As the least understood mode of alternative splicing, Intron Retention (IR) is emerging as an interesting area and has attracted more and more attention in the field of gene regulation and disease studies. Existing methods detect IR exclusively based on one or a few predefined metrics describing local or summarized characteristics of retained introns. These metrics are not able to describe the pattern of sequencing depth of intronic reads, which is an intuitive and informative characteristic of retained introns. We hypothesize that incorporating the distribution pattern of intronic reads will improve the accuracy of IR detection. Here we present DeepRetention, a novel approach for IR detection by modeling the pattern of sequencing depth of introns. Due to the lack of a gold standard dataset of IR, we first compare DeepRetention with two state-of-the-art methods, i.e. iREAD and IRFinder, on simulated RNA-seq datasets with retained introns. The results show that DeepRetention outperforms these two methods. Next, DeepRetention performs well when it is applied to third-generation long-read RNA-seq data, while IRFinder and iREAD are not applicable to detecting IR from the third-generation sequencing data. Further, we show that IRs predicted by DeepRetention are biologically meaningful on an RNA-seq dataset from Alzheimer’s Disease (AD) samples. The differential IRs are found to be significantly associated with AD based on statistical evaluation of an AD-specific functional gene network. The parent genes of differential IRs are enriched in AD-related functions. In summary, DeepRetention detects IR from a new angle of view, providing a valuable tool for IR analysis.

Intron Retention (IR) is an alternative splicing mode through which introns are retained in mature RNAs rather than being spliced in most cases. IR has been gaining increasing attention in recent years because of its recognized association with gene expression regulation and complex diseases. Continuous efforts have been dedicated to the development of IR detection methods. These methods differ in their metrics to quantify retention propensity, performance to detect IR events, functional enrichment of detected IRs, and computational speed. A systematic experimental comparison would be valuable to the selection and use of existing methods. In this work, we conduct an experimental comparison of existing IR detection methods. Considering the unavailability of a gold standard dataset of intron retention, we compare the IR detection performance on simulation datasets. Then, we compare the IR detection results with real RNA-Seq data. We also describe the use of differential analysis methods to identify disease-associated IRs and compare differential IRs along with their Gene Ontology enrichment, which is illustrated on an Alzheimer’s disease RNA-Seq dataset. We discuss key principles and features of existing approaches and outline their differences. This systematic analysis provides helpful guidance for interrogating transcriptomic data from the point of view of IR.