IntroductionCircular RNAs (circRNAs) are a class of covalently closed single-stranded RNAs generated in cells mainly through back-splicing. Although discovered more than 30 years ago, only recently has it been recognized that circRNAs are abundant and evolutionarily conserved among eukaryotes with cell-specific expression patterns1,2. The circRNAs were thought to function as noncoding RNAs by regulating different steps of gene expression, including transcription, splicing, mRNA stability, translation, and signaling pathways through interactions with their target DNA, RNA, and proteins3,4,5. However, certain circRNAs can also be translated into functional proteins6,7,8,9,10,11, showing promise as a new type of mRNA therapeutics due to their higher cellular stability12,13. In addition to the superior stability, the properly purified circRNAs have lower immunogenicity even without base modification14,15,16,17, providing a wider expression window for therapeutic proteins as the next generation of RNA therapy18,19.Canonical translation of linear mRNA is initiated with the recognition of the m7G cap structure by eukaryotic initiation factor 4E (eIF4E) that is in synergy with eIF4G and eIF4A, forming the eIF4F complex and further recruiting other translation initiation factors and the ribosome20. Under stress conditions (e.g., hypoxia, apoptosis, starvation, and viral infection), the canonical translation is downregulated, while the proteins essential for cell survival and stress recovery are selectively translated21. A major mechanism involved in this noncanonical translation is through the use of IRES, which are often assisted by certain RBPs known as IRES trans-acting factors (ITAFs), in a cap-independent manner22. The IRES-mediated translation has also been observed in naturally occurring10,23,24 and synthetically designed circRNAs6,7,11,25,26. However, the efficiency of IRES-driven translation is generally low27, posing a challenge for leveraging circRNAs as therapeutic tools. In addition, the activity of IRES is highly dependent on the expression and activity of specific ITAFs, resulting in cell-type-specific regulation of IRES-mediated translation22.Despite their biomedical promise, the trans-acting factors that enhance cap-independent circRNA translation (i.e., functional equivalents of ITAFs) have not been thoroughly characterized28. Since certain RBPs, such as PABPC1 and IGF2BP1, can promote linear mRNA translation29,30, identifying trans-acting factors that specifically regulate circRNA translation is critical and could pave the way for the development of novel translation manipulation tools and therapeutic applications involving circRNAs. To study trans-acting factors that regulate circRNA translation, we employed a luciferase-based circRNA reporter system coupled with a tethered RBP library to systematically identify putative proteins that can promote circRNA translation (termed translation activators herein). Remarkably, we found that a large number of RBPs tested could facilitate noncanonical translation when tethered to the circRNAs. By engineering the cognate binding motifs of these activators into circRNA, endogenous factors can be recruited to drive circRNA translation. Moreover, we further developed a machine learning (ML) model using experimentally tested proteins as the training set to predict additional translation activators from all known RBPs. Both experimentally identified translation activators and predicted activators were shown to stimulate IRES-mediated circRNA translation, suggesting they may function as novel ITAFs for specific IRES elements. Overall, our study extended the scope of current knowledge on translation regulatory RBPs and uncovered an inherent connection between RBP sequence and their translation-promoting activity, which may hold promise for applications in RNA-based therapeutics.ResultsEngineer a reporter system to assess RBP activities in regulating circRNA translationWe have previously developed a circEGFP reporter system to identify regulatory cis-elements that drive cap-independent translation of circRNAs7,9,31. Building upon a similar design principle, we engineered a luminescence-based system that allows specific detection of translation products from circRNAs vs. linear mRNAs. In this system, the coding sequence (CDS) of firefly luciferase (Fluc) was split into two fragments in reverse order, flanked by two introns with complementary elements. Through back-splicing, it generates a circular Fluc (cFluc) RNA with the intact open reading frame (ORF) that can be translated into Fluc protein (Fig. 1A). As a control, we incorporated the Renilla luciferase (Rluc) gene into an inverse orientation, which can be used to calibrate for the effect of cell transfection and gene expression from linear mRNA (Fig. 1A).Fig. 1: Design and validation of circular reporter system to examine RBP activities.Full size imageA Schematic of the circular luciferase reporter. The Fluc gene was split into two reverse-ordered fragments and placed in a single exon, allowing back-splicing into a circular RNA via the complementary sequences in the flanking introns. Two tandem PUF binding sites were inserted between the start and stop codons of cFluc for tethering artificial PUF factors. An Rluc gene was included as a transfection control. B Experimental and data analysis workflow. Luciferase activity and RNA levels were measured from the same sample to assess back-splicing and translation efficiency. Measurements were normalized to the values obtained from the PUF-only control. C Validation of the reporter system by tethering two translation initiation factors, eIF4E and eIF4G2. Left panel, the schematic of the design of the indicated constructs. Right panel, translation efficiency of indicated constructs on the cFluc reporter are determined, using the luciferase assay described in (B). Data are presented as mean ± SD of n = 3 biologically independent transfections. P- values were calculated by one-way ANOVA and Dunnett’s test.Next, we used a modified Pumilio and FBF homology (PUF) domain that specifically binds an 8-nucleotide (nt) RNA sequence (UUGAUAUA) to tether candidate RBPs onto the circRNA32,33. The PUF-RBPs were co-transfected with the circular RNA reporter in human embryonic kidney 293 T (HEK 293 T) cells and analyzed at 36 h post-transfection (Methods). As a control, we utilized a PUF domain fused only with a FLAG tag (PUF-only). We quantified the luciferase activities and RNA levels of cFluc and Rluc, and normalized to the samples co-expressed with the PUF-only control (Fig. 1B).Validation with known translation initiation factors eIF4E and eIF4G2 confirmed the system’s functionality. eIF4E is a canonical cap-binding protein in synergy with eIF4A and eIF4G, recruiting other initiation factors to activate translation20 while eIF4G2 is a reported IRES-dependent or m6A-driven translation factor34. We fused the ORFs of eIF4E and eIF4G2 with the FLAG epitope and the PUF binding domain at the N terminus (Fig. 1C, left). As expected, tethering eIF4E and eIF4G2 to upstream of the circRNA start codon both enhanced the translation efficiency (TE) of circRNA, whereas eIF4E/eIF4G2 without PUF did not activate the expression of cFluc (Fig. 1C, right), suggesting that direct tethering through the PUF domain is essential for RBPs’ activity. Although eIF4E has not been reported to promote circRNA translation, our findings suggest that the recruitment of eIF4E by an artificial system can promote circRNA translation, probably via the association of eIF4E with the translation initiation apparatus. Collectively, this tethering system may serve as a robust tool for identifying trans-acting factors that promote circRNA translation.Systematic screen of RBPs that affect circRNA translationPrevious study demonstrated that a large number (414) of mouse proteins are associated with ribosomes and functionally enriched in the ribosome assembly process, and thus may play a role in regulating translation initiation35. We compared them with all identified human RBPs (N = 1542) and obtained 261 ribosome-associated RBPs (Fig. 2A). This list was further refined through manual curation based on literature, removing most core ribosomal components, nuclear proteins, transporters, or proteins with well-identified functions in other RNA regulatory pathways, resulting in the final list of 110 putative RBPs (Fig. 2A and “Methods”). We hypothesized that these proteins may have the potential to modulate cap-independent circRNA translation. Gene Ontology (GO) analysis of the selected 110 RBPs revealed a significant enrichment in core pathways related to RNA splicing, translation initiation, and ribosomal biogenesis (Supplementary Fig. 1A). Moreover, most of these RBPs are highly expressed as judged by RNA-seq and ribosome profiling data36 (Supplementary Fig. 1B), consistent with the notion that high-abundance proteins often sustain broad cellular functions such as translation.Fig. 2: Screen for RBPs that promote circRNA translation.Full size imageA Selection of 110 putative RBPs and construction of the PUF-RBP fusion protein library. B