CrossNN 中枢神经系统肿瘤 DNA 甲基化分类器的临床评估Clinical evaluation of the CrossNN DNA methylation classifier for central nervous system tumors.

Abstract

DNA methylation profiling is an integral diagnostic tool in the classification of central nervous system (CNS) tumors. While the Heidelberg CNS Tumor Methylation Classifier is widely used to support CNS tumor diagnostics, new classifiers such as CrossNN are emerging. However, their clinical performance and added value within routine diagnostic workflows remain insufficiently explored. In this study, we evaluated the diagnostic performance of the CrossNN classifier in a real‐world CNS tumor cohort and compared it with the established Heidelberg classifier to assess its potential as both a non‐inferior alternative and a complementary tool to improve diagnostic accuracy. A retrospective cohort of CNS tumors profiled using Illumina Human Methylation 930k EPIC v2 BeadChip arrays was analyzed. Classifier outputs were compared with integrated WHO CNS5 (2021) diagnoses. In addition, CrossNN and Heidelberg outputs were harmonized to WHO CNS5 (2021) tumor type levels and evaluated both individually and within sequential and parallel diagnostic workflows. The proposed workflows were subsequently assessed in an independent prospective validation cohort. Among 205 samples, CrossNN correctly classified 88.8% of cases and demonstrated 86.8% concordance with the Heidelberg classifier. CrossNN demonstrated non‐inferior classification performance compared with the Heidelberg classifier. Combining both classifiers increased the number of clinically informative and correct classifications by nearly 10%. This finding was confirmed in an independent validation cohort of 41 samples. In conclusion, these results demonstrate the complementary strength of the CrossNN and Heidelberg classifiers as a dual‐classifier strategy to improve diagnostic confidence and accuracy in routine CNS tumor diagnostics.

1 INTRODUCTION

Accurate classification of central nervous system (CNS) tumors is imperative for optimal patient management. Although formerly determined by histological assessment alone, the classification of these tumors is increasingly supported by molecular parameters. The 5th edition of the World Health Organization (WHO) classification of CNS tumors in 2021 introduced the use of DNA methylation profiling for molecular classification of CNS tumors and resulted in major advances in this field [1, 2].

DNA methylation is a stable epigenetic modification preserved during cell division and plays an important role in tissue‐specific gene regulation [3]. As a result, DNA methylation patterns are highly cell type‐specific. In neoplastic cells, these patterns are characterized by global hypomethylation in combination with focal hypermethylation of promoter regions of tumor suppressor genes, while retaining information on the cell of origin [4]. These tumor‐specific methylation profiles allow discrimination between neoplastic and non‐neoplastic cells, as well as enabling molecular classification of CNS tumors. Furthermore, these methylation signatures remain preserved in formalin‐fixed paraffin‐embedded (FFPE) samples and are unaffected by sample processing or storage conditions, making them suitable for routine diagnostic applications [5].

Building on these principles, machine‐learning‐based classifiers have been developed to categorize CNS tumors based on genome‐wide methylation profiles [6, 7, 8, 9]. Among these, the Heidelberg CNS Tumor Methylation Classifier (hereafter, the Heidelberg classifier) is the most widely used methylation‐based classifier for CNS tumor diagnostics and an important component of current molecular classification approaches. The current classifier is based on the Illumina Human Methylation 930 k EPIC v2 BeadChip array. This random forest‐based algorithm is trained on 7495 array‐based methylation profiles and can distinguish up to 184 CNS subclasses with a 95% subclass‐level accuracy in its latest version (v.12.8; https://epignostix.com/) [10]. For each sample, the classifier generates a prediction with a corresponding calibrated score, indicating the degree of similarity between the sample and the reference profiles in the training database. Sill et al. defined predictions with the Heidelberg classifier as confident when the calibrated score is ≥0.9. Capper et al. described predictions with scores between 0.3 and 0.9 as indicative, and highlighted the need for strong correlation with clinicopathological findings [11]. Predictions with a calibrated score <0.3 are non‐informative. In addition, the EpignostiX report includes a copy number profile and MGMT promoter methylation status prediction. Alternative approaches for whole‐genome methylation profiling have been described, including faster and more cost‐effective methods, such as low‐coverage whole‐genome nanopore sequencing, which show diagnostic performance comparable to array‐based data for CNS tumor classification [8].

