this point, model estimates lost stability after approximately 9 years, making extrapolation to fixed 5- or 10-year intervals unreliable. We therefore report results within the period where the model assumptions remained valid and estimates were statistically robust.

Methylation Classifier Is Validated in Independent NFPitNET Cohort

Using the cluster-specific DNA methylation signatures from our discovery cohort, we developed a prognostic predictive model that classified our independent validation cohort according to k-cluster membership with high accuracy (AUC = 0.97) (Figure 3a). To assess external validity, we applied the model to 3 publicly available PitNET methylome datasets: Silva-Junior et al.(n = 77)¹³ (Figure 3a), Neou et al.(n = 46)¹⁴ (Figure 3b), and Belakhoua et al.(n = 118)¹⁵ (Figure 3c), all of which included SF1-lineage NFPitNETs. The classifier reproduced the 5-cluster structure identified in our study, including the 4 SF1-lineage clusters and the single TPIT/PIT1-lineage cluster, with 97.5% overall accuracy (Figure 3a-c). Predicted clusters were highly correlated with tumor lineage across all cohorts. In the Belakhoua et al.dataset,¹⁵ recurrence was reported in 10.2% (12/118) of cases, distributed across k2 (n = 3), k3 (n = 2), and k4 (n = 7) (Figure 3c).

Figure 3.

(a-c) DNA methylation heatmaps displaying the 562 most differentially methylated probes (DMPs; β-values) across assigned PitNET k-means clusters in publicly available data taken from (a) Silva-Junior et al.(n = 77); (b) Neou et al.(n = 46); and (c) Belakhoua et al.(n = 118). Samples are sorted into methylation-based clusters and annotated with clinicopathological/molecular features. Vertical tracks (right): cluster-specific DMP group comparisons.

To further test reproducibility, we applied the model to prospectively collected paired FFPE and fresh-frozen samples from 10 patients undergoing surgery for NFPitNET. All but one sample (ID 270) clustered in the same subgroup as its paired sample, and all samples clustered according to lineage (Supplementary Figure S1d). Notably, when the model trained on FFPE samples was applied to paired fresh-frozen material, 90% (9/10) were assigned to the same cluster as their FFPE counterparts, matching tumor lineage.

Distinct Methylation Patterns and Gene Promoter Associations for NFPitNET Clusters

Cluster-specific DMPs revealed distinct methylation patterns across the clusters. Cluster k1 and k2 were predominantly hypomethylated (93 hypo- and 15 hypermethylated probes for k1; 79 hypo- and 45 hypermethylated probes for k2), while k3, 4, and 5 were mostly hypermethylated (84 hyper- and 14 hypomethylated for k3; 65 hyper- and 45 hypomethylated for k4; 73 hyper- and 45 hypomethylated for k5) (Figure 1a, Supplementary File S2). Analysis of promoter regions, which regulate gene expression, identified a few cluster-specific DMPs. In k1, we discovered a hypermethylated promoter CpG site related to the CCDC141 gene. In k2, 6 promoters were identified: 5 hypomethylated and 1 hypermethylated, associated with genes such as RP11-204E9.1, RSPH6A, RP11-260M19.2, RP11-219G17.4, and ATP1A3. In k3, 4, and 5, we found 5, 2, and 4 promoter regions, respectively, which were associated with significant DMPs linked to genes RP11-497G19.3, ANK2, and CASC15, while k4 was associated with ATP2C1 and HDC-promoters, and k5 had promoters linked to NCKAP5 and GUCA1C (Supplementary File S2).

Sparse CNVs Suggest Nonstructural Drivers of Cluster Differences Across SF1 Subgroups

Targeted CNV profiling of relevant genes previously linked to pituitary tumorigenesis and/or prognosis in the literature and cluster-defining DMPs revealed sparse and cluster-restricted alterations across the all lineage subgroups (including SF1-lineage subgroups). For instance, we observed cluster-specific CNV enrichment (amplification or deletions) across comparisons, further detailed in Supplementary Figure S3a. For example, BMP5 amplification was more frequent in k2 (40.6%) compared to other subgroups (15.4%), while DPH6 amplification was higher in k3 (48.1%) compared to others (22.9%).

