Endocrine Abstracts

Adrenocortical tumors are common, occuring in 5-7% of the population. Adrenocortical carcinoma (ACC) is rare (0.7-2/million/year) and it has a poor prognosis with a five-year survival of less than 30% in advanced stages. The histological differentiation of benign and malignant adrenocortical tumors is challenging.

Objectives: To explore the diagnostic utility of multiple microRNAs in various combinations as markers of adrenocortical malignancy by using artificial intelligence methods, based on machine learning and neural networks.

Materials and Methods: 63 formalin-fixed, paraffin-embedded (FFPE) adrenocortical tissues were studied. The discovery cohort included 10 adrenocortical adenoma (ACA) and 10 ACC samples. An independent validation cohort encompassed another 21 ACC and 22 ACA samples. 16 microRNAs shown to be differentially expressed based on literature data were included. MicroRNA expression was studied by a 2-step TaqMan RT-qPCR. RNU48 was used as an internal, alongside with cel-miR-39 as an external control. Normalization of microRNAs was performed with the ΔCt method using R package NormqPCR. The order of microRNAs for the grouping of ACA and ACC samples was determined by the random forest classification method. The possibility of automatic classification of samples into ACA or ACC groups was tested by machine learning methods (R packages nnet and caret). Only models with more than 90% classification capability were selected for RT-qPCR validation and subsequent artificial intelligence-based classification. The best performing microRNA combinations (statistical models) were selected by neural network-based, 90-10% random learner-tester cross validation. 24 microRNA models were included in the validation performed in a blind manner.

Results: Hsa-miR-195, hsa-miR-375, hsa-miR-483_3p, hsa-miR-483_5p and hsa-miR-503 were the best 5 microRNAs revealed by random forest algorithm to correctly classify the previously unkown samples. The following three, best performing statistical models were selected out of the former microRNAs: hsa-miR-195 + hsa-miR-210 + hsa-miR-503, hsa-miR-210 + hsa-miR-375 + hsa-miR-503 and hsa-miR-210 + hsa-miR-483-5p + hsa-miR-503 with sensitivity and specificity of 90.91-90.48; 90.91-90.48 and 90.91-95.24 %, respectively. The diagnostic performance of these three models was clearly superior over that of individual microRNAs.

Conclusion: We have established three microRNA combinations with outstanding diagnostic performance using artificial intelligence-based methods. These biomarker combinations can help histological analysis, and their use in small amount preoperative biopsy samples might also be envisaged.