A machine learning framework for accurately recognizing circular RNAs for clinical decision-supporting

• Main Authors: Yidan Wang, Xuanping Zhang, Tao Wang, Jinchun Xing, Zhun Wu, Wei Li, Jiayin Wang
• Format: Article
• Language: English
• Published: BMC 2020-07-01
• Series: BMC Medical Informatics and Decision Making
• Subjects: RNA-seq data analysis; Circular RNA; Detection method; Machine learning; High precision
• Online Access: http://link.springer.com/article/10.1186/s12911-020-1117-0

Abstract Background Circular RNAs (circRNAs) are those RNA molecules that lack the poly (A) tails, which present the closed-loop structure. Recent studies emphasized that some circRNAs imply different functions from canonical transcripts, and further associated with complex diseases. Several computational methods have been developed for detecting circRNAs from RNA-seq data. However, the existing methods prefer to high sensitivity strategies, which always introduce many false positives. Thus, in clinical decision-supporting system, a comprehensive filtering approach is needed for accurately recognizing real circRNAs for decision models. Methods In this paper, we first reviewed the detection strategies of the existing methods. According to the features from RNA-seq data, we showed that any single feature (data signal) selected by the existing strategies cannot accurately distinguish a circRNA. However, we found that some combinations of those features (data signals) could be used as signatures for recognizing circRNAs. To avoid the high computational complexity of the combinational optimization problem, we present CIRCPlus2, which adopts a machine learning framework to recognize real circRNAs according to multiple data signals captured from RNA-seq data. By comparing multiple machine learning frameworks, CIRCPlus2 adopts a Gradient Boosting Decision Tree (GBDT) framework. Results Given a set of candidate circRNAs, reported by any existing detection tool(s), the features of each candidate are extracted from the aligned reads. The GBDT framework can be trained by a training dataset. By applying the selected features on the framework, the predictions on true/false positives are reported. To verify the performance of the proposed approach, we conducted several groups of experiments on both real RNA-seq datasets and a series of simulation datasets with different preset configurations. The results demonstrated that CIRCPlus2 clearly improved the specificities, while it also maintained high levels of sensitivities. Conclusions Filtering false positives is quite important in RNA-seq data analysis pipeline. Machine learning framework is suitable for solving this filtering problem. CIRCPlus2 is an efficient approach to identify the false positive circRNAs from the real ones.