Codon optimality modulates cellular stress and innate immune responses triggered by exogenous RNAs

mRNAs with optimal sequences produce more protein

To increase reporter protein expression from exogenous mRNAs, we developed and used a new codon optimality formula to re-engineer the coding sequences of NLuc, RLuc, and FLuc. For ease of detection, FLuc and NLuc were engineered with 3x FLAG tags (with codon distributions similar to their fused reporter protein) at their C-terminus. Overall, the optimal reporters have increased GC content, reduced U content, and a lower minimum free energy suggesting a more stable tertiary structure (Figure S1A) [42]. In accordance with previously published studies, our resulting optimal reporters have substantially higher GC-ending optimal codons (GC3) and lower AU-ending non-optimal codons (AU3) compared to their native counterparts (Figure S1A) [31, 46]. While there is unanimity regarding the superiority of optimal sequences, there is a stark disagreement as to precisely which codons are most optimal. Exemplifying this disagreement as shown in Table S2, with ~220 and ~300 different codon selections merely by count, our codon optimality formula generates starkly different sequences for 550-codon open reading frame of FLuc than the formulas from two other well-regarded teams [47, 48]. Critically, the codon selection differences shown in the table only account for the most favored codon used and do not account for the positional choices for where each codon was employed [47, 48]. To further illustrate the point, the CAI scores of poorly and highly optimal FLuc mRNA in Barrington et al. are 0.62 and 0.75 (compared to ours at 0.73 and 0.97) respectively [47]. Since these sequence designs are so divergent, although they encode the identical protein, it is exceedingly unlikely that the cellular RNA surveillance machinery will survey and act upon our coding sequences as was observed in those earlier studies [47, 48].

We in vitro transcribed both non-optimal and optimal linear mRNAs encoding the individual reporters with standard ribo-nucleotides (Figure S1). To compare the expression of the reporter mRNAs in vivo, we injected non-optimal and optimal FLuc reporter mRNAs encapsulated in lipid nanoparticles (LNP) into the left and right flanks of BALB/c mice respectively. Although this design risked diffusion of the signal, we reasoned that the reduction in animals used outweighed this possible limitation. Mice were imaged at 24, 48, and 72 hours after administration using an In Vivo Imaging System (IVIS) (Figures 1A and S2). The luminescence from optimal FLuc mRNA was about an order of magnitude higher than the signal from non-optimal FLuc mRNA at every time point, and there were no concerns about signal diffusion between the injections (Figures 1A and S2). Although there is no consensus regarding a unified codon optimality score, we used the human CAI score as an approximate measure of codon optimality [35, 37]. Depending upon the reporter, the resulting codon selections for each luciferase reporter corresponded to a CAI score of 0.98, 0.97, and 0.97 for the highly optimal coding sequences of NLuc (ΔCAI = +0.25), RLuc (ΔCAI = +0.31), and FLuc (ΔCAI = +0.24), respectively (Figure S1A).

To further test the generalizability of our optimality framework, we redesigned the coding sequence of the chikungunya virus (CHIKV) glycoprotein and compared its expression against the native viral sequence and a sequence generated by a commercially available codon optimization tool. Our optimized sequence showed superior expression compared to the other sequences (Figure S3), demonstrating that our codon optimality formula can be applied to therapeutically relevant coding regions.

The optimality of test mRNAs affects protein output from both test and control mRNAs in vitro

We co-transfected reporter mRNAs with control mRNAs into multiple cell lines for dual luciferase assays. To our surprise, while the optimal RLuc and FLuc reporters had 6-10-fold increases in luciferase activity when we normalized the signals to the co-transfected control RNA, the optimal NLuc reporters showed no difference in luciferase activity in HEK293 cells and only had a marginal increase in luciferase activity in U2OS cells (Figure 1B). However, upon closer examination of the raw, non-normalized data, we noticed that all optimal reporters showed significant increases in luciferase activities when compared to non-optimal mRNAs in all four (HEK293, U2OS, NIH/3T3, and BJ) cell types assayed (Figures 1C, left Y-axis, and S4A). Intriguingly, we found that the co-transfected control reporter in non-optimal test reporter-transfected cells showed substantially lower levels of luciferase activity driven from the control mRNA compared to the same mRNA in cells transfected with optimal test reporters (Figure 1C, right Y-axis). In other words, the optimality of the test mRNA was exerting an effect on and changing the signal returned by the control mRNA. This made normalizing the data difficult since the control RNA is assumed to remain constant. Further, the distortions of signals from the control RNA held true regardless of which pair of luciferase transcripts were used. To rule out that this effect was caused by experimental artifacts within dual luciferase assays, we validated this observation by co-transfecting the optimal or non-optimal luciferase reporters with mRNA encoding an EGFP reporter protein (Figure S4B). In agreement with the dual luciferase assays, cells co-transfected with non-optimal luciferase mRNAs showed lower EGFP fluorescence as measured by an Incucyte live cell imaging system than cells transfected with optimal luciferase reporters (Figure S4B).

To further exclude the possibility that this observation was caused by the effects of mRNA codon optimality on cell health, viability of the transfected cells was evaluated using CellTiter-Glo assay. Cells transfected with non-optimal or optimal RLuc mRNAs showed similar viability at 24 and 48 hours after transfection (Figure S5). Together, these data show that the codon optimality of experimental mRNAs influences translation from the co-transfected control mRNAs.