Cluster-Specific SF1-NFPitNET Displays Inverse Methylation-Expression Patterns in Cell-Cycle and Immune Genes

To gain biological insight into the mechanisms underlying the clinical differences observed across methylation-defined subgroups, we integrated DNA methylation and transcriptomic data. Given that the SF1-lineage k3 subgroup showed distinct clinical features and a higher recurrence risk compared to other methylation clusters, we focused our analysis on this subgroup. To identify tumors corresponding to k3 and examine their transcriptional correlates, we applied our methylation-based classifier to the Neou cohort,¹⁴ which includes paired methylation and transcriptomic profiles. The 10 most significant DEGs associated with k3 compared to the other SF1-lineage NFPitNETs were DHRS9, FBLN5, KCNJ2, FAM20A, FJX1, GNB4, PLA2G4A, KCNH3, TSPAN1, and TAGLN2 (Supplementary File S3). Gene set enrichment analysis revealed a comprehensive network-plot with differential expressions of relevant genes such as GALNT14, MEG3, and FAS1 (Figure 4a). The top 20 significantly enriched canonical pathways and gene ontologies are presented in main pathway headings, which primarily involve immune regulation pathways and mitotic cell cycle regulations (Figure 4b). We identified genes whose expressions were negatively correlated with the methylation levels of their regulatory regions (eg, ABR, RFS12, NCS1) and presented differential expression and methylation trends across SF1 NFPitNET k3-specific comparisons (Supplementary Figure S2a-e).

Figure 4.

(a) Network plot displaying differentially expressed genes (DEGs) derived through supervised analysis across expression data. Discovered genes are associated with significantly enriched ontological categories and are size coded according to total number of genes attributed to each term. (b) Top 20 canonical pathways from the gene ontology (GO) database related to the derived DEGs.

Discussion

This study represents the largest microarray-based DNA methylation analysis of NFPitNETs with longitudinal imaging follow-up to date. Our findings revealed 5 distinct DNA methylation subgroups, including 4 within the SF1-lineage subtype and 1 enriched with PIT1/TPIT-lineages, each characterized by unique methylation patterns, which were recapitulated and validated in external cohorts (Figure 3a-c).^13–15 Strikingly, the methylation subgroups were associated with significant differences in PFS and postoperative tumor regrowth in our cohort (Figure 2a-d), corroborating our previous pilot study findings.²²

Multiple studies suggest DNA methylation profiling as a standalone modality for diagnostic classification and/or recurrence risk stratification across PitNETs (functioning and nonfunctioning) and other CNS tumors^13–15,21. Multiomic analyses developed in some of these studies showed strong concordance between methylation-defined subgroups and those defined by other molecular layers. For instance, in PitNETs, Neou et al.⁴ demonstrated alignment of epigenetic classes with gene expression programs and mutational profiles. Across CNS tumors, Capper et al.⁰ established a methylation-based diagnostic classifier enabling reliable classification, complementary to histopathology, even without concurrent multiomic data. Jotanovic et al.² showed that genome-wide methylation patterns distinguished between aggressive, metastatic, and benign PitNETs, suggesting their prognostic implication. These findings collectively support our results and reinforce the value of DNA methylation profiling as a clinically informative and relevant framework for PitNET classification and prognosis. Future integrative studies combining methylation, transcriptomic, and proteomic data will be essential to further refine the biological underpinnings of the identified subgroups.

Within this framework, we observed that not all SF1-lineage NFPitNETs share the same clinical trajectory. Specifically, subgroups k3 and k5 exhibited recurrence rates comparable to the TPIT-enriched cluster (k4), typically associated with more aggressive behavior,^31,32 despite similar postoperative growth dynamics across subgroups (Figure 2c). These observations suggest that clinical variability may reflect epigenetic diversity within the SF1 lineage, underscoring the value of methylation-based classification for risk stratification. When analyzed independently, k3 showed significantly greater tumor expansion after 6 years. Thus, DNA methylation profiling reveals clinically meaningful heterogeneity among SF1-NFPitNETs, not discernible through histology or immunohistochemistry.

