Artificial neural networks for multi-omics classifications of hepato-pancreato-biliary cancers: towards the clinical application of genetic data

Purpose: Several multi-omics classifications have been proposed for hepato-pancreato-biliary (HPB) cancers, but these classifications have not proven their role in the clinical practice and been validated in external cohorts. Patients and methods: Data from whole-exome sequencing (WES) of The Cancer Genome Atlas (TCGA) patients were used as an input for the artificial neural network (ANN) to predict the anatomical site, iClusters (cell-of-origin patterns) and molecular subtype classifications. The Ohio State University (OSU) and the International Cancer Genome Consortium (ICGC) patients with HPB cancer were included in external validation cohorts. TCGA, OSU and ICGC data were merged, and survival analyses were performed using both the ‘classic’ survival analysis and a machine learning algorithm (random survival forest). Results: Although the ANN predicting the anatomical site of the tumour (i.e. cholangiocarcinoma, hepatocellular carcinoma of the liver, pancreatic ductal adenocarcinoma) demonstrated a low accuracy in TCGA test cohort, the ANNs predicting the iClusters (cell-of-origin patterns) and molecular subtype classifications demonstrated a good accuracy of 75% and 82% in TCGA test cohort, respectively. The random survival forest analysis and Cox’ multivariable survival models demonstrated that models for HPB cancers that integrated clinical data with molecular classifications (iClusters, molecular subtypes) had an increased prognostic accuracy compared with standard staging systems. Conclusion: The analyses of genetic status (i.e. WES, gene panels) of patients with HPB cancers might predict the classifications proposed by TCGA project and help to select patients suitable to targeted therapies. The molecular classifications of HPB cancers when integrated with clinical information could improve the ability to predict the prognosis of patients with HPB cancer.