Recently, a novel classifier, CrossNN, has been introduced that enables classification of DNA methylation profiles generated across multiple platforms [7]. This classifier is based on a neural network framework designed to handle variable and sparse input data while still being fully explainable. The classifier was trained on the Heidelberg brain tumor classifier v11b4 reference dataset, including 2801 samples across 82 tumor types [6]. At the methylation class family level, CrossNN achieved an overall accuracy of 95.6%, ranging between accuracies of 97% for EPICv2 array data and 89.5% for targeted sequencing data. To enable cross‐platform compatibility, methylation data are binarized prior to classification. For array‐based data, β values greater than 0.6 are encoded as 1, β values of 0.6 or lower as −1, and missing CpG sites as 0. Based on the binarized input, the classifier generates a list of predictions with a confidence score reflecting the similarity between the sample and the reference profiles. Yuan et al. defined confidence cutoffs of 0.4 for microarray‐based data and 0.2 for sequencing‐based data. To date, however, no independent external validation of CrossNN using whole‐genome methylation microarray data has been reported.

In this study we evaluated the performance of the CrossNN classifier in a real‐world CNS tumor cohort and compared it with the established Heidelberg classifier (v12.8). We demonstrate that the performance of CrossNN is non‐inferior to the Heidelberg classifier, supporting its use as an alternative diagnostic approach. Moreover, applying both classifiers simultaneously improves diagnostic accuracy in CNS tumor diagnostics.

2 MATERIALS AND METHODS

2.1 Sample selection

Methylation profiling of CNS tumors has been implemented as part of routine diagnostic practice at the Antwerp University Hospital (UZA, Belgium) since 2022. For this study, tumor samples profiled consecutively between August 2024 and June 2026 were included. Ethical approval for the use of these retrospective samples was obtained from the UZA Ethics Committee (Project ID 7381; EDGE 004111). Histopathological diagnosis and tumor cell content were assessed by two experienced neuropathologists (M.A. and A.S.). Tumor percentages ranged from 10% to 95%. An overview of included tumor entities is provided in Table S1.

2.2 DNA extraction and methylation profiles

DNA was extracted from FFPE tumor tissue using the QIAamp DNA Mini QIAcube Kit (Qiagen, Hilden, Germany) according to the manufacturer's protocol. Bisulfite conversion of 40–250 ng DNA, as quantified with a Qubit dsDNA BR assay kit (Thermo Fisher Scientific), was carried out using the EpiTect Fast Bisulfite Conversion Kits (Qiagen, Hilden, Germany), followed by DNA restoration with the Infinium HD FFPE DNA Restore Kit (Illumina, San Diego, CA, USA). Genome‐wide DNA methylation profiles were obtained using the Illumina Infinium MethylationEPICv2 (EPICv2) BeadChip and iScan device (Illumina, San Diego, CA, USA) according to the manufacturer's protocol with adaptations for FFPE samples. All samples underwent standard quality control as specified by the manufacturer (Illumina, San Diego, CA, USA). Samples failing these metrics or without a definitive diagnosis were excluded from further analysis. All processing steps were performed in accordance with routine diagnostic workflows under ISO 15189 accreditation at UZA.

2.3 Heidelberg classifier prediction

To generate methylation class and class family predictions with the Heidelberg classifier, paired raw intensity array files were processed with the current classifier version (v.12.8) via https://epignostix.com/. For each sample the classifier provided a predicted methylation (super)family and, where applicable, a corresponding (sub)class, along with a calibrated score ranging from 0.1 to 0.99. Superfamily predictions with a score <0.3 were not reported.

2.4 CrossNN prediction

Raw intensity array files were processed in R (v.4.4.1) using the minfi Bioconductor package (v.1.52.1). Beta values were generated with the preprocessIllumina function and subsequently binarized as described by Yuan et al. for input into the CrossNN classifier [7, 12]. The resulting binary matrices were processed locally using the currently available version of the CrossNN brain tumor model (version as of January 23, 2024), available at https://gitlab.com/euskirchen-lab/crossNN. For each sample, the classifier returned a list of predicted methylation classes or class families, with their associated scores ranging from near zero to 0.99. As for the Heidelberg classifier, these scores reflect the degree of similarity between the sample's methylation profile and the reference profiles in the training database.

2.5 Data analysis

For all analyses, the reference (true) tumor type was defined by the integrated histopathological and molecular diagnosis determined by two experienced neuropathologists (M.A. and A.S.), in accordance with WHO CNS5 (2021). As these diagnoses were established with knowledge of the Heidelberg classifier output, potential bias toward this classifier cannot be excluded. To address this, a targeted post‐hoc review was conducted for cases in which CrossNN generated incorrect predictions that differed from the Heidelberg classifier. This post‐hoc review re‐integrated histopathological and molecular data, incorporating CrossNN predictions instead of Heidelberg predictions. None of the 15 reviewed cases led to a change in the final diagnosis (Table S2).

In this study, all results were compared at the WHO CNS5 (2021) tumor type level, as most subclasses currently do not carry direct clinical implications. However, for astrocytoma, IDH‐mutant, the distinction between low‐grade and high‐grade disease was retained caused by its clinical relevance. Additionally, medulloblastoma, SHH‐activated, TP53‐mutant and TP53‐wildtype were grouped as a single tumor type (medulloblastoma, SHH‐activated), as neither classifier differentiates between these subtypes.

