A recent study published in eBioMedicine has demonstrated a neural network classifier that can effectively differentiate between intrahepatic cholangiocarcinomas (iCCAs), intrahepatic metastases of pancreatic ductal adenocarcinoma (PAAD), and normal bile duct tissue with high accuracy.

The proposed tool could also be used for improving the diagnosis of pancreato-biliary cancers of the liver and can serve as the basis for developing more rapid intraoperative classifiers that can have a real-time impact on surgical decision making, the authors noted.

Read more about CCA therapies

In the present study, the research team presented a DNA methylation-based classifier running on machine-learning algorithms capable of differentiating between iCCAs and PAAD. The researchers collected genome-wide DNA methylation data from publicly available sources for 259 patients with iCCA, 70 patients with normal bile ducts, and 431 patients with PAAD.

The cohorts were classified into a reference set of 399 patients and a validation set of 361 patients. Three machine learning models, including support vector machine, random forest, and neural network, were trained using a reference cohort to differentiate between those entities. Moreover, the researchers assessed and validated those classifiers with an internal cohort of 72 samples and performed rejection class analysis to predict sample class.

The study results indicated that the neural network, random forest classifiers, and support vector machine on the validation cohort obtained accuracies of 97.68%, 96.5%, and 95.62%, respectively. Furthermore, anomaly detection and threshold filtering improved the accuracy to 99.07%, 100%, and 96.22%, for the neural network, random forest classifier, and support vector machine. The neural network classifier with applied filters was used to test the in-house cohort due to its higher accuracy and more predictable cases, obtaining an accuracy of 95.45%.

In addition, comparison results of the neural network classifier with that of an established 2 marker immunohistochemistry-based (IHC) classifier showed that the earlier performed considerably better than the later on differentiating PAAD metastases from iCCA (neural network accuracy 94.28%, versus IHC accuracy 75%). Furthermore, the neural network classifier successfully excluded the samples that were not iCCA, PAAD, or normal bile.

The study also confirmed the presence of 2 important subtypes of iCCA: group 1 likely being the molecular counterpart to small-duct type iCCA, and group 2 being the molecular counterpart to large-duct type iCCA. “To the best of our knowledge, our study is by far the largest research to characterize the methylation landscape of pancreato-biliary tumours,” the authors added.

Differentiating iCCA from a hepatic metastasized PAAD remains a challenging surgical pathology diagnosis due to their similar morphological and immunohistochemical pattern with multiple driver mutations.

 A number of mechanisms have been put forward to solve this challenging issue, including the use of biopsies to differentiate iCCA and PAAD; however, due to limited tissue, the distinction is still not well understood. Recently, genome-wide DNA methylation patterns have been observed to help differentiate between classes of tumors and predict their origin using different machine-learning techniques. Nonetheless, choosing the best approach is vital.


Dragomir MP, Calina TG, Perez E, et al. DNA methylation-based classifier differentiates intrahepatic pancreato-biliary tumours. eBioMedicine. Published online June 20, 2023. doi:10.1016/j.ebiom.2023.104657.