Hierarchical clustering of luciferase activity with Ward’s minimum variance method. CircRNA amount, back-splicing efficiency and translation efficiency (TE) for the 110 RBPs were calculated according to the workflow in Fig. 1B (n = 3 biologically independent transfections). The SD values for all data points are provided in Supplementary Data 2. C Correlation between luciferase activity and circRNA amount for 110 RBPs tested. The mean ± SD is shown for all 110 RBPs (n = 3 biologically independent transfections). Pearson correlation coefficients (R) are indicated, with p = 1.7 × 10−15. D Correlation between translation efficiency and circRNA amount for 110 RBPs tested. The mean ± SD is shown for all 110 RBPs (n = 3 biologically independent transfections). Pearson correlation coefficients (R) are indicated, with p = 3.0 × 10−4. E Scatter plot of TE values for all the tested RBPs. The RBPs with TE ≥ 1.3 were identified as translation activators. F Functional network of the identified activators. The activators were analyzed using the STRING database and clustered into three main groups by the CytoScape MCODE tool.Next, full-length CDS of the 110 candidate RBPs were cloned into the PUF tethering constructs (Fig. 2A) and transfected in triplicate into HEK 293 T cells, with the PUF-only control being used for normalization. All 110 RBPs were included in the screening, with expression of each fusion protein verified by Western blot (Supplementary Fig. 1C). Since PUF-RBPs may affect cFluc signals by regulating back-splicing efficiency and/or cap-independent translation, we quantified both linear and circular mRNA levels, as well as Fluc and Rluc activities, to dissect their effects on both processes (Fig. 1B). Subsequently, the translation efficiency, luciferase activity, splicing efficiency, and circRNA amount affected by PUF-RBPs were clustered and plotted using a heatmap (Fig. 2B). Interestingly, we observed a general positive correlation between circRNA amount and luciferase activity (R = 0.67, Fig. 2C) and an even stronger correlation with the back-splicing efficiency (R = 0.79, Supplementary Fig. 2A). However, such positive correlation disappeared when we compared the circRNA amount to the TE after normalization by considering the effect from back-splicing (Fig. 2D). These results necessitated the correction for the effect of circRNA biogenesis (i.e., back-splicing) when calculating TE (Fig. 1B). As expected, a subset of RBPs were found to affect circRNA back-splicing, leading to changes in circRNA levels (Supplementary Fig. 2B).Characteristics of identified RBPs that activate circRNA translationWe further focused on RBPs that increased relative translation efficiency (TE) of circRNAs by at least 30% (i.e., using a relative TE cutoff of 1.3). Of the 110 RBPs examined, 68 were identified as translation activators (Fig. 2E). We first investigated the underlying biological functions of these candidate activators using GO analysis, and found strong functional enrichments in translation initiation (e.g., eukaryotic initiation factors and ribosomal proteins), RNA processing (e.g., hnRNPs) and ribosomal biogenesis (e.g., DDX protein family) (Fig. 2F). These functional enrichments were consistent with our screening design and objectives, which further reassured the reliability of our findings.Notably, many identified translation activators, such as hnRNPUL1, RPS2, IGF2BP3 and SRSF1, were previously reported to bind IRES-like short elements in a circular GFP reporter system11. In addition, hnRNPM was reported as an ITAF to induce translational switch under hypoxia to promote colon cancer development37. Interestingly, we identified more than 10 members from the DEAD-box helicase family as translation activators (Fig. 2F), some of which had been previously shown to possess ITAF activities. For example, DDX3X was documented to engage with a number of 40S ribosomal components and facilitate IRES-driven translation in FMDV38. In addition, eIF4A3, a DEAD‑box helicase that constitutes a key subunit of the exon–junction complex (EJC), was recently reported to facilitate circRNA translation by directly binding to eIF3G and facilitating internal ribosomal entry39.To examine whether these identified activators of cap-independent translation exhibit distinct functions in linear vs. circular mRNAs, we utilized a standard bicistronic reporter40. We introduced two consecutive PUF binding motifs between the start codon of linear Fluc and the stop codon of linear Rluc in the reporter construct (Supplementary Fig. 3A). A total of 15 tethered RBPs with a range of TE values were tested by co-transfecting with the reporter into HEK 293 T cells. The results revealed a strong positive correlation between the translation enhancing activities of RBPs measured by this bicistronic reporter system and their activities observed in the circRNA context (Supplementary Fig. 3B), suggesting potential general roles of certain RBPs in promoting cap-independent translation.Given the significance of RBP intracellular localization in modulating RNA metabolism41, we clustered our translation activators based on their subcellular localization information from the UniProt database. This analysis uncovered that the majority of the 68 translation activators locate in both the nucleus and cytoplasm (Supplementary Fig. 3C), consistent with previous findings that most known ITAFs regulate IRES-dependent translation by shuttling between the two cellular compartments22. Interestingly, of the 68 translation activators, 27 were associated with stress granules that contain many specific translation initiation factors, while 12 were mapped to P bodies that are enriched with factors in mRNA degradation and decay42. Furthermore, quantification of relative levels of cytoplasmic circRNAs revealed no significant change upon expression of RBP‑fusion proteins (Supplementary Fig. 3D), suggesting that the observed changes in circRNA translation may not be due to indirect effects of altered RNA localization.Translation activators exhibit ITAF-Like features and functionTo our knowledge, only a limited number of trans-acting factors were reported to drive circRNA translation11,39. As described in the introduction, the translation activators that we identified to enhance cap-independent circRNA translation were functionally equivalent to ITAFs. Accordingly, we surveyed the literature and found that approximately 55 ITAFs have been documented to modulate cellular IRESes under different physiological conditions, predominantly functioning as activators22. In light of this, we compared the newly identified activators with the 55 reported cellular ITAFs (Supplementary Fig. 4A). Among the 68 identified activators, 14 were previously reported as ITAFs (eIF4A1, eIF4G1, eIF4G2, eIF5B, FUS, hnRNPK, hnRNPM, NCL, PA2G4, PTBP1, RACK1, RPL11, RPL26, RPL38), with the remaining 54 representing novel translation activators (Fig. 3A). Notably, many members of the La-related protein family (LARP1, LARP4, LARP4B, LARP7) and heterogeneous nuclear ribonucleoproteins (hnRNPA3, hnRNPF, hnRNPH1, hnRNPU, hnRNPUL1) emerged as potential ITAFs to facilitate circRNA translation.Fig. 3: Activators can enhance circRNA translation through binding to endogenous RNA targets.Full size imageA Venn diagram showing overlap between identified activators and reported ITAFs, with 14 shared proteins, 54 unique to activators, and 33 unique to reported ITAFs. B Peak distribution of reported ITAFs and identified activators based on eCLIP data in HepG2 cells, with yellow, dark blue, and pink blocks representing 5’UTR, CDS, and 3’UTR, respectively. C Counts of IRESes containing eCLIP peaks of different RBPs within a 150-nt window upstream and downstream of each IRES in HepG2 cells. D Diagram of three representative LARP7 binding peaks on UBAP2, HSPA8, and TMEM107 transcripts, based on eCLIP data from HepG2 cells. E RIP-qPCR assay compared the enrichment of immunoprecipitated target transcripts under overexpression of Flag-tagged LARP7 versus mock control. F RIP-qPCR assay demonstrated the enrichment of circFluc expression in reporters containing indicated LARP7 binding motifs versus the non-targeting control (NC). G Assessment of translation enhancement of LARP7 on circFluc containing the indicated binding motif. TE values upon LARP7 overexpression were normalized to the mock control. H RIP-qPCR assay detecting enrichment of CVB3-circFluc containing the indicated LARP7 binding motifs versus the non‑targeting control (NC). I Measurement of translation efficiency of Fluc generated from CVB3-circFluc reporter containing indicated LARP7 binding sites, relative to the non-targeting control (NC). J Assessment of translation efficiency of Fluc generated from CVB3-circFluc reporter containing indicated LARP7 binding sites upon LARP7 knockdown. Left panel, confirmation of LARP7 knockdown by western blot and qPCR. Right panel, relative change of TE upon LARP7 knockdown. The TE values for all groups were normalized to the non-targeting scrambled shRNA control. In Figure (E–J), data are presented as mean ± SD of n = 3 biologically independent