Machine learning-based multiple cancer detections with circulating miRNA profiles in the blood.

Abstract

3037 Background: Early detection of cancer is one of the most important unmet clinical demands. A wide variety of circulating microRNAs (miRNAs) that specifically indicate many types of cancer have been identified, and their miRNA expression profiles are considered as potential biomarkers. Therefore, circulating miRNAs may serve as a non-invasive liquid biopsy diagnostic tool for early detection of many types of cancer. Here, a novel blood-based diagnostic method combined with machine learning techniques is developed using the entire circulating miRNA expression repertoire in serum without prior selection of miRNA marker sets. Methods: To validate this diagnostic method, clinical serum samples from cancer patients with five types of cancer (breast cancer(272), colorectal cancer(239), lung cancer(223), stomach cancer(221) and pancreatic cancer(100)) and 289 non-cancer volunteers were collected. Serum samples were immediately processed and their small RNAs were extracted. The entire miRNA expression profile is analyzed via next generation sequencers. The resulting total miRNA expression profile was used to train machine learning models, including deep learning techniques, without prior selection of miRNAs by human intervention. The machine learning model was trained with a training set to test set ratio of 4:1 and was carefully monitored by 5-fold cross-validation to avoid overfitting. Results: The diagnostic model provided 88% accuracy for all five cancer types (mean). The overall average AUROC was 0.954. Especially for breast cancer, the machine learning model provided 90% accuracy and 91 % sensitivity at 90% specificity. The overall AUROC was 0.966. High sensitivity was obtained regardless of the stage of the cancers, indicating that the possibility of early detection of cancer is kept high. Conclusions: Circulating miRNAs can be informative biomarkers for the earliest cancer detection in combination with machine learning. Unlike other cancer diagnostic methods where only a handful number of biomarkers are considered, this novel miRNA diagnostic platform method that uses machine learning reads a large set of miRNA expression profiles and automatically extracts the specific patterns of miRNA expression for early detection of multiple cancer types. In addition, the main advantage of miRNA-based cancer diagnosis is that they are more sensitive even in the early stages of cancer, compared to other diagnostic methods, such as cell-free DNA diagnostics, where the sensitivity of many types of cancer in the early stages still remains low. This approach could be easily expanded to other cancer types. Given the potential value of early detection in fatal malignancies, further validation studies are justified in future population-based studies. Many cancer research institutes are currently conducting further clinical trials to validate this early cancer diagnosis based on miRNA expression profiles.