CORD-19:03ea3a614b56409d3f099c9ad764864293132540 JSONTXT 8 Projects

Live-cell single RNA imaging reveals bursts of translational frameshifting Abstract Ribosomal frameshifting during the translation of RNA is implicated in both human disease and viral infection. While previous work has uncovered many mechanistic details about single RNA frameshifting kinetics in vitro, very little is known about how single RNA frameshift in living systems. To confront this problem, we have developed technology to quantify live-cell single RNA translation dynamics in frameshifted open reading frames. Applying this technology to RNA encoding the HIV-1 frameshift sequence reveals a small subset (~8%) of the translating pool robustly frameshift in living cells. Frameshifting RNA are preferentially in multi-RNA "translation factories," are translated at about the same All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint 2 rate as non-frameshifting RNA (~2 aa/sec), and can continuously frameshift for more than four rounds of translation. Fits to a bursty model of frameshifting constrain frameshifting kinetic rates and demonstrate how ribosomal traffic jams contribute to the persistence of the frameshifting state. These data provide novel insight into retroviral frameshifting and could lead to new strategies to perturb the process in living cells. All rights reserved. No reuse allowed without permission. Frameshifting is a fundamental biological process in which a ribosome translating an RNA slips by +/-1 nucleotides, resulting in the translation of an entirely different peptide sequence from that point forward. While frameshifting is generally detrimental to protein fidelity (Belew et al., 2014; Choi et al., 2015) , the process effectively creates two distinct proteins from a single RNA (Clark et al., 2007; Meydan et al., 2017; Yordanova et al., 2015) . Viruses exploit this aspect of frameshifting to minimize their genomes and to successfully replicate in host cells (Brierley and Dos Ramos, 2006; Brierley et al., 1989; Caliskan et al., 2015; Cardno et al., 2015; Mouzakis et al., 2013) . A prototypical example is HIV, which utilizes frameshifting to translate the gag-pol proteins from a single viral RNA (Guerrero et al., 2015) . Although frameshifting has been extensively studied in vitro and in bulk assays (Caliskan et al., 2014; Charbonneau et al., 2012; Chen et al., 2014; Lopinski et al., 2000; Mouzakis et al., 2013; Ritchie et al., 2017) , the process has never been observed at the single molecule level in living cells. This leaves many basic questions about frameshifting unresolved. In particular, it is not clear how heterogenous frameshifting is from one RNA to another, nor is it clear if single RNA continuously frameshift in a constitutive fashion or if instead they frameshift in definable bursts, as has been observed for transcription (Lionnet and Singer, 2012) and 0-frame translation (Wu et al., 2016) . Finally, the localization of frameshifting has never been investigated, so it is not clear if frameshifting occurs all throughout the cell or is instead preferentially localized to specific sub-cellular regions. In the case of HIV-1 gag-pol, for example, previous assays have shown that 5-10% of translated protein product is frameshifted (Brierley and Dos Ramos, 2006; Dulude et al., 2002; Grentzmann et al., 1998; Mouzakis et al., 2013) . This is thought to occur when ribosomes translate through a specialized frameshift sequence (FSS) containing a stem loop structure that slows down incoming ribosomes and causes them to slip back one nucleotide on a slippery sequence preceding the stem loop. Whether or not this occurs constitutively and with equal probability on all HIV-1 RNA or if instead it occurs on a specialized subset that are in the right place, at the right time, and with the right factors remains to be determined. To directly address these sorts of questions, we have developed technology to visualize and quantify single RNA frameshifting dynamics in living cells. Using multi-frame repeat epitopes, complementary high-affinity fluorescent probes that selectively bind the epitopes, and multicolor single-molecule microscopy, we are able to simultaneously monitor the translation of single RNA into two unique nascent polypeptide chains encoded in shifted open reading frames. Application to the HIV-1 FSS uncovers unexpected heterogeneity in the production of frameshifted product and implicates a novel bursty frameshifting mechanism. Besides frameshifting, our technology can now be used to examine other translational regulatory dynamics, including upstream open reading frame selection, non-canonical initiation, and ribosomal shunting. We anticipate multi-frame nascent chain tracking will be a powerful new tool to dissect complex translational regulatory dynamics in living cells and organisms. A multi-frame tag to monitor single RNA translation in two reading frames simultaneously We created a multi-frame (MF) tag to monitor, in