AI-powered histopathology classifiers identify transcriptomic biomarkers associated with cell cycle regulation and increased mitotic activity in poorly differentiated neuroendocrine carcinoma

Abstract

Background

Neuroendocrine neoplasms (NENs) are a biologically heterogeneous group of epithelial malignancies arising from neuroendocrine cells across multiple organs. Although considered rare in comparison to other solid tumors, their incidence has increased more than six-fold in recent decades. A major clinical challenge is the accurate stratification between aggressive, poorly differentiated (PD) neuroendocrine carcinomas (NECs) and slow growing well-differentiated (WD) neuroendocrine tumors (NETs), which differ significantly in prognostic outcomes and treatment response. Conventional histopathology is often insufficient for reliable classification or determination of tumor origin, motivating the need for molecularly informed diagnostic approaches.

Methods

We performed transcriptomic profiling of 36 neuroendocrine neoplasms (NENs) including 21 well-differentiated neuroendocrine tumors (WD-NETs) and 15 poorly differentiated neuroendocrine carcinomas (PD-NECs) from 11 distinct anatomical sites. We developed an interpretable AI-based histology classifier achieving an area under the receiver operating characteristic curve (AUC-ROC) of 0.87–0.93.

Results

PD-NECs exhibited upregulation of genes involved in cell-cycle regulation and DNA damage response, including SFN, CHEK1, E2F1, CDC6, and TTK. In contrast, neuroendocrine lineage–associated genes such as CAMK2B and PAK3, which are preferentially expressed in neural tissues, were enriched in WD-NETs. AI-derived transcriptomic biomarkers distinguishing WD-NET and PD-NEC (PAK3 and EZH2, respectively) were orthogonally validated by immunohistochemistry. Our findings were further supported by independent validation in an publicly available NEN DNA methylation dataset. Moreover, DNA methylation–based mitotic age analysis demonstrated significantly higher mitotic activity and proliferative potential in lung NECs compared with lung NETs, consistent with our transcriptomic findings. Our analysis is publicly available at https://nen-pd-wd-hist.epigenomeocean.org/.

Conclusions

This study establishes an integrative multi-omics framework for molecular classification of NENs that complements traditional histopathological assessment. By combining transcriptomic profiling, interpretable machine learning, and epigenetic validation, our approach enables improved tumor stratification, identification of mixed or misclassified cases and supports the development of biomarker-driven diagnostic and therapeutic strategies.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s40246-026-01032-8.

Background

Neuroendocrine neoplasms (NENs) arise from specialized epithelial cells that secrete amine or peptide hormones in response to neural and endocrine stimuli. These cells are most abundant in the gastrointestinal (GI) tract, pancreas, and lungs but occur across many organ systems [1–6]. Although NENs remain less common than many solid tumors, their incidence has risen substantially over recent decades. In the United States, rates increased from 1.09 per 100,000 in 1973 to 6.98 per 100,000 in 2012 (6.4-fold) [7], with similar increases in the United Kingdom and Taiwan [8].

Accurate histological classification is essential for prognosis and therapy. However this remains challenging due to key feature overlap across NEN subtypes [9]. Differentiation status, as assessed by morphology and proliferation markers, is the strongest prognostic determinant [10], however tumors with similar histology can follow divergent clinical courses. For instance poorly differentiated neuroendocrine carcinomas (PD-NECs) are highly aggressive with rapid progression and limited treatment durability [11] with pancreatic PD-NECs having a median overall survival of ~11–12 months [12, 13]. In contrast, well-differentiated neuroendocrine tumors (WD-NETs) are typically indolent, with median survival exceeding 5–10 years depending on primary site and grade. For example, pancreatic WD-NETs demonstrate a median survival of 8–12 years, while small-intestinal NETs frequently exceed 65–75% 5-year survival [8, 10, 14].

The World Health Organization (WHO) framework relies heavily on histomorphology and proliferation thresholds (mitotic count >20 per 2mm² and Ki-67 >20%) to classify PD-NECs as high-grade tumors [15, 16]. While clinically essential, this classification schema has significant limitations arising from histopathological overlap between PD-NEC and WD-NET. Notably, WD-NETs span a broad grade spectrum (G1–G3), and high-grade WD-NETs with morphologic features overlapping those of PD-NEC have been reported, underscoring the diagnostic challenges inherent to grade-based classification alone [17, 18]. Furthermore classical immunohistochemical markers such as chromogranin A and synaptophysin, and aberrant p53/RB1 patterns, are useful but not definitive and frequently overlap in high-grade tumors [15, 19]. A recent multi-institutional study reported poor interobserver agreement in high-grade NEN classification, with experts failing to reach consensus in nearly two-thirds of cases [15]. This highlights the limitations of morphology-only approaches and motivates the need for improved quantitative molecular classifiers.

Over the past several decades, advancements in next-generation sequencing (NGS) technologies have substantially advanced our understanding of the molecular landscape of NENs and the complex transcriptional and epigenetic overlap between histologic groups [11, 20]. For instance, although typically slow growing, WD-NETs may in some cases exhibit high proliferation and TP53 alterations typical of PD-NECs, limiting diagnostic specificity of these mutational markers [21]. Additionally, PD-NEC drivers remain incompletely characterized due to rarity and substantial intra- and inter-tumoral heterogeneity [22–24]. NENs also have low tumor mutational burden (TMB) compared with many solid tumors [6, 22], constraining genomics-only approaches.

Several molecular studies have also sought to further refine NEN classification beyond morphology. Protein expression analyses have shown loss of DAXX/ATRX in WD-NETs and aberrant p53/RB1 in PD-NECs [19]. Moreover, epigenetic profiling has begun to show support for distinct DNA methylation landscapes within NEN [25]. However many studies of these studies are limited to pancreatic NENs or emphasize tissue-of-origin rather than histologic differentiation. Large methylation studies have also shown the ability to infer primary tissue sites across organs, although these do not yet resolve the molecular distinctions of PD-NECs versus WD-NETs [26–28]. Transcriptomic profiling offers a complementary framework by capturing active cellular programs (lineage identity, proliferation, stress responses). Whether these differentiation states represent conserved, tissue-agnostic molecular features in NENs remains unclear.