To better understand the biological mechanisms underlying these differences, we integrated DNA methylation and transcriptomic data, focusing on the higher regrowth k3. Pathway analysis of putative methylation-regulated genes revealed enrichment in cell-cycle regulation, cellular adhesion, and immune-related processes (Figure 4a). Several of these genes overlapped with those reported by Jotanovic et al.² (eg, GALNT14, ITGAM, WNT6) linking cell-cycle-associated methylation changes to tumor regrowth, further supporting the prognostic value of the k3 subgroup (Figure 4b). Together, these findings suggest that epigenetic dysregulation may influence tumor behavior by modulating transcriptional programs governing proliferation and immune signaling, highlighting candidate biomarkers, particularly k3-DMPs associated with G1/S transition and immune checkpoint pathways, warranting functional and cohort-level validation (eg, targeted bisulfite assays, reporter or CRISPR perturbation).

To assess whether structural variants contributed to these subgroup differences, we inferred CNVs from methylation data. CNV alterations were sparse and cluster-restricted, indicating that large-scale genomic instability is unlikely to drive the observed epigenetic and clinical divergence (Supplementary Figure S3a). For instance, BMP5, which modulates pituitary hormone secretion and somatostatin receptor signaling³³ was enriched in k2; DPH6, linked to translation and stress responses,³⁴ emerged as amplified in k3. This pattern aligns with Jotanović et al. who reported limited CNV burden in benign PitNETs,¹² suggesting primarily epigenetically driven subgroup differences. Although orthogonal validation (eg, FISH, SNP-array, low-pass WGS) remains warranted, these findings reinforce the integrative value of methylation and CNV analyses for pinpointing potential modulators of tumor behavior.

Interestingly, one SF1 tumor clustered with the TPIT/PIT1-enriched group, likely reflecting mixed-lineage differentiation rather than misclassification (Figure 1b). Recent reports of PIT1/SF1 co-expressing PitNETs recognized in the WHO 2022 classification,^33,34 support this interpretation, underscoring the biological sensitivity of methylation-based clustering in capturing hybrid or transitional phenotypes beyond histopathological resolution.

Building on these findings, we developed and validated a prognostic machine learning classifier to assign NFPitNET methylation subgroups, achieving a classification accuracy of 97.5% to assign cluster membership across 3 independent external cohorts.^13–15 A shared limitation between these external studies is the restricted availability of long-term outcome data, which precludes definitive validation of the prognostic value of our methylation clusters. Notably, Belakhoua et al.⁵ reported recurrences in 10.2% of cases (12/118), primarily involving SF1-lineage (n = 5), silent TPIT (n = 6), and one unknown NFPitNETs. When mapped with our classifier, these cases were distributed across k2 (n = 3), k3 (n = 2), and k4 (n = 7) (Figure 3c), demonstrating a preliminary enrichment of recurrence within the TPIT-enriched k4 subgroup. While these findings align with other’s and our observation of heightened aggressive behavior in TPIT-lineage tumors,^31,32 the limited number of recurrent cases and absence of follow-up highlight the necessity for larger, prospectively followed cohorts to confirm cluster-specific prognostic associations. Furthermore, our model maintained 90% accuracy between paired FFPE and frozen samples, supporting technical robustness across tissue preservation methods, as shown previously.¹⁶

Our work complements prior prognostic frameworks grounded in histopathological and risk factor-based classification.^6,10,35 The Trouillas system, which combines proliferative and invasive markers, has shown predictive validity across PitNET types.³⁵ However, within the SF1-lineage, markers such as Ki-67 and p53 show limited and inconsistent prognostic value.³⁶ In our cohort, Ki-67 and p53 staining data were limited. To overcome this limitation, we applied our methylation classifier to NFPitNETs of the Belakhoua et al.cohort,¹⁵ which included more comprehensive Ki-67 data. Consistent with their own findings, Ki-67 indices in the Belakhoua cohort did not correlate with recurrence or our methylation subgroup assignment (Figure 3c).³⁷ Similarly, Jotanovic et al.² reported that IHC markers have limited discriminative capacity among benign PitNETs. Together, these observations highlight the need for molecular tools that capture biological heterogeneity invisible to routine histopathology.