Classifier outputs were categorized as either correct, higher‐level, or incorrect classifications. A prediction was considered correct if it matched the WHO CNS5 (2021) tumor type of the integrated reference diagnosis. Predictions at a more specific subclass level within the correct WHO CNS5 (2021) tumor type were also considered correct. Predictions were defined as a higher‐level classification if the Heidelberg classifier did not reach the WHO CNS5 (2021) tumor type level but assigned a correct superfamily prediction (e.g., true tumor type: glioblastoma, IDH‐wildtype; classifier output: adult‐type diffuse glioma). Predictions that not met any of these criteria were considered incorrect.

To account for tumor types that are not represented in the CrossNN reference cohort, certain predefined biologically and clinically related predictions were considered higher‐level classifications rather than incorrect classifications. This approach was intended to avoid penalizing biologically meaningful predictions that could not be resolved more specifically because of limitations in the reference cohort. Specifically, CrossNN glioblastoma, IDH‐wildtype predictions were considered higher‐level predictions for diffuse pediatric‐type high‐grade gliomas, H3‐wildtype and IDH‐wildtype (HGG, PAED), as both entities represent H3‐ and IDH‐wildtype diffuse high‐grade gliomas with overlapping therapeutic options and prognosis. Likewise, CrossNN ganglioglioma predictions were considered higher‐level predictions for polymorphous low‐grade neuroepithelial tumors of the young (PLNTY), as both entities share overlapping methylation‐based features. Finally, because CrossNN does not distinguish MYCN‐amplified spinal ependymoma from other spinal ependymomas, CrossNN spinal ependymoma predictions were considered higher‐level classifications.

The Heidelberg classifier generates predictions across multiple hierarchical levels. To enable direct comparison with the CrossNN classifier, outputs were harmonized to the WHO CNS5 (2021) tumor type level using predefined mapping rules (Table S3). The highest‐scoring prediction at the WHO tumor type level was used as final output. All predictions with a calibrated score <0.3 were classified as non‐informative. Heidelberg predictions with calibrated scores ≥0.9 were considered confident, whereas those with scores between 0.3 and <0.9 were considered unconfident, in accordance with Capper et al. [11].

CrossNN does not generate hierarchical predictions. Therefore, for the primary analysis, the highest‐scoring prediction was used as classifier output. For the comparison with the Heidelberg classifier, WHO‐level outputs were derived from the top three highest‐scoring predictions. When multiple top‐three predictions mapped to the same WHO tumor type (Table S3), their scores were cumulatively summed to reflect the total probability mass assigned to that tumor type. This approach reflects the distribution of classifier confidence across biologically related subclasses rather than introducing additional predictive information, consistent with the presentation of predictions in the graphical user interface of CrossNN (https://crossnn.charite.de) and the Heidelberg classifier. If this combined score exceeded the score of the highest‐ranking individual prediction, the combined score and associated WHO tumor type were used as the final classifier output. Otherwise, the highest‐ranking prediction and its score were retained. A sensitivity analysis was performed to assess the robustness of the harmonization strategy across different values of k (top 1–5 predictions). CrossNN predictions were considered confident at scores ≥0.4 and unconfident at scores <0.4. No minimum cutoff threshold was applied to CrossNN predictions, as none was defined by Yuan et al. [7]. All CrossNN analyses adhered to the classifier‐specific threshold.

2.6 Statistical analysis

All statistical analyses and visualizations were performed using the R software (version 4.2.2). Data distribution was assessed using the Shapiro–Wilk test and visual inspection of quantile‐quantile plots. Differences in confidence scores and tumor purity were assessed using the Wilcoxon rank‐sum test, with effect sizes reported as rank‐biserial correlation coefficients (r). Classifier performance was evaluated using overall accuracy with 95% confidence intervals (CI; Wilson method). Agreement between classifiers was assessed using Cohen's kappa, and differences in paired classification outcomes between CrossNN and the Heidelberg classifier were evaluated using McNemar's test.

Non‐inferiority of CrossNN was assessed using the paired difference in overall accuracy (CrossNN—Heidelberg), with a predefined non‐inferiority margin of −3%, chosen to represent a clinically acceptable difference in classification performance. A one‐sided 95% bootstrap CI was estimated based on 10,000 resamples. Non‐inferiority was concluded if the lower bound of this confidence interval exceeded the predefined margin. All statistical analyses were two‐sided unless otherwise specified, and a p‐value ≤0.05 was considered statistically significant.