To address these limitations, we developed an interpretable transcriptome-based machine learning framework to distinguish PD-NECs from WD-NETs across diverse anatomical sites. Using targeted transcriptomic profiling, we trained supervised classifiers (random forest and logistic regression) that explicitly model histopathology differentiation while enabling direct interrogation of feature importance. This approach identified activation of cell-cycle, replication stress, and DNA damage response programs in PD-NECs versus preservation of neuroendocrine lineage and neuronal signaling programs in WD-NETs. We validated these signatures using orthogonal immunohistochemical markers and independent DNA methylation datasets and demonstrating increased mitotic activity in PD-NECs. Together, this framework provides a quantitative, biologically interpretable basis for resolving histologic ambiguity and highlights transcriptomic state as a robust indicator of differentiation-related aggressiveness.

Methods

Study design and oversight

This was a single-center, retrospective study conducted at Fox Chase Cancer Center (FCCC) to develop and validate molecular classifiers of neuroendocrine neoplasm (NEN) histology. Formalin-fixed, paraffin-embedded (FFPE) tumor specimens were profiled by targeted transcriptomics; selected markers were validated by immunohistochemistry (IHC) on tissue microarrays (TMAs) from the same cohort. The study was exploratory in nature; sample size was determined by tissue availability for the rare PD-NEC subtype rather than a priori power calculations. As a fixed sample size bounds the smallest effect reliably detectable regardless of outcome, we report minimum-detectable-effect-size (MDES) and cross-validation stability analysis for each primary endpoint below (see Statistical analyses, Methods). To contextualize statistical power given the realized sample sizes, we computed Cohen's d, achieving 80% power at the relevant significance threshold for each primary endpoint (two-sample t-test, unequal-n noncentral-t formulation), alongside the observed effect sizes for significant findings. Independent validation was performed using external public DNA methylation datasets (GSE211483) and gene expression data (GSE211485). All procedures were approved by the FCCC Institutional Review Board, and written informed consent was obtained from all participants.

Statistical analyses

Given the realized group sizes (WD-NET n = 21, PD-NEC n = 15; 730 endogenous genes), the study had 80% power to detect Cohen's d ≥ 1.85 at the Bonferroni-corrected threshold (α = 0.05/730) and d ≥ 1.12 at the effective FDR<0.05 threshold. The 203 genes meeting our significance criteria had a median observed |d| of 1.53 (IQR 1.08–2.03; max 4.06), indicating the design is sufficiently powered for the large effects that dominate the reported hits, with reduced sensitivity to small/moderate effects. All statistical analyses were performed in R (version ≥ 4.1) and Python (version ≥ 3.10). Unless otherwise noted, two-sided tests were used with FDR control by Benjamini–Hochberg. Continuous variables are summarized as median (IQR) unless specified. Group comparisons for IHC H-scores used Wilcoxon rank-sum tests with FDR adjustment; correlations used Pearson’s r. Plots were generated using ggplot2, ComplexHeatmap, matplotlib, and seaborn. Batch effect was assessed using PERMANOVA (Euclidean distance, scaled log-CPM, 9,999 permutations). The same test was used to assess histology alone and, separately, primary site, age, and sex as co-factors using marginal (type-III) tests so each term is assessed after adjusting for the others. Reproducibility was supported by scripted analyses, version-controlled code, and fixed random seeds where applicable.

Cohort and tissue microarrays

Tumor specimens were obtained from patients with well-differentiated (WD) neuroendocrine tumors (NETs) and PD neuroendocrine carcinomas (NECs) from multiple primary sites (Table S1). Clinicopathologic data were de-identified prior to analysis. TMAs were constructed from 36 FFPE blocks (21 WD, 15 PD). Hematoxylin-and-eosin–stained sections were reviewed by a board-certified pathologist to mark representative tumor regions. Duplicate cores (diameter 1 mm) were taken per case using a precision arrayer (Beecher Instruments, Silver Spring, MD, USA) and placed on separate recipient blocks to mitigate positional bias. Tissue had been fixed in 10% phosphate-buffered formaldehyde (24–48 h), processed, and embedded per standard protocols. Each TMA core was scored individually; duplicate core values were averaged per case.

Transcriptomic profiling

FFPE sections were reviewed by a trained pathologist to enrich for invasive tumor areas prior to RNA extraction. Total RNA was extracted using the High Pure FFPET RNA Isolation Kit (Roche), following the manufacturer's protocol. RNA was hybridized overnight to probes of the NanoString nCounter PanCancer Immune Profiling Panel (784 probes targeting 730 endogenous genes (MAPK, STAT, PI3K, RAS, etc.), 40 housekeeping, 6 positive and 8 negative control probes), strategically capturing 13 canonical cancer pathways, including Cell Cycle, Apoptosis, Hedgehog, Wnt, DNA Damage Control, Transcriptional Regulation, Chromatin Modification, and TGF-β) [29], then purified and digitally counted on the nCounter platform (NanoString Technologies). The NanoString nCounter Analysis System is specifically designed to handle clinical-grade samples, including degraded FFPE-derived RNA. Its enzyme-free, direct hybridization workflow eliminates amplification bias and inherently minimizes RNA quality degradation and batch-related variability, supporting the reliability of the gene expression data generated from these samples. Manufacturer-supplied positive, negative, and housekeeping controls were used to assess RNA quality and run performance. While RNA extraction was performed using a high-quality FFPE-specific kit and all samples met NanoString nCounter platform input-quality thresholds—providing indirect reassurance against gross RNA-quality confounding—we acknowledge that the absence of directly reported DV200 values and block-age data represents a limitation. Raw reporter code count (RCC) files were used for all downstream analyses (Supplementary Data 1).