The recently proposed PANOMEN-3 model in Ho et al.represents a step forward toward a standardized, evidence-based framework for pituitary adenoma prognosis.¹⁰ Although this 9-tiered system has shown successful recurrence risk stratification in their retrospective study, we observed no significant differences in PANOMEN-3 grade distributions across methylation subgroups (P = .72). This suggests that while clinical models capture disease extent, epigenetic profiling adds substantial and complementary biological resolution, especially valuable in histologically uniform subtypes such as SF1-NFPitNETs. Together, these approaches could form a multilayered risk assessment model combining clinical, pathological, and molecular data for precision prognostication.

Machine learning-based frameworks have also advanced PitNET risk prediction. Hussein et al.underscored the capacity of machine learning models to identify the most informative features for predicting clinical outcomes in a large cohort NFPitNET with a long follow-up (2-15 years).³⁸ The authors reported machine-learning models (SVM and decision tree) trained on clinical, surgical, and imaging variables achieved a good accuracy for predicting remission, stability, and regrowth (AUC-ROC ≈ 0.78-0.91); however, predictors performed no better than random chance for recurrence prediction following complete tumor resection (AUC-ROC ≈ 0.5 across all time points). These findings underscore the limitations of models relying exclusively on perioperative or radiologic parameters to capture the biological determinants of recurrence. Recognizing these constraints, the authors emphasized the need to incorporate other data (eg, operative video features) to enhance model performance. In this context, methylation profiling offers an added layer of granularity through capturing intrinsic tumor biology and epigenetic diversity. Our classifier could complement clinical models by integrating molecular and clinical dimensions to refine PitNET prognostication.

Finally, although our analysis was conducted using the Illumina MethylationEPIC v1 array, now discontinued, most probes are retained in the v2 version. Future revalidation on EPIC v2 data using probe conversion and cross-platform normalization will confirm compatibility and ensure sustained predictive performance.

In conclusion, considering their heterogeneity and that more than half of patients will not experience regrowth, a uniform follow-up strategy may not be appropriate for NFPitNETs, especially considering some benefit from less intensive surveillance. DNA methylation profiling, particularly within SF1-positive tumors, captures molecular subtypes with distinct biological and clinical behavior, supporting more tailored follow-up approaches. While methylation-based classifiers are already established for diagnosis, grading, and recurrence risk assessment in CNS tumors, including PitNETs,^30,39,40 the prognostic application to PitNETs remains in early stages and requires prospective multicenter validation. Integrating such molecular tools into clinical practice (eg, PANOMEN-3 classification) could advance NFPitNET management toward a precision medicine framework.

Supplementary Material

Supplementary material is available online at Neuro-Oncology (https://academic.oup.com/neuro-oncology).

Author Contributions

M.W.M. contributed to study conceptualization, data curation, formal analysis, investigation, methodology, project administration, resources, software, validation, visualization, and writing of the original draft, as well as review and editing of the manuscript. G.A.H. contributed to study conceptualization, data curation, formal analysis, methodology, resources, software, validation, visualization, and writing of the original draft, as well as review and editing of the manuscript. M.S.A. contributed to study conceptualization, investigation, methodology, resources, supervision, and review and editing of the manuscript. B.H. contributed to investigation, resources, supervision, validation, and review and editing of the manuscript. C.B.P. contributed to investigation, resources, and review and editing of the manuscript. H.B.B. contributed to data curation, investigation, resources, and review and editing of the manuscript. J.K.P. contributed to investigation, resources, supervision, validation, and review and editing of the manuscript. C.P contributed to data curation, formal analysis, methodology, validation, visualization, as well as review and editing of the manuscript. A.V.C. contributed to study conceptualization, data curation, formal analysis, investigation, methodology, resources, software, supervision, validation, and writing of the original draft, as well as review and editing of the manuscript. F.R.P. contributed to study conceptualization, funding acquisition, investigation, methodology, project administration, resources, supervision, and writing of the original draft, as well as review and editing of the manuscript.

Conflict of Interest Statement We declare no competing interests.

Funding

This study was funded by Novo Nordisk Foundation, Brødrene Hartmanns Fond, Tornøes og Høyrups Fond, Beckett-Fonden, Aase og Ejnar Danielsens Fond, and Odense University Hospital.

Data Availability

M.W.M. and G.A.H. accessed and verified the underlying data. All authors had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Data Sharing

Preprocessed and de-identified DNA methylation data will be deposited in Mendeley Data, alongside the methylation-based classifier. All relevant codes will be released on GitHub.