Circular RNAs (circRNAs) are covalently closed RNA molecules generated through back-splicing. They originate from diverse genomic regions, including exonic, intronic, and intergenic loci (Conn et al., 2024). Compared to linear mRNAs, circRNAs exhibit greater exonuclease resistance, extended half-lives, and lower expression levels (Jeck et al., 2013). Due to their low abundance and circular structure, circRNAs were initially regarded as aberrant splicing byproducts or non-functional ‘splicing noise’. However, circRNAs are now recognized as ubiquitous regulators (Liu et al., 2017). Recent advances reveal their conservation across species and functional roles in metabolic pathways and disease processes in both animals and plants (Ivanov et al., 2015; Zhang and Dai, 2022). The most well-documented function involves acting as miRNA sponges to post-transcriptionally regulate gene expression (Zhang et al., 2018; Zhang X. et al., 2020; Hashemi et al., 2025; Marei et al., 2025). Additionally, circRNAs interact with RNA-binding proteins (RBPs), modulate gene expression through RNA-RNA interactions, and can undergo cap-independent translation to produce functional peptides or proteins with biological activity (Zhou et al., 2020; Hossain et al., 2023; Yi et al., 2024).

Compared to proteins translated from linear mRNAs, circRNA-derived proteins are generally shorter (typically 100–500 amino acids) yet play significant roles in development, disease progression, and tumorigenesis (Peng et al., 2021; Wei et al., 2025). For instance, Peng et al. identified the functional peptide GSPT1-238aa encoded by circGSPT1, which interacts with the vimentin/Beclin1/14-3-3 complex and regulates autophagy through the PI3K/AKT/mTOR signaling pathway in gastric cancer cells (Hu et al., 2022). Beyond disease regulation, circRNAs are increasingly recognized as a promising platform for vaccine development (Zhang et al., 2024), with most circRNA-based vaccines designed around their encoded therapeutic peptides/proteins (Chen et al., 2023). As vaccine vectors, circRNAs achieve higher expression at lower doses than linear mRNA counterparts, underscoring their therapeutic potential. Hence, Further investigation into circRNA translation functions and applications is warranted (Lokras et al., 2024).

Due to their covalently closed structure, circRNAs lack 5′ caps and 3′ poly(A) tails, resulting in a translation initiation mechanism distinct from linear mRNA (Hwang and Kim, 2024). Canonical mRNA translation employs the 5′ m⁷G cap structure and initiation factors—including eIF4E, eIF4G, and eIF4A—to recruit ribosomes (Gentry et al., 2025). In contrast, circRNAs rely cap-independent mechanisms, including: internal ribosome entry site (IRES)-mediated translation, N⁶-methyladenosine (m⁶A)-mediated initiation (also referred to as MIRES), and other forms of internal initiation (Nott et al., 2004; Yang and Wang, 2019; Zhang G. et al., 2023). These pathways employ diverse regulatory elements whose full functional repertoire remains incompletely characterized and warrants further investigation.

Advances in circRNA functional research are critically dependent on bioinformatics tools (Drula et al., 2024). Diverse computational methodologies have been applied to circRNA prediction and functional characterization, including detection tools (e.g., CIRI2) (Gao et al., 2018), sequence assembly tools (e.g., JCcirc) (Zhang J. et al., 2023), function prediction tools (e.g., iCRBP-LKHA, MSMCDA, CircDA) (Niu et al., 2024; Yuan et al., 2024; Zhang et al., 2025). For translation-specific prediction, specialized algorithms have emerged. Experimental evidence-based approaches integrate high-throughput data—such as ribosome profiling (Ribo-seq) and mass spectrometry (MS)—to validate translation events. Representative tools include MStoCIRC (Cao and Li, 2022), CircCode (Sun and Li, 2019), and CircPro(Meng et al., 2017). Complementarily, sequence-driven methods leverage machine learning and deep learning frameworks to predict translational potential from intrinsic circRNA features, exemplified by cirCodAn and CICADA (Barbosa et al., 2024; Fan et al., 2025). In conclusion, these methodologies enable a systematic and accurate assessment of circRNA translation potential.