3 RESULTS

3.1 Study cohort characteristics

The study cohort included 266 FFPE samples that underwent routine DNA methylation profiling at UZA between August 2024 and June 2026. Of these, 5/266 failed quality metrics and 15/266 did not receive a definitive diagnosis and were excluded. Median age at diagnosis was 58 years (range 2–85), including 16 pediatric and 230 adult cases (Table 1). The cohort included 133 male and 113 female patients. Adult‐type diffuse gliomas were most common (45.1%, 111/246), followed by meningiomas (31%, 75/246). Tumor grades were available for 97% (239/246) of cases, with 31% (73/239) WHO grade 1, 20.1% (48/239) WHO grade 2, 4.2% (10/239) WHO grade 3, and 45.2% (108/239) WHO grade 4. Clinicodemographic characteristics are provided in Table 1. The retrospective study cohort comprised samples 1–205, while samples 206–246 were used for independent validation of the proposed diagnostic workflows (Table S4).

3.2 Evaluation of CrossNN performance

Using the WHO CNS5 (2021) diagnosis of the integrated pathology report as reference standard, the CrossNN classifier correctly classified 88.3% of samples (181/205, 95% CI: 83.2%–92.0%, Figure 1A). In 75.1% (154/205) of cases, CrossNN produced a confident prediction (α ≥ 0.4), of which 95.5% (147/154) were correctly classified. In the remaining 24.9% of cases (51/205), accuracy was 66.7% (34/51).

Figure 1.

In five samples, CrossNN did not identify the exact WHO tumor type, as these were not represented in the classifier's reference cohort. Instead, the classifier generated biologically meaningful higher‐level predictions that may support the final diagnosis when interpreted alongside histopathological findings. Specifically, two HGG, PAED samples were classified as midline glioblastomas IDH‐wildtype with low confidence; two spinal ependymoma samples were assigned to spinal ependymoma with high confidence, although MYCN amplification status could not be resolved; and one PLNTY sample was classified as ganglioglioma with low confidence.

Across tumor entities, highest‐ranking predictions were correct for all meningioma (MNG) cases (65/65, Figure 1B). Several entities, however, showed misclassifications against the integrated diagnosis in the pathology report. Among glioblastomas (GBM), 12.2% (9/74) were predicted as control tissue (CONTR), while 2.7% (2/74) and 1.4% (1/74) were assigned to diffuse midline glioma H3 K27M‐mutant (DMG, K27) and low‐grade glioma, dysembryoplastic neuroepithelial tumor (LGG, DNT), respectively. In pilocytic astrocytoma (PA), 22.2% (2/9) of samples were misclassified as LGG, DNT. For high‐grade astrocytoma, IDH‐mutant (A IDH, HG), the classifier predicted 1/5 samples as PA and 1/5 as low‐grade astrocytoma, IDH‐mutant (A IDH). One A IDH was predicted as control tissue. No correct highest‐ranking predictions were generated for ganglioglioma (GG) or rosette‐forming glioneuronal tumor (LGG, RGNT).

Confidence scores differed significantly between correct and incorrect predictions, with a moderate effect size (p < 0.001, r = 0.353, Figure 1C). Correct classifications were associated with higher confidence scores (median 0.66 vs. 0.24). Tumor purity showed a similar pattern, with a small effect size (p < 0.001, r = 0.243, Figure 1D) and a higher median tumor purity in correct compared to incorrect predictions (75% vs. 55%). Among confident predictions, all but 5/46 (10.9%) GBM cases were predicted correctly. These five cases were misclassified as control tissue (Figure S1). Histopathological review showed that these samples were characterized by extensive necrosis or low tumor purity.

Exploration of alternative confidence thresholds did not improve the trade‐off between the proportion of confident predictions and accuracy within confident predictions (Figure S2). Increasing the threshold reduced the number of confident predictions, while accuracy within this subset remained stable at approximately 95%.

3.3 CrossNN and Heidelberg classifier performance at the WHO entity level output

To enable direct comparison between the CrossNN and Heidelberg classifiers, CrossNN subclass predictions were harmonized to the WHO CNS5 (2021) tumor type level by mapping lower‐level (sub)class predictions and their associated confidence scores to their corresponding WHO tumor type (Table S3). Using this strategy, 10 samples were reclassified from previously “correct & unconfident” to “correct & confident”, and two predictions that were previously “incorrect & unconfident” became “incorrect & confident”. To assess the robustness of this harmonization strategy, a sensitivity analysis was performed using values of k ranging from 1 to 5. WHO‐level classifications remained stable across all k‐values (Figure S3). Compared with the primary analysis (k = 3), increasing k to 5 reclassified only two cases from “correct & unconfident” to “correct & confident”, without changing any tumor type predictions.