transfections. P-values are calculated by a two-sided Student’s t test followed by Bonferroni correction (adjusted α = 0.05).To gain insight into how these newly identified translation activators may interact with endogenous mRNA, we analyzed the RNA-binding pattern of both identified activators and reported ITAFs using enhanced cross-linking immunoprecipitation (eCLIP) data generated by the ENCODE consortium in HepG2 and K562 cells43. We first employed peak density as a metric to categorize RBPs by determining their binding preferences across the 5’ untranslated region (5’UTR), CDS and 3’ untranslated region (3’UTR) region of annotated transcripts. The majority of known ITAFs and newly identified activators bound primarily to the 5’UTR and 3’UTR. By contrast, DDX3X, eIF3D, EFTUD2, and U2AF1 showed binding peaks in the CDS as well as the 5’UTR in HepG2 cells (Fig. 3B). The binding pattern was slightly different in the K562 cell line, likely attributable to the influences of cell specificity and variations in eCLIP data quality (Supplementary Fig. 4B).To further explore the potential involvement of these translation activators in the IRES-dependent translation, we examined the correlation between eCLIP peaks of different activators and the known IRES regions using the 693 human IRES sequences from IRESbase44. We counted the number of IRESes that contain nearby eCLIP peaks (within a 150 nt window) of different RBPs. Among these activators, DDX3X and LARP4 were ranked as the top two RBPs with overlapped IRES counts (Fig. 3C). DDX3X was reported to function as ITAF to promote IRES-dependent translation45. Interestingly, the potential roles of La-related proteins (LARPs, identified as new translation activators in our study) in regulating cap-independent translation have not been previously investigated in detail.Since the main objective of our study is to advance circRNA applications, we hypothesize that incorporating predefined binding sites for known translation activators during circRNA design may recruit the endogenous factors to enhance IRES-dependent circRNA translation. To test this hypothesis, we focused on LARP7 and inserted its binding motif into the circRNA reporter. LARP7 was selected because it was among the most effective RBPs in promoting translation efficiency in our screen, and also because other members of the LARP family are well-documented regulators of RNA metabolism, including translation, making LARP7 an attractive candidate for further investigation.We then sought to validate the functional relevance of the identified LARP7 binding sites. Three prominent eCLIP peaks located in the 5′UTRs of UBAP2 (Ubiquitin Associated Protein 2), 3’UTR of HSPA8 (Heat shock 70 kDa protein 8) and TMEM107 (Transmembrane Protein 107) (Fig. 3D). These genes were selected for experimental confirmation with RIP-qPCR, confirming that LARP7 directly interacts with these endogenous target RNAs (Fig. 3E).To assess the translational impact of LARP7, each validated binding site was cloned into a cFluc reporter and co-transfected with a FLAG-LARP7 expression plasmid. The RIP-qPCR assay confirmed that circular Fluc with UBAP2- and HSPA8-derived binding sequence can coimmunoprecipitate with LARP7 at significantly higher levels (Fig. 3F). Overexpression of LARP7 significantly increased the TE of reporters containing the UBAP2-, HSPA8-, or TMEM107-derived binding sequences, compared to a mock control (Fig. 3G). These results indicate that the recruitment of LARP7 to specific RNA motifs is sufficient to enhance translation, supporting its potential role in promoting cap-independent translation.In practical applications, an IRES is usually incorporated into circular RNA to improve protein yield. To further test whether incorporating LARP7 binding motifs can recruit endogenous LARP7 and enhance an IRES-directed circRNA translation, we introduced a frequently used IRES, the coxsackievirus B3 (CVB3), into the circRNA reporter with UBAP2- and HSPA8-derived binding sequence. Similar to the results of the previous reporter, RIP-qPCR confirmed that circular Fluc with CVB3-IRES and LARP7 binding motifs also coimmunoprecipitated with LARP7 (Fig. 3H). As expected, the incorporation of LARP7 binding motifs significantly enhanced translation of CVB3-IRES reporter (Fig. 3I). Furthermore, we constructed two independent LARP7 knockdown cell lines (LARP7-sh1 and LARP7-sh2; Fig. 3J, left) to determine whether translation of this circRNA reporter is dependent on LARP7. Luciferase assays showed that knockdown of endogenous LARP7 specifically reduced translation of circFluc reporters containing LARP7-binding sites compared with the scrambled shRNA control (Fig. 3J, right). Together, these results indicate that motif-directed recruitment of LARP7 can stimulate IRES-dependent circRNA translation.Functional modules in the newly identified translation activatorsEncouraged by our previous investigation, we sought to conduct a systematic dissection of the functional domains within these translation activators. Many RBPs have modular configurations, with the RNA-binding domains to specifically recognize their targets and the functional domains to manipulate RNA metabolism46,47. Notably, specific functional domains or unstructured fragments of RBPs can mimic the regulatory effects of the intact protein on regulating alternative splicing48,49, back-splicing50 and translation51,52 when directed to their RNA targets. Building on these insights, we aimed to identify the particular domains or fragments of the activators that serve as functional modules to govern the translation of circRNAs.We initially analyzed the annotated domains of the identified translation activators (Supplementary Fig. 5A) and found many of them contain RNA helicase domains in addition to canonical RNA-binding domains (such as RRM and KH domains). However, the majority of the translation activators do not have well-defined domains that may serve as functional modules. Based on domains and region annotations from the UniProt database, we selected several representative activators and tested the activity of their putative functional truncations using the tethering assay (Fig. 4A). We first sought to examine activators with well-characterized domains, including helicase domain containing proteins (DDX21, eIF4A1 and SKIV2L2) and RRM containing proteins (FUS, RBM28 and LARP7). For translation activators without well-defined domains, such as eIF3D, eIF5B and SRRM1, we took intrinsically disordered regions (IDRs) into consideration. Additionally, we examined the translation regulatory patterns for NMD3, which contains a specific region essential for the nuclear export of the 60S ribosomal subunit53,54. hnRNPUL1 and PES1 were also selected in our study due to their high TE values (> 2) in our tethering assay. In total, the N-terminal and C-terminal truncations of 12 distinct translation activators were fused to the PUF domain, using the same backbone and conditions as our previously established PUF-RBPs library. These constructs were co-transfected with the cFluc reporter into the HEK 293 T cell line to evaluate their TE values independently.Fig. 4: Domain analysis of identified activators reveals sequence similarity.Full size imageA Diagram of Domain configurations of tested activators. The split position between N-terminal (yellow) and C-terminal (blue) fragments is indicated. B TE values obtained for full-length protein (gray), N-terminal truncations (yellow), and C-terminal truncations (dark blue) of the 12 activators upon tethering in HEK 293 T. Data are presented as mean ± SD of n = 3 biologically independent transfections. The dashed line indicates the TE cutoff of 1.3. C Multiple amino acid sequence alignment of 110 RBPs using ClustalW, followed by clustering with MEGA X (neighbor-joining method). Red dots represent activators (TE ≥ 1.3), and gray dots represent others (TE 1.3; Fig. 6D). Distinct patterns emerged for the group II IRESes: only DDX21 and DEK significantly enhanced translation of the HCV IRES (Fig. 6E), while most RBPs activated the CSFV IRES (Fig. 6F). Notably, the intended positive control eIF4G2 stimulated only the CVB3 and EMCV IRESes (groups IV and III), highlighting the context specificity of RBP-mediated translation activation. Thus, translation activators identified via tethering screening or ML can promote IRES-driven circRNA translation, but their efficacy is strongly dependent on the specific IRES.DiscussionOptimizing translation efficiency is crucial for leveraging circRNAs as a next-generation RNA therapeutic platform. Previous studies have primarily focused on the design and refinement of cis-elements (e.g., IRES, UTRs)11,25,26, while investigations into the role of trans-acting regulatory proteins have been relatively limited, although they play critical roles in circRNA translation9,28,61. In this study, we systematically investigated potential RBPs that can enhance cap-independent circRNA translation by screening an RBP tethering library. This work represents a comprehensive characterization and predictive identification of potential ITAFs for circRNA translation, significantly