Text-based Integration of Mutational Profiles for Classification and Biomarker Identification of Intraductal Papillary Mucinous Neoplasms of the Pancreas

Abstract

AbstractMotivationIntraductal Papillary Mucinous Neoplasms (IPMNs) are a common cystic precursor for pancreatic ductal adenocarcinoma (PDAC). Detecting these pre-malignant lesions poses a challenge for diagnostic tools due to their relatively low occurrence rate. However, a better understanding of the lesions’ composition could enable effective decision-making, risk assessment, treatment selection, and, most importantly, prevention.MethodsIn this work, we introduce a new framework for integrating information from mutational profiles using transformer-based models for stratification and biomarker identification in IPMNs vs. PDAC. We show that the numerical descriptor vectors can be used to construct highly predictive Artificial Neural Networks for disease classification. The derived mutational representations can be supported by other data types (here, mRNA) and further improve the accuracy of the classifiers. Besides the AI-driven methodology for biomarker discovery in cancer research, we also propose methods to maximize AI’s utility by recycling its knowledge to facilitate our limited understanding of the disease. We propose Natural Adversary Analysis – an AI-driven inference to detect IPMNs with a high probability of progression to malignancy.ResultsThe proposed model supports 12 clinically relevant genetic biomarkers with high mutation rates (such asKRAS, GNAS, ARID1A, andCDKN2A) and suggests biomarkers not yet recognized (such asRADIL, TTN, andZNF287). We broaden the study’s scope by investigating rarely mutated genes and reveal 14 biomarkers with potential clinical importance. Several genes with low mutation rates, includingTMPRSS1, CDH22, CCND2, CYFIP2, CBLL1, andOPCML, are also addressed as potential biomarkers by our framework. Finally, the predictive robustness of the identified biomarker set is validated externally on the patient data from the Moffitt Cancer Center study, including six pairs of matched tumor and normal IPMN samples. We show that the presented mutational profile (MP-derived) gene panel has equivalent predictive power to clinically driven panels.ConclusionsHere, we show the proof-of-concept that AI can serve the clinic and discover biomarkers beyond clinically known regimes. In line with that, we propose a translational AI-based approach for 1) disease stratification (IPMNs vs. PDAC), 2) biomarker identification, and 3) transferring the model knowledge to predict cysts’ risk of progression.