Using the harmonization strategy based on the top three CrossNN predictions, CrossNN correctly identified 88.8% (182/205, 95% CI: 83.7%–92.4%) of cases (Figure 2A). Confident predictions were obtained in 81% (166/205) of cases, with an accuracy of 94.6% (157/166) within this group. Correct predictions remained associated with higher confidence scores than incorrect predictions, with a moderate effect size (median 0.74 vs. 0.31, p < 0.001, r = 0.343, Figure S4). As this strategy increased the number of “correct & confident” predictions, despite a small decrease in accuracy, it was used for all subsequent analyses.

Figure 2.

At the WHO tumor type level, CrossNN achieved an accuracy of 88.8% (182/205, 95% CI: 83.7%–92.4%), compared with 86.3% (177/205, 95% CI: 81%–90.4%) for the Heidelberg classifier. The observed paired accuracy difference was 2.5%. No significant difference in accuracy was observed between both classifiers (McNemar test, p = 0.359). Furthermore, CrossNN met the predefined criterion for non‐inferiority, as the lower bound of the one‐sided 95% bootstrap confidence interval (10,000 resamples) for the paired accuracy difference of −0.97% exceeded the predefined non‐inferiority margin of −3%.

3.4 Heidelberg and CrossNN as complementary tools

The Heidelberg and CrossNN classifiers showed an almost perfect agreement at the WHO tumor type level (Cohen's κ = 0.836). Concordant predictions, excluding higher‐level CrossNN outputs, were observed in 85.8% (176/205) of samples, of which 96.6% (170/176) were correct (Figure 2B,C). Among the correct and concordant cases, both classifiers were confident in 79.4% (135/170) of cases. Discordant predictions occurred in 4.9% (10/205) of cases. Within this group, CrossNN correctly classified four cases and Heidelberg four different cases. In two cases, neither classifier was correct. Notably, higher confidence scores did not consistently associate with correct classification among discordant cases. An overview of all predictions is provided in Table S4.

Next, we evaluated whether one classifier could provide clinically useful information when the other generated an unconfident, higher‐level, or non‐informative result. The Heidelberg classifier generated confident predictions (α ≥ 0.9) in 76.6% (157/205) of the study cohort, of which 93.6% (147/157) were concordant with the integrated pathology diagnosis. Among the 34 unconfident Heidelberg predictions, CrossNN generated 20 “correct & confident” classifications, two “incorrect & confident” classifications and one “higher‐level & confident” classification by correctly identifying spinal ependymoma without MYCN amplification status (Figure 3A). In accordance with the predefined evaluation criteria, this prediction was considered a higher‐level classification. Of the five Heidelberg superfamily predictions, CrossNN correctly identified three cases. However, none reached a confidence score above 0.4 and were therefore categorized as “correct & unconfident” (Figure 3A). Among the nine non‐informative (α < 0.3) Heidelberg outputs, CrossNN produced one “correct & confident” and one “incorrect & confident” prediction.

Figure 3.

CrossNN generated 39 unconfident predictions (Figure 3B). Within this subset, the Heidelberg classifier generated 16 confident predictions, of which 11 were correct. Among the two spinal ependymoma predictions, which were considered higher‐level as CrossNN does not distinguish MYCN amplification status, Heidelberg produced one “correct & confident” classification. Other predefined higher‐level CrossNN outputs could only be recognized retrospectively after comparison with the reference diagnosis and were therefore not considered higher‐level predictions in this analysis.

3.5 Clinical implementation of the Heidelberg and CrossNN classifier

To assess the clinical utility of combining both classifiers, we evaluated parallel and sequential diagnostic workflows and predefined two outcome categories: a pass‐group, in which the classifier output provides a high diagnostic weight, and a cautious group in which the pathologist would receive two potential diagnoses.

In the parallel workflow, both classifiers were applied simultaneously. Here, classification was based solely on concordance between the two classifiers, independent of confidence. Cases were assigned to the pass‐group when both classifiers agreed at the WHO tumor type level. All remaining cases were assigned to the cautious group.

In the sequential workflow, the second classifier was applied only if the first generated an unconfident, higher‐level, or non‐informative result. Cases were assigned to the pass‐group if the first classifier produced a confident prediction. Once a second classifier was applied, cases were assigned to the pass‐group if the second classifier agreed at the WHO CNS5 (2021) tumor type level with the first classifier, independent of confidence. All discordant results, including superfamily predictions, were assigned to the cautious‐group.

When both classifiers were applied simultaneously, at least one correct classification was provided in 92.2% (189/205, Figure 4A). The pass‐group contained 86.8% (178/205) of cases and achieved an accuracy of 96.1% (171/178).

Figure 4.

In the sequential workflow, when the Heidelberg classifier was used first, 48/205 cases required additional CrossNN analysis. This approach yielded at least one correct classification in 90.3% (185/205). The pass‐group comprised 90.7% (186/205) of cases, with an accuracy of 93% (173/186). When CrossNN was applied first, 41/205 cases proceeded to Heidelberg testing. This approach resulted in at least one correct classification in 91.7% (188/205) of cases, with 89.3% (183/205) assigned to the pass‐group and an accuracy of 94.5% (173/183).

