A circulating miRNA-based AI prediction system to identify multiple types of cancer.

Abstract

3050 Background: Cancer is a major problem for human health. Development of an early diagnostic tool can increase the survival of cancer patients. Liquid biopsies have many advantages over traditional tumor tissue biopsies. Circulating microRNAs (miRNAs) are one type of liquid biopsies in part because they regulate the expression and thus functions of their target genes. Circulating miRNAs are stable, non-invasive and changes in their expression are detectable in the early stage of cancer progression, often before clear evidence of tissue biopsy/image tests. To date, there are few liquid biopsy-based tools for multiple-cancer diagnosis and their performance is unsatisfactory. The development of a non-invasive, effective early detection system for cancers is urgently needed. Methods: We integrated and investigated circulating miRNA expression data of 5046 non-cancer samples along with 3856 cancer samples of 6 major cancer types downloaded from publicly available databases. We used these expression data along with gender to establish a multiple cancer type AI prediction system. Furthermore, we built comprehensive interaction networks (miRNA-drug, miRNA-target gene) and performed functional enrichment analysis. Results: We constructed high-performance AI prediction model that can detect and differentiate 6 cancer groups from one non-cancer group. A median of sensitivity of 93.84% in test data was achieved for the multiple cancer classification task. A panel of gender and 15 most important circulating miRNAs was further shown to achieve excellent performance (sensitivity = 90.44%), with just a bit of decrease in the sensitivity of using the full set (gender and 2565 miRNAs). The 15 key circulating miRNAs worked well for the early stage (stage 1: sensitivity = 88%), much better than other liquid-biopsy results reported in the literature. This is important because these miRNAs and the AI system can be used to significantly decrease clinical cost and increase efficiency of early diagnosis, not to mention it is non-invasive. Finally, we constructed comprehensive interaction networks (drug/ target gene) for these key miRNAs to explore potential therapeutic strategies and understand the underlying biological mechanisms. Conclusions: In the study, we constructed the multiple-cancer prediction AI system to classify groups of normal individuals and cancer patients of multiple types, while finding key circulating miRNAs. As several key circulating miRNAs were shown to be potential drug targets or serve as diagnosis biomarkers to fulfill the aim of cancer precision medicine, this work represents a significant step toward achieving the goal of developing a non-invasive tool for early diagnosis of cancers.