Cytologic and histopathologic diagnosis of non-ductal pancreatic neoplasms can be challenging in daily clinical practice, whereas it is crucial for therapy and prognosis. The cancer methylome is successfully used as a diagnostic tool in other cancer entities. Here, we investigate if methylation profiling can improve the diagnostic work-up of pancreatic neoplasms. DNA methylation data were obtained for 301 primary tumors spanning 6 primary pancreatic neoplasms and 20 normal pancreas controls. Neural Network, Random Forest, and extreme gradient boosting machine learning models were trained to distinguish between tumor types. Methylation data of 29 nonpancreatic neoplasms (n = 3708) were used to develop an algorithm capable of detecting neoplasms of non-pancreatic origin. After benchmarking 3 state-of-the-art machine learning models, the random forest model emerged as the best classifier with 96.9% accuracy. All classifications received a probability score reflecting the confidence of the prediction. Increasing the score threshold improved the random forest classifier performance up to 100% with 87% of samples with scores surpassing the cutoff. Using a logistic regression model, detection of nonpancreatic neoplasms achieved an area under the curve of >0.99. Analysis of biopsy specimens showed concordant classification with their paired resection sample. Pancreatic neoplasms can be classified with high accuracy based on DNA methylation signatures. Additionally, non-pancreatic neoplasms are identified with near perfect precision. In summary, methylation profiling can serve as a valuable adjunct in the diagnosis of pancreatic neoplasms with minimal risk for misdiagnosis, even in the pre-operative setting.

中文翻译：

---

DNA 甲基化分析可实现非导管原发性胰腺肿瘤的准确分类

非导管胰腺肿瘤的细胞学和组织病理学诊断在日常临床实践中可能具有挑战性，但对于治疗和预后至关重要。癌症甲基化组已成功用作其他癌症实体的诊断工具。在这里，我们研究甲基化分析是否可以改善胰腺肿瘤的诊断检查。获得了 301 个原发性肿瘤的 DNA 甲基化数据，其中包括 6 个原发性胰腺肿瘤和 20 个正常胰腺对照。训练神经网络、随机森林和极端梯度增强机器学习模型来区分肿瘤类型。 29 个非胰腺肿瘤 (n = 3708) 的甲基化数据被用来开发一种能够检测非胰腺起源肿瘤的算法。在对 3 个最先进的机器学习模型进行基准测试后，随机森林模型以 96.9% 的准确率成为最佳分类器。所有分类都会收到反映预测置信度的概率分数。提高分数阈值将随机森林分类器的性能提高了 100%，其中 87% 的样本分数超过了截止值。使用逻辑回归模型，非胰腺肿瘤的检测曲线下面积>0.99。活检标本的分析显示其分类与其配对的切除样本一致。胰腺肿瘤可以根据 DNA 甲基化特征进行高精度分类。此外，非胰腺肿瘤的识别精度近乎完美。总之，甲基化分析可以作为胰腺肿瘤诊断的一个有价值的辅助手段，即使在术前情况下，误诊的风险也很小。