Pre-processing and differential expression analysis

Raw counts were normalized and transformed in R. Background correction and positive/negative control normalization were applied as per NanoString recommendations. Library size normalization and variance modeling were performed with the voom (variance modeling at the observational level) function from the limma R/Bioconductor package (Supplementary Data 2), and log-counts per million (logCPM) values were computed (edgeR [30]). Although experiments were conducted across three batches, batch effects were systematically assessed using PCA and PERMANOVA. No significant batch effect was detected (Figure S1). Differential expression was assessed with linear modeling and empirical Bayes moderation (Supplementary Data 3). Multiple testing corrections used the Benjamini–Hochberg false discovery rate (FDR); unless stated otherwise, significance thresholds were FDR < 0.05 and log2FC ≥ 1. Principal component analysis (PCA) was performed on scaled logCPM matrices.

Immunohistochemistry and quantitative image analysis

IHC was performed on 5 µm TMA sections. Slides were deparaffinized, rehydrated, and subjected to heat-induced epitope retrieval in 0.01 M citrate buffer (pH 6.0) for EZH2, H3K27me3, and PAK3. Endogenous peroxidase activity was quenched in 3% H₂O₂. Sections were incubated overnight at 4 °C with primary antibodies: EZH2 (D2C9, rabbit, 1:50; Cell Signaling Technology, #5246), Tri-methyl-Histone H3 (Lys27) (C36B11, rabbit, 1:50; Cell Signaling, #9733), and PAK3 (A-20, rabbit, 1:500; Invitrogen, PA5-15,118). Detection used the EnVision+ polymer system (Dako/Agilent) with 3,3′-diaminobenzidine (DAB) chromogen. Slides were counterstained with hematoxylin, dehydrated, cleared in xylene, and mounted. Negative controls used matched isotype control IgG.

Whole-slide images were acquired on a PerkinElmer CRi Vectra 2 platform. Multispectral images were unmixed and analyzed in inForm (Akoya Biosciences) using a trainable workflow: (i) tissue segmentation to identify neuroendocrine tumor regions and (ii) cell segmentation to delineate hematoxylin-positive nuclei and DAB-positive subcellular compartments. Staining intensity was binned (0/1+/2+/3 +), and H-scores were computed as 1x(%1 +) + 2x(%2 +) + 3x(%3 +) (range 0–300). Nuclear and cytoplasmic H-scores were derived as appropriate. Antibody specificity for EZH2 was confirmed by loss of signal following EZH2 depletion in retinal pigment epithelial (RPE) cells.

Machine-learning classification of histology

Histologic class (PD vs WD) was modeled using logistic regression (LR) and a random forest classifier (RFC) implemented in scikit-learn (Python). Features comprised normalized transcript levels for the 730 endogenous genes (housekeeping and positive/negative control probes were excluded from the feature set). Data were stratified to preserve class balance. Hyperparameters were tuned via RandomizedSearchCV with internal cross-validation, reporting accuracy, balanced accuracy, precision/recall, F1, and ROC-AUC. To assess variability of performance at this sample size, stratified 5-fold cross-validation was repeated 100 times with independent fold partitions. Out-of-fold performance was stable across repeats (LR: AUC 0.995 ± 0.009, accuracy 0.946 ± 0.028; RF: AUC 0.992 ± 0.011, accuracy 0.937 ± 0.026; mean ± SD), and both classifiers discriminated significantly above a label-permuted null (null AUC ≈ 0.50; empirical P = 0.001, 1,000 permutations).

Feature importance and model interpretability

Global and local feature attributions were derived from: (i) LR coefficients (L1/L2/elastic-net regularization), (ii) RFC Gini importance, and (iii) SHAP (Shapley Additive exPlanations) values computed on the RFC using mean absolute class-specific SHAP. Robust markers were identified by intersecting features meeting pre-specified thresholds (β ≥ 0.04 for LR; Gini ≥ 0.001; mean SHAP (PD class) ≥ 0.001). Gene-level summaries and per-sample SHAP plots are provided in Supplementary materials.

Pathway enrichment analysis

Pathway analysis was performed in Cytoscape (v3.10.1) using gene sets from LR (positive/negative coefficients under L1/L2/elastic-net) and from RFC SHAP/Gini (PD- and WD-enriched features analyzed separately). Enrichment used the WikiPathways and Reactome pathway databases, with network summarization via EnrichmentMap using similarity metrics to cluster related pathways. Reported statistics include similarity scores and FDR-adjusted p-values. Cell-cycle modules (WP179, G1/S, G2/M, ATR) and PI3K–Akt/RB pathways were prioritized based on effect size and recurrence across models.

Primary-site inference for tumors of unknown origin

For five tumors lacking annotated primary sites, site-of-origin inference was performed using k-nearest neighbors (k = 5) in PCA space. Each unannotated sample was assigned the primary site of its five nearest labeled neighbors, and inferred assignments were visualized on the PCA plot. Pairwise Pearson correlations across all samples were summarized in a heatmap, and quantile–quantile (QQ) plots were generated to visualize transcriptomic concordance for selected high- and low-similarity pairs.

Independent validation using external DNA methylation and gene expression data

Independent validation was performed on public lung NEN methylation data (Illumina Infinium MethylationEPIC; GSE211483). Unsupervised analyses (PCA and hierarchical clustering) were performed on filtered β-matrices without variance preselection for the primary figures; sensitivity analyses using variable CpGs were also evaluated. Polycomb target–based stemness (pcgtAge) scores were computed to infer mitotic activity. Where matched RNA data (GSE211485) were available, methylation–expression relationships were assessed by Wilcoxon tests with Benjamini–Hochberg adjustment.

Software and versions

Key software included R (≥ 4.1) with Bioconductor (minfi, ChAMP, limma, edgeR), Python (≥ 3.10) with scikit-learn (≥ 1.3) and shap (≥ 0.43), and Cytoscape (v3.10.1). Exact package versions and command-line invocations are documented in the repository.

