Ribosome profiling provides a global snapshot of translation by sequencing ribosome-protected mRNA fragments (RPFs) (Ingolia, 2014). Relative to the coding sequence (CDS), ribosome footprints accumulate at both start and stop codons. We recently developed Ezra-seq, a high-resolution ribosome profiling method with exceptional 5′ end precision (Mao et al., 2023). Owing to its superior 3-nt periodicity, Ezra-seq enables sensitive detection of start codon-associated ribosome frameshifting (Mao et al., 2023). We reasoned that Ezra-seq would also allow detailed characterization of terminating ribosomes. Indeed, when mRNAs are aligned to their annotated stop codons, we observed sharply defined boundaries of termination-associated footprints (Figure 1A, right panel). Unlike initiating ribosomes, which position the AUG codon in the P site, terminating ribosomes place the stop codon in the A site, as indicated by a prominent 5′ end peak at –15 nt relative to the stop codon. Consistent with previous reports (Ingolia et al., 2011), terminating ribosomes protected a longer stretch of mRNA (Figure 1A, bottom panel). Compared with elongating ribosomes, terminating ribosomes displayed two distinct footprint populations: long (30–31 nt) and short (20–23 nt) reads, a feature not observed for initiating ribosomes (Figure 1A, bottom panel). These shorter footprints likely represent terminating ribosomes with an unoccupied A-site.

Closer inspection of stop codon-associated footprints revealed an additional 5′ end peak at –12 nt, which became more prominent among shorter reads (Figure 1B). Previous toe-printing assays have shown that eRF1 induces a forward movement of terminating ribosomes, shifting the leading edge from +13 nt to +15 nt (Pisarev et al., 2007). Moreover, single-molecule analyses have identified distinct pre- and post-termination phases catalyzed by eRF1 (Lawson et al., 2021). Together, these observations suggest that the two 5′ end peaks correspond to pre- and post-terminating ribosome states, with the latter likely adopting a rotated conformation. However, we cannot exclude the possibility that a subset of these ribosomes transiently positions the stop codon in the P-site prior to ribosome disassembly.

To probe the kinetics of termination pausing, we performed cycloheximide (CHX) chase experiments in HEK293 cells. CHX stalls elongating ribosomes on mRNAs but does not arrest terminating ribosomes (Shen et al., 2021). Whereas a 5 min CHX treatment substantially reduced footprint density at stop codons, a 30 min pre-treatment nearly abolished the stop codon-associated peaks (Figure 1—figure supplement 1A and B). These results indicate that terminating ribosomes undergo transient pausing rather than stable stalling at the stop codon. In line with this notion, we observed little accumulation of upstream ribosomes, suggesting that termination pausing rarely triggers ribosome collisions.

Despite the overall presence of termination pausing, ribosome density at individual stop codons varied over several orders of magnitude. To quantify this variability, we calculated a stop codon pausing index for each mRNA by normalizing read density at the stop codon to average CDS occupancy, thereby accounting for differences in mRNA abundance and translation efficiency (Figure 1C). To identify sequence elements associated with termination pausing, we compared mRNAs with high and low pausing indices. Although no significant motifs were detected downstream of stop codons, a GA-rich motif upstream of the stop codon was strongly enriched among mRNAs exhibiting robust termination pausing (E=2.1×10^–23, Figure 1C). These findings indicate that certain coding sequences upstream of the stop codon play a critical role in modulating termination dynamics.

For mRNAs lacking termination pausing, ribosomes may either dissociate rapidly from the mRNA or undergo stop codon readthrough. These scenarios are expected to produce distinct ribosome occupancy patterns in 3′ UTR. Strikingly, we observed higher 3′ UTR ribosome density on mRNAs lacking termination pausing (Figure 1D). This effect could not be explained by downstream sequence bias, as the identity of the +4 nt had minimal impact on 3′ UTR translation (Figure 1—figure supplement 1C). Moreover, 3′ UTR-associated footprints exhibited poor 3 nt periodicity, in sharp contrast to the well-phased CDS reads (Figure 1—figure supplement 1D). These observations suggest that canonical stop codon readthrough alone cannot account for the elevated 3′ UTR ribosome density observed in the absence of termination pausing.

To directly interrogate the dynamics of translation termination, we developed a terminating ribosome profiling strategy by isolating ribosomes associated with eRF1 (Figure 2A). HEK293 cells were first crosslinked by formaldehyde, followed by cell lysis and RNase I digestion. Monosomes were isolated by sucrose gradient centrifugation, and eRF1-bound ribosomes were subsequently enriched by immunoprecipitation, as confirmed by immunoblotting (Figure 2—figure supplement 1A). In the absence of crosslinking, ribosomal proteins were minimally recovered by the eRF1 antibody, consistent with the transient nature of eRF1-ribosome interactions. Deep sequencing of ribosome-protected fragments from eRF1-bound ribosomes (eRF1-seq) revealed a pronounced accumulation of reads precisely at the annotated stop codons (Figure 2A, right panel). Remarkably, eRF1-seq retained single-nucleotide resolution, enabling unambiguous identification of termination sites on endogenous mRNAs (Figure 2B). Similar to conventional Ribo-seq, eRF1-seq also observed the forward-shifted footprints characteristic of post-termination ribosomes (Figure 2C). With excellent reproducibility between biological replicates (Figure 2—figure supplement 1B), eRF1-seq robustly and selectively profiles terminating ribosomes.

Consistent with our Ribo-seq analyses, eRF1-seq revealed substantial heterogeneity in termination pausing, as not all mRNAs displayed prominent eRF1 peaks at their annotated stop codons (Figure 2B). To identify determinants of this variability, we compared mRNAs with and without eRF1 peaks. Neither 3′ UTR length nor downstream sequence features correlated with eRF1 peak height (Figure 2—figure supplement 1C). In contrast, sequences upstream of the stop codon were significantly enriched for a GA-rich motif among mRNAs exhibiting strong eRF1 peaks (Figure 2D). This observation mirrors the motif identified by Ribo-seq analysis (Figure 1C), reinforcing the conclusion that termination pausing is an intrinsic, sequence-encoded property of individual mRNAs. Notably, all three stop codons exhibited similar pausing features and shared upstream sequence motifs (Figure 2—figure supplement 1D and E).