Codon optimality does not differentially affect the stability of exogenous mRNAs

The reduced protein output from non-optimal mRNAs could conceivably reflect accelerated degradation of the non-optimal transcript itself [41, 49]. To test this directly, we quantified the relative abundance of exogenous FLuc CDS and the exogenous 3’ UTR by RNA-sequencing at 3, 6, and 24 hours post-transfection, normalizing to the 3-hour time point to obtain a measure of transcript stability over the experimental window. RNA-seq quality assessment via Pearson correlation analysis showed strong agreement among biological replicates and principal component analysis (PCA) revealed consistent grouping of samples by treatment and time point (Figures S6A and S6B). Non-optimal and optimal FLuc transcripts showed broadly similar rates of degradation across both the CDS and 3’ UTR regions (Figure S6C). The comparable decay kinetics of the two transcripts suggest that codon optimality-mediated mRNA degradation does not account for the observed differences in protein output. This is consistent with steady-state mRNA turnover dominating over optimality-dependent surveillance or decay mechanism under these experimental conditions. Together, these data indicate that differential mRNA stability does not account for the observed differences in protein output between non-optimal and optimal mRNA-transfected cells.

Unmodified linear mRNAs with poorly optimal codons drive temporally distinct waves of cellular stress

The suppression of co-transfected control mRNA expression by non-optimal codon mRNAs suggested that codon optimality of reporter mRNAs may affect transcriptome-wide changes in gene expression. To gain mechanistic insight into the cellular response triggered by non-optimal mRNA, we conducted whole-transcriptome analysis of HEK293 cells at 3, 6, and 24 hours after transfection with FLuc-encoding unmodified linear mRNAs. This analysis revealed temporally distinct waves of stress signaling (Figures 2 and S7). At 3-6 hours, we observed rapid induction of immediate early genes (IEGs) including JUN, FOS, EGR1, and DUSP1 which are canonical targets of ZAKα-p38/JNK-AP-1 signaling reminiscent of activation of RSR by ribosome collision events (Figures 2 and S7A) [50, 51]. Notably, DUSP1, a dual-specificity MAP kinase phosphatase and direct transcriptional target of p38/JNK, was induced as early as 3 hours post-transfection, providing an intrinsic negative-feedback mechanism, thereby explaining the resolution of the JUN/FOS/EGR1 signature by 24 hours (Figure S7A). In contrast, canonical ATF4-driven ISR target genes, including DDIT3, TRIB3, CHAC1, and PPP1R15A, showed a delayed, progressively building induction that was more prominent at 24 hours than at 6 hours, consistent with the requirement for sustained eIF2α phosphorylation and preferential ATF4 translation via its upstream open reading frames before ATF4 transcriptional targets can accumulate [45, 52]. ATF3 was unique in showing early induction that was maintained through 24 hours, reflecting its dual regulation by both AP-1 (early, RSR-driven) and ATF4 (late, ISR-driven) inputs. Together, these data indicate that poorly optimal codon usage sequentially activates the RSR and ISR, with distinct transcriptional signatures associated with ZAKα and GCN2 activation, respectively.

Unmodified linear mRNAs with poor codon optimality activate innate immunity genes

Further analysis of the transcriptome profiles provided additional mechanistic insight into the innate immune responses elicited by mRNA transfection. The expression of IFNB1 peaked at 6 hours and declined substantially by 24 hours, a pattern characteristic of an acute and transient primary IRF3-driven induction (Figures 2B and S7B) [53]. Similarly, CXCL10, a chemokine with dual IRF3 and NF-κB regulatory inputs, peaked at 6 hours and declined substantially by 24 hours, mirroring IFNB1 kinetics (Figure 2) [54]. In contrast, downstream ISGs including IFIT1, IFIT2, IFIT3, ISG15, RSAD2, DDX60, and OAS3, exhibited progressive upregulation peaking at 24 hours, consistent with ISGF3-driven transcriptional amplification downstream of the IFN-β autocrine loop [55]. The early immune response was also characterized by robust upregulation of chemokines (CCL5 and CXCL11) and innate immune regulators (IRF9, IRF1, STAT1), consistent with the establishment of a secondary interferon-driven transcriptional program. Notably, EIF2AK2 (protein kinase R; PKR) transcript levels were markedly elevated at 24 hours but not at earlier time points, consistent with its classification as an ISG transcriptionally driven by the ISGF3 autocrine loop rather than as an early primary dsRNA sensor at the transcript level. At the pathway level, Gene Set Enrichment Analysis (GSEA) further confirmed significant enrichment of IFN-α/γ response and TNF-α signaling via NF-κB in cells transfected with non-optimal mRNA compared with optimal mRNA at both time points (Figure S7C).

To independently validate these global transcriptomic changes and confirm that reduced immunogenicity is directly driven by optimal codon usage, we next performed RT-qPCR analysis on a targeted panel of genes known to be activated by different innate immune pathways. Consistent with our bulk RNA-sequencing data, transfection of three independent non-optimal reporter mRNAs in U2OS cells induced robust innate immune activation relative to mock transfection (Figure 3). Notably, optimal mRNAs resulted in a significant suppression of innate immunity gene activation compared to the non-optimal reporters. Together, these findings demonstrate that poor codon optimality enhances the immunogenicity of exogenous mRNAs, thereby contributing to the activation of innate immune pathways.