Results

Principal component analysis reveals a conserved differentiation axis separating PD-NEC and WD-NET

We profiled 36 neuroendocrine neoplasms (NENs) from the Fox Chase Cancer Center (FCCC) spanning 11 primary anatomical sites (esophagus, lung, breast, stomach, pancreas, small bowel, cecum, colon, rectum, ileum) and an additional five tumors of unknown origin (Fig. 1A; Table S1). Independent pathological review classified 21 tumors (58%) as well-differentiated neuroendocrine tumors (WD-NETs) and 15 (42%) as poorly differentiated neuroendocrine carcinomas (PD-NECs). Unsupervised PCA revealed clear separation of PD-NECs and WD-NETs along principal component 1 (Fig. 1B).

Fig. 1.

Transcriptomic clustering distinguishes well-differentiated NETs from poorly differentiated NECs in Clinical Cohort. A Sources of tumors from FCCC Cohort. Illustration shows the anatomical distribution of the 36 formalin-fixed paraffin-embedded (FFPE) tumor samples profiled using the NanoString PanCancer Pathways panel, including 21 well-differentiated (WD) neuroendocrine tumors (NETs; grades 1–3; blue) and 15 poorly differentiated (PD) neuroendocrine carcinomas (NECs; grade 3; orange). B Principal component analysis (PCA) plot of normalized RNA expression values demonstrates linear separation between PD NECs and WD NETs along principal component 1 (PC1), with PD samples appearing more dispersed. Each point represents one tumor sample; PD NECs are shown in orange, WD NETs in blue. Two PD NEC samples (615 and 4709) cluster closely to WD NETs whilst two PD samples (1005 and 3067) separate from PD NEC samples. C Volcano plot showing differentially expressed genes (DEGs) between PD and WD tumors. Genes with |log2FC|≥ 1 and FDR < 0.05 are highlighted. Selected genes with higher expression in PD NECs (e.g., CHEK1, TTK, EZH2) and WD NETs (e.g., CACNB2, CAMK2B, RASA4, PAK3) are labeled. Upregulated genes are shown in red, downregulated genes in blue and no significant genes are shown in grey

Importantly, separation was observed despite substantial heterogeneity in tumor primary site, suggesting the presence of conserved, tissue-agnostic transcriptional features linked to differentiation state [23, 31–33]. Primary site is a recognized source of transcriptomic variation in several cancer types [34]. As our FCCC cohort spans 11 sites, we performed permutational multivariate analysis of variance (PERMANOVA) with primary site (and age, sex) as co-factors using marginal (type-III) permutation tests. Histology remained the dominant and only significant driver of transcriptomic variance after adjustment (R² = 0.12, P = 1×10⁻⁴), while the marginal effect of primary site was not significant (P = 0.074); the WD/PD axis dominated PC1 (t-test P ≈ 10⁻¹³) orthogonally to site (Figure S1).

Our findings indicate that transcriptomic conservation associated with histopathological differentiation is detectable even within a heterogeneous multi-site cohort. To define genes contributing to this separation, differential expression analysis was performed (log2FC > 1, FDR < 0.05). PD-NECs exhibited increased expression of genes involved in DNA replication and checkpoint control, including CDC6, CHEK1, SFN, and the epigenetic regulator EZH2 (Fig. 1C). These genes collectively represent a replication-competent, checkpoint-activated state consistent with high mitotic activity and replication stress adaptation. In contrast, WD-NETs showed higher expression of genes associated with neuronal signaling and regulated secretory function, including PAK3, CAMK2B, and RASA4, indicating a possible preservation of neuroendocrine lineage identity.

Unsupervised clustering reveals molecular substructure and an intermediate differentiation state

To identify which transcriptomic variations distinguish the histopathological subtypes NET and NEC, we performed unsupervised hierarchical clustering of normalized expression profiles. Clustering identified three groups (Fig. 2A). Group 1 (n = 20) was composed primarily of WD-NETs and was characterized by elevated expression of genes associated with neuroendocrine lineage identity and calcium-dependent signaling, including CACNB2, ZBTB16, GRIA3, and WNT4 (Figure S2A–D). These features suggest a transcriptional program consistent with regulated secretory identity and neuronal differentiation. Group 2 (n = 12) was enriched for PD-NECs and exhibited increased expression of genes involved in cell-cycle regulation, chromatin remodeling, and oncogenic signaling, including CDKN2A and EZH2 (Suppl. Data 2) and, CTNNB1, MYB, FANCA, and NOTCH1 (Figure S2E–H). The coordinated upregulation of cell-cycle and epigenetic regulators suggests a proliferative and transcriptionally reprogrammed state aligned with poorly differentiated carcinoma biology. Group 3 (n = 4) comprised tumors of mixed histology (2 WD-NETs, 2 PD-NECs) and displayed attenuated expression of both Group 1 and Group 2 marker sets, along with increased GATA3 (Figure S2I). Two cases harbored TP53 mutations without concurrent RB1 alteration. These tumors were more dispersed in PCA space (Fig. 2B), indicating greater transcriptional heterogeneity.

Fig. 2.

Unsupervised hierarchical clustering identifies a third transcriptional group. A Heatmap of normalized gene expression (voom + limma) for the top variably expressed genes across 36 NEN tumor samples profiled by NanoString (~ 800 cancer-related genes). Rows represent genes (z-score scaled), and columns represent tumors, annotated by clinical and genomic features including age, gender, race, grade, stage, Ki-67 index, primary site, and selected mutations (e.g., TP53, KRAS, PIK3CA). Hierarchical clustering reveals three major groups: Group 1 (predominantly WD NETs), Group 2 (predominantly PD NECs), and Group 3 (intermediate/mixed subtype). B Principal component analysis (PCA) of RNA expression by subtype clustering. Tumors from Group 3 (e.g., cases 3067 and 1005) occupy an intermediate transcriptional space between canonical WD and PD clusters, suggesting a molecularly distinct subgroup. Tumor groups are colored by clustering assignment: Group 1 (blue), Group 2 (orange), and Group 3 (purple)