Despite the prominent peak at annotated stop codons, individual eRF1 peaks were broadly distributed across transcripts, spanning the 5′ UTR, CDS, and 3′ UTR (Figure 3A). The presence of eRF1 peaks in the 5′ UTR is not unexpected, as ~50% of human mRNAs harbor upstream open reading frames (uORFs) (Orr et al., 2020). Indeed, ~30% of eRF1 located in the 5′ UTR coincided with termination sites of uORFs previously identified by Ribo-seq (Figure 3—figure supplement 1A). By revealing additional, previously unannotated termination sites in the 5′ UTR, eRF1-seq substantially expands the catalog of uORFs. Nearly all A-site codons underlying 5′ UTR eRF1 peaks corresponded to canonical stop codons (Figure 3B), and Ribo-seq revealed a reduction in read density downstream of those sites (Figure 3B, right panel), consistent with bona fide translation termination. In addition to the 5′ UTR, we identified a total of 807 eRF1 peaks in the 3′ UTR that shared similar features (Figure 3—figure supplement 1B). Compared with conventional Ribo-seq, which typically yields sparse 3′ UTR reads, eRF1-seq provided an improved signal-to-noise ratio for detecting translation events in the 3′ UTR.

Even under stringent peak-calling criteria, eRF1-seq detected a substantial number of peaks within CDS. Many of these CDS-associated eRF1 peaks positioned their A sites at in-frame sense codons (Figure 3—figure supplement 1C). Notably, the majority of those codons were ‘stop-like’ in that they corresponded to frameshifted versions of canonical stop codons. For example, CUG represents a –1 frameshifted UGA, whereas GAG, GAA, and AAG represent +1 frameshifted UGA or UAA. Despite robust eRF1 occupancy, those sense codons did not trigger translation termination, as Ribo-seq revealed no decrease in downstream read density (Figure 3—figure supplement 1C). These observations suggest that eRF1 can transiently compete with cognate A-site tRNAs during elongation, resulting in failed termination attempts.

In addition to in-frame sense codons, a large fraction of CDS-associated eRF1 peaks corresponded to out-of-frame canonical stop codons, particularly UGA (Figure 3C). Supporting authentic termination at these sites, Ribo-seq revealed reduced ribosome occupancy downstream (Figure 3C, right panel). Interestingly, most out-of-frame termination events occurred near the beginning of CDS. Beyond overlapping uORFs, this pattern further supports the occurrence of start codon-associated ribosome frameshifting that we reported recently (Mao et al., 2023). Out-of-frame eRF1 peaks were also enriched near the 3′ end of CDS (Figure 3C and D), consistent with migration of terminating ribosomes at stop codons to search for upstream start codons (Figure 3D). To directly validate stop codon-associated reinitiation, we inserted a 9 nt sequence derived from CASQ2, which contains an out-of-frame AUG codon upstream of a UAG stop codon, between GFP and HiBiT (Figure 3E). Consistent with previous reports (Gould et al., 2014), mutation of the UAG stop codon abolished reinitiation and eliminated out-of-frame HiBiT expression (Figure 3E), confirming that reinitiation is dependent on termination at the stop codon.

Given the wide range of termination pausing observed at individual stop codons, we next sought to identify the sequence determinants of termination pausing using a massively parallel reporter assay (MPRA). In contrast to analysis of endogenous genes, whose sequences are constrained by evolutionary selection, MPRA leverages fully randomized sequences to enable unbiased identification of sequence elements governing translation behavior. We previously employed a uORF-based MPRA to assess start codon usage with randomized sequence contexts (Jia et al., 2020). In this system, mRNA variants with efficient uORF translation are retained in the monosome fraction, whereas efficient translation of downstream GFP relocates the mRNA to the polysome. Thus, the monosome/polysome ratio serves as a quantitative readout of uORF translation efficiency.

To interrogate stop codon usage, we modified the uORF reporter by replacing the uORF stop codon with a 9-nt randomized sequence (Figure 4A). Introduction of an in-frame stop codon within the insert would terminate uORF translation and restrict the mRNA to the monosome fraction. To eliminate transcriptional variation associated with plasmid-based expression, we synthesized the reporter library by in vitro transcription. Following transfection of the mRNA pool into HEK293 cells, monosome and polysome fractions were separated by sucrose gradient centrifugation and analyzed by deep sequencing. Among mRNAs recovered from the monosome fraction, all three canonical stop codons (UGA, UAG, UAA) were strongly enriched (Figure 4A). Importantly, these stop codons were preferentially positioned in-frame within the randomized insert (Figure 4—figure supplement 1A and B). Codons enriched in alternative reading frames were also informative; for example, codons enriched in frame 2 predominantly belong to NUA and NUG, consistent with frameshifted presentations of in-frame stop codons. Together, these results establish MPRA as a robust platform for dissecting stop codon-associated sequence features.

Ribosomal pausing at uORF stop codons is expected to retain mRNAs in the monosome fraction by suppressing both leaky scanning and stop codon readthrough. We, therefore, leveraged the monosome/polysome ratio as a proxy for termination pausing. We first examined the impact of downstream sequences by placing a 9-nt randomized sequence immediately after a fixed UAG stop codon (Figure 4—figure supplement 1C). Despite prior reports that a cytosine (C) at the +4 nt position enhances readthrough (Wangen and Green, 2020), mRNA variants containing C-rich downstream sequence were depleted from both monosome and polysome fractions. This depletion likely reflects accelerated mRNA turnover caused by translation into 3′ UTR (Kesner et al., 2023; Müller et al., 2023). Consistent with our Ribo-seq and eRF1-seq analyses, no specific downstream sequences were enriched in the monosome fraction (Figure 4—figure supplement 1D), indicating that sequences following the stop codon exert minimal influence on termination pausing.

We next assessed whether upstream sequences modulate stop codon fidelity by inserting a 9-nt randomized sequence immediately before the UAG stop codon (Figure 4B). Strikingly, G-rich sequences were strongly enriched in the monosome fraction (Figure 4B, Figure 4—figure supplement 1E). Notably, this enrichment lacked reading frame information, indicating that nucleotide composition rather than codon identity governs monosome enrichment. This feature is reminiscent of G-quadruplex formation that hinders ribosome movement. In contrast, C-rich sequences were selectively depleted from the monosome fraction (Figure 4B), suggesting that the C-rich upstream sequence contexts promote stop codon readthrough and downstream translation. These findings closely mirror the sequence features revealed by Ribo-seq and eRF1-seq.