expanding the scope of regulatory factors in cap-independent translation.Our study used a tethering system that relies on localized recruitment of RBPs to the circRNA rather than global protein abundance, thus even a moderate expression of functional RBPs is usually sufficient to detect translational activity. There is a commonly accepted limitation in such a screen, where normalizing the expression levels across hundreds of proteins is technically challenging, and thus it is standard practice to use consistent plasmid amounts for all RBP constructs in this type of work32,49,62. As expected, we noted the variability in protein expression in large-scale screens involving diverse RBPs with different sizes, stability, and translation efficiency (Supplementary Fig. 1C). However, we did not observe any correlation between the protein expression levels and their activity in regulating circRNA translation, and even moderate expression of some RBPs (such as EIF5B, DDX24, or DDX46, Supplementary Data 2) was sufficient to activate reporter circRNA translation, suggesting this system can tolerate a certain degree of expression variability.Despite the limitation of the tethering system, it also has the benefit as a convenient approach for the initial identification of candidate factors that may function as ITAFs to promote IRES-dependent translation in cells. We further tested the activity of these putative ITAFs in the context of circRNAs containing different IRESes, and found that their activity is dependent on the specific IRES. Such IRES-specific regulation is a fundamental feature of cap-independent translation, and our study provides a roadmap for leveraging this specificity to optimize circRNA therapeutics. While the tethering screen is an artificial system to examine IRES-ITAF compatibility, the results may offer a design principle to optimize circRNA translation by choosing matched IRES-ITAF pairs or by inserting cognate motifs of certain ITAFs. Additionally, such observations also reflect the extra layer of regulatory complexity in cap-independent translation.Compared to the tethering system using the MS2 coat protein, our strategy using a PUF domain offers more flexibility due to the programmable RNA-binding specificity of PUF, as it can be engineered to recognize endogenous sequences without inserting multiple MS2 hairpins into the target. Previously, engineered PUF domains have been combined with various effector domains to manipulate different RNA processing pathways63, including mRNA splicing33 and translation64. In addition, Cao et al. engineered a light-inducible system to activate translation using the PUF scaffold with eIF4E65. Our work has generated various artificial PUF factors with different functional domains, which may be useful in specifically modulating circRNA translation. Further optimization could adapt the PUF domain to directly target cis-elements within endogenous circRNAs without changing their sequences. Moreover, the PUF domain may be further modified by including additional repeats to recognize longer RNA sequences (e.g., 10–12 nt), which can effectively increase its specificity and reduce off-target effects for in vivo applications66,67.Our screen was carried out with an IRES-free system in order to identify all the general activators for cap-independent translation. Subsequently, these newly identified activators were examined for their activity in promoting IRES-driven translation (Fig. 6), and the results showed that the ITAF activities of these activators were generally dependent on the specific IRES context. This result is consistent with earlier findings from several ITAF examples, suggesting that the context-dependent activity of ITAFs seems to be a general feature of IRES-driven translation. This variability may be due to the structural diversity of different IRES elements and their distinct requirements for translation initiation factors. Although the IRES-free reporter system may not be directly used in most applications, it will provide a useful framework in the design of circRNAs. Since many of these activators (RBPs) have cell-specific expression, we could select appropriate combinations of IRES-ITAFs based on the IRES types and expression profiles of the translation activator, achieving more efficient and specific translation of circRNAs. Such information may also advance the field of circRNA research by offering new insights into the functional interactions between different IRESes and their trans-acting factors.We found that 68 out of the 110 tested RBPs demonstrated the capacity to enhance the translation efficiency of circRNAs (Fig. 2E). This seemingly high positive rate is probably because we pre-selected RBPs that were found to bind ribosomes. In addition, we found that RPL5, RPL18, and other ribosomal proteins may function as potential circRNA translation activators (Fig. 2F). The regulatory role of these core ribosomal proteins may partially be explained by the specialized ribosomes hypothesis, where the heterogeneity in ribosomal protein composition endows ribosomes with differential selectivity for translating specific subpools of transcripts, thereby altering the translation of particular mRNAs68,69. In line with this, many ribosomal proteins have been reported to act as regulators in the selective translation of viral transcripts in an IRES-dependent manner70. Our findings support this hypothesis and expand its implications to circRNA biology.We also noted that 7 out of 68 activators (RBM5, FUS, U2AF1, TRMT1L, HNRNPK, RBM17, and PTBP1) with luciferase activity below 0.7. All of these activators exhibited splicing efficiency below 0.5 (Fig. 2). Consequently, the elevated TE values were obtained when we calculated the TE values from the relative ratio of luciferase activity to circRNA amount. Notably, all 7 proteins are involved in RNA splicing regulation. Thus, we speculate that these proteins may have multifunctional roles, acting as potential negative regulators for back-splicing while serving as positive regulators for circRNA translation. Although these factors may not increase the translation output of endogenous circRNAs that are generated from back-splicing, we speculate that they may still have practical applications in the context of in vitro–circularized RNAs, where circRNA abundance remains constant without requiring a back-splicing process. In such cases, these trans-acting factors could potentially enhance the translation output of in vitro synthesized circRNA drugs.While almost all known ITAFs were studied in the context of linear mRNA, our results provide evidence that a subset of RBPs can enhance cap-independent translation for both circRNA and linear mRNA (Supplementary Figs. 3A, B). One possible explanation is that these proteins may play multifunctional roles by recruiting ribosomes and core translation initiation factors in synergy with IRES elements under cellular stress. It’s also worth noting that, even for linear mRNAs, translation predominantly occurs in a circularized form via head-to-tail RNA looping mediated by the interaction between cap-binding proteins and eIF4G71. Therefore, the study on circRNA translation may also shed light on the regulation of linear mRNA. Moreover, elucidating the precise mechanisms by which RBPs regulate translation, and how to distinguish the IRES and ITAF functions in circRNA vs. linear RNA, necessitates further exploration.Our results also showed that certain fragments of the RBPs can promote circRNA translation when tethered to circRNAs (Fig. 4B). This discovery has two important implications. First, it suggests that truncated proteins retain the ability to recruit translation machinery and promote translation initiation for circRNAs. Second, our approach offers a safer and more practical way to control translation with engineered proteins in mammalian cells, as the overexpression of a full-length RBP might interfere with certain cellular processes by binding to its endogenous targets. Interestingly, we found that the N-terminal IDR fragment of eIF5B may have independent activities in promoting translation. This observation aligns with in vitro findings that the disordered N-terminal region of eIF5B can stimulate the IRES activities in endogenous human genes by facilitating the formation of cytoplasmic granules72. Importantly, activation domains within RBPs have been extensively studied in other critical gene regulation processes, including transcription73,74,75 and alternative splicing48,49. Given the conceptual similarities among RNA regulatory processes, we anticipate that future screenings will be developed to identify additional translation activation domains.We further developed an ML approach to predict RBP function in controlling translation directly from their amino acid sequences. AI-based approaches like AlphaFold have made great advances in predicting protein functions from specific structures76,77. However, predicting the functions of unstructured peptides remains challenging78,79. Our approach differs from the conventional functional predictions in that it predicts the activities