Unmodified linear mRNAs with poor codon optimality induce eIF2α phosphorylation in a predominantly GCN2-dependant manner

In Neurospora, reporter mRNAs with non-optimal codons can induce the phosphorylation of translation initiation factor eIF2α in a CDS length-dependent manner [56]. The effect that codon usage has on eIF2α phosphorylation resembles the effect of ribosome stalling induced by translation elongation inhibitors [29, 56, 57]. Phosphorylated eIF2α leads to global reduction in translation initiation and is an integral part of integrated stress response pathways [32, 43]. We compared levels of eIF2α phosphorylation in HEK293 cells individually transfected with unmodified linear mRNAs encoding non-optimal or optimal reporters by western blot analysis. Interestingly, cells transfected with non-optimal reporters exhibited significant increases in eIF2α phosphorylation over optimal mRNAs (Figures 4A and S8A). The increased levels of phosphorylated eIF2α are consistent with globally reduced (but not blocked) translation initiation and could explain the reduced control reporter expression observed in cells transfected with non-optimal mRNA.

To identify whether dsRNA contamination generated during IVT, which can activate the eIF2α kinase PKR, contribute to eIF2α phosphorylation, we measured the levels of dsRNA in non-optimal and optimal IVT mRNAs. Non-optimal mRNAs have slightly higher dsRNA content than their optimal counterparts (Figure S8B). Cellulose purification, which selectively removes dsRNA byproducts, reduced the levels of phosphorylated PKR and improved reporter protein output from non-optimal mRNAs compared to unpurified preparations (Figure S8C) [58]. The level of eIF2α phosphorylation was partially reduced while the control RNA expression was increased in cells transfected with purified non-optimal mRNAs (Figures S8C and S8D). However, expression from cellulose-purified non-optimal mRNA remained substantially lower than that from optimal mRNA prepared under identical conditions, demonstrating that dsRNA contamination contributes to, but does not fully account for, the translational suppression associated with non-optimal codon usage.

To elucidate which stress-induced kinase(s) is/are responsible for the elevated levels of phosphorylated eIF2α, we pre-treated HEK293 cells with the inhibitors of PKR (C16), PKR-like endoplasmic reticulum kinase (PERK) (AMG-44), and GCN2 (A-92) prior to transfection of the non-optimal or optimal FLuc reporter mRNAs and conducted western blot analysis to assay eIF2α phosphorylation (Figure 4B, top) [59–61]. The blots show that the PKR inhibitor had a minimal effect, though we note that interpretation of C16 experiments requires caution. Recent work has shown that C16 can activate GCN2, which could partially counterbalance the reduction in PKR-mediated eIF2α phosphorylation and thereby underestimating the contribution of PKR under our experimental conditions [60]. The PERK inhibitor caused no change in eIF2α phosphorylation, excluding the accumulation of misfolded proteins and/or ER stress as drivers of the observed stress response (Figures 4B and S8E). In fact, the eIF2α phosphorylation observed in cells transfected with non-optimal mRNAs was only suppressed by A-92. Further, as shown before, optimal mRNAs cause a small increase in eIF2α phosphorylation, but even that is suppressed by A-92 treatment. These data suggest that, in our assays, GCN2 is the major stress-induced kinase triggered by non-optimal mRNA (Figures 4B and S8E).

As eIF2α phosphorylation inhibits global protein synthesis, and we saw that A-92 and C16 (to a much smaller extent) reduced eIF2α phosphorylation, we reasoned that the inhibitors could influence protein production from the transfected mRNAs. We used western blotting with anti-FLAG antibodies to assay translation from non-optimal or optimal FLuc RNA in the presence or absence of the three inhibitors. In cells transfected with non-optimal RNA, treatment with A-92 resulted in a noticeable increase in FLuc protein levels (Figures 4B, bottom and S8E), indicating that GCN2 inhibition boosted translation from our non-optimal mRNA. We independently confirmed the effect of each inhibitor by measuring the luciferase activity of cells transfected with either non-optimal or optimal RLuc mRNAs. As expected, A-92 treatment led to a substantial increase in luciferase activity from non-optimal mRNA compared to DMSO control (Figure 4C). In agreement with previous experiments, A-92 treatment also led to an increase in control RNA expression when co-transfected with non-optimal mRNA (Figure S8D). This effect was specific to the A-92 and non-optimal mRNA combination, as no increase in protein expression was observed with the other inhibitors, or when the inhibitors were added to cells transfected with optimal mRNAs.