Orthogonal validation of differentiation-linked markers by immunohistochemistry

To validate transcriptomic markers at the protein level, we performed quantitative immunohistochemistry (IHC) on a tissue microarray (TMA) constructed from the discovery cohort. Based on differential expression, we selected EZH2 as a PD-NEC associated marker and PAK3 as a WD-NET associated marker (Fig. 3A–C). Consistent with transcriptomic findings, PD-NECs demonstrated significantly higher nuclear expression of EZH2 and its associated repressive chromatin mark H3K27me3 (Figure S3). EZH2 was selected due to robust mRNA upregulation (FC = 4.2x, p.adj = 9.38 × 10^–5) and its established role in Polycomb-mediated epigenetic repression and proliferative signaling (Fig. 3B) [35]. Increased H3K27me3 further supports the presence of a chromatin-repressed, dedifferentiated transcriptional state in PD-NECs (Figure S3). In contrast, WD-NETs exhibited significantly higher PAK3 expression (FC = 7.7, adjusted p = 2.67 × 10^–8), consistent with retention of neuronal and neuroendocrine lineage programs (Fig. 3D). Nuclear H-scores (0–300) confirmed subtype-specific protein expression patterns, demonstrating concordance between RNA and protein levels. Antibody specificity for EZH2 antibody (CS5246) was validated by depletion experiments in retinal pigment epithelial (RPE) cells, which abrogated signal intensity (Figure S4).

Fig. 3.

Immunohistochemical validation of subtype-specific markers. A Representative EZH2 immunohistochemistry (IHC) staining in a PD NEC tumor shows strong nuclear expression. B Representative PAK3 IHC staining in a WD NET tumor shows increased cytoplasmic localization. C Boxplots summarizing IHC scores for EZH2 and PAK3 D EZH2 shows significantly higher expression in PD NECs (FDR-adjusted p = 9.38e−05) across FCCC NEN cohort (PD NEC, n = 15; WD NET, n = 21), while PAK3 expression is significantly elevated in WD NETs (FDR-adjusted p = 2.67e−08). P-values were calculated using the two-tailed Wilcoxon rank-sum test and corrected using the Benjamini–Hochberg method

Interpretable machine learning resolves transcriptomic biomarkers associated with NEN differentiation

Normalized NEN transcriptional expression profiles (730 endogenous genes) supported accurate separation of PD-NECs from WD-NETs using interpretable classifiers (logistic regression and random forest). Stratified cross-validation yielded a ROC-AUC of 0.87–0.93 for the logistic-regression and random-forest classifiers (Fig. 4A), consistent with a differentiation-linked signal in the transcriptome. A prior classifier trained on a restricted small-intestinal NET panel reported 71–78% accuracy [31], though differences in cohort, assay, and metric preclude direct benchmarking.

Fig. 4.

Machine learning classifiers identify discriminative transcriptomic biomarkers and subtype-specific gene programs. A Receiver operating characteristic (ROC) curves showing performance of logistic regression (LR; dashed/solid green lines) and random forest classifier (RFC; dashed/solid lines) models on both training and test datasets. Curves represent classification of NEC vs NET using transcriptomic profiles; area under the curve (AUC) is annotated for each model. B Barplot of top ten and bottom ten genes ranked by LR model coefficients. LR coefficients are shown in blue, representing higher weight toward NEC and NET classification, respectively. C Barplot showing SHAP (SHapley Additive exPlanations) feature importance values derived from the RFC model. Green bars represent the relative contribution of each gene to model predictions. D Venn diagram showing the overlap of top-ranking features (top 50 genes) across three importance ranking methods: LR coefficient, SHAP importance, and RFC Gini importance. CAMK2B is labeled as one of the genes identified by all three approaches. E Expression profile of CAMK2B across adult human tissues based on RNA-seq from the GTEx Portal. Color shading represents tissue-level expression (TPM, transcripts per million). F Barplot of select genes upregulated in WD NETs in FCCC cohort based on DE analysis (log2FC > 1, FDR < 0.05). G Barplot of select genes upregulated in PD NECs under the same criteria. H Pathway enrichment analysis of genes with positive coefficients from the LR model. The most significantly enriched pathway is WP179 Cell Cycle pathway (p = 8.49 × 10⁻³¹) based on Reactome similarity scoring. Positive LR coefficients and PD-associated SHAP features are marked by open triangles, respectively. I Schematic representation of the cell cycle pathway, with genes upregulated in the PD-NEC discovery cohort marked by red triangles. Barplot shows significantly enriched biological processes ranked by −log10(FDR-adjusted p value)

Logistic regression identified SFN, ZIC2, FANCA, HIST1H3B, CHEK1, E2F1, CDC6, LAMC2, and DKK1 as the top positively associated biomarkers for PD-NEC, whereas CAMK2B, PACRG, and FGF14, among others, were negatively associated and instead emerged as the strongest biomarkers for WD-NET (Fig. 4B, Tables S2-4). To further evaluate machine learning classifier outputs at the individual sample level, we computed SHAP (Shapley Additive exPlanations) values from the trained random forest model. SHAP values estimate the marginal contribution of each feature to a specific prediction, offering interpretable insights into the decision boundaries for PD NECs and WD NETs for non-linear RF models (Fig. 4C). Aggregated SHAP importances across PD NEC predictions, double confirmed by LR weight analysis, revealed FANCA, HIST1H3B, CDC6, and TTK. FANCA and CDC6 are essential for replication fork stability, while TTK encodes a dual-specificity kinase required for centrosome duplication and chromosome alignment [36]. HIST1H3B, a histone variant, and SFN, a 14-3-3 checkpoint protein, also showed elevated importance, reinforcing their role in PD NEC pathogenesis. Notably, several SHAP-prioritized genes (FANCA, HIST1H3B, CDC6, SFN) overlapped with top positive-weight features from logistic regression, underscoring their consistency across models (Fig. 4B, C).

WD-NET–associated signals were captured by features with negative SHAP values and negative LR coefficients, highlighting lineage-preserved programs. The strongest WD-NET features included CAMK2B (–0.0058) and PAK3 (–0.0027), along with neuroendocrine/neuronal signaling genes such as GRIA3, consistent with the synaptic and calcium-channel programs enriched in WD-NETs. This pattern indicates that the classifiers distinguish not only proliferative PD-NEC programs but also differentiation-linked WD-NET biology.