of proteins or protein domains without clearly defined structures. In fact, we found that some IDRs (e.g., those of SRRM1 and eIF5B) are sufficient to promote noncanonical translation (Fig. 4B). These poorly defined unstructured regions may still contribute critically to translation activation through dynamic interactions with RNA or the translation machinery. This interpretation is further supported by our sequence-based ML analysis, which shows that amino acid composition offers greater predictive value than a conserved structural fold for predicting translation activator activity. Consistent with this idea, the sequence composition was also found to play a critical role in determining the RBP activity in regulating alternative splicing32. We speculate that the clustering of specific residues may facilitate the recruitment of specific binding partners, leading to the formation of protein complexes that regulate translation. A limitation of our ML model is the risk of false positive predictions, and the validation was restricted to a few RBPs. Future studies should explore additional RBPs involved in circRNA translation regulation, which may also consider additional parameters in refining the model. Despite this limitation, our study offers the possibility to predict translation activators based on their sequence compositions, expanding the scope of translation regulatory RBPs and allowing for the engineering of novel activators.In summary, we conducted a systematic investigation with a tethering assay to identify and characterize the trans-acting factors that promote the cap-independent translation of circRNA. A limitation of the current study is that most analyses were conducted with synthetic reporters; therefore, future validations using endogenous circRNAs with more mechanistic details remain an important direction. In addition to expanding current knowledge on the regulation of circRNA translation, this study may contribute to the application of circRNAs as a platform technology in RNA drug development. The specific expression profiles of these translation activators may be used in controlling IRES-mediated translation of circRNAs under different cellular contexts. For example, the context-dependent translation activation by binding of endogenous LARP7 to its cognate motifs could offer a strategy to enhance circRNA translation without overexpression of exogenous factors, thereby improving therapeutic safety and efficacy. Another application relies on the usage of expression preference of translation activators in different cell types and their cognate binding sites to engineer IRESs that can be switched on in certain cells, which is especially valuable in designing circRNAs with T-cell specific translation for cancer therapy. We anticipate that this line of research will offer a viable strategy for manipulating circRNA translation, thereby enhancing their applications in RNA therapy.MethodsConstruction of circular luciferase reportersTo generate circular RNA reporters, the firefly luciferase gene was split into two fragments in reverse order, consistent with the split cFluc plasmid11. Complementary elements flanked by introns were included to facilitate back-splicing by base pairing80. Cognate binding sites of the PUF domain were inserted into the middle of reverted luciferase fragments flanked by restriction endonuclease sites (EcoRI, NEB, catalog Number: R3101V; SpeI, NEB, catalog Number: R3133V). The construct was verified by Sanger sequencing, and primer information was listed in Supplementary Data 1.Construction of RBP tethering libraryTo generate different tethered RBPs, we first modified the expression construct for PUF-RBP fusion protein by replacing the fragment encoding the SR domain with a multiple cloning site and GS linker sequence that can be used for different functional domains32.To generate PUF tethered constructs for full-length RBPs and C-terminal/N-terminal fragments (i.e., effector domains), we used primers containing suitable restriction sites (XbaI, NEB, catalog number: R0145V; NotI NEB, catalog number: R3189V) and human cDNAs for amplification (SuperScript III, Invitrogen, catalog number: 18080044). The amplified DNA fragments were then cloned into pGL-PUF backbone, flanking the corresponding restriction endonucleases by homologous recombination (ClonExpress@ Entry One Step Cloning Kit, Vazyme, catalog number: C112-01). A glycine-rich linker was inserted between the effector protein and PUF to enhance the conformational flexibility and stable expression of the fusion proteins81. The sequences of all cDNA clones were verified, and all primers and relevant information were included in Supplementary Data 1.Cell culture and transfectionHEK 293 T cells (ATCC, CRL-3216) were cultured in DMEM medium supplemented with 10% FBS. One day before transfection, HEK 293 T cells were seeded into 24-well plates, and the transfection density was controlled at 35-45%. For each transfection, 1.1 μL Lipofectamine 3000 (Invitrogen, catalog number: L3000075) and 0.72 μg plasmids (comprising 120 ng of reporter gene plasmid and 600 ng of PUF-RBP plasmid) were mixed with 50 μL Opti-MEM (Invitrogen, catalog number: 51985091) in separate tubes according to the manufacturer’s instructions. After gentle mixing and incubation for 15 min at room temperature, the solutions were added to a 24-well plate and incubated in a 37 °C incubator. The medium was changed after 8–12 h, and cells were harvested 36 h post-transfection.Dual luciferase assayCells in each well of the 24-well plate were lysed with 200 μL Passive Lysis Buffer (Progema, catalog number: E1910) at room temperature for 30 min. After overnight freezing at − 80 °C, half of the lysed cells were used for luciferase activity detection and western blotting, while the remaining half was reserved for RNA extraction. For luciferase activity detection, 10 μL of cell lysate was transferred to a 384-well plate. In the first reaction, 20 μL LARII (Progema, catalog number: E1910) was added to assess the activity of firefly luciferase, followed by 20 μL Stop & Glo for Renilla luciferase with a parallel signal quenching of luciferase signal. The final luciferase activity was measured as the ratio of firefly luciferase activity normalized by internal Renilla luciferase activity.RNA-ImmunoprecipitationTo examine the binding of LARP7 protein to candidate eCLIP-annotated sites, HEK 293 T cells were transfected using Lipofectamine 3000 (Invitrogen, catalog number: L3000075) under two experimental conditions: (1) Transfection with Flag-tagged LARP7 or PUF-LARP7 plasmid alone, to assess binding to endogenous target sites within mRNAs of HSPA8, UBAP2, and TMEM107; (2) Co-transfection of the same LARP7 constructs with dual-luciferase reporter plasmids carrying inserted LARP7-binding motifs, to evaluate binding to the synthetic target sequences. 48 h after transfection, cells were collected and lysed in RIP lysis buffer (Sigma, catalog number: 17-700). 1/10 of each lysate was saved as the Input control. The remainder was incubated overnight at 4 °C with magnetic beads pre‑coated with either anti‑Flag antibody (Sigma, catalog number: B3111) or control IgG (CST, catalog number: 2729S). Beads were then collected, washed 3 times with RIP wash buffer (Sigma, catalog number: 17-700), and digested with proteinase K. The immunoprecipitated RNA and the input RNA were each spiked with spike-in RNA before extraction with TRIzol (Thermo Fisher). RNA was reverse-transcribed with the PrimeScript RT Reagent Kit (TaKaRa, catalog number: RR092A) using a mixture of oligo(dT) and random hexamer primers, according to the manufacturer’s protocol. The resulting cDNA was analyzed by quantitative PCR on a Roche LightCycler 480 instrument with 2× Universal SYBR Green Fast qPCR Mix (Abclonal, catalog number: RK21203), using gene-specific primers or primers targeting the cFluc reporter gene. For quantification, RNA levels were first normalized to spike‑in RNA and then expressed relative to IgG‑immunoprecipitated RNA. Detailed information for related antibodies was provided in Supplementary Data 4.RNA extraction, semi-quantitative RT-PCR and real-time PCRTotal RNA was extracted from harvested transfected cells using TRIZOL reagent. For semi-quantitative PCR, 1 μg of total RNA was treated with gDNA Eraser to remove genomic DNA, and then reverse transcribed with PrimeScript RT Reagent Kit (Takara, catalog number: RR092A). For real-time PCR, one-tenth of the RT product was used as template using AceQ qPCR SYBR Green Master Mix (Vazyme, catalog number: Q111-02) on a Roche LC480 system, following the manufacturer’s instructions. The amount of circRNA and circRNA back-splicing efficiency were measured by RT-PCR using designed primers (listed in the Supplementary Data 1).Gene knockdown with lentiviral shRNAThe oligo pairs, designated LARP7-sh1and LARP7-sh2, were first phosphorylated at the 5ʹ-end using T4 Polynucleotide Kinase (NEB, catalog number: M0201V) and then annealed to form double-stranded oligonucleotides. These annealed oligos were ligated into the pLKO.1 vector that had been digested with AgeI (NEB, catalog