Optimal codons increase reporter protein expression from linear and circular RNAs

mRNA engineering approaches, including chemical modification of nucleotides and RNA circularization, have also been shown to reduce immunogenicity while increasing the efficiency and durability of protein expression [8, 14]. We next evaluated the level and duration of protein expression from our optimality reporters delivered using three different RNA therapy platform technologies (unmodified mRNA, modified mRNA, and circRNA). Consistent with the mRNA vaccines for COVID-19, modified linear mRNAs were made by complete substitution of U with m^1Ψ [39, 40]. circRNAs containing a CVB3 IRES driving the translation of an optimal or non-optimal reporter CDS were generated by adapting a previously published permuted intron-exon (PIE) strategy [14]. Transfection into HEK293 and U2OS cells showed that optimal reporter mRNAs exhibited higher luciferase activities compared to non-optimal counterparts across both RNA types (Figures 5A and 5B). We also observed different expression profiles and durability of protein expression when comparing linear and circular RNAs in which circRNAs predominantly showed lower peak level protein expression but maintained protein expression longer than linear mRNAs. The expression level and durability of expression also varied depending upon the reporter protein, but optimal sequences always outperformed non-optimal sequences for both parameters. Further, non-optimal circRNAs generally had similar expression profiles and RNA half-lives as non-optimal mRNAs (Figures 5A and 5B).

Since incorporating modified nucleotides into mRNAs and circularization have been shown to reduce the immunogenicity of exogenous RNAs, we also tested to see how modified mRNA and circRNAs affected the expression of the transfection control mRNA, eIF2α phosphorylation, and innate immunity genes. The co-transfection of different optimality reporters encoding RLuc with an EGFP control mRNA in HEK293 and U2OS cells revealed that the reduction of EGFP signals observed in non-optimal unmodified linear mRNA transfection was abrogated when modified linear mRNA or circRNA were used (Figure S9A). Next, we analyzed the level of eIF2α phosphorylation in HEK293 cells individually transfected with unmodified linear, modified linear, or circRNAs encoding non-optimal or optimal RLuc reporters. Western blot analysis shows that the induction of eIF2α phosphorylation by non-optimal reporters was diminished when the identical CDS was delivered as either modified linear mRNA or circRNA (Figure S9B). The immunogenicity of different RNA platforms encoding optimality reporters was then compared by RT-qPCR analysis of transfected U2OS cells. The expression of innate immunity genes induced by non-optimal sequences was significantly blunted when delivered by circRNAs or nucleoside-modified mRNAs (Figure S9C). While m^1Ψ substitutions abolished immunogenicity of linear mRNAs, non-optimal circRNAs remained somewhat immunogenic, possibly linking GCN2 driven responses to RNA immunogenicity. Another possibility is that trace amounts of circRNA precursors or processing intermediates (that are below the detection threshold of FlashGels) (Figure S1) drive the immune activation observed in our circRNA samples. Regardless, these data are consistent with each RNA therapy platform conferring unique immunostimulatory, or immune-evasive, properties.

RNA circularization increases the lifespan of optimal reporter transcripts

circRNAs have been touted as a promising next-generation RNA therapy platform since they are resistant to cellular exonucleases and confer longer in vivo lifespan than mRNA [14]. Further, costly modified nucleotides and capping reagents are not required for circRNA production [15]. To de-couple mRNA lifespan from the half-life of their encoded reporter proteins, we fused CL1/PEST degron motifs to the C-terminus of each reporter protein to stimulate targeted degradation by proteasome [62, 63]. We estimated the half-life of all three luciferase reporter proteins in multiple cell lines by measuring luciferase activities of cells transfected with unmodified linear mRNA. Half the cells were treated with the translation inhibitor cycloheximide (CHX), and luciferase signals were measured 0, 1, 2, and 3 hours after CHX addition. Consistent with previous reports, luminescence data showed that the reporters bearing degron motifs had ~8-fold lower signals than non-degron containing versions, and that the half-life of degron-bearing luciferase reporter proteins was ~1 hour or less (Figures 6A and S11A) [62]. Notably, at the protein level, both non-optimal and optimal mRNA-encoded reporter proteins showed essentially identical half-lives, validating that the codon optimality of the RNA does not affect the stability of the encoded reporter protein. We then used an extended time-course to compare the luciferase activity of each destabilized reporter from unmodified linear, modified linear, and circRNAs. As predicted, circRNAs encoding optimal degron-tagged reporters have longer half-lives than their linear mRNA counterparts (Figures 6B, S10, and S11B). We also observed the differences in the half-lives of RNAs encoding different reporters that correlated with the length of the CDS, with RNAs encoding shorter ORFs being longer-lived. Notably, we did not assay RNA levels directly, but since the half-life of the degron-containing proteins are under an hour, we know that the RNA encoding them must be both present and actively translated to observe signals as many as four days later.

Codon optimality of test mRNAs does not affect proteasome-mediated degradation

A potential alternative explanation for the reduced co-transfected control mRNA expression observed in non-optimal mRNA-transfected cells is that non-optimal mRNA enhances proteasome-mediated degradation of co-expressed bystander proteins. To test this directly while excluding the confounding effect of translational suppression, which would also reduce control reporter levels but through a different mechanism, we employed a sequential transfection strategy. Degron-tagged control reporter mRNAs were transfected into HEK293 and U2OS cells 16 hours prior to the transfection of non-optimal or optimal test mRNAs, allowing a stable pool of control reporter protein to accumulate. Cells were then treated with the proteasome inhibitor MG-132 or DMSO simultaneously with test mRNA transfection for 6 hours. Regardless of the reporter transfected, we did not observe significant differences in luciferase signals, which we use as a proxy for protein degradation rates, of the control reporter proteins after transfection with either non-optimal or optimal mRNAs (Figure S12). These results demonstrate that non-optimal mRNA does not enhance proteasome-mediated degradation of co-expressed proteins, and that enhanced protein turnover does not contribute to the reduced control reporter expression we observe in non-optimal mRNA-transfected cells. We note that the stability of the degron-tagged reporter protein themselves was confirmed separately by CHX chase experiments (Figures 6A and S11A), which demonstrate essentially identical protein half-lives for both non-optimal and optimal reporter proteins, confirming that codon optimality does not influence the intrinsic stability of the encoded protein.