It was not immediately clear whether translation in the 3′ UTR arose from stop codon readthrough or stop codon-associated reinitiation. To distinguish these possibilities, we investigated the nature of downstream translation triggered by upstream C-rich sequences using individual mRNA reporters. Specifically, we placed a C-rich sequence immediately upstream of the uORF stop codon UAG (Figure 4C). Because many C-rich triplets encode proline, an imino acid known to affect ribosome dynamics, we deliberately selected non-proline C-rich codons to avoid confounding effects. As a control, we used a GA-rich sequence identified from eRF1-seq as being associated with strong termination pausing. To sensitively detect translation downstream of the stop codon, we inserted a HiBiT coding sequence immediately after UAG and HiBiT signals can be detected with superior sensitivity (Jia et al., 2020). The canonical start codon of HiBiT was omitted to exclude leaky scanning-mediated translation. In the absence of the stop codon, the GA-rich sequence produced higher HiBiT signals than the C-rich sequence, likely reflecting difference in the amino acids encoded by the inserts. Strikingly, in the presence of the UAG stop codon, the C-rich sequence nearly doubled HiBiT signals compared with the GA-rich sequence (Figure 4C, bottom panel). This result is consistent with the MPRA data, indicating that C-rich coding sequences preceding the stop codon not only attenuate termination pausing but also actively promote downstream translation. Importantly, this phenomenon was not restricted to uORFs, as similar results were obtained when the uORF was replaced by a GFP coding sequence (Figure 4—figure supplement 2A).

Canonical stop codon readthrough is expected to proceed in-frame and generate a fusion protein with a C-terminal extension. To test this model, we placed the HiBiT sequence in different reading frames relative to the stop codon. Unexpectedly, robust HiBiT signals were detected from reporters in both frame 1 and frame 2 (Figure 4D). In contrast, the GA-rich sequence preceding the stop codon strongly suppressed downstream HiBiT translation. This frame-independent effect was reproducible across multiple sequence variants (Figure 4—figure supplement 2B). To directly examine the translational products, we conducted immunoblotting of whole cell lysates using HiBiT antibodies (Figure 4E). GFP-fusion proteins (~30 kDa) were readily detected in C-rich reporters irrespective of the HiBiT reading frames, ruling out reinitiation as the underlying mechanism. Notably, levels of non-fusion GFP were comparable across reporters, indicating that the differential HiBiT signals were not attributable to altered translation initiation or differences in the amino acids encoded by the upstream inserts. Together, these results demonstrate that C-rich coding sequences promote ribosome readthrough at the stop codon, enabling translation into the 3′ UTR in all three reading frames.

We next examined the positional dependence of the C-rich effect by introducing C triplets at varying distances upstream of the stop codon. HiBiT reporter assays revealed that a C triplet positioned immediately adjacent to the stop codon was the least effective at promoting 3′ UTR translation (Figure 4—figure supplement 2C). This finding suggests that the sequence determinant governing ribosome sliding resides upstream of the E site, rather than directly at the termination codon itself.

ABCE1 (Rli1 in yeast) is a conserved ABC-type protein that plays a crucial role in translation termination and ribosome recycling (Pisarev et al., 2010). Previous studies have shown that depletion of ABCE1 leads to ribosome stalling at stop codons and increased ribosome occupancy within 3′ UTRs (Annibaldis et al., 2020). To investigate whether ABCE1 exhibits sequence preference towards terminating ribosomes, we knocked down ABCE1 in HEK293 cells using shRNA (Figure 4—figure supplement 3A). As expected, ABCE1 silencing impaired cell proliferation and reduced global protein synthesis (Figure 4—figure supplement 3B). We next performed ribosome profiling in ABCE1 KD and control cells using Ezra-seq. After normalizing ribosome occupancy across CDS, loss of ABCE1 resulted in a modest but reproducible increase in ribosome density at stop codons (Figure 4—figure supplement 3C). A closer inspection of footprint distribution revealed that ABCE1 depletion selectively increased ribosome density at the –15 nt position, especially for shorter footprints, whereas the forward-shifted –12 nt peak was largely unaffected (Figure 4—figure supplement 3D). These results suggest that ABCE1 primarily facilitates late-stage termination or ribosome splitting, and its absence delays pre-termination progression. A prior study reported that ABCE1 knockdown in HeLa cells promotes translation into 3′ UTR in all three reading frames (Annibaldis et al., 2020). We were unable to robustly detect this phenomenon by Ribo-seq, likely due to the limited sequencing depth and/or incomplete ABCE1 knockdown in HEK293 cells. To overcome these limitations, we employed a sensitive HiBiT-based 3′ UTR reporter assay, which showed increased 3′ UTR translation irrespective of the sequence context upstream of the stop codon (Figure 4—figure supplement 3E). Therefore, the ribosome recycling factor ABCE1 plays a generic role in translation termination independent of mRNA sequence context.

For a terminating ribosome, the mRNA sequence preceding the stop codon is positioned upstream of the E-site. Given the narrow dimensions of the mRNA channel, it is unlikely that external protein factors, such as ABCE1 directly regulate ribosome behavior in a sequence-specific manner. Instead, prior crosslinking studies have shown that mRNA near the exit site interacts with the 3′ terminus of 18 S rRNA (Pisarev et al., 2008). Such proximity is also evident in recent cryoEM structures of mammalian initiating ribosomes (Brito Querido et al., 2020), although stable base pairing was not explicitly resolved (Figure 5A). A similar juxtaposition between mRNA and the 3′ end of 18 S rRNA has been observed in elongating ribosomes during translocation (Flis et al., 2018). Structural modeling based on available cryo-EM data suggests that, from early to late translocation intermediates (POST-1 to POST-3), the distance between the mRNA segment (–9 to –3) and the 3′ end of 18 S rRNA progressively decreases (Figure 5B). These observations support a model in which mRNA undergoes post-decoding scanning by the 3′ terminus of 18 S rRNA as it exits the ribosome. Notably, the 3′ end of 18 S rRNA harbors a highly conserved U-rich sequence (GAUCAUUA). We propose that GA-rich mRNA sequences can engage in transient U:A and U:G base pairing with 18 S rRNA near the exit site (Figure 5A, Figure 5—figure supplement 1A), thereby slowing mRNA passage and promoting termination pausing. In contrast, C-rich sequences would evade this rRNA-mediated checkpoint, resulting in faster mRNA transit and reduced termination pausing.