number: R3552S) and EcoRI (NEB, catalog number: R3101V). The resulting plasmid—either the control pLKO.1-scramble or the shRNA constructs (pLKO.1-shRNA)—was transfected into HEK 293 T cells together with the packaging plasmids psPAX2 and pMD2.G at a ratio of 4:3:1, using Lipofectamine 3000 (Thermo Fisher, catalog number: L300007). The lentiviral supernatant was harvested 48 h after transfection. HEK 293 T cells were then infected with either the control or shRNA-containing lentivirus for 48 h, followed by selection with 5 μg/ml puromycin.Western blottingCells were lysed in 1 × S-PAGE loading buffer (a mixture of 5×SDS protein loading buffer (Beyotime, catalog number: P0280), RIPA buffer and protease inhibitor), heated at 98 °C for 10 min. The samples were separated by 4-20% SurePAGE™ Gel (GenScript, catalog number: M00657) for 1 h at 140 V and then transferred to PVDF membranes for 2 h at 75 V. Commercial primary antibodies and HRP-linked secondary antibodies were used for blotting with detailed information listed in Supplementary Data 4. The blots were visualized using an enhanced chemiluminescence detection kit (NCM Biotech, catalog number: P10300) and followed by ChemiDoc Touch Imaging System (Image Lab, Bio-Rad).Analysis of meta-transcripts peak density from eCLIP-seq dataTo analyze the distribution of RBP binding sites across different transcript regions, we first downloaded all RBP idr (irreproducible discovery rate) peak files (GENCODE version 19) from eCLIP-seq data on the ENCODE project website (https://www.encodeproject.org), covering 103 RBPs from the HepG2 cell line and 120 RBPs from the K562 cell line. Transcript abundance in the cell lines was derived from ENCODE’s rRNA-depleted RNA-seq data (HepG2: ENCFF000FVT, ENCFF000FVU; K562: ENCFF000HJF, ENCFF000HJX). We filtered for highly expressed (TPM ≥ 1) transcripts to serve as representative transcripts. To prevent redundancy from multiple isoforms of the same gene, we selected the transcript with the highest expression level per gene, yielding 10,722 transcripts in HepG2 and 9624 transcripts in K562 for subsequent analysis.We further calculated the peak density of 5’UTR, CDS, and 3’UTR among exon regions of each transcript. Then the three regions of each transcript were divided into bins, and the average numbers of peaks for the exonic genomic locus were calculated within each bin.To evaluate the appropriate number of bins, we calculated the median lengths of the 5’UTR, CDS, and 3’UTR regions in transcripts with TPM > 10. Based on these medians, the transcripts in HepG2 were divided into 182 bins (20 in 5’UTR, 100 in CDS, and 67 in 3’UTR) and the transcripts in K562 were divided into 175 bins (20 in 5’UTR, 100 in CDS, and 55 in 3’UTR).Finally, we normalized the peak coverage of each bin in each transcript by dividing the sum of coverage of all transcripts. The sum of the normalized coverages on all transcripts represented the distribution of specific RBP binding sites, and we defined these values as the meta-transcript densities. For visualization, we calculated the z-score value of the average density across bins for each RBP and represented it using the R package pheatmap.Analysis of the overlapping eCLIP-seq peaks with IRES regionsTo seek the translation regulation evidence via IRES elements, we calculated the overlaps between eCLIP peaks and IRES regions. We first downloaded the 693 human IRES sequences from IRESbase44 and then defined the upstream and downstream exonic 150 nt around IRES as IRES regions. The overlapping sequences were extracted using an in-house Python script.Sparse partial least squares discriminant analysis (sPLS-DA)Existing supervised machine learning models, such as random forests and support vector machines, face limitations when processing feature matrices characterized by high multicollinearity and sparsity. To address this issue, we employed a supervised machine learning algorithm, a sparse version of Partial Least Squares Regression (sPLS-DA) to classify the translation activators according to their activities.The sPLS-DA algorithm incorporates Lasso regularization into the PLS loading vector during the singular value decomposition (SVD) process82, and allows for dimensionality reduction, feature extraction, and discriminant analysis of sparse matrices48,83. In our analysis, we structured the predictor dataset, X, as an n × p matrix, where n represents the number of observations and p denotes the number of variables. The response dataset, Y, was organized as an n × q matrix with n observations and q classes (where q = 2). This response dataset was recoded into a dummy matrix suitable for classification purposes in PLS. The PLS algorithm constructs a set of orthogonal components designed to maximize the sample covariance between the response and the linear combination of the predictor variables:$${arg }{max }_{{u{\prime} }_{h}{u}_{h}=1,{v{\prime} }_{h}{v}_{h}=1}{{cov}}^{2}({{u}^{{\prime} }}_{h}X,{v{\prime} }_{h}Y)$$where \({u}_{h}\) and \({v}_{h}\) represent the left and right singular vectors for the SVD of \({X}^{T}Y\) respectively in each iteration or dimension h of the PLS (where h = 1… H, H is the chosen number of dimensions). The original approach was based on SVD and the optimization problem of loading vectors when \({{{\rm{M}}}}_{{{\rm{h}}}}={X}_{h}^{T}{Y}_{h}\) is:$${min }_{{u}_{h},{v}_{h}}{{||}{{{\rm{M}}}}_{{{\rm{h}}}}-{{{\rm{u}}}}_{{{\rm{h}}}}{{{\rm{v}}}}_{{{\rm{h}}}}^{{\prime} }{||}}_{{{\rm{F}}}}^{2}+{P}_{\lambda }\left({u}_{h}\right)_{{u}_{h},{v}_{h}}$$where \({P}_{\lambda }\left({u}_{h}\right)\) is the soft thresholding function that approximates Lasso penalty function.In the discrimination step, the PLS model can be formulated as \(Y=\,X\beta+E\), where \(\beta\) is the matrix of the regression coefficients and \(E\) is the residual matrix. The coefficient matrix β is calculated as \(\beta=\,{W}^{*}{V}^{T}\), where V is the matrix of right singular vectors derived from the SVD decomposition (columns v1,…, vH) and \({W}^{*}={W({u}^{T}W)}^{-1}\). Here, W represents the matrix of regression coefficients from the regression of X on the latent variable \({t}_{h}={v{\prime} }_{h}Y\), and \(u\) is the matrix of left singular vectors (columns u1,…, uH).In the predictive step, a new set of samples is calculated as \({Y}_{{test}}={X}_{{test}}\beta\), where β is the same matrix of the regression coefficients, \({X}_{{test}}\) is the predictor matrix for test samples. The classification of each test sample (each row in \({Y}_{{test}}\)) assigns the column index of the element with the highest predicted value within that row.sPLS-DA Model constructionBased on data from the initial screening, we classified the 110 RBPs into two groups, activator and others, using a translation efficiency (TE) threshold of 1.3. To represent the amino acid composition of each full-length protein, we then extracted the frequencies of short-peptide features, which include mono-peptide, di-peptide, and tri-peptide (see Fig. 5A). Features with high sparsity ( > 0.95) were excluded from the analysis, and the remaining features were then input into the sPLS-DA model, which was implemented using the R package mixOmics. Model performance was evaluated through 7-fold cross-validation, conducted with the R package crossval. Short-peptide frequencies were calculated by an in-house Python script. The figures were generated using ggplot2. The machine learning prediction values for the reported RBPs can be found in Supplementary Data 3.Declaration of AI-assisted technologies in the writing processDuring the preparation of this work, we used ChatGPT to correct the grammar and wording, and subsequently reviewed and edited the content as needed. We take full responsibility for the content of the publication.Reporting summaryFurther information on research design is available in the Nature Portfolio Reporting Summary linked to this article.Data availabilityThe data supporting the findings of this study are available from the corresponding authors upon request. Source data for the figures and Supplementary Figs. are provided as a Source Data file. Source data are provided in this paper.Code availabilityThe scripts used in this study are available on GitHub (https://github.com/CikiyWang/CITActivator) and Zenodo (https://doi.org/10.5281/zenodo.21123139).ReferencesPatop, I. L. & Wust, S. Kadener S. Past, present, and future of circRNAs. EMBO J. 38, e100836 (2019).Article PubMed PubMed Central Google Scholar Kristensen, L. S., Andersen, M. S., Stagsted, L. V. W., Ebbesen, K. K., Hansen, T. B. & Kjems, J. The biogenesis, biology and characterization of circular RNAs. Nat. Rev. Genet. 20, 675–691 (2019).Article CAS PubMed Google Scholar Liu, C.-X. & Chen, L.-L. Circular RNAs: Characterization, cellular roles, and applications. Cell 185, 2016–2034 (2022).Article CAS PubMed Google Scholar Kristensen, L. S., Jakobsen, T., Hager, H. & Kjems, J. The emerging roles of circRNAs in cancer and oncology. Nat. Rev. Clin. Oncol. 