To identify robust biomarkers across modeling strategies, we intersected feature rankings from three independent metrics (1) Logistic regression coefficients ≥|0.04|, (2) Random forest Gini importance ≥ 0.001 and (3) SHAP importance for PD NEC class ≥ 0.001 (Tables S5-6). Amongst 28 genes, eight genes met all three criteria: CAMK2B, BRIP1, PAK3, CACNB2, CDC6, HIST1H3B, DKK1, and FANCA (Fig. 4D). These genes span DNA repair (e.g., BRIP1, FANCA), chromatin regulation (HIST1H3B), cell cycle checkpoints (CDC6), and neuroendocrine signaling (CAMK2B, PAK3). CAMK2B, a prominent marker of WD-NET, produces serine/threonine kinases in in calcium signaling, essential for neuronal development, synaptic plasticity, and memory formation, and is predominantly expressed in brain-related tissues, suggesting that WD-NET preserves these functional elements (Fig. 4E). CAMK2B exhibits elevated expression in WD-NET, whereas CHEK1 is preferentially expressed in PD-NEC (Fig. 4F, G).

Pathway enrichment analysis confirms a replication–lineage polarity in PD-NEC

Pathway analysis of PD-weighted genes revealed significant enrichment of cell-cycle programs (WP179; FDR=10^–31), G1/S and G2/M transitions, ATR checkpoint signaling, Fanconi anemia, PI3K–Akt–mTOR signaling, and RB pathway dysregulation (Fig. 4H; Tables S7-10). The core module centers on CHEK1, SFN, CDC6, E2F1, TTK, and MCM family members, forming a tightly connected replication stress and mitotic checkpoint axis (Fig. 4I). In contrast, WD-NET–associated genes mapped to neuronal development and synaptic signaling programs such as calcium channel subunits (CACNB2, CACNA1D) and glutamatergic signaling genes (GRIA3) (Tables S11-14). Together, these results define a replication-driven PD-NEC state versus a neuronal lineage-preserved WD-NET state.

Transcriptome-guided inference of tumor origin

NENs frequently lack canonical driver mutations, limiting the utility of genomic data including in determining tumor site of origin [33]. For instance, whole exome sequencing of 48 small intestinal NENs revealed a median of just 0.1 somatic SNVs per megabase (range: 0–0.59) [19]. Similarly, pancreatic NENs exhibit lower TMB (0.82/Mb) compared to their exocrine counterparts, with mutations in MEN1, DAXX, and ATRX paradoxically associated with favorable outcomes [37]. When examining mutation patterns we find similar trends, with TP53 mutation occurring at a higher frequency in PD-NECs compared to WD-NETs, consistent with recent series reporting TP53 alteration in ~79% of NECs versus retained wild-type TP53 in the large majority of grade−3 NETs [21]. However, our data demonstrates that the expression differences between WD-NET and PD-NEC are modest in those genes, so unlikely to serve as reliable histopathological markers (Figure S5A-F) and tissue-of-origin markers. Inherently, transcriptional profiling should offer a more robust strategy for inferring tumor origin and histopathological classification than genomic mutations.

Five samples in our cohort had no documented known primary site (Fig. 5A): four were poorly differentiated (PD_0001, PD_0009, PD_0014, PD_0015) and one well-differentiated (WD_0021). To explore whether transcriptomic proximity could nominate a likely origin, we correlated each unknown-primary tumor against the expression profiles of tumors with documented sites (Fig. 5B). PD_0009 most closely resembled a small-bowel NEC (r = 0.92), while PD_0014 and PD_0015 most closely resembled a pancreatic NEC (r = 0.88 and 0.86, respectively). For the remaining two tumors, inference was inconclusive (Tables S15). WD_0021 matched pancreatic and pulmonary references almost equally (r = 0.87 and 0.86) and is therefore reported as pancreas-or-lung rather than a single origin. PD_0001 showed only weak, non-specific similarity to any reference (top correlation r = 0.70, to an esophageal NEC, with closely spaced and biologically inconsistent runner-up candidates); we therefore do not assign it a confident primary site and flag it as falling below the threshold at which nearest-neighbor inference is reliable. Pairwise transcriptomic correlations were highest among WD-NETs (r = 0.97, between WD_0006 and WD_0017), intermediate among PD-NECs (r = 0.92), and lowest between subtypes between PD-NEC (PD_001) and WD-NET (WD_001) (r = 0.48) consistent with the expected transcriptional divergence across differentiation grades (Fig. 5C–E). While exploratory, these findings suggest transcriptomic proximity may assist in origin inference when appropriate reference cohorts exist.

Fig. 5.

Transcriptomic similarity and primary site inference based on gene expression. A Principal component analysis (PCA) of RNA expression across NEN tumor samples, colored by inferred primary tumor site. Sites were assigned using a K-nearest neighbors (KNN) classifier trained on samples with known origins. Shape denotes histological subtype: squares indicate poorly differentiated (PD) NECs, circles indicate well-differentiated (WD) NETs. Fill denotes data partition: closed black shapes represent test set tumors (n = 6), open shapes represent training set tumors (n = 30). B Heatmap of pairwise Pearson correlation coefficients (r) for all tumor samples; correlation strength is visualized on a continuous color scale. C Quantile–quantile (QQ) plot for the tumor pair with the highest correlation (WD_0006 vs WD_0017). D QQ plot for the tumor pair with the highest correlation (PD_0008 vs PD_0009). E QQ plot for the tumor pair with the lowest transcriptomic correlation (PD_0001 vs WD_0001)

Independent validation via DNA methylation profiling from external dataset

To determine whether the transcriptional differences between poorly differentiated neuroendocrine carcinomas (PD-NECs) and well-differentiated neuroendocrine tumors (WD-NETs) were reflected at the epigenetic level, we analyzed an independent lung NEN cohort profiled using the EPIC DNA methylation array (GSE211483). Principal component analysis of genome-wide methylation β-values demonstrated separation of PD-NECs from WD-NETs, with PC1 explaining 19.2% of the variance (Fig. 6A). Unsupervised clustering of representative differentially methylated CpG sites distinguished normal lung, WD-NET, and PD-NEC samples, supporting the existence of distinct histopathologic subtype-associated DNA methylation profiles (Fig. 6B).