In this review, we systematically introduce the information related to circRNA translation, as shown in Figure 1. First, we outline circRNA biogenesis and translation mechanisms, establishing the theoretical foundation and necessity for specialized analysis strategies. We then summarized experimental techniques and databases supporting computational methodology development. Building on this foundation, we systematically reviewed existing circRNA translation analysis tools and their underlying strategies, with particular emphasis on practical applicability across biological contexts. Finally, we benchmark sequence-based methods using a unified dataset and propose actionable directions for future methodology development in circRNA translation prediction.

Experimental strategies form the cornerstone of circRNA research. For studying circRNA translation, these methods are essential to determine coding capacity and elucidate initiation mechanisms. The resulting data provide critical support for developing computational methods to predict translatable circRNAs. This section describes two categories of experimental approaches for studying circRNA translation (Table 1): high-throughput screening for genome-wide identification of translated circRNAs, and targeted validation for mechanistic investigation of individual circRNAs.

Growing interest in functional circRNA analysis has established translation studies as a key focus, driving the accumulation of diverse translation evidence. Consequently, this information is now integrated into major circRNA databases, including specialized resources dedicated to circRNA translation. This outlines the core data types, features, and applications of these repositories. As cataloged in Table 2, circRNA translation-related databases are categorized as: general circRNA databases and specialized circRNA translation databases. These resources support mechanistic research while providing valuable datasets for the development, training, benchmarking, and validation of computational tools designed for circRNA translation analysis.

(i) CircRNA translation databases curate evidence related to the coding potential of circRNAs. These databases typically provide core details such as genomic loci and host genes, though they often lack comprehensive functional characterization. riboCIRC and TransCirc are two databases integrating computational predictions and experimental evidence to investigate circRNA translation (Huang et al., 2021; Li et al., 2021).

riboCIRC specializes in systematically identifying ribosome-associated circRNAs from public Ribo-seq datasets. It also predicts circRNA-specific ORFs and annotates key translation initiation elements, including IRESs and m⁶A modification sites. Furthermore, riboCIRC integrates evidence from MS and provides a manually curated catalog of experimentally validated translatable circRNAs from the literature. TransCirc aggregates seven types of evidence to infer circRNA coding potential. These include: (1) ribosome and polysome profiling data; (2) experimentally mapped translation initiation sites (TISs); (3) IRES elements; (4) m⁶A modifications; (5) circRNA-specific ORFs length; (6) machine learning-derived sequence composition scores; and (7) direct MS detection of peptides spanning back-splice junctions. By integrating the direct and indirect lines of evidence, TransCirc provides a comprehensive platform for evaluating translation capacity and putative products of human circRNAs.

(ii) Comprehensive circRNA databases systematically curate extensive information on circRNAs, encompassing basic annotations such as genomic coordinates, nucleotide sequences, and host genes, as well as functional annotations (e.g., miRNA/RBP binding sites), and coding potential assessment.

circRNADb (Chen et al., 2016) offers comprehensive circRNA annotations and evaluates coding potential through multiple approaches. It predicts IRES elements using the VIPS algorithm and identifies ORFs exceeding 300 nt. In addition, it annotates functional protein domains and post-translational modification sites using SMART and other established tools. The database also integrates known human proteins (UniProt 2013_07) with circRNA-encoded sequences and validates peptides through MS data from PRIDE. CircInteractome integrates publicly available circRNA, miRNA, and RBP datasets to facilitate the bioinformatic analyses of circRNA binding sites and their potential regulatory roles (Panda et al., 2018). In the context of circRNA translation, researchers use experimentally validated IRES sequences from the IRESite database to align with mature circRNA sequences to identify putative IRES within circRNAs. Another representative resource, C2CDB, provides basic annotations for circRNAs, including sequence and expression profiles, and incorporates functional features such as miRNA and RBP interactions, predicted coding potential, and secondary structure information (Zuo et al., 2024). To assess translational potential, circRNA sequences are annotated with translational elements based on data from IRESbase and m6A-Atlas, and putative ORFs are predicted with in-house algorithms. Moreover, C2CDB integrates publicly available translatome datasets to provide in vivo evidence of circRNA translation. circAtlas 3.0 is another large-scale resource that collected circRNA data from 33 tissue types across 10 vertebrate species (Wu et al., 2024). To explore the coding capacity of circRNAs, IRESfinder and ORFfinder are employed to identify potential IRES elements and coding sequences within circRNAs.