Taken together, our data indicate that exogenous mRNAs enriched in non-optimal codons preferentially trigger a temporal cascade of cellular stress and innate immune pathways, including an early RSR, a transient primary IFN induction, and a sustained ISR and ISG program. These responses were partly driven by dsRNA contamination from IVT reactions and a codon-intrinsic ribosome stalling signal that activates GCN2. eIF2α phosphorylation downstream of GCN2 activation globally suppresses translation initiation, explaining the reduced expression of co-transfected control reporters in non-optimal mRNA-transfected cells. This effect is attributable to translational suppression rather than to differential mRNA stability, protein degradation, or effects on cell viability. These responses were fully or partially abrogated by pre-treatment with A-92, a GCN2 inhibitor, or when nucleoside-modified or circular RNAs were used. Finally, we show that circularizing an optimal CDS enhances both RNA lifespan and the durability of protein expression from exogenous RNAs.

Plasmid construction

The template plasmids used for in vitro transcription of linear and circular mRNAs encoding non-optimal luciferase reporters were constructed using e-Zyvec’s (Polyplus-Sartorius) whole plasmid synthesis service. The design of the degron-containing luciferase reporters was patterned after those reported here [62]. The linear mRNA template consists of a T7 promoter, 5’ UTR, NcoI and XhoI cloning sites for inserting the ORFs, 3’ UTR, 90-nt poly(A) tail, followed by BspQI restriction site. The circRNA template begins with a T7 promoter and ends with a BspQI restriction site. Between the T7 promoter and the restriction site, homology regions, spacers and permuted Anabaena pre-tRNA group I intron-exon sequences on the 5’ and 3’ ends flank a CVB3 IRES and the indicated coding region that was adapted from the literature [14, 15]. The optimal reporter gene fragments were synthesized by Twist Bioscience and subcloned into linear mRNA and circRNA vectors by ligation and HiFi DNA assembly, respectively. To generate optimal linear mRNA templates, the linear mRNA backbone and gene fragments were digested by NcoI and XhoI followed by ligation. The circRNA backbone was PCR amplified using primers CS31 and CS32 (see Table S4 for all primer sequences). Gene fragments were PCR amplified using primers CS33 and CS34. The amplified products were gel purified and assembled by NEBuilder HiFi DNA Assembly master mix. All linear mRNA and circRNA vectors were transformed into NEB stable competent E. coli and NEB 5-alpha competent E. coli, respectively. The plasmids were miniprepped with Monarch Plasmid Miniprep Kit and sequence-verified with whole-plasmid sequencing. For linear mRNA templates, the length of poly(A) tail was verified by Sanger sequencing. All novel plasmids will be provided in a timely manner upon request.

In silico RNA Sequence assessment

The minimum free energy (MFE) was calculated from the predicted RNA secondary structure using RNAFold v2.7.0 at the default settings [101]. The codon adaptation index (CAI) was calculated using CAI calculator on https://www.biologicscorp.com/tools/CAICalculator/ with default settings [36].

Optimal coding sequence design

First, we acknowledge that coding sequence possibilities are effectively limitless even for short proteins (~1.9 x 10¹⁰⁹ for 238 amino acid EGFP). As such, our formula aims to meet a target distribution of selected codons across the entire coding region rather than producing a single reproducible output. The formula’s most selected codons are shown in Table S2. When present, second and third choice codons are placed to minimize direct repetition of individual codons. Furthermore, UUUU and UUUC sequences are universally avoided in the coding region to eliminate frameshifting [102].

mRNA synthesis and purification

The sequences for all mRNAs and circRNAs used in the study, and also the sequence of a representative circRNA precursor prior to circularization are shown in Table S5. Unmodified or modified linear mRNAs were synthesized by in vitro transcription from a linearized plasmid DNA template and co-transcriptionally capped using a HiScribe^® T7 mRNA Kit with CleanCap Reagent AG Kit according to the manufacturer’s instructions. For modified linear mRNAs, UTP was replaced by N¹-methylpseudo-UTP. circRNA precursors were synthesized by in vitro transcription from a linearized plasmid DNA template using a HiScribe^® T7 High Yield RNA Synthesis Kit according to the manufacturer’s instructions. The in vitro transcription reactions were treated with DNase I to remove residual DNA template and RNAs were purified by Monarch Spin RNA Cleanup Kit. We modified an earlier method to make circRNA [14, 103] . Briefly, GTP was added to a final concentration of 2 mM to the purified circRNA precursor (see Table S5) in circularization buffer containing 50 mM Tris-HCl, 10 mM MgCl₂, 1 mM DTT, pH 7.5. The RNA was then heated at 55°C for 8 min. The circRNA products were further purified using a Monarch Spin RNA Cleanup Kit. To remove leftover linear RNA species, the purified circRNA products were treated with 0.5 U of RNase R per 1 μg of circularized RNA at 37°C for 1 h. The quality and quantity of RNA were evaluated by Denovix DS-11 FX+ Spectrophotometer and FlashGel RNA Cassettes (Figure S1B–D).