Fig. 6.

DNA methylation–based biomarkers distinguish NETs from NECs and reflect transcriptional and epigenetic subtype-specific programs. A Principal component analysis of genome-wide methylation β-values from Infinium MethylationEPIC arrays (GSE211483; n = 33 lung samples) demonstrates separation among normal lung, well-differentiated neuroendocrine tumors (NETs), and poorly differentiated neuroendocrine carcinomas (NECs) along PC1 and PC2. B Heatmap of the top 20 differentially methylated CpG sites across normal lung, NET, and NEC samples. Colors represent methylation β-values, with blue indicating lower methylation and red indicating higher methylation. C–F Violin plots showing methylation levels of candidate CpG biomarkers across normal lung, NET, and NEC groups. C SFN (cg07786675; TSS200/first exon) and D FANCA (cg08211068; TSS1500/5′ UTR) exhibit NEC-associated hypomethylation. E CAMK2B (cg05610097; gene body/enhancer) exhibits NET-associated hypomethylation. F FGF14 (cg22579071; TSS200/promoter) shows lower methylation in NETs than in normal lung and NECs. G pcgtAge scores derived from Polycomb-group target CpGs increase from normal lung and NETs to NECs, indicating greater mitotic activity in NECs. H MKI67 expression is significantly elevated in NECs relative to NETs. I–L External RNA-sequencing analysis (GSE211485) demonstrates significantly higher expression of I EZH2 and J PRC1 in NECs and significantly higher expression of K PAK3 and L GRIA3 in NETs. Methylation β-values were obtained from GSE211483, and expression values are presented as log2-transformed RNA-sequencing measurements from GSE211485. Statistical significance is indicated by adjusted p-values or significance symbols, as shown

We examined CpG sites associated with key transcriptomic features distinguishing PD-NECs from WD-NETs. Significant subtype-specific methylation differences were observed at CpG sites annotated to SFN, FANCA, CAMK2B, and FGF14 (Fig. 6C–F). SFN and FANCA exhibited substantially lower methylation in PD-NECs, consistent with their increased transcriptional activity in the poorly differentiated subtype (Fig. 6C, D). In contrast, CAMK2B and FGF14 displayed distinct WD-NET-associated methylation patterns (Fig. 6E, F). Together, these findings provide independent epigenetic support for genes identified through the transcriptomic analyses.

PD-NECs also exhibited significantly higher pcgtAge scores than WD-NETs and normal lung samples, indicating increased mitotic activity in the poorly differentiated tumors (Fig. 6G). Consistent with this finding, MKI67 expression was significantly elevated in PD-NECs relative to WD-NETs (Fig. 6H), further supporting the highly proliferative phenotype of PD-NECs. Finally, analysis of an independent external gene-expression dataset (GSE211485) showed significantly higher expression of EZH2 and PRC1 in lung NECs, whereas PAK3 and GRIA3 were expressed at higher levels in lung NETs (Fig. 6I–L). These independent methylation and gene-expression findings were concordant with our primary transcriptomic analyses and further support the biological distinction between PD-NECs and WD-NETs.

Discussion

We demonstrate that transcriptomic profiling enables accurate classification of neuroendocrine neoplasm (NEN) subtypes using interpretable machine learning approaches. PD-NECs are characterized by upregulation of cell-cycle and replication stress regulators (SFN, CHEK1, E2F1, CDC6, TTK), consistent with high proliferation. WD-NETs showed enrichment of neuroendocrine lineage and synaptic genes (CAMK2B, GRIA3, PAK3, FGF14), aligning with a more differentiated phenotype, though functional validation is needed to link these genes to NET biology. Many PD-NEC-associated features regulate DNA replication, checkpoint control, and chromatin state, suggesting a shared proliferative program, while WD-NET features reflect lineage and neuronal signaling programs. EZH2 and H3K27me3 were higher in PD-NECs, while PAK3 was higher in WD-NETs. External methylation data from lung NENs supported the epigenetic signature of dedifferentiation in NECs (promoter hypermethylation, higher mitotic activity, and methylation–expression shifts at MKI67 and SFN).

These results are concordant with observations that high-grade NECs share programs with small-cell lung cancer [24]. This is consistent with reports that, when primary and metastatic GEP-NEN samples are compared directly, tumor grade and differentiation, rather than anatomical origin, are the principal determinants of the transcriptome [21, 25]. Extrapulmonary NECs (EPNECs) have been shown to stratify by neuroendocrine lineage-specific transcription factors and NECs across tissue sites resolve into molecularly distinct subtypes defined by a small set of master regulators (ASCL1, NEUROD1, HNF4A, POU2F3, and YAP1) [38], an architecture analogous to that described in pulmonary large-cell neuroendocrine carcinoma (P-LCNEC). The majority of these master regulators are absent from the NanoString nCounter panel by design, limiting further investigation.

The models were trained on a targeted NanoString panel rather than whole-transcriptome profiling. This targeted panel interrogates tumor-cell-intrinsic signaling pathways and, by design, contains only a few immune-lineage markers (12/152 immune/ stromal signature genes present, 7.9%; Figure S6), and therefore we could not formally deconvolve immune or stromal fractions (as shown in MCP-counter, quanTIseq and EPIC; Figure S7); a minor component of the observed differences could therefore reflect microenvironmental composition, and gene-level interpretations should be read with this caveat.

Still, several of these convergent markers carry potential translational value, with several PanNEN studies supporting further investigation the epigenetic regulator EZH2 inhibitors like CPI-1205 and next-generation dual EZH1/2 inhibitors such as Tulmimetostat, which has entered early-stage development in other malignancies [39]. Similarly, CHEK1 and TTK enrichment nominates testable replication-stress and mitotic-checkpoint dependencies within the highly proliferative PD-NEC compartment. CHK1 inhibitors such as SRA737 [40] and TTK inhibitors such as CFI-402257 [41] have reached early-phase solid-tumor trials but remain investigational, with limited single-agent efficacy, treatment-related toxicity, and uncertain predictive biomarkers constraining clinical translation. In contrast, concordant RNA and protein enrichment currently supports PAK3 primarily as a candidate WD-NET lineage and classification biomarker, rather than as a therapeutic target, because no clinically validated PAK3-selective therapy is currently available.