To directly test whether putative mRNA: rRNA base pairing contributes to termination pausing, we sought to perturb the 3′ end sequence of 18 S rRNA. Because the presence of hundreds of rDNA copies precludes genome editing approaches, such as CRISPR/Cas9, we instead used a previously described 18 S rRNA expression system that enables incorporation of exogenously expressed 18 S rRNA into ∼15% of 40 S subunits in transfected cells (Burman and Mauro, 2012; Wan et al., 2018). We engineered an 18 S rRNA mutant by substituting the last two U residues with G (GAUCAGGA), a change predicted to weaken interactions with GA-rich mRNA while favoring pairing with C-rich sequence. Overexpression of the mutant 18 S in HEK293 cells resulted in polysome profiles comparable to those observed with wild-type 18 S rRNA (Figure 5—figure supplement 1B), and ribosome profiling by Ezra-seq revealed similar ribosome occupancy across coding regions (Figure 5—figure supplement 1C). However, when mRNAs were stratified by the sequence motif upstream of the stop codon, expression of the mutant 18 S reduced the differential termination pausing between GA-rich and C-rich sequences (Figure 5C). Specifically, mRNAs containing GA-rich motifs exhibited reduced stop codon peaks, whereas those bearing C-rich motifs showed increased ribosome accumulation at stop codons. The reciprocal shift in termination pausing in response to mutation of the 18 S rRNA 3′ end provides strong evidence that sequence-specific termination pausing is mediated by interaction between mRNA and 18 S rRNA.

We further validated this model using HiBiT-based reporters to assess 3′ UTR translation in cells expressing WT or mutant 18 S rRNA. Expression of the mutant 18 S rRNA attenuated 3′ UTR translation from reporters containing C-rich upstream sequences (Figure 5D). Conversely, reporters bearing GA-rich sequences exhibited increased HiBiT signals, indicative of enhanced 3′ UTR translation. Together with the swapped termination pausing profiles observed by ribosome profiling, these results establish a crucial role for the 3′ end of 18 S rRNA in enforcing termination fidelity. Notably, the 3′ terminal sequence of 18 S rRNA is highly conserved across eukaryotes (Figure 5—figure supplement 1D). Annotated stop codons in the human genome are preferentially preceded by GA-rich sequences (Figure 5—figure supplement 1E). In contrast, out-of-frame stop codons show a higher prevalence of C-rich upstream contexts. These patterns suggest an evolutionary advantage for termination pausing at annotated stop codons, likely serving to minimize stop codon slippage and aberrant readthrough while preserving proteome integrity.

Having established sequence-specific termination pausing in cultured cell lines, we next explored its physiological relevance across tissues. To this end, we conducted Ribo-seq on a diverse panel of mouse tissues, including liver, brain, kidney, heart, and testis (Figure 6A). These tissues displayed distinct polysome profiles and global ribosome occupancy across coding regions (Figure 6A, Figure 6—figure supplement 1A), reflecting substantial differences in translational states. Strikingly, ribosome density at start and stop codons varied most prominently among tissues and did so in a reciprocal manner. Testis exhibited relatively modest initiation peaks but pronounced ribosome accumulation at stop codons (Figure 6A). In contrast, liver, heart, and brain showed robust initiation peaks with little or no detectable termination pausing. These differences were not attributable to tissue-specific gene expression, as the same reciprocal pattern persisted when analyses were restricted to genes commonly expressed or uniquely expressed in liver and testis (Figure 6—figure supplement 1B). Consistent with sequence-dependent termination pausing, mRNAs with prominent stop codon peaks in testis were enriched with GA-rich sequences upstream of the stop codon (Figure 6—figure supplement 1C). Notably, these same mRNAs exhibited minimal termination pausing in liver. Moreover, the attenuated stop codon peak in mouse liver has been independently reported (Gobet et al., 2020), excluding the possibility of technical bias in Ezra-seq.

The unexpected tissue-specific variation in termination pausing points to an additional regulatory layer beyond mRNA sequence context. One candidate is the eukaryote-specific ribosomal protein S26 (Rps26), which contacts mRNA upstream of the E-site (Pisarev et al., 2008; Sharifulin et al., 2012). Recent cryo-EM structures revealed that both Rps26 (AA62 – 70) and 18 S rRNA (nt1857 – 1863) jointly shape the mRNA path near the ribosomal exit site (Figure 5B). Because Rps26 lies between the mRNA and the 3′ end of 18 S rRNA, its absence could enhance mRNA:rRNA interactions and thereby strengthen termination pausing. Supporting this notion, a recent study demonstrated that Rps26 can dissociate from fully assembled 80 S ribosomes under stress conditions, generating ribosome heterogeneity (Yang and Karbstein, 2022). To assess whether Rps26 abundance varies across tissues, we compared ribosomal proteins in tissue homogenates. When normalized to β-actin, core ribosomal proteins, such as Rpl4 were present at comparable levels across tissues, whereas Rps26 levels were markedly reduced in testis relative to other tissues examined (Figure 6B). Differential Rps26 expression between liver and testis is also evident in Human Protein Atlas (https://www.proteinatlas.org/) (Uhlén et al., 2015). We further examined Rps26 distribution across polysome fractions in liver and testis. In liver, Rps26 and the control ribosomal protein RACK1 exhibited similar polysome association. In contrast, testis showed pronounced depletion of Rps26 from polysomes (Figure 6C), accompanied by accumulation of Rps26 in ribosome-free fractions. Whether this reflects active dissociation of Rps26 from translating ribosomes in testis remains to be determined. These findings nevertheless suggest Rps26 stoichiometry contributes to tissue-specific termination pausing, likely by modulating access of the mRNA to the 3′ end of 18 S rRNA.

Based on available ribosome structures containing bound mRNA, Rps26 is positioned between the exiting mRNA and the 3′ end of 18 S rRNA (Figure 7A). We hypothesize that loss of Rps26 would increase the accessibility of the mRNA to the 3′ terminus of 18 S rRNA, thereby facilitating mRNA:rRNA interactions and promoting termination pausing. Consistent with this model, normal mode analysis (NMA) using anisotropic network models (Bahar et al., 2007) predicts that removal of Rps26 permits coordinated twisting and closer approximation between the mRNA segment (–3 to –9) and the 3′ end of 18 S rRNA (Figure 7B). This enhanced spatial parity would be further stabilized when the mRNA sequence is complementary to the rRNA. Given the U-rich nature of the 18 S rRNA 3′ end, such interactions are predicted to be most favorable for GA-rich mRNA sequences.