To remove dsRNA byproducts, the in vitro transcribed mRNAs were subjected to cellulose-based purification as described previously [58]. The cellulose fibers were prewashed by suspending in chromatography buffer containing 10 mM HEPES (pH 7.2), 0.1 mM EDTA, 125 mM NaCl, and 16% ethanol at a concentration of 0.2 g cellulose/mL and incubated for 10 min under vigorous shaking. Next, 700 μL cellulose slurry was transferred to a microcentrifuge spin column and centrifuge at 14,000 × g for 1 min. The flowthrough was discarded, and 500 μL chromatography buffer was added to the spin column and shaken vigorously for 5 min to resuspend the cellulose within the spin column. After centrifugation, the flowthrough was discarded and 100 - 500 μg IVT mRNA in 500 μL chromatography buffer was added to the spin column containing the prewashed cellulose. The IVT mRNA and cellulose slurry were shaken vigorously at room temperature for 30 min. After centrifugation, the unbound single-stranded IVT mRNA in the flowthrough was purified by Monarch Spin RNA Cleanup Kit.

Generation of dsRNA

The DNA template was first generated by PCR using forward and reverse primers that contain T7 RNA polymerase promoter sequences to amplify an optimal mScarlet open reading frame as a template (Table S4). The resulting product has T7 RNA polymerase promoter sequence at both ends, was validated with agarose gels and purified with Zymo DNA clean and concentrator-5 and used as a template for IVT. The dsRNA was synthesized using HiScribe^® T7 High Yield RNA Synthesis Kit overnight at 37°C. in these conditions, the IVT reaction produces two complementary single-stranded RNAs, which spontaneously anneal to form a 160 bp dsRNA. After DNaseI-XT treatment, dsRNA was column-purified using Monarch Spin RNA cleanup kit.

Encapsulation of RNA into lipid nanoparticles (LNP)

LNP containing optimal and non-optimal FLuc RNA were prepared as previously described [104]. Briefly, microfluidic mixing by NanoAssemblr BT Microfluidics Mixing System (Precision NanoSystems, Vancouver, CA) of an ethanolic phase containing 0.8 mMol 1,2-distearoyl-sn-glycero-3-phosphocholine (DPSC), 3.85 mMol cholesterol, 0.15 mMol 1,2-distearoyl-sn-glycero-3-phosphoethanolamine-N-[methoxy(polyethylene glycol)-2000] (ammonium salt) (DSPE-PEG(2000)), and 5.2 mMol heptadecan-9-yl 8-((2-hydroxyethyl)(6-oxo-6-(undecyloxy)hexyl)amino)octanoate (SM-102) with an aqueous phase containing 150 μ RNA (optimal or non-optimal) in 0.1 M Citrate Buffer (pH 5) at 1:3 ratio at a flow rate of 10 mL/min was performed. To remove residual ethanol and unencapsulated RNA, LNP were dialyzed using D-Tube Dialyzer Maxi, 12-14 kDa with 1X PBS Buffer. LNP were then characterized for hydrodynamic diameter, polydispersity index and zeta potential by dynamic light scattering (Malvern Zetasizer, Malvern Instruments). RNA content was assessed by RiboGreen RNA Assay as described by the manufacturer. The LNP had a diameter in the range of 70-110 nm, neutral zeta potential and encapsulation efficiency of >90%. Final working concentrations were diluted in 1X PBS before animal injections.

In vivo assessment of FLuc RNA expression

All animal experiments were conducted in accordance with protocols approved by the institutional animal care and use committee (IACUC) at the Houston Methodist Research Institute. Female BALB/c mice (4-5-week-old) were purchased from Jackson Laboratories. The mice were housed with controlled temperature (25 °C), 12:12h lighting cycle, and access to standard diet and water.

Mice received flank injections with 5 μg (in 100 μL PBS) of RNA encapsulated in LNP. To reduce intra-animal variability, each mouse (n=5) received an intramuscular injection with LNP containing optimal FLuc RNA LNP (right flank) and non-optimal FLuc RNA LNP (left flank). Mice were imaged daily for 3 days to assess bioluminescent signal following intraperitoneal injection of 150mg/kg luciferin using IVIS System (Perkin Elmer). IVIS bioluminescence imaging was performed for whole animal (dorsal view), left flank, and right flank positions. After imaging at 72h (3 days), the animals were sacrificed according to IACUC-approved protocols.