19, 188–206 (2022).Article CAS PubMed Google Scholar Conn, V. M., Chinnaiyan, A. M. & Conn, S. J. Circular RNA in cancer. Nat. Rev. Cancer 24, 597–613 (2024).Article CAS PubMed Google Scholar Chen, C. Y. & Sarnow, P. Initiation of protein synthesis by the eukaryotic translational apparatus on circular RNAs. Science 268, 415–417 (1995).Article CAS PubMed Google Scholar Wang, Y. & Wang, Z. Efficient backsplicing produces translatable circular mRNAs. RNA 21, 172–179 (2015).Article CAS PubMed Google Scholar Abe, N. et al. Rolling Circle Translation of Circular RNA in Living Human Cells. Sci. Rep. 5, 16435 (2015).Article CAS PubMed PubMed Central Google Scholar Yang, Y. et al. Extensive translation of circular RNAs driven by N(6)-methyladenosine. Cell Res. 27, 626–641 (2017).Article CAS PubMed PubMed Central Google Scholar Pamudurti, N. R. et al. Translation of CircRNAs. Mol. Cell 66, 9–21 e27 (2017).Article CAS PubMed PubMed Central Google Scholar Fan, X., Yang, Y., Chen, C. & Wang, Z. Pervasive translation of circular RNAs driven by short IRES-like elements. Nat. Commun. 13, 3751 (2022).Article CAS PubMed PubMed Central Google Scholar Wei, H.-H., Zheng, L. & Wang, Z. mRNA therapeutics: New vaccination and beyond. Fundam. Res. 3, 749–759 (2023).Article CAS PubMed PubMed Central Google Scholar Perenkov, A. D., Sergeeva, A. D., Vedunova, M. V. & Krysko, D. V. In vitro transcribed RNA-based platform vaccines: past, present, and future. Vaccines 11, 1600 (2023).Article CAS PubMed PubMed Central Google Scholar Wesselhoeft, R. A., Kowalski, P. S. & Anderson, D. G. Engineering circular RNA for potent and stable translation in eukaryotic cells. Nat. Commun. 9, 2629 (2018).Article PubMed PubMed Central Google Scholar Wesselhoeft, R. A., Kowalski, P. S., Parker-Hale, F. C., Huang, Y., Bisaria, N. & Anderson, D. G. RNA Circularization diminishes immunogenicity and can extend translation duration. Mol. Cell 74, 508–520 (2019).Article CAS PubMed PubMed Central Google Scholar Petkovic, S. & Müller, S. RNA circularization strategies in vivo and in vitro. Nucleic Acids Res. 43, 2454–2465 (2015).Article CAS PubMed PubMed Central Google Scholar Wei, H. et al. A flexible, efficient, and scalable platform to produce circular RNAs as new therapeutics. Preprint at https://doi.org/10.1101/2022.05.31.494115 (2022).Niu, D., Wu, Y. & Lian, J. Circular RNA vaccine in disease prevention and treatment. Signal Transduct. Target. Ther. 8, 341 (2023).Article CAS PubMed PubMed Central Google Scholar Holdt, L. M., Kohlmaier, A. & Teupser, D. Circular RNAs as therapeutic agents and targets. Front. Physiol. 9, 1262 (2018).Article PubMed PubMed Central Google Scholar Hinnebusch, A. G. The scanning mechanism of eukaryotic translation initiation. Annu Rev. Biochem. 83, 779–812 (2014).Article CAS PubMed Google Scholar Kwan, T. & Thompson, S. R. Noncanonical Translation Initiation in Eukaryotes. Cold Spring Harb. Perspect. Biol. 11, a032672 (2019).Article CAS PubMed PubMed Central Google Scholar Godet, A. C. et al. IRES Trans-Acting Factors, Key Actors of the Stress Response. Int. J. Mol. Sci. 20, 924 (2019).Article CAS PubMed PubMed Central Google Scholar Legnini, I. et al. Circ-ZNF609 is a circular RNA that can be translated and functions in myogenesis. Mol. Cell 66, 22–37 (2017).Article CAS PubMed PubMed Central Google Scholar Yang, Y. et al. Novel role of FBXW7 circular RNA in repressing glioma tumorigenesis. J. Natl. Cancer Inst. 110, 304–315 (2018).Article CAS PubMed PubMed Central Google Scholar Chen, C.-K. et al. Structured elements drive extensive circular RNA translation. Mol. Cell 81, 4300–4318 (2021).Article CAS PubMed PubMed Central Google Scholar Chen, R. et al. Engineering circular RNA for enhanced protein production. Nat. Biotechnol. 41, 262–272 (2023).Article CAS PubMed Google Scholar Feng, X.-Y., Zhu, S.-X., Pu, K.-J., Huang, H.-J., Chen, Y.-Q. & Wang, W.-T. New insight into circRNAs: characterization, strategies, and biomedical applications. Exp. Hematol. Oncol. 12, 91 (2023).Article CAS PubMed PubMed Central Google Scholar Deviatkin, A. A. et al. Cap-Independent Circular mRNA Translation Efficiency. Vaccines 11, 238 (2023).Article CAS PubMed PubMed Central Google Scholar Passmore, L. A. & Coller, J. Roles of mRNA poly(A) tails in regulation of eukaryotic gene expression. Nat. Rev. Mol. Cell Biol. 23, 93–106 (2022).Article CAS PubMed Google Scholar Zhu, T. Y., Hong, L. L. & Ling, Z. Q. Oncofetal protein IGF2BPs in human cancer: functions, mechanisms and therapeutic potential. Biomark. Res. 11, 62 (2023).Article PubMed PubMed Central Google Scholar Yang, Y. & Wang, Z. IRES-mediated cap-independent translation, a path leading to hidden proteome. J. Mol. Cell Biol. 11, 911–919 (2019).Article CAS PubMed PubMed Central Google Scholar Mao, M. et al. Modeling and Predicting the Activities of Trans-Acting Splicing Factors with Machine Learning. Cell Syst. 7, 510–520 (2018).Article CAS PubMed PubMed Central Google Scholar Wang, Y., Cheong, C.-G., Tanaka Hall, T. M. & Wang, Z. Engineering splicing factors with designed specificities. Nat. Methods 6, 825–830 (2009).Article CAS PubMed PubMed Central Google Scholar Marash, L. et al. DAP5 promotes cap-independent translation of Bcl-2 and CDK1 to facilitate cell survival during mitosis. Mol. Cell 30, 447–459 (2008).Article CAS PubMed Google Scholar Simsek, D. et al. The Mammalian Ribo-interactome Reveals Ribosome Functional Diversity and Heterogeneity. Cell 169, 1051–1065 (2017).Article CAS PubMed PubMed Central Google Scholar Calviello, L. et al. Detecting actively translated open reading frames in ribosome profiling data. Nat. methods 13, 165–170 (2016).Article CAS PubMed Google Scholar Chen, T.-M. et al. hnRNPM induces translation switch under hypoxia to promote colon cancer development. EBioMedicine 41, 299–309 (2019).Article PubMed PubMed Central Google Scholar Han, S. et al. Ribosomal protein L13 promotes IRES-driven translation of foot-and-mouth disease virus in a helicase DDX3-dependent manner. J. Virol. 94, 01679–01619 (2020).Article Google Scholar Chang, J. et al. An interaction between eIF4A3 and eIF3g drives the internal initiation of translation. Nucleic Acids Res. 51, 10950–10969 (2023).Article CAS PubMed PubMed Central Google Scholar Carter, M. S. & Sarnow, P. Distinct mRNAs that encode La autoantigen are differentially expressed and contain internal ribosome entry sites. J. Biol. Chem. 275, 28301–28307 (2000).Article CAS PubMed Google Scholar Benoit Bouvrette, L. P. et al. RBP Image Database: A resource for the systematic characterization of the subcellular distribution properties of human RNA binding proteins. Nucleic Acids Res. 51, D1549–D1557 (2022).Article Google Scholar Riggs, C. L., Kedersha, N., Ivanov, P. & Anderson, P. Mammalian stress granules and P bodies at a glance. J. Cell Sci. 133, jcs242487 (2020).Article CAS PubMed PubMed Central Google Scholar Van Nostrand, E. L. et al. A large-scale binding and functional map of human RNA-binding proteins. Nature 583, 711–719 (2020).Article PubMed PubMed Central Google Scholar Zhao, J. et al. IRESbase: a comprehensive database of experimentally validated internal ribosome entry sites. Genomics Proteom. Bioinform. 18, 129–139 (2020).Article Google Scholar Su, Y.-S., Tsai, A.-H., Ho, Y.-F., Huang, S.-Y., Liu, Y.-C. & Hwang, L.-H. Stimulation of the internal ribosome entry site (IRES)-dependent translation of enterovirus 71 by DDX3X RNA helicase and viral 2 A and 3 C proteases. Front. Microbiol. 9, 382467 (2018).Article Google Scholar Lunde, B. M., Moore, C. & Varani, G. RNA-binding proteins: modular design for efficient function. Nat. Rev. Mol. Cell Biol. 8, 479–490 (2007).Article CAS PubMed PubMed Central Google Scholar Pereira, B., Billaud, M. & Almeida, R. RNA-binding proteins in cancer: old players and new actors. Trends Cancer 3, 506–528 (2017).Article CAS PubMed Google Scholar Mao, M. et al. Modeling and predicting the activities of trans-acting splicing factors with machine learning. Cell Syst. 7, 510–520 (2018).Article CAS PubMed PubMed Central Google Scholar Schmok, J. C. et al. Large-scale evaluation of the ability of RNA-binding proteins to activate exon inclusion. Nat. Biotechnol. 42, 1429–1441 (2024).Article CAS PubMed PubMed Central Google Scholar Qi, Y. et al. Engineering circular RNA regulators to specifically promote circular RNA production. Theranostics 11, 7322–7336 (2021).Article CAS PubMed PubMed Central Google Scholar Reynaud, K., McGeachy, A. M., Noble, D., Meacham, Z. A. & Ingolia, N. T. Surveying the global landscape of post-transcriptional regulators. Nat. Struct. Mol. Biol. 30, 740–752 (2023).Article CAS PubMed PubMed Central Google Scholar Villa, N. & Fraser, C. S. Human eukaryotic initiation factor 4 G directly binds the 40S ribosomal subunit to promote efficient translation. J. Biol. Chem. 300, 107242 (2024).Article CAS PubMed PubMed Central Google Scholar Trotta, C. R., Lund, E., Kahan, L., Johnson, A. W. & Dahlberg, J. E. Coordinated nuclear export of 60S ribosomal subunits and NMD3 in vertebrates. EMBO J. 22, 2841–2851 (2003).Article CAS PubMed PubMed Central Google Scholar Lee, A. S., Kranzusch, P. J., Doudna, J. A. & Cate, J. H. eIF3d is an mRNA cap-binding protein that is required for specialized translation initiation. Nature 536, 96–99 (2016).Article CAS PubMed PubMed Central Google Scholar Gerstberger, S., Hafner, M. & Tuschl, T. A census of human RNA-binding proteins. Nat. Rev. Genet. 