living cells, the translation of single RNAs with overlapping open reading frames (ORFs). The tag builds off earlier technology to visualize translation using repeat FLAG or SunTag epitopes labeled by fluorescent Fab or scFv, respectively (Lyon and Stasevich, 2017; Morisaki and Stasevich, 2018) . In the MF tag, FLAG epitopes in the 0 frame are separated from one another by SunTag epitopes in the -1 frame. With this arrangement, single RNAs with ribosomes translating the 0 frame will produce FLAG epitopes labeled by Fab, while those with ribosomes translating the -1 frame will produce SunTag epitopes labeled by scFv. Thus, depending on the chosen frame(s), polysomes will appear all green (all ribosomes translating the 0 frame), all blue (all ribosomes translating the -1 frame), or some combination of the two (Fig. 1A) . To ensure both frames of the MF tag encode functional proteins and have similar ribosomal occupancy, we inserted the first exon of the human GNAS1 locus downstream of the FLAG and SunTag epitopes. The GNAS1 locus contains two overlapping ORFs of roughly equivalent lengths that encode peripheral membrane proteins in the 0 and -1 frames: XXLb1 and AlexX (Abramowitz et al., 2004; Aydin et al., 2009) (Fig. 1A ). In combination with the epitopes, this arrangement has several advantages: 1) epitope densities are maximal since -1 frame epitopes act as linkers for the 0 frame and vice versa; 2) signals are digital, so frameshifted and non-frameshifted species are marked by two distinct probes/colors, and 3) epitopes are placed in nearly equivalent positions, so signals appear at roughly the same time and with similar amplification when translated with similar kinetics. As a first application of the MF tag, we focused on -1 programmed ribosomal frameshifting caused by the HIV-1 frameshift sequence (FSS). We inserted the FSS upstream of our MF tag and transiently transfected the resulting construct into U-2 OS cells. The FSS contains a slippery poly-U stretch nine nucleotides upstream of a stem loop. Two to ten hours post transfection, we observed cells with tens or hundreds of individual RNA diffusing throughout the nucleus and cytoplasm (Fig. 1B) . Nascent Chain To confirm these RNA were active translation sites, or polysomes, we performed two experiments. First, we re-imaged cells 12-24 hours after transfection. At these later time points, Fab and scFv began to accumulate in the cell membrane ( Fig. S1 and Movie this frame encoding a functional protein, as demonstrated by shifting the sequence by one nucleotide into the 0 frame ( Fig. S1 and Movie S2, right panels). Second, we treated cells with the translational inhibitor puromycin. Just minutes after treatment, we observed a dramatic decrease in the number of Fab-and/or scFv-labeled RNA, consistent with the premature release of nascent chains ( Fig. 1D and Movie S3). Together, these data provide strong evidence that we are able to detect single RNA frameshifting dynamics with the MF tag. The HIV-1 FSS structure has been previously shown to produce frameshifted protein with an efficiency of 5-10% based on the dual luciferase assay and similar bulk assays (Brierley and Dos Ramos, 2006; Dulude et al., 2002; Grentzmann et al., 1998; Mouzakis et al., 2013) . However, it remains unclear how this percentage is established. One possibility is that all RNA behave more or less the same and their ribosomes frameshift with 5-10% probability. On the opposite end of the spectrum, it is possible that RNA display a high degree of heterogeneity, so that just 5-10% of RNA have ribosomes that frameshift with nearly 100% probability. A third possibility is that frameshifting is common on all RNA, but frameshifted proteins are less stable and degraded faster than nonframeshifted proteins. As a first step to quantify frameshifting dynamics at the single-molecule level, we tracked thousands of individual RNA approximately two to ten hours post transfection with and without the FSS sequence present (±FSS; Fig. 2A -C). In the +FSS cells, we found 92 ± 1.3% of translation sites were translating the canonical 0 frame alone, while 6.2 ± 1.1% were translating both the 0 and -1 frames. Only rarely did we observe translation All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint sites translating just the -1 frame (1.6 ± 0.5%, Fig. S2 and Movie S4). To ensure these results were not influenced by the multi-frame tag, we reversed the FLAG and SunTag epitopes in the tag and repeated experiments. This gave nearly the same percentages, confirming the tag order and/or epitope positioning did not bias measurements ( Fig. S3 and Movie S5). We then repeated experiments in cells transfected with the control -FSS construct. In this case, we observed virtually no frameshifting sites (0.9 ± 0.7%) ( Fig. 2C and Movie S6). Taken together, these data suggest the FSS alone causes ~8% of translation sites to frameshift. To further characterize the efficiency of frameshifting, we quantified the