Cell culture and transfection

HEK293 and NIH/3T3 cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM) plus 10% FBS; U2OS cells in McCoy’s 5A plus 10% FBS; BJ cells in Eagle’s Minimum Essential Medium (EMEM) plus 10% FBS. All cells were maintained in a 37°C incubator with 5% CO₂ and passaged approximately every 3 days. mRNAs were transfected with Lipofectamine MessengerMAX Transfection Reagent at equimolar concentrations unless stated otherwise along with an internal control mRNA per the manufacturer’s protocols. The total amount of RNA transfected in each experiment ranged between 20 and 70 ng/well of a 96 well tissue culture dish. The transfection reagent was removed, and fresh medium was added to the cells 6 h after transfection.

In vitro luciferase assays

The reagents for the dual luciferase assays were obtained from Promega, and the assays were performed according to the manufacturer’s instructions. Briefly, cells were lysed with Passive Lysis Buffer at the indicated time points after transfection. For Firefly/Renilla assays, luciferase activity was measured using Dual-Glo Luciferase Assay System. For Firefly/NanoLuc assays, luciferase activity was measured using Nano-Glo Dual-Luciferase Reporter Assay System. Luminescence was measured on a white 96-well flat-bottom microplate using a SpectraMax iD5 Microplate Reader (Molecular Devices).

For protein half-life determination, 24 h after transfection, cells were either lysed with the passive lysis buffer and frozen at −20°C (for T0 control) or treated with 100 μ/mL cycloheximide (CHX) for the indicated amounts of time before lysis/freezing. Luciferase assays were performed on thawed samples using the appropriate dual-reporter assay system mentioned above.

For determining proteasome-mediated protein degradation, cells were transfected with degron-tagged control reporter mRNAs for 16 h. Subsequently, cells were transfected with the optimality reporter mRNAs and treated with DMSO (for control) or 10 μ/mL MG-132. After 6 h, cells were lysed with passive lysis buffer, and luciferase assays were performed using the corresponding dual-reporter assay system.

Live fluorescence microscopy imaging

An Incucyte S3 Live-Cell Analysis System (Sartorius) was used to measure and quantify the levels of EGFP control mRNA expression. Four independently transfected wells were used to measure each time point and condition. For each well, five images from different positions within each well were acquired at time points indicated after transfection for each signal determination. The signals for each image were quantified using the Incucyte S3 software (Incucyte 2022B Rev2), averaged for each well, and are shown as the mean fluorescence intensity.

To evaluate the effects of the inhibitors A-92 and C16 over time, live-cell fluorescence imaging was performed using an Incucyte S3 Live-Cell Analysis System (Sartorius). Circular EGFP control mRNA expression was monitored and quantified at the indicated time points after transfection. For each condition and time point, three independently transfected wells were analyzed. Three images were acquired from different positions within each well for each fluorescence signal measurement. Fluorescence signals from each image were quantified using Incucyte S3 software (Incucyte 2025B), averaged for each well, and reported as mean fluorescence intensity.

Cell viability assay

HEK293 cells were transfected with unmodified non-optimal or optimal RLuc reporter mRNAs. After 24 h and 48 h, cell viability was assayed using CellTiter-Glo Luminescence Cell Viability Assay according to the manufacturer’s protocol. Briefly, a volume of the reagent equal to the cell culture medium was added to each well. After 10-minute incubation at room temperature, luminescence was measured on a white 96-well flat-bottom microplate using a SpectraMax iD5 Microplate Reader (Molecular Devices).

RNA extraction and bulk RNA sequencing

HEK293 cells were transfected with linear FLuc mRNAs at equimolar concentrations using Lipofectamine MessengerMAX Transfection Reagent according to the manufacturer’s instructions, together with a EGFP circular RNA (circGFP-3xFLAG) to serve as transfection control. Cells transfected with circGFP-3xFLAG alone (mock) served as negative control. The transfection medium was removed, and fresh complete medium was added to the cells 6 h after transfection. Cells were harvested in Trizol LS Reagent at 3, 6, and 24 h post-transfection.

Total RNA was extracted using the Monarch Spin RNA Isolation Kit according to the manufacturer’s instructions. RNA quality and integrity were assessed using an Agilent 4200 TapeStation (Agilent Technologies). Purified RNA samples were submitted to Plasmidsaurus, Inc. (Eugene, OR) for bulk RNA sequencing. Library preparation and sequencing were performed by Plasmidsaurus using their standard protocols. Raw FASTQ files, together with vendor-generated preliminary analysis reports and figures, were returned for downstream analysis. Analysis was performed with tools provided by Plasmidsaurus. The RNA-sequencing data generated in this study have been deposited in the NCBI Gene Expression Omnibus (GEO) under accession number GSE338825.

Raw sequencing reads were evaluated for quality using FastQC v0.12.1 and processed with fastp v0.24.0. Processing included poly-X tail trimming, 3’ quality-based trimming, filtering for a minimum Phred quality score of 15, and removal of reads shorter than 50 bp. Quality-filtered reads were aligned to the reference genome using STAR v2.7.11, with non-canonical splice junctions removed and unmapped reads retained. Aligned reads were coordinate-sorted using samtools v1.22.1. PCR and optical duplicates were removed by UMI-based deduplication using UMIcollapse v1.1.0.