To test whether reduced Rps26 levels affect ribosome dynamics at stop codons, we depleted Rps26 in HEK293 cells using shRNA (Figure 7—figure supplement 1A). Rps26 silencing led to an increased 60 S peak and reduced polysome formation (Figure 7—figure supplement 1B), consistent with its known role in 40 S subunit maturation (Plassart et al., 2021). We then conducted ribosome profiling in parallel with RNA-seq. Remarkably, Rps26 depletion resulted in reduced ribosome density at start codons but increased ribosome accumulation at stop codons (Figure 7—figure supplement 1C). This reciprocal change mirrors the tissue-specific differences observed between initiation and termination (Figure 6A). Importantly, the elevated termination pausing occurred predominantly at stop codons preceded with GA-rich sequences (Figure 7C). To assess the functional consequences of altered termination dynamics, we employed HiBiT-based 3′ UTR reporter assay. Because Rps26 silencing globally reduced protein synthesis (Figure 7—figure supplement 1B), HiBiT signals were normalized to upstream GFP expression. Under these conditions, only reporters containing GA-rich sequences, but not C-rich sequences, exhibited reduced HiBiT translation in Rps26-depleted cells (Figure 7D). Therefore, Rps26 specifically modulates termination fidelity at GA-rich stop codons by constraining mRNA:rRNA interactions.

To further substantiate this conclusion, we overexpressed Rps26 in HEK293 cells (Figure 7—figure supplement 1D), which did not perturb polysome formation (Figure 7—figure supplement 1E). Ribosome profiling revealed that Rps26 overexpression lowered ribosome occupancy from initiation through termination (Figure 7—figure supplement 1F). When mRNAs are stratified by upstream sequence context, Rps26 overexpression caused a pronounced (>50%) reduction in ribosome density at stop codons preceded by GA-rich sequences (Figure 7E). Notably, ribosome density upstream of these stop codons was also diminished, suggesting that increased Rps26 levels broadly restrain mRNA:rRNA engagement during elongation. In contrast, mRNAs harboring the C-rich sequence upstream of the stop codon displayed relatively enhanced termination peaks. To directly link stop codon pausing with termination fidelity, we conducted HiBiT reporter assays in cells with Rps26 overexpression. In line with reduced termination pausing, Rps26 overexpression increased 3’ UTR translation (Figure 7F), confirming that diminished mRNA:rRNA interaction compromises termination fidelity. Collectively, these results establish Rps26 as a key modulator of sequence-specific termination pausing.

Although MPRA is informative in identifying sequence motifs, it is subject to sequencing bias that could alter the sequence enrichment. Introducing mutations into 18 S rRNA is technically challenging due to the presence of hundreds of rDNA copies. We could not determine the fraction of ribosomes bearing 18 S rRNA mutations in transfected cells. Finally, the ribosome structural modeling is speculative and the role of Rps26 in mRNA:rRNA interactions requires more direct evidence.

HEK293-K^b cells and Lenti-X 293T cells are maintained in Dulbecco’s Modification of Eagle’s Medium (Corning, 10–013-CV) with 10% fetal bovine serum (Sigma, 12,306 C). All cells were grown at 37 °C with 5% CO₂. Their original identities have been authenticated by ATCC and tested negative for mycoplasma contamination.

C57BL/6 mice were obtained from the Jackson Laboratory. All animals (1–6 mice per cage) were housed in a 12 hr light/dark cycle in the Weill Hall animal facility at Cornell University with the supervision of the Center for Animal Resources and Education (CARE) breeding program. All animals used in this study were handled in accordance with federal and institutional guidelines, under a protocol approved by the Cornell University Institutional Animal Care and Use Committee, protocol 2017–0035. Mice were housed under specific pathogen-free conditions in an Association for the Assessment and Accreditation of Laboratory Animal Care International-accredited facility and cared for in compliance with the Guide for the Care and Use of Laboratory Animals.

The following antibodies were used at their indicated experimental concentrations: anti-GFP (Proteintech, 50430–2-AP, 1:1000), anti-eRF1 (Santa Cruz Biotechnology, sc-365686, 1:200), anti-eRF3 (Santa Cruz Biotechnology, sc-515615, 1:200), anti-Rps26 (Proteintech, 14909–1-AP, 1:500), anti-Rpl4 (Proteintech, 11302–1-AP, 1:1000), anti-ABCE1 (Abcam, ab185548, 1:1000), anti-myc (Cell Signaling, 2272 S, 1:1000), anti-RACK1 (BD Transduction Laboratories, 610177, 1:1000), and anti-β-Actin (Sigma-Aldrich, A5441, 1:5000). anti-mouse IgG horseradish peroxidase (HRP)-conjugated secondary antibody (Sigma-Aldrich, A0168, 1:10,000) or anti-rabbit IgG secondary antibody conjugated to peroxidase (Sigma-Aldrich, A9169, 1:10,000).

The templates of EGFP-HiBiT reporters were PCR-amplified from the pcDNA3-EGFP vectors using reverse primers containing the desired sequence. In some cases, the template of reporters based on HiBiT-EGFP was generated using a two-step PCR amplification approach. First, the full length of HiBiT and EGFP was amplified from pcDNA3-EGFP to generate HiBiT-EGFP. The resulting PCR product was used as a template to produce the full-length reporter using a second forward primer containing the respective sequence. After column purification (QIAGEN), the DNA template (1–2 μg) was utilized to generate mRNAs suitable for transfection. For exogenous Rps26, the full-length coding sequence of human Rps26 was cloned into pcDNA3.1(myc-His B) using BamH I and Hind III restriction sites. To create the 18 S rRNA mutant, site-directed mutagenesis was performed using Q5 Site-Directed Mutagenesis Kit (New England Biolabs) according to the manufacturer’s manual. The mutation was confirmed by Sanger DNA sequencing. DNA sequences of all primers used in this study are listed in the Key resources table.

To prepare mRNA reporters, 1–2 μg PCR products described above were utilized as templates to generate mRNAs suitable for transfection. In vitro transcription was performed for 1  hr at 37°C using mMESSAGE mMACHINE T7 Transcription Kit (Invitrogen) followed by poly(A) tailing (Ambion) at 37°C for 30 min. The resulting RNAs were purified using RNA Clean & Concentrator (Zymo Research) following the manufacturer’s instructions and stored at –80 °C.

For mRNA reporter transfection, cells were transfected with in vitro transcribed mRNA (1 μg) in Opti-MEM (125  µl) using Lipofectamine MessengerMAX (1 µl) in Opti-MEM (125  µl), unless stated otherwise. The cells were incubated with the mRNA/Lipofectamine MessengerMAX mixture for 4 hr followed by immunoblotting, HiBiT assay or polysome profiling. For Rps26 or 18 S rRNA overexpression, 2 μg of plasmids were mixed with 4 μl Lipofectamine 2000 (Invitrogen) followed by incubation with cells for at least 24 hr, unless stated otherwise.

