Abstract
Circular RNA is a novel class of endogenous non-coding RNAs that have been largely discovered in eukaryotic transcriptome. The circular structure arises from a non-canonical splicing process, where the donor site backsplices to an upstream acceptor site. These circular form of RNAs are conserved across species, and often show tissue or cell-specific expression. Emerging evidences have suggested its vital roles in gene regulation, which are further associated with various types of diseases. As the fundamental effort to elucidate its function and mechanism, numerous efforts have been devoted to predicting circular RNA from its primary sequence. However, statistical learning methods are constrained by the information presented with explicit features, and the existing deep learning approach falls short on fully exploring the positional information of the splice sites and their deep interaction. We present an effective and robust end-to-end framework, JEDI, for circular RNA prediction using only the nucleotide sequence. Our framework first leverages the attention mechanism to encode each junction site based on deep bidirectional recurrent neural networks and then presents the novel cross-attention layer to model deep interaction among these sites for backsplicing. Finally, JEDI is capable of not only addressing the task of circular RNA prediction but also interpreting the relationships among splice sites to discover the hotspots for backsplicing within a gene region. Experimental evaluations demonstrate that JEDI significantly outperforms several state-of-the-art approaches in circular RNA prediction on both isoform-level and gene-level. Moreover, JEDI also shows promising results on zero-shot backsplicing discovery, where none of the existing approaches can achieve.
Introduction
Circular RNAs (circRNAs) represent a significant class of non-coding RNAs characterized by a covalently closed loop structure. This structure is formed through a non-canonical splicing process known as "backsplicing," where a downstream donor site is joined to an upstream acceptor site. While circRNAs are conserved across species and play vital roles in gene regulation and disease, their identification from primary sequences remains a challenge.
Existing approaches generally fall into two categories: statistical learning methods that require heavy manual feature engineering, and generic deep learning models that often fail to capture the specific spatial relationships and deep interactions between donor and acceptor sites. JEDI (Junction Encoder with Deep Interaction) aims to bridge this gap by focusing specifically on the interaction between these sites using only the raw nucleotide sequence.
Methods
The JEDI framework is an end-to-end deep learning model designed to predict circRNA by modeling the complex interactions between potential splice sites.
The architecture consists of three primary components:
Attentive Junction Encoders
Instead of processing the entire gene sequence, JEDI extracts small "flanking regions" around potential splice sites. These regions are represented as $k$-mer embeddings and processed through Bidirectional Gated Recurrent Units (Bi-GRUs).
An attention mechanism is then applied to these hidden states to highlight the most important nucleotides near the junction, resulting in a fixed-length vector representation for each acceptor and donor site.
Cross-Attention Layer
The core innovation of JEDI is the cross-attention layer. This layer models the "deep interaction" between all potential acceptor and donor sites within a sequence. By calculating attention scores between pairs of sites, the model effectively simulates the biological competition and pairing logic inherent in the backsplicing process.
Final Prediction
The final representations are aggregated and passed through a fully connected layer to produce a binary prediction, indicating whether a circRNA exists at either the isoform or gene level.
Results and Discussion
JEDI was evaluated on human and mouse benchmarks, significantly outperforming eight state-of-the-art baseline methods, including circDeep and DeepCirCode.
Performance and Efficiency
JEDI achieved a state-of-the-art accuracy of 98.99% on human isoform-level prediction and 96.46% on the more complex gene-level task. Furthermore, JEDI is highly efficient, completing training for the isoform task in roughly 2.75 minutes, compared to over 24 hours for some baseline methods on the same hardware.
Interpretability and Zero-Shot Discovery
A major advantage of JEDI is its interpretability. Because the model uses attention weights to make decisions, researchers can identify which specific pairs of sites the model considers most likely to form a circular bond. In zero-shot tests, JEDI achieved a Mean AUC of 0.7992 for discovering specific backsplice sites without being explicitly trained on those annotations.
Biological Insights
Hyper-parameter analysis revealed that nucleotides very close to the junction (a flanking length of 4) are the most informative, as adding too much distant sequence data introduced noise. Additionally, the model's strong performance on mouse data after being trained on human data suggests it captures fundamental, evolutionarily conserved rules of RNA splicing.
Conclusion
JEDI presents a robust and interpretable framework for circular RNA prediction. By leveraging attentive junction encoders and a novel cross-attention mechanism, it effectively models the deep interactions necessary for backsplicing. While currently reliant on known exon-intron boundaries, JEDI's ability to perform zero-shot discovery and its high efficiency make it a powerful tool for transcriptome research.
Summary updated