precise number of frameshifted versus non-frameshifted ribosomes per translation site. To do this, we re-imaged the MF tag at high laser powers so we could simultaneously visualize single mature XXLb1 or AlexX proteins in comparison to their nascent chains within translation sites (Fig. S4 ). After renormalizing the fluorescence, we found frameshifting sites with translation in both the 0 and -1 frames had 7.8 ± 0.7 ribosomes total, 1.9 ± 0.3 of which were frameshifted, while 0-frame only translation sites had 3.2 ± 0.1 ribosomes and -1-frame only translation sites had 3.4 ± 1.4 frameshifted ribosomes (Fig. 2B,D) . Thus, although just ~8% of translation sites contained frameshifted ribosomes, within these subset of sites, anywhere between 30 to 100% of ribosomes were frameshifted. These data support a heterogenous RNA model in which frameshifting occurs on a small subset of RNA with high probability. Given the heterogeneity of observed frameshifting, we hypothesized that the frameshifting state could be stimulated by a specific sub-cellular environment. To test this All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint hypothesis, we performed a statistical analysis of all tracks to see if any biophysical parameters correlated with frameshifting. This revealed frameshifting sites diffuse more slowly than other translation sites or RNA (Fig. 2E ) and were also slightly more likely to be near the nuclear periphery (Fig. 2F) . These results are consistent with a slight preference for frameshifting in the perinuclear endoplasmic reticulum, where RNA are less mobile and more efficiently translated (Voigt et al., 2017; Wang et al., 2016) . Of the parameters we quantified, one of the strongest correlates of frameshifting was RNA intensity. Specifically, sites translating the -1 frame had an average RNA intensity that was nearly 30% brighter than RNA only spots ( Fig. 2B and S5A ). In addition, these brighter frameshifting sites were translation-dependent, dissociating upon puromycin treatment (Movie S7). We wondered if this might be an artifact due to the aggregation of probes. Self-aggregation of Fab was unlikely as this would cause 0-frame translation sites to also be brighter, which was not the case (Fig. 2B) . Similarly, it was unlikely to be due to self-aggregation of scFv, as this would cause 0-frame translation sites to be brighter in the reverse MF tag, which was also not the case (Fig. S5B ). This left the possibility that Fab aggregate with scFv. To rule this out, we reimaged the MF tag without Fab. As frameshifting translation sites remained brighter than other sites ( Fig. S5C ), we conclude the brighter RNA signal is not a tagging artifact, but instead represents a propensity for frameshifting sites to associate with other translating RNA in higher-order complexes reminiscent of "translation factories" (Morisaki et al., 2016; Pichon et al., 2016) . We did not observe factories in cells transfected with the control -FSS construct ( Fig. S5D ). We therefore reasoned the FSS itself might facilitate frameshifting in factories. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https: //doi.org/10.1101/478040 doi: bioRxiv preprint To test this model, we co-transfected a short oligo RNA containing the FSS alone into cells expressing the +FSS MF tag. This led to a significant increase in the fraction of frameshifting sites translating just the -1 frame, from 1.6 to 5.6% when 1 g oligo was added, and up to 9.5% when 4 g oligo was added (Fig. 3A,B) . The significant increase in the -1 only sites suggests RNA can exist in a state in which all or nearly all ribosomes are frameshifted. Moreover, frameshifting continued to occur preferentially within factories, having RNA signals from 30 to 50% brighter than RNA-only spots (Fig. 3C) , even though the overall fraction of RNA within factories remained unchanged (Fig. S6 ). Importantly, in control cells expressing the -FSS construct, we did not see an increase in frameshifting upon oligo RNA co-transfection. Collectively, these data provide further support for a model in which the HIV-1 FSS itself facilitates frameshifting in factories. The ~8% of frameshifted translation sites we observed is consistent with previous measurements of 5-10% frameshifted protein product (Brierley and Dos Ramos, 2006; Dulude et al., 2002; Grentzmann et al., 1998; Mouzakis et al., 2013) . All else equal, this implies that frameshifting alone can explain the steady-state levels of frameshifted protein, without the need for other regulatory mechanisms, such as protein degradation. To test this hypothesis, we measured the elongation rate of frameshifted ribosomes compared to non-frameshifted ribosomes. Fits to the linear portion of run-off curves yielded similar post-tag run-off times of ~386 s and ~475 s, corresponding to average elongation rates of 2.7 ± 0.3 aa/sec and 2.2 ± 0.6 aa/sec, respectively ( Fig. 4A and Fig. S7 ). Fluorescence recovery after photobleaching experiments further confirmed