15, 829–845 (2014).Article CAS PubMed PubMed Central Google Scholar Zhou, M. et al. Mass spectrometry reveals modularity and a complete subunit interaction map of the eukaryotic translation factor eIF3. Proc. Natl. Acad. Sci. USA 105, 18139–18144 (2008).Article CAS PubMed PubMed Central Google Scholar Smith, E. A. et al. The nuclear DEK interactome supports multi-functionality. Proteins Struct. Funct. Bioinform. 86, 88–97 (2018).Article CAS Google Scholar Berlanga, J. J., Baass, A. & Sonenberg, N. Regulation of poly (A) binding protein function in translation: Characterization of the Paip2 homolog, Paip2B. RNA 12, 1556–1568 (2006).Article CAS PubMed PubMed Central Google Scholar Kieft, J. S. Viral IRES RNA structures and ribosome interactions. Trends Biochem. Sci. 33, 274–283 (2008).Article CAS PubMed PubMed Central Google Scholar López-Ulloa, B., Fuentes, Y., Pizarro-Ortega, M. S. & López-Lastra, M. RNA-Binding proteins as regulators of internal initiation of viral mRNA translation. Viruses 14, 188 (2022).Article PubMed PubMed Central Google Scholar Sinha, T., Panigrahi, C., Das, D. & Chandra Panda, A. Circular RNA translation, a path to hidden proteome. Wiley Interdiscip. Rev. RNA 13, e1685 (2022).Article CAS PubMed Google Scholar Luo, E. C. et al. Large-scale tethered function assays identify factors that regulate mRNA stability and translation. Nat. Struct. Mol. Biol. 27, 989–1000 (2020).Article CAS PubMed PubMed Central Google Scholar Wei, H. & Wang, Z. Engineering RNA-binding proteins with diverse activities. Wiley Interdiscip. Rev. RNA 6, 597–613 (2015).Article CAS PubMed Google Scholar Cooke, A., Prigge, A., Opperman, L. & Wickens, M. Targeted translational regulation using the PUF protein family scaffold. Proc. Natl. Acad. Sci. USA 108, 15870–15875 (2011).Article CAS PubMed PubMed Central Google Scholar Cao, J., Arha, M., Sudrik, C., Schaffer, D. V. & Kane, R. S. Bidirectional regulation of mRNA translation in mammalian cells by using PUF domains. Angew. Chem. Int. Ed. Engl. 53, 4900–4904 (2014).Article CAS PubMed Google Scholar Zhao, Y.-Y. et al. Expanding RNA binding specificity and affinity of engineered PUF domains. Nucleic Acids Res. 46, 4771–4782 (2018).Article CAS PubMed PubMed Central Google Scholar Han, W., Huang, W., Wei, T., Ye, Y., Mao, M. & Wang, Z. Programmable RNA base editing with a single gRNA-free enzyme. Nucleic Acids Res. 50, 9580–9595 (2022).Article CAS PubMed PubMed Central Google Scholar Genuth, N. R. & Barna, M. The discovery of ribosome heterogeneity and its implications for gene regulation and organismal life. Mol. Cell 71, 364–374 (2018).Article CAS PubMed PubMed Central Google Scholar Shi, Z. et al. Heterogeneous ribosomes preferentially translate distinct subpools of mRNAs genome-wide. Mol. Cell 67, 71–83 (2017).Article CAS PubMed PubMed Central Google Scholar Dong, H.-J., Zhang, R., Kuang, Y. & Wang, X.-J. Selective regulation in ribosome biogenesis and protein production for efficient viral translation. Arch. Microbiol. 203, 1021–1032 (2021).Article CAS PubMed Google Scholar Wells, S. E., Hillner, P. E., Vale, R. D. & Sachs, A. B. Circularization of mRNA by eukaryotic translation initiation factors. Mol. Cell 2, 135–140 (1998).Article CAS PubMed Google Scholar Harris, M.T. & Marr, M.T. The intrinsically disordered region of eIF5B stimulates IRES usage and nucleates biological granule formation. Cell Rep.42, https://doi.org/10.1016/j.celrep.2023.113283 (2023).Tycko, J. et al. High-throughput discovery and characterization of human transcriptional effectors. Cell 183, 2020–2035 (2020).Article CAS PubMed PubMed Central Google Scholar Alerasool, N., Leng, H., Lin, Z.-Y., Gingras, A.-C. & Taipale, M. Identification and functional characterization of transcriptional activators in human cells. Mol. Cell 82, 677–695 (2022).Article CAS PubMed Google Scholar Ravarani, C. N., Erkina, T. Y., De Baets, G., Dudman, D. C., Erkine, A. M. & Babu, M. M. High-throughput discovery of functional disordered regions: investigation of transactivation domains. Mol. Syst. Biol. 14, e8190 (2018).Article PubMed PubMed Central Google Scholar Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).Article CAS PubMed PubMed Central Google Scholar Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024).Article CAS PubMed PubMed Central Google Scholar Li, M. Q. C. et al. Advantages and Limitations of AlphaFold in Structural Biology: Insights from Recent Studies. Protein J. 45, 22–38 (2026).Article PubMed Google Scholar Omidi, A., Bui, J. M. & Gsponer, J. Predicting protein interfaces in the age of AlphaFold: Why dynamics and disorder remain a challenge. Cell Syst. 17, 101508 (2026).Article CAS PubMed Google Scholar Chen, C., Yang, Y. & Wang, Z. Study of circular RNA translation using reporter systems in living cells. Methods 196, 113–120 (2021).Article CAS PubMed Google Scholar Reddy Chichili, V. P., Kumar, V. & Sivaraman, J. Linkers in the structural biology of protein–protein interactions. Protein Sci. 22, 153–167 (2013).Article CAS PubMed PubMed Central Google Scholar Rohart, F., Gautier, B., Singh, A. & Lê Cao, K.-A. mixOmics: An R package for ‘omics feature selection and multiple data integration. PLoS Comput. Biol. 13, e1005752 (2017).Article PubMed PubMed Central Google Scholar Lê Cao, K.-A., Boitard, S. & Besse, P. Sparse PLS discriminant analysis: biologically relevant feature selection and graphical displays for multiclass problems. BMC Bioinform. 12, 1–17 (2011).Article Google Scholar Download referencesAcknowledgementsWe thank members and alumni of the Wang lab, in particular F.X.J., L.J.F. and M.M.W. for advice and support. We thank Dr. Xiaowei Song for sharing a dozen plasmids containing full-length RBPs.FundingThis work is supported by the National Natural Science Foundation of China (NSFC grant 32530065 to Z.W., NSFC grant 32171294 to H. Y. W.), the National Key Research and Development Program of China (2025YFA0923103 and 2021YFA1300503 to Z.W.), the BMGF-NSFC Joint Project on Vaccine Research and Development (W2412022 and 2024VTJP1001 to Z.W.).Author informationAuthor notesThese authors contributed equally: Qian-Yun Lu, Si-Qi Wang, Yan-Wen Ye.Authors and AffiliationsShanghai Institute of Nutrition and Health, University of Chinese Academy of Sciences, Chinese Academy of Sciences, Shanghai, ChinaQian-Yun Lu, Si-Qi Wang, Yan-Wen Ye, Huanhuan Y. Wei & Zefeng WangDepartment of Systems Biology, Innovative Center of RNA Medicine, School of Life Sciences, Southern University of Science and Technology, Shenzhen, ChinaYan-Wen Ye, Tong Wei & Zefeng WangKey University Laboratory of RNA Medicine of Guangdong, Souther University of Science and Technology, Shenzhen, ChinaYan-Wen Ye, Tong Wei & Zefeng WangShanghai CirCode Biomedicine Co. Ltd, Shanghai, ChinaYun Yang & Zefeng WangGuangming Advanced Research Institute, Southern University of Science and Technology, Shenzhen, ChinaZefeng WangAuthorsQian-Yun LuView author publicationsSearch author on:PubMed Google ScholarSi-Qi WangView author publicationsSearch author on:PubMed Google ScholarYan-Wen YeView author publicationsSearch author on:PubMed Google ScholarTong WeiView author publicationsSearch author on:PubMed Google ScholarYun YangView author publicationsSearch author on:PubMed Google ScholarHuanhuan Y. WeiView author publicationsSearch author on:PubMed Google ScholarZefeng WangView author publicationsSearch author on:PubMed Google ScholarContributionsConceptualization: Z. Wang, Q-Y. Lu, Y. Yang. Funding acquisition: Z. Wang, Y. Yang, H. Y. Wei. Resources: Z. Wang, H. Y. Wei. Project administration: Z. Wang, Q-Y. Lu, H. Y. Wei. Methodology: Q-Y. Lu, S-Q, Wang, Y-W. Ye. Data curation: Q-Y. Lu, S-Q, Wang, Y-W. Ye, H. Y. Wei. Formal analysis: Q-Y. Lu, S-Q, Wang, Y-W. Ye. Investigation: Q-Y. Lu, S-Q, Wang, Y-W. Ye, H. Y. Wei, T. Wei. Software: S-Q, Wang. Supervision: Z. Wang, Q-Y. Lu, Y. Yang, H. Y. Wei. Validation: Q-Y. Lu, S-Q, Wang, Y-W. Ye, H. Y. Wei, T. Wei. Visualization: Q-Y. Lu, S-Q, Wang, Y-W. Ye, H. Y. Wei. Writing-original draft: Q-Y. Lu, Z. Wang, S-Q, Wang, Y-W. Ye. Writing-review and editing: Q-Y. Lu, Z. Wang, H. Y. Wei, Y.-W. Ye, S-Q, Wang.Corresponding authorsCorrespondence to Huanhuan Y. Wei or Zefeng Wang.Ethics declarationsCompeting interestsThe authors declare no competing interests.Peer reviewPeer review informationNature Communications thanks Bin Shao, and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.Additional informationPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Supplementary informationSupplementary Information (download PDF )Description of Additional Supplementary Files (download PDF )Supplementary Dataset 1 (download XLSX )Supplementary Dataset 2 (download XLSX )Supplementary Dataset 3 (download XLSX )Supplementary Dataset 4 (download XLSX )Reporting Summary (download PDF )Transparent Peer Review file (download PDF )Source dataSource Data (download XLSX )Rights and permissionsOpen Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.Reprints and permissionsAbout this article