Cells grown in a 35 mm dish were transfected with mRNA reporters (1 μg) described above. Transfected cells were washed with PBS and then lysed using a Nano-Glo HiBiT Lytic Detection System (Promega) according to the manufacturer’s instructions. HiBiT signals were measured using a Luminometer (Atto).

Cells grown in a 6-well plate were transfected with transcribed mRNA reporters (1 μg) described above. Transfected cells were washed twice with ice-cold PBS and lysed on ice by adding SDS-PAGE sample buffer (50 mM Tris pH 6.8, 100 mM DTT, 2% SDS, 0.1% bromophenol blue, 10% glycerol), followed by heating for 10 min at 95 °C. Protein samples were separated on SDS-PAGE gels followed by transferring to PVDF membranes (Thermo Fisher Scientific). Membranes were blocked in 5% non-fat milk (Bio-Rad) in TBS containing 0.1% Tween-20 (TBST) for 1 hr, followed by incubation overnight with primary antibodies at 4 °C. After 3×10 min washes in TBST, membranes were incubated with anti-mouse IgG horseradish peroxidase (HRP)-conjugated secondary antibodies at room temperature for 1 hr. Membranes were then washed 3×10 min in TBST at room temperature and visualized using chemiluminescence by exposing to ECL film (GE Healthcare).

To prepare tissue lysates, mouse tissues were dissected and snap-frozen in liquid nitrogen. Frozen tissues were thawed and homogenized on ice with a homogenizer (U.S. Solid) in ice-cold RIPA buffer (150  mM NaCl, 50  mM Tris-HCl pH 8.0, 1% Triton X-100, 0.5% sodium deoxycholate, 0.1% SDS) with 1×protease inhibitors (Roche). After centrifugation at 16,000 g for 20  min at 4 °C, supernatants were collected for immunoblotting as described above. Proteins in mouse tissue fractions collected from sucrose density gradients were precipitated by trichloroacetic acid (TCA) and resuspended in 2× SDS-containing sample buffer. To detect HiBiT-tagged proteins by the Nano-Glo HiBiT blotting system (Promega), cell lysates were resolved by SDS-PAGE and transferred to PVDF membranes as described above. The membrane was incubated with 1× TBST for 1 hr at room temperature, followed by incubation in the LgBiT/buffer solution (50 μl of LgBiT protein in 10 ml of Nano-Glo blotting buffer) at room temperature for 1 hr. Twenty microliters of substrate was added to the incubation solution for an additional 5 min. The membrane was exposed to ECL film in the same manner as the immunoblot analysis.

shRNAs targeting Rps26 and ABCE1 were designed from the BROAD RNAi consortium database and subcloned into DECIPHER pRSI9-U6-(sh)-UbiC-TagRFP-2A-Puro (Cellecta). A scrambled shRNA was used as a control. Lentiviral particles were produced using Lenti-X 293T cells (Clontech). The supernatants containing viral particles were harvested at 48 hr after transfection and filtered through a 0.45 μm Millex-HP filter unit (Millipore). HEK293 cells were transduced with shRNA lentivirus for 48 hr followed by selection with 2 µg/ml puromycin. Knockdown efficiency was detected by immunoblotting using indicated antibodies. The oligonucleotide sequences are listed in the Key resources table.

A total of four plates (10 cm) of HEK293 cells grown to 80% confluency were washed with cold PBS and lysed in the polysome lysis buffer (10 mM HEPES, pH 7.4, 100 mM KCl, 5 mM MgCl₂, 100 μg/ml cycloheximide with 1% Triton X-100). The nuclei were pelleted by spinning at 14,000 rpm for 10 min at 4 °C. For mouse tissues, 100 mg of frozen samples were homogenized on ice using a Dounce homogenizer in 1 ml polysome lysis buffer. Homogeneous lysates were cleared by centrifugation at 14,000 rpm for 10 min at 4 °C. 500 µl of lysates were loaded onto a 15–45% (wt/vol) sucrose density gradient freshly prepared in a SW41 ultracentrifuge tube (Beckman) using a Gradient Master (BioComp Instruments). Samples were centrifuged at 180,000 g for 2 hr 30 min at 4 °C in a Beckman SW41 rotor. Polysome profiles were recorded at A254 using the Brandel Gradient Fractionation System and an ISCO UA-6 UV/Vis detector.

The Ezra-seq has been described previously (Mao et al., 2023). In brief, an aliquot of ribosome fractions representing monosomes or polysomes was collected followed by digestion with E. coli RNase I (Ambion, 750 U per 100 A260 units) by incubation at 4 °C for 1 hr. RNA was extracted using TRIzol LS reagent (Invitrogen) followed by ethanol precipitation. The ribosome-protected mRNA fragments (RPFs) were separated on a 15% polyacrylamide TBE-urea gel (Invitrogen) and visualized using SYBR Gold (Invitrogen). Selected regions in the gel corresponding to 25–35 nt were excised and dissolved by soaking in 400 μl RNA elution buffer (300 mM NaOAc pH 5.2, 1 mM EDTA, 0.1 U/μl SUPERase·In) for 10  min at 70°C. The gel debris was removed using a Spin-X column (Corning), followed by ethanol precipitation. 14 μl RNAs (10–200 ng) was mixed with 1 μl T4 PNK (NEB), 20 U SUPERase·In in 1× T4 PNK buffer and incubated at 37 °C for 30 min followed by 65 °C for 20 min. After ethanol precipitation, 10 μl dissolved RNA was mixed with 1 μl homemade Ezra enzyme, 1 μl Poly(A) Polymerase (NEB), and 20 U SUPERase·In in 7 μl Ezra buffer. After incubation at 37 °C for 30 min followed by 65 °C for 20 min, 1 μl of 1 μM 5′end adaptor, 1 μl of 1 μM biotinylated reverse transcription primer, and 20 U SUPERase·In were added and incubated at 70 °C for 3 min followed by slowly cooling down (3 °C/min) to 25 °C. The hybridized RNA sample was mixed with 10 μl of pre-washed streptavidin beads and incubated at room temperature for 10 min. Beads were washed and re-suspended in 10 μl nuclease-free water. Ligation was performed for 60 min at 25 °C by mixing beads with a 10 μl reaction mixture (1× T4 Rnl2 reaction buffer, 20 U SUPERase·In, 15% PEG8000 and 200 U T4 RNA ligase 2 truncated KQ (NEB)). After washing once with 2× SSC, beads were re-suspended in 12 μl nuclease-free water and mixed with 8 μl cDNA synthesis mixture (5× first strand buffer, 0.1 M DTT, 10 mM dNTP, RNaseOUT, and SuperScript III) followed by incubation at 50 °C for 30 min. After washing once with 2× SSC, beads were resuspended in 10 μl nuclease-free water and incubated at 95 °C for 2 min, then immediately placed on the ice for 1 min. After placing on a magnet stand for 1 min, the supernatant cDNA was amplified by PCR using barcoded sequencing primers. PCR was performed by mixing 1× HF buffer, 0.5 mM dNTP, 0.25 μM PCR primers and 0.025 U Phusion polymerase. PCR was carried out under the following conditions: 98 °C, 30 s; (98 °C, 5 s; 68 °C, 15 s; 72 °C, 10 s) for 12 cycles; 72 °C, 3 min. PCR products were separated on an 8% polyacrylamide TBE gel (Invitrogen). DNA products with the expected size of 180 bp were excised and recovered from DNA elution buffer (300 mM NaCl, 1 mM EDTA). After quantification by Agilent BioAnalyzer DNA 1000 assay, equal amounts of barcoded samples were pooled and sequenced using NextSeq 500 (Illumina). The oligonucleotide sequences are listed in the Key resources table.