Alignment quality, strand specificity, and read distribution across genomic features were assessed using RSeQC v5.0.4 and Qualimap v2.3. Quality control metrics were summarized using MultiQC v1.32. Gene-level expression quantification was performed using featureCounts from the Subread package v2.1.1, with strand-specific counting, fractional assignment of multi-mapping reads, and exons and 3’ UTRs specified as features grouped by gene ID. Gene counts were annotated with gene biotype and additional metadata extracted from the reference GTF file.

Normalized expression values were calculated using trimmed mean of M-values (TMM) normalization. Sample-to-sample relationships were evaluated using Pearson correlation and visualized by correlation heatmaps and principal component analysis (PCA) (Figures S6A and S6B). Differential expression analysis was performed using edgeR v4.0.16 following standard workflows, including filtering of lowly expressed genes with edgeR::filterByExpr using default parameters. Where applicable, functional enrichment analysis was conducted by gene set enrichment analysis (GSEA) using gseapy v0.12 and the MSigDB Hallmark gene sets. Figures related to the RNA-seq analysis (Figures 2, S6, and S7) were generated using custom R scripts.

Western blot analysis

Sixteen hours after transfection, media was removed, and cells were washed with phosphate-buffered saline (PBS). Cells were collected by scraping with a cell lifter and lysates prepared using NP-40 lysis buffer (50 mM Tris, pH 8.0, 150 mM NaCl, 1% NP-40) with protease and phosphatase inhibitors. Proteins were collected from the supernatant after centrifugation at 16,000 x g for 20 min. Protein concentrations were determined using the Pierce BCA protein assay kit using a bovine serum albumin standard curve. Protein samples were then prepared in Laemmli sample buffer containing 50 mM DTT and heated at 95°C for 5 min. Samples were run on a 12% or 4-20% Mini-PROTEAN TGX Precast Protein Gel with Tris/Glycine/SDS buffer. The protein was transferred to a 0.2 μM PVDF membrane using Trans-blot Turbo Transfer System according to the manufacturerαs instructions. The membrane was blocked with 1x TBST with 5% BSA and sustained rocking at room temperature for 1 h. The following antibodies were used: phosphorylated eIF2α (Ser51) (1:1000 dilution), eIF2α (1:1000 dilution), α-tubulin (1:20000 dilution), β-actin HRP-conjugated (1:5000 dilution), FLAG Tag (1:5000 dilution),HRP-conjugated goat anti-rabbit IgG (1:5000 dilution), HRP-conjugated horse anti-mouse IgG (1:5000 dilution). Primary antibodies were diluted in 1x TBST with 5% BSA, and blots were incubated overnight at 4°C with sustained rocking, except for HRP-conjugated β-actin, which was incubated for 1 h at room temperature. Membranes were washed three times with 1x TBST and incubated with secondary antibodies with rocking at room temperature for 1 h.

For lysates of cells treated with kinase inhibitors, membranes were blocked in Everyblot (Bio-Rad) for 20 min with rocking at room temperature. Primary and secondary antibodies were subsequently diluted in Everyblot blocking buffer as well. Membranes were washed three times with 1x TBST, and protein signal was detected with the Clarity Western ECL Substrate and imaged on the ChemiDoc MP Imaging System (Bio-Rad).

For eIF2α kinase inhibition experiments, cells were pre-treated prior to mRNA transfection with individual kinase inhibitors at concentrations and durations that were demonstrated effective in the literature [59, 60]. For PERK inhibition, cells were pre-treated with AMG-44 at 20 μM for 2.5 h [60]. For GCN2 inhibition, cells were pre-treated with A-92 at 1 μM for 5 h [59]. For PKR inhibition, cells were pre-treated with C16 at 1 μM for 1 h [59].

RT-qPCR

Total RNA was extracted from cultured cells by adding TRI Reagent to the cells and RNA was extracted by phase separation and isopropanol precipitation per the manufacturer’s instructions. One μg of total cellular RNA was treated with DNase I and reverse transcribed with random hexamer primers and ProtoScript II Reverse Transcriptase. RT-qPCR was performed with SSOAdvanced Universal SYBR Green Supermix on a CFX96 Touch Real-Time PCR Detection System. Each sample was measured in technical triplicates, and three biological replicates were performed per experiment. All primers used are listed in Table S4. Relative gene expression was calculated by the 2^−ΔΔCt method using β-actin as an internal control.

Measurement of dsRNA levels

Double-stranded RNA (dsRNA) levels in non-optimal and optimal RNA samples, either untreated or purified by cellulose treatment, were quantified using the EasyAna dsRNA (Modified) Quantitative Detection Kit (ELISA) 2.0 (Vazyme, Cat. No. DD3509EN-01), according to the manufacturer’s instructions. All RNA samples were normalized to the same concentration before preparing 1:100 and 1:900 dilutions. Three independent biological replicates were analyzed for each experimental condition, with two technical replicates per biological replicate at each dilution. The colorimetric reaction was terminated using Stop Solution (Thermo Fisher Scientific, Cat. No. N600), and absorbance was measured at 450 nm using a SpectraMax iD5 Multi-Mode Microplate Reader (Molecular Devices).

A dsRNA standard curve was generated using serial dilutions ranging from 0.0234 to 1.5 ng/mL. Sample dsRNA concentrations were interpolated from the standard curve and corrected for the corresponding dilution factor. Only measurements that fell within the quantitative range of the standard curve were included in the analysis.