3.6 Prospective validation of the clinical decision workflow

To evaluate the clinical applicability of the proposed diagnostic workflows, we performed a prospective validation in an independent cohort of 41 samples (samples 206–241, Table S4). When both classifiers were applied simultaneously, 87.8% (36/41) of cases showed concordant predictions and were therefore assigned to the pass group (Figure 4B). Within this group, all predictions were concordant with the integrated pathology diagnosis. In the Heidelberg‐first sequential workflow, eight cases proceeded to CrossNN testing. This resulted in 90.2% (37/41) of cases being assigned to the pass group, with an accuracy of 100% within this group. In the CrossNN‐first sequential workflow, nine cases required additional Heidelberg analysis, yielding 87.8% (36/41) pass group cases with an accuracy of 100%.

4 DISCUSSION

DNA methylation profiling is an essential component of tumor classification in neuro‐oncology, improving diagnostic accuracy, particularly for diagnostically challenging CNS tumors. With the increasing availability of methylation‐based classifiers, understanding how their outputs should be interpreted and how different classifiers may complement each other has become increasingly important. In this study, we evaluated the CrossNN classifier in an independent validation cohort of real‐world FFPE samples and examined its complementarity with the Heidelberg classifier in a clinical setting. CrossNN can be implemented locally without licensing requirements, facilitating its use in routine clinical workflows. Overall, CrossNN demonstrated non‐inferiority to the Heidelberg classifier. In our cohort, its classification accuracy aligned closely with the performance reported by Yuan et al. for EPICv2 data, supporting the robustness of the model across cohorts [7]. When both classifiers were applied simultaneously, concordant predictions were predominantly correct, indicating that inter‐classifier agreement can serve as a marker of diagnostic confidence. Notably, CrossNN also performed well in pediatric tumors, a subgroup in which molecular profiling is often indispensable. In this subset, 14 of 16 cases were classified correctly, matching the performance of the Heidelberg classifier. The two incorrect classifications involved (i) a ganglioglioma predicted as control tissue and (ii) a PLNTY predicted as ganglioglioma.

Our results confirm a strong association between CrossNN confidence scores and classification accuracy. High confidence predictions were generally correct, demonstrating their value as indicators of diagnostic reliability. This mirrors the interpretation of the calibrated score use in the Heidelberg classifier, where confidence metrics determine whether a result is considered diagnostically informative [10, 11]. Accordingly, CrossNN confidence scores can be incorporated into clinical decision‐making, particularly in diagnostically challenging cases. Furthermore, unconfident CrossNN outputs should be considered non‐informative when used as a standalone tool, as approximately one‐third of unconfident predictions were incorrect within our cohort.

Harmonization of classifier outputs to WHO CNS5 (2021) tumor type levels with cumulative summation of confidence scores further enhanced diagnostic confidence. Although this approach resulted in a small decrease in accuracy within the confident prediction subset, it increased the absolute number of correct and confident predictions, translating into additional patients receiving a reliable molecular classification. Importantly, all confident but incorrect predictions exclusively involved tumors misclassified as control tissue.

Misclassification of CNS tumors as control tissue reflects biological limitations rather than a failure of the classification algorithm. Each case was characterized by extensive necrosis or low tumor purity, conditions in which degraded or non‐neoplastic methylation signals dominate the methylation profile. Similar observations have been reported previously, highlighting the impact of tissue composition on classifier performance [8, 13]. These findings emphasize that a control tissue prediction should lead to careful reassessment of sample quality and tumor content rather than being interpreted against a neoplastic process, particularly when histopathological or radiological findings suggest a malignancy. Emerging approaches such as methylation‐based deconvolution may overcome this limitation, as demonstrated by Wu et al., by improving classifier performance in low‐purity samples [13].

CrossNN demonstrated added value as a complementary tool to the Heidelberg classifier. The use of both classifiers resulted in nearly 10% more correctly classified cases in the pass group than either classifier alone, consistent with observations reported for the Bethesda v3 classifier [14]. Although the resource‐intensive strategy of applying both classifiers simultaneously requires dual analysis, it resulted in fewer incorrect predictions, remaining below 5%. Discordant results were observed in a small subset of cases. However, these often involved closely related entities, such as glioblastoma and HGG, PAED, or non‐informative and control tissue outputs, thereby still providing an informative top‐2 prediction to support diagnostic decision‐making. A less resource‐demanding sequential approach, in which unconfident, higher‐level, or non‐informative outputs from one classifier were re‐analyzed using the second classifier, produced slightly more correct classifications within the pass group. However, this strategy also increased the number of incorrect predictions within this group. Practical implementation of the dual‐classifier approach will depend on local infrastructure and resource availability. While simultaneous application of both classifiers increases computational and analytical workload, a sequential workflow may offer a more feasible alternative by limiting second‐classifier analyses to cases with unconfident initial predictions, thereby reducing turnaround time and potential future licensing costs while retaining most of the diagnostic benefit observed in this study. Importantly, prospective evaluation in an independent cohort yielded similar findings, supporting the clinical applicability of these workflows. However, the prospective cohort was limited in size, requiring larger prospective studies to confirm these observations. Based on these findings, simultaneous application of both classifiers represents the most robust strategy with the lowest risk of misclassification, while the sequential method may serve as a pragmatic alternative in resource‐limited settings.