Computational tools for predicting circRNA translation are indispensable for advancing circRNA functional studies. In recent years, numerous bioinformatics methods have been developed to assess the translational potential of circRNAs, including both novel circRNA-specific algorithms and linear RNA predictors adapted for circRNAs. To systematically categorize these methods, we classify them into two groups: (1) Translational Regulatory Element Predictors, (2) Integrated circRNA Translation Potential Predictors.

In the previous section, we introduced various tools developed to predict the translational potential of circRNAs. While some of these tools can directly assess whether a circRNA is translated into a protein, others are limited to predicting the presence of translational regulatory features within the sequence. However, a unified benchmark for evaluating the performance of circRNA translation prediction methods is currently lacking.

circRNA translation has emerged as a frontier in RNA biology, with increasing evidence revealing its pivotal roles in cellular regulation and disease pathogenesis. Several circRNA-encoded proteins have shown significant promise as biomarkers and therapeutic targets in cancer and other major diseases (Lu et al., 2021; Wen et al., 2022). In this review, we first present a comprehensive overview of circRNA biogenesis, translation mechanisms, and relevant databases that provide the theoretical framework and data foundation for method development. Subsequently, we introduce existing circRNA translation analysis tools along with their underlying methodologies, and conduct a systematic performance comparison of various sequence-based prediction methods using a unified benchmark dataset.

Despite recent advances, key limitations constrain the accuracy and generalizability of circRNA translation prediction. Current bioinformatic tools predominantly rely on high-throughput sequencing data (e.g., Ribo-seq, mass spectrometry), which—while powerful—introduce substantial costs and analytical complexity. Simultaneously, machine learning approaches for direct sequence-based prediction face multifaceted constraints.

A limitation stems from inaccurate full-length circRNA sequences generated through short-read assembly (Huang et al., 2021; Li et al., 2021). Most entries in translational databases derive from short-read assembly, where misassembly sequences may propagate biases into downstream predictions. Long-read sequencing technologies address this limitation by capturing complete circRNA isoforms (Hou et al., 2023), enabling precise identification of BSJ-spanning ORFs and regulatory motifs essential for translation initiation.

The environmental dependency of circRNA translation further complicates predictions: translation occurs exclusively during cellular stress, differentiation, or disease states in many cases. Bulk sequencing methods ignore such heterogeneity by averaging signals across cell populations, masking rare translational events. Integrated single-cell RNA sequencing and proteomics will resolve this by detecting cell-type-specific circRNA-derived peptides with unprecedented resolution (Peng et al., 2023). In addition, low-abundance circRNA peptides often evade detection in conventional bulk MS/Ribo-seq datasets, this approach will provide high-quality training data for predictive models.

The high cost of experimental validation further limits access to high-quality, well-annotated circRNA datasets. In terms of data, in silico data augmentation offers a practical solution (Shorten et al., 2021): by exploiting the covalently closed circular topology of circRNAs, the circular sequence can be virtually broken at any position along the loop, generating multiple linear sequence variants of the same molecule. This approach artificially expands training samples for both coding and non-coding circRNAs, enhancing model robustness without incurring experimental costs. In terms of model architecture, transfer learning represents a compelling strategy to enhance model generalizability. In this framework, LLMs can be pre-trained on extensive linear RNA datasets to capture generalizable sequence and structural features, and subsequently fine-tuned on circRNA-specific tasks to exploit shared biological characteristics (Thirunavukarasu et al., 2023). Furthermore, implementing multi-task learning architectures that jointly optimize auxiliary objectives—such as ORF identification or the detection of cis-regulatory elements (e.g., IRES motifs and m6A modification sites)—can promote the learning of robust, shared representations (Yang, 2024). Such approaches have the potential to substantially improve prediction accuracy and model robustness across diverse biological contexts.

In summary, circRNA translation analysis remains challenging in several aspects, including data integration and model optimization. However, with ongoing advancements in computational approaches and experimental technologies, more comprehensive and accurate identification of circRNA translation events is anticipated, facilitating deeper insights into disease mechanisms and broader clinical applications.