A total of four plates (10 cm) of HEK293 cells with 80% confluence were washed three times with ice-cold DPBS. Cells were fixed in ice-cold formaldehyde solution (0.5% in DPBS) for 10  min at 4°C on a rocker. After washing with ice-cold DPBS three times, cells were quenched in ice-cold buffer (50  mM Glycine, 50  mM Tris-HCl, pH 7.5, in nuclease-free water) for 10  min at 4°C on a rocker. Cells were then washed with polysome buffer and lysed in the 400  μl of polysome lysis buffer with 1% Triton-X-100 on ice. Cell debris was removed by centrifugation at 15,000  rpm for 10  min at 4°C. The supernatant was digested with RNase I (Ambion, 750 U per 100 A260 units) for 1 hr at 4°C. Digested supernatant was loaded onto sucrose gradients for polysome profiling as described above. The 80 S fraction was collected (~200  μl total) and mixed with 10  μg eRF1 antibody and 0.5 U/μl SUPERase·In (Invitrogen) followed by incubation under gentle rotation at 4°C for 5  hr. Protein A/G beads were added into the mixture and rotated at 4°C overnight. Beads were washed three times and then resuspended in 600  μl of polysome buffer. RNA was extracted from resuspended beads in polysome buffer. Briefly, samples were brought to room temperature and then adjusted to 10  mM Tris-HCl pH 7.4, 10  mM glycine, 1% (w/v) SDS, and 10  mM EDTA pH of 8.0 followed by incubation at 65°C for 5  min. One volume of acidic phenol/chloroform solution was added and vortexed at maximum speed for 2  min. Mixtures were then placed into a thermomixer and shaken at 1400  rpm for 20  min at 65°C to reverse the cross-links. After centrifugation at 14,000  rpm for 5  min at 4°C, the aqueous phase was precipitated with ethanol. Purified RNA was used for cDNA library construction and high-throughput sequencing as described above.

From the PCR product of HiBiT-EGFP described above, a second PCR was conducted using the pooled oligonucleotide library (IDT) and a primer containing the T7 promoter. The DNA template (1–2 μg) was utilized to generate the mRNA library via in vitro transcription as described above. Cells with 80% confluence were transfected with 6  µg of mRNA library using Lipofectamine MessengerMAX. Cells were lysed 4  hr after transfection, followed by polysome profiling as described above. Fractions of 500  µl corresponding to monosomes or polysomes were collected for RNA extraction using TRIzol LS. RNA was purified using RNA Clean & Concentrator and eluted with 11  µl of nuclease-free water. The purified RNA was reverse transcribed using SuperScript III and gene-specific primers. In brief, RNA samples were mixed with 1 μl of 10 mM dNTP, 2 pmol reverse primer, and incubated at 65 °C for 5 min, then immediately placed on ice for 1 min. The reverse transcription was carried out by incubating with the 7 μl reaction mixture (5× first strand buffer, 0.1 M DTT, RNaseOUT, and SuperScript III) at 50 °C for 60 min followed by heating at 70 °C for 15 min. The products were then amplified with Illumina-based sequencing primers with a barcode. PCRs were performed by mixing 1× HF buffer, 0.5 mM dNTP, 0.25 μM PCR primers, and 0.025 U Phusion polymerase. The PCR was initiated at 98 °C, 30 s; then (98 °C, 5 s; 68 °C, 15 s; 72 °C, 10 s) for 12 cycles; 72 °C, 3 min. The PCR products with the expected size of 190 bp were excised from an 8% polyacrylamide TBE gel. The DNA products were recovered from DNA elution buffer, followed by quantification using Agilent BioAnalyzer DNA 1000 assay. Equal amounts of barcoded samples were pooled for sequencing using NextSeq 500 (Illumina). The oligonucleotide sequences are listed in the Key resources table.

To gain insights about the relative orientations and interaction likelihood between the 5′ end of the mRNA (site –13 to –3 related to the A-site) and the 3′ end of the 18 S rRNA along the ribosome translocation, a structural analysis was carried out using the respective PDBs 6GZ3 (with a 3.60 Å resolution), 6GZ4 (3.60 Å), 6GZ5 (3.50 Å), and 6yal (3.00 Å). The PDBs 6GZ3, 6GZ4, and 6GZ5 encompass three respective intermediate snapshots between the PRE and the POST translocation steps for the eukaryotic 80 S ribosome (hereafter referred simply as PRE and POST states). These three structures (hereafter referred to as translocation intermediates POST 1–3, or simply TI-POST 1–3 for the respective PDBs 6GZ3 – 5) were solved, described, and discussed previously (Flis et al., 2018). The mRNA environment sequentially described by these three PDBs provides a reasonable glimpse about the changes occurring after the recognition of the A site codon and along the movement of the ribosome to the next codon. The PDB 6YAL, in turn, is the Homo sapiens 48 S late-stage initiation complex (Simonetti et al., 2020). Due to the higher resolution of the PDB 6YAL, its mRNA structure is solved until a higher 5′ extension compared to the 6GZ3–5 PDBs (The mRNA 5′ in 6YAL starts from the –18 nt related to the site A, while in 6GZ3 it starts from the site –7 and from the site –6 in 6GZ4 and 6GZ5). In this way, we have made use of the 5′ fragment of the 6YAL mRNA structure to build a rigid body model of the same extension starting from the –13 nt at each one of the PDBs 6GZ3–5, using structural alignment of the respective backbone atoms in each oligonucleotide extremity in PyMOL [Schrodinger, LLC. 2010. The PyMOL Molecular Graphics System, Version 2.5].