Several limitations of the CrossNN classifier are important for clinical implementation. Tumor misclassification largely reflects limitations of the reference dataset. Rare entities (e.g., rosette‐forming glioneuronal tumors) are underrepresented, and several tumor types included in the Heidelberg classifier are absent in CrossNN. As a result, tumors not represented in the CrossNN reference cohort are assigned to the closest available class. When interpreted alongside histopathological and molecular findings, such outputs can still provide meaningful clinical guidance. For example, a CrossNN prediction of ganglioglioma as PLNTY may still direct the pathologist toward the correct diagnosis through morphological assessment. Further misclassification occurred among biologically related entities within the same superfamily or class. One remaining misclassification of A IDH, HG as PA could not be attributed to biological relatedness. However, this case also yielded a non‐informative output with the Heidelberg classifier. Beyond misclassification, CrossNN provides limited additional molecular information, such as MYCN amplification. As copy number profiles and MGMT promoter methylation status are clinically relevant, and in some entities essential for diagnosis, prognosis, and treatment planning, this represents a clinical gap. This limitation can be overcome by deriving complementary copy number profiles directly from the array data. Finally, although CrossNN identified meningiomas with 100% accuracy, it does not provide subclassification, which remains important for clinical decision‐making. In these cases, histological assessment combined with copy number profiling remains essential for tumor grading.

Certain limitations of this study should be considered when interpreting the results. First, classifier outputs were benchmarked against the integrated pathology diagnosis, which incorporates the Heidelberg classifier and therefore introduces an inherent reference bias in favor of Heidelberg. Although discordant cases were re‐evaluated, such bias cannot be fully eliminated, preventing definitive claims regarding the superiority of one classifier over the other. Importantly, despite this potential bias, CrossNN achieved a comparable overall accuracy and met the predefined criterion for non‐inferiority. As any residual reference bias would be expected to favor Heidelberg, these findings support the robustness of the observed CrossNN performance. Second, the cohort reflects real‐world diagnostic case distribution and was therefore dominated by common entities such as adult‐type diffuse gliomas and meningiomas, while rare and diagnostically challenging tumor types were underrepresented. Although this enhances clinical relevance, it restricts reliable assessment of classifier performance for rare tumor types and prevents definitive conclusions regarding performance in these specific diagnostic cases. Future studies including larger numbers of rare CNS tumor types are needed to further evaluate its performance across the full spectrum of CNS tumors. Third, this study only evaluated data generated using the Illumina Human Methylation 930k EPIC v2 array. Although platform‐independence and robustness to sparse data have been reported as key advantages of CrossNN, our study did not directly assess these properties. Therefore, generalizability to other methylation platforms or to datasets with reduced CpG coverage remains to be assessed in future studies.

In conclusion, the CrossNN classifier proved to be non‐inferior to the Heidelberg classifier in an independent cohort at the WHO CNS5 (2021) tumor entity level, supporting its potential use as an alternative, license‐free diagnostic tool. Confidence scores showed to be an important indicator for clinical interpretability when used as a standalone classifier. Moreover, combining CrossNN and the Heidelberg classifier increased the number of correct and clinically informative classifications, highlighting the potential value of a dual‐classifier strategy for improving diagnostic confidence in routine practice.

AUTHOR CONTRIBUTIONS

Conceptualization: J.D., L.v.K., M.A., A.S., and K.Z. Methodology: J.D., L.v.K., T.M., B.F., S.K., M.A., A.S., and K.Z. Investigation: J.D., L.v.K., M.A., A.S., and K.Z. Writing: J.D., L.v.K., S.K., K.O.d.B., T.M., B.F., M.A., A.S., and K.Z. Funding Acquisition: L.v.k., K.O.d.B., S.K., K.Z. Resources: L.v.K., S.K., K.O.d.B., T.M., B.F., M.A., A.S., and K.Z. Supervision: L.v.K., M.A., A.S., and K.Z.

FUNDING INFORMATION

Kom op tegen Kanker (Stand up to Cancer), the Flemish cancer society (KOTK 13571).

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

ETHICS STATEMENT

Ethical approval for the use of retrospective samples was obtained from the UZA Ethics Committee (Project ID 7381; EDGE 004111).