Rare cases of tumor progression from a low-grade NET to a high-grade and NEC-like phenotype have also been documented [21, 25]. For instance, a series of G1/G2 NETs evolving into G3 neoplasms with acquired TP53 and RB1 co-mutations have been reported, blurring the distinction between G3 NET and true NEC. Such cases are exceedingly uncommon (progression from NET to NEC is considered very rare however this highlights that subtype transitions, while infrequent, are possible. More commonly, what appears as “NET to NEC” transformation in the clinic may turn out to be a misclassification or a mixed tumor scenario [24]. Additionally, outside the GI/Lung context, some non-neuroendocrine carcinomas can transdifferentiate into a neuroendocrine phenotype under therapy pressure—a phenomenon known as neuroendocrine lineage plasticity [42]. This is well documented in prostate cancer treated with androgen blockade, where emergent neuroendocrine prostate cancer (NEPC) shares molecular programs with small-cell lung cancer [43].

The existence of NE lineage plasticity in other epithelial cancers suggests that epithelial tumor cells can switch on a neuroendocrine program (often via epigenetic reprogramming [44], effectively mimicking a PD-NECs. In the context of NENs, however, most PD-NECs appear to arise de novo rather than from pre-existing NETs and further investigation is required to clearly elucidate mechanisms. Our data support this paradigm where the clear bifurcation in gene expression and DNA methylation profiles between WD and PD classes implies separate evolutionary trajectories. Monitoring high-grade NETs for acquisition of NEC-like molecular traits (such as a surge in cell-cycle gene expression or loss of endocrine markers) might help flag patients at risk of aggressive transformation.

Several limitations warrant acknowledgment. The discovery cohort is modest in size and heterogeneous with respect to primary organ site. Histology and grade are collinear by definition (WD-NET: G1/G2; PD-NEC: G3); consequently, the classifier may partly capture proliferative activity rather than differentiation state alone. Furthermore, the external expression and methylation datasets offer cross-modal biological support rather than direct validation of the classifier. While we used cross-validation and external datasets, the analysis remains retrospective and has limitations; thus larger and prospectively collected, multi-center cohorts with harmonized assays will be required to define clinically actionable thresholds, evaluate generalizability across sites, and test whether classifier-guided decisions improve outcomes. More broadly, this framework illustrates how interpretable transcriptomic models may potentially uncover conserved cellular states across various anatomical sites. Similar differentiation proliferation features have been reported in solid epithelial cancers with lineage plasticity, suggesting the approaches may generalize beyond PD-NECs and WD-NETs [44, 45].

Conclusions

This study demonstrates that neuroendocrine neoplasms (NENs) can be robustly stratified into well-differentiated NETs and poorly differentiated NECs using transcriptomic data and interpretable machine learning algorithms. By integrating logistic regression, random forest classifiers, and SHAP-based feature attribution, we identified reproducible gene signatures, including SFN, FANCA, HIST1H3B, CHEK1, CDC6 for PD-NECs and CAMK2B, PAK3, GRIA3 for WD-NET. These markers reflect underlying biological differences, such as cell cycle activation in PD-NECs and neuronal identity in WD-NETs. We validated these findings across multiple levels: (1) orthogonal protein expression by immunohistochemistry, (2) inference of unknown tumor primaries via transcriptomic proximity, and (3) independent external validation using DNA methylation data from a separate lung NEN cohort. The concordant separation of subtypes across expression, methylation, and machine learning predictions underscores the robustness of our discoveries in molecular signatures. These results highlight the potential for AI-guided classifiers to augment histopathological diagnosis, and resolve ambiguous NEN cases.

Author contributions

H.L. and N.V. conceptualized the study. P.F.E. designed the transcriptome study. S.P. and J.M. performed transcriptome experiments. G.Y. and C.Q. conducted histological analysis. T.K., K.C., E.E., J.Q. and H.L. performed computational analyses and developed machine learning models. G.D.A. conducted validation. H.L. implemented the visualization application. T.K., H.L., S.P., K.S.C., J.R.W., I.A., and N.V. contributed to manuscript writing. H.L. and N.V. supervised the study. All authors reviewed and approved the final manuscript.

Funding

This research was supported by new faculty funds from Fox Chase Cancer Center and the Cancer Center Support Grant P30CA006927. Work related to this study is supported by R35GM144131 (J.R.W.) and NIH/NCI Cancer Center Support grant P30 CA006927 (J.R.W.).

Data availability

Formalin-fixed, paraffin-embedded (FFPE) tumor tissues were collected under IRB-approved protocols at Fox Chase Cancer Center and are not publicly available due to patient confidentiality requirements; additional access may be granted upon reasonable request, subject to institutional review and execution of a material transfer agreement (MTA). Gene expression data generated in this study are provided in Supplementary Data 1. Publicly available datasets used in this study include GEO accessions GSE211483 and GSE211485. Analysis scripts, trained models, and visualization tools are deposited in the project GitHub repository (NEN_hist_classifier_ge; https://github.com/comp-epi-lee-lab/NEN_hist_classifier_ge) and are publicly accessible.

Declarations

Ethics approval and consent to participate

All procedures involving human participants were conducted in accordance with the Declaration of Helsinki. The Institutional Review Boards (IRB) of Fox Chase Cancer Center approved the study.

Consent for publication

Not applicable.

Competing interests

J.R.W. at Fox Chase Cancer Center has served or is serving as a consultant or advisor for Qsonica, Daiichi Sankyo, Inc., and Vyne Therapeutics in past year. J.R.W. received an honorarium from Arima Genomics and is a co-investigator on a sponsored project. J.R.W also received funding from Oryzon Genomics. K.S.C. received funding from Janssen Pharmaceuticals.