The relative local fluctuations of the exit mRNA channel at the ribosome structure and the consequent proximity likelihood between the mRNA 5′ extension and the rRNA 3′ end was estimated in the presence and absence of Rps26 by normal mode analysis (NMA) using anisotropic network modeling (ANM) (Atilgan et al., 2001). For this analysis, we used the atoms from the PDB 6YAL around a 30 Å region centered on Rps26 (depicting the mRNA channel exit) both in the presence and absence of this protein. The PDB 6YAL was chosen due to its higher resolution as a whole, besides longer extension of the mRNA 5′ end compared to the other three PDBs structurally analyzed in this study. Furthermore, the region around 30 Å from Rps26 (hereafter called Rps26 site) encompass a symmetric sphere composed basically of residues from the 40 S subunit (including the 18 S 3′ extension), mRNA (sites –13 to +5 nt related to the A-site), and the anti-codon loop from the P-site tRNA, common to all the four structures here analyzed. Finally, this region in 6YAL presents a relatively small global root mean square deviation (RMSD), considering the protein and RNA backbone, related to both PDBs 6GZ3 and 6GZ5 (1.478 Å in the two cases). In this way, the PDB 6YAL was considered an accurate approximation of the general environment of the Rps26/mRNA 5′ end/rRNA 3′ end triad in ribosome, despite portraying a pre-initiation structure.

The ANM was carried out using the ProDy tools (Bakan et al., 2014; Bakan et al., 2011; Zhang et al., 2021). A Hessian matrix was built upon the backbone atoms of the Rps26 site for both proteins (Cα) and RNA (P, C4’, and C2) with and without Rps26 (hereafter referred to as +Rps26 and –Rps26). The mRNA was considered from the –10 to the +5 nt, once the first three residues (−13 to –11) are more distant from the rRNA 3′ end and free of direct contacts with the neighborhood as a whole, which makes their movements dominate the NMA if they are considered on the ANM (not shown). The Hessian matrix was configured using the default parameters of distance cutoff and gamma function, with the respective values of 15 Å and 1.0 kcal/(mol.Å2). Initially, the first 50 normal modes of each system were estimated by obtaining their respective covariance matrices by diagonalizing the Hessian matrices. The modes simultaneously containing the largest possible eigenvalues and higher mRNA fluctuations at the –10 to –3 extension (directly parallel to the rRNA 3′ end and separated from it by Rps26), as well as the fluctuations of the last 10 residues at the 18 S 3′ end, were selected. In this way, the 15 first ANM normal modes estimated from the +Rps26 and –Rps26 systems were taken for conformational sampling analysis.

The +Rps26 and –Rps26 structures of the Rps26 site from above were taken for sampling alternate conformations along the global fluctuations described by their respective 15 first ANM modes using ProDy (Bakan et al., 2014; Bakan et al., 2011; Zhang et al., 2021). Basically, an extended ANM model containing all the protein and rRNA atoms for each structure was built from the original coarse-grain model containing only the backbone atoms used to build the Hessian matrix. All the atoms in the extended model still obey the movements dictated by the selected 15 first normal modes, with each side chain atom moving in the same direction that the backbone atoms of the residue to which they belong. A set of 70 conformations symmetrically distributed along the fluctuation governed by the 15 first normal modes and with an average RMSD of 2.5 Å related to the input structure was sampled for each one of the +Rps26 and –Rps26 models. Finally, from each original 70-conformer set, a subset containing only the sterically feasible structures (i.e. without significant clashes or conformational distortions) was taken for analysis. Although higher refinements would be necessary to take these final conformers to rigorous molecular dynamics or free energy calculation studies, they provide enough insights about the Rps26 influence on the mRNA 5′ end/ 18 S 3′ end fluctuations and interaction distance likelihood at the ribosome context.

The adaptor of sequencing reads was clipped by Cutadapt, using parameters: -a AAAAAA --max-n 0.1-m 15. The clean reads were then aligned to human transcriptome (GRCh38.81), which contains the protein-coding transcripts with the longest CDS, using STAR with default parameters. To avoid ambiguity, reads mapped to multiple positions or with >2 mismatches were disregarded for further analysis. Ribosome P-site was defined as the positions of 12^th, 13^th, and 14^th from 5′ end of the read (position 0). A-site was defined as the positions of 15^th, 16^th, and 17^th. To generate an aggregation plot around the start and stop codons, for each mRNA, the aligned reads at individual sites were normalized by the mean reads of the CDS. mRNAs with total reads in CDS <16 or the CDS sites covered by footprints <10% were excluded. The normalized values of the sites with the same distance relative to the start codon or stop codon were averaged across the transcriptome.

To identify termination peaks, all reads of eRF1-seq were assigned to individual sites on mRNAs. The mRNAs with <10 total reads from eRF1-seq were excluded. A 120-nt sliding window was used to scan along the mRNA, the sites with terminating reads 10-fold higher than the average reads within the sliding window were defined as the termination peaks.

For each mRNA, all possible uORFs starting with AUG were first extracted. A Wilcoxon test was applied to test whether the in-frame reads are significantly higher than the other two frames. The two p-values were then combined to a single p-value using Stouffer’s method. uORFs with a false discovery rate (FDR)<0.05 were defined as the uORFs with robust translation.

A 30-nt sliding window was used to scan 3′ UTR. For each window, the minimum fold free energy (MFE) was calculated by ViennaRNA [PMID: 22115189] using default parameters.

For each raw sequencing file, the adaptors at both ends were removed by cutadapt. The trimmed reads with length unequal to nine nucleotides were excluded from analysis. The remaining trimmed reads were counted and then an RPM value (reads per million) was obtained by dividing the resultant read count by the total count.

Data are presented as mean ± SEM, unless otherwise stated. At least three independent biological replicates have been performed for each experiment. The number of independent experiments is indicated. Statistical tests used and specific p-values are indicated in the figure legends.