Supporting information

ACKNOWLEDGEMENTS

Project funded by Kom op tegen Kanker (Stand up to Cancer), the Flemish cancer society (KOTK 13571).

During the preparation of this work the first author used ChatGPT (v.5.2) in order to edit language. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content. No final decision‐making processes relied on AI tools.

DATA AVAILABILITY STATEMENT

The Illumina Human Methylation 930k EPIC v2 BeadChip array data have been deposited in the NCBI's Gene Expression Omnibus (GEO) database (GSE325845).

References

1. 1 Louis DN , Perry A , Wesseling P , Brat DJ , Cree IA , Figarella‐Branger D , et al. The 2021 WHO classification of tumors of the central nervous system: a summary. Neuro‐Oncology. 2021;23(8):1231–1251. 10.1093/neuonc/noab106 34185076 PMC8328013
2. 2 Aldape K , Capper D , von Deimling A , Giannini C , Gilbert MR , Hawkins C , et al. cIMPACT‐NOW update 9: recommendations on utilization of genome‐wide DNA methylation profiling for central nervous system tumor diagnostics. Neuro‐Oncol Adv. 2025;7(1):vdae228. 10.1093/noajnl/vdae228 PMC1178859639902391
3. 3 Galbraith K , Snuderl M . DNA methylation as a diagnostic tool. Acta Neuropathol Commun. 2022;10(1):71. 10.1186/s40478-022-01371-2 35527288 PMC9080136
4. 4 DNA methylation and cancer. Advances in genetics [Internet]. USA: Academic Press; 2010 [cited 2026 Apr 8]. p. 27–56. 10.1016/B978-0-12-380866-0.60002-2 20920744
5. 5 Davalos V , Esteller M . Cancer epigenetics in clinical practice. CA Cancer J Clin. 2023;73(4):376–424. 10.3322/caac.21765 36512337
6. 6 Capper D , Jones DTW , Sill M , Hovestadt V , Schrimpf D , Sturm D , et al. DNA methylation‐based classification of central nervous system tumours. Nature. 2018;555(7697):469–474. 10.1038/nature26000 29539639 PMC6093218
7. 7 Yuan D , Jugas R , Pokorna P , Sterba J , Slaby O , Schmid S , et al. crossNN is an explainable framework for cross‐platform DNA methylation‐based classification of tumors. Nat Cancer. 2025;6(7):1283–1294. 10.1038/s43018-025-00976-5 40481322 PMC12296554
8. 8 Vermeulen C , Pagès‐Gallego M , Kester L , Kranendonk MEG , Wesseling P , Verburg N , et al. Ultra‐fast deep‐learned CNS tumour classification during surgery. Nature. 2023;622(7984):842–849. 10.1038/s41586-023-06615-2 37821699 PMC10600004
9. 9 Singh O , Abdullaev Z , Aldape K . PATH‐01. The NCI/Bethesda DNA methylation classifier: an important diagnostic tool for CNS tumors. Neuro‐Oncol Pediatr. 2025;1(Supplement_1):wuaf001.280. 10.1093/neuped/wuaf001.280
10. 10 Sill M , Schrimpf D , Patel A , Sturm D , Jäger N , Sievers P , et al. Advancing CNS tumor diagnostics with expanded DNA methylation‐based classification. Cancer Cell. 2025;44(2):340. 10.1016/j.ccell.2025.11.002 41349541
11. 11 Capper D , Stichel D , Sahm F , Jones DTW , Schrimpf D , Sill M , et al. Practical implementation of DNA methylation and copy‐number‐based CNS tumor diagnostics: the Heidelberg experience. Acta Neuropathol (Berl). 2018;136(2):181–210. 10.1007/s00401-018-1879-y 29967940 PMC6060790
12. 12 Fortin JP , Triche TJ Jr , Hansen KD . Preprocessing, normalization and integration of the Illumina HumanMethylationEPIC array with minfi. Bioinformatics. 2017;33(4):558–560. 10.1093/bioinformatics/btw691 28035024 PMC5408810
13. 13 Wu Z , Abdullaev Z , Pratt D , Chung HJ , Skarshaug S , Zgonc V , et al. Impact of the methylation classifier and ancillary methods on CNS tumor diagnostics. Neuro‐Oncol. 2021;24(4):571–581. 10.1093/neuonc/noab227 PMC897223434555175
14. 14 Brandenburg C , Starzetz T , Luger A , Singh O , Aldape KD , Schweizer L . PATH‐63. Analysis of histological, molecular and preanalytic features of unclassifiable CNS specimens: comparison of the Heidelberg v12.8 and Bethesda v3 methylation‐based brain tumor classifiers. Neuro‐Oncol. 2025;27(Supplement_5):v255. 10.1093/neuonc/noaf201.1015