these measurements without the use of drugs ( Fig. S8 and Movie S8). Based on these similar All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint elongation rates for both non-frameshifting and frameshifting sites, a single round of translation would take ~9 minutes. Accounting for the number of ribosomes per translation site and their relative fractions, we calculate a cell with 100 RNA would produce ~115 frameshifted protein per hour compared to ~2,220 canonical proteins. In other words, frameshifted proteins would account for ~5% of the total, in agreement with earlier measurements. Thus, the dynamics of the FSS sequence alone can be sufficient to account for the steady-state levels of frameshifted protein in living cells, without the need for additional regulatory mechanisms. Despite the similar elongation rates, we noticed a slight (~60 s) delay in the run-off response at frameshifting translation sites (Sun sites Fig 4A) . Given earlier work showing the potential of the FSS to pause ribosomes (Dulude et al., 2002) , we envisioned this delay could be due to a queue or traffic jam of ribosomes upstream of the FSS within frameshifting sites. As the backed-up ribosomes clear the traffic jam, they replenish the loss of ribosomes running-off. Only after the traffic jam is fully cleared does the number of ribosomes beyond the FSS (with labeled nascent chains) begin to decay. To test this possibility, we added a 10x HA epitope repeat upstream of the FSS (Fig. 4B ). This served two purposes: first, it allowed us to monitor both ribosomes upstream of the FSS (translating HA epitopes) and ribosomes downstream (translating either 0 frame FLAG or -1 frame SunTag epitopes); second, the arrangement more closely mimicked the natural placement of the FSS between the gag-pol polyproteins. In particular, the additional sequence space upstream of the FSS could accommodate longer ribosomal traffic jams, should they occur. We hypothesized that longer ribosomal All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint traffic jams would lead to longer run-off delays. Consistent with this, ribosomes within frameshifting sites took much longer to run-off (Fig. 4B) , with frameshifted ribosome levels fluctuating but remaining overall steady for upwards of 3000 s, despite an overall ribosome loss (Figs. S9 and S10). We observed this trend even at the single-molecule level, where the total number of ribosomes (marked by HA) dropped, but the number of Such a release can be seen in the single-molecule track at the ~1000 s time point, when the frameshift signal gets significantly brighter. Interestingly, this increase in brightness coincided with the association of an additional non-translating +FSS RNA (gray arrows in Fig. 4C ). This is consistent with the hypothesis that multi-RNA factories stimulate frameshifting. Although difficult to capture, we observed this type of stimulated frameshifting burst in another single frameshifting RNA track as well (Fig. S11 ). To quantify the kinetics of frameshifting, we developed two candidate models and attempted to fit each to our three main observations: (i) the percentages of frameshifting versus non-frameshifting translation sites (Fig. 4C ), (ii) ribosomal occupancies (Fig 4D) , and (iii) run-off kinetics (Figs. 4E,F). Both models include initiation of ribosomes, codondependent elongation of proteins along the RNA template, and ribosomal exclusion to block ribosomes from passing one another or occupying the same place on the RNA. The only difference in the two models is the treatment for how ribosomes shift from the 0 to the -1 frame. The first model assumes constitutive frameshifting, in which all ribosomes frameshift at the FSS with equal probability. This model could capture either observation (i) or (ii), but not both simultaneously; frameshifting either led to excessively large fractions of frameshifting sites or excessively small ribosomal loading, in disagreement with our observations that a relatively small fraction of RNA frameshift with relatively high ribosomal occupancies. Even with addition of distinct pauses in elongation at the FSS in both frames, the constitutive model was unable to fit our data (Fig. S12 ). The second model is inspired by two-state gene models that are commonly used to describe heterogeneous transcription (Munsky et al., 2012) . In this `bursty' model, RNA stochastically switch between non-frameshifting and frameshifting states in which either 0% or 100% of ribosomes produce frameshifted proteins (Fig. 5A ). In the bursty model, the RNA frameshift state is assumed to switch ON and OFF at rates and , respectively, and the steady state fraction of RNA in the ON state is given by = /( + ). We estimated f from the observed fraction of frameshifting translation sites (Fig. 2C ). To estimate the timescale of and , we tracked translation sites for longer periods of time. To achieve this tracking, we doubled the number of epitopes in the multiframe reporter (creating the +FSS 2x multi-frame tag) and changed our imaging strategy to sample the RNA signal at all time points and the 0 and -1 translation signals once every fifth time point. This arrangement substantially reduced photobleaching and allowed us to continuously track and monitor the translational status of single translation sites in 3D for nearly an hour. Fig. 5B shows the frameshifting state survival times for the seven translation sites we tracked in this manner, including one site that frameshifted for longer All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint than 40 minutes ( Fig. S13 and Movie S10), representing at least four rounds of translation at our measured elongation rate of ~2.4 aa/sec. Remarkably, this frameshifting translation site associated with another for a large part of the 40-minute imaging window, which supports our hypothesis that the FSS facilitates frameshifting in multi-RNA sites. From the frameshift state survival times, we fit the rate of to be ~0.0013 sec -1 (Fig. 5B S12 ). To account for the different run-off delays seen at frameshifting and non-frameshifting sites (seen in Fig. 4A ,B), the model required elongation pauses of 1/ ~ 48 s at non-frameshifting sites and 1/ * ~ 90 s at frameshifting sites. We explored if codon usage could also explain the differences in run-off times. However, according to the codon adaptation index (Gorgoni et al., 2016; Sharp and Li, 1987) , which is related to the speed at which each codon is translated in the simulation, there is no notable difference between the 0 and -1 frames ( Fig. S14) . Moreover, the distinct pauses predicted by the model are comparable to those previously measured using in vitro and in vivo bulk assays (Lopinski et al., 2000) . All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint Because the estimated average pause time, 1/ * , is greater than the average initiation time, 1/ , ribosomes could initiate faster than they clear the FSS with an excess rate of − * = 0.009 per second and create upstream traffic jams in frameshifting sites. These traffic jams would continue to build for as long as the RNA remains in the frameshifted state, or approximately 1/ = 770 s on average. This would allow for the accumulation of approximately ( − * ) / ~ 7 ribosomes (on the HA multi-frame tag, Table 1 ). Occasionally, traffic jams can extend all the way back to the start codon, as seen in a sample simulation (Movie S11). However, because these traffic jams occur upstream of the FSS within the HA epitope repeats, the nascent chain intensity at translation sites is only partially correlated with ribosomal occupancy. As a result, the model predicts frameshifting sites are slightly brighter than non-frameshifting sites ( Fig. 5G ), which we also observe experimentally (Fig. S15 ). Moreover, due to the long time it can take to clear traffic jams, frameshifting can persist for hours following the global shut down of translation initiation (with harringtonine, for example), as can be seen in simulations of the best-fit model (Movie S12) and consistent with what we observed in Figs. 4C and S11. Thus, the final bursty model suggests a mechanism by which frameshifting can persist for long periods in the absence of translation initiation. Frameshifting is a common tactic used by viruses to minimize their genomes for faster, more efficient replication in host cells, effectively getting two viral proteins for the price of one viral RNA. While the general architecture of frameshift sequences is well characterized and the dynamics of frameshifting have been measured with single-All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint nucleotide precision in vitro, until now frameshifting had not been directly observed in a living system. To achieve this, we created a multi-frame repeat epitope tag that can light up single RNA translation sites in different colors depending on which open reading frame is being translated. Together with sensitive single-molecule microscopy and computational modeling, we have demonstrated five novel aspects of HIV-1 frameshifting: (i) frameshifted proteins originate from a small subset of RNA that frameshift with high efficiency; (ii) frameshifting occurs preferentially in multi-RNA translation factories that are facilitated in part by the frameshift sequence; (iii) frameshifting occurs in bursts on single RNA that can last for several rounds of translation; (iv) ribosomes that frameshift are paused for longer at the frameshifting sequence than ribosomes that do not frameshift; and (v) pauses at the frameshift sequence induce ribosomal traffic jams that can maintain the production of frameshifted protein despite global inhibition of translation. Fig. 6 summarizes our findings. In contrast to constitutive frameshifting on any RNA, our data indicate that frameshifting occurs in bursts on a subset of RNA. Bursty expression has been demonstrated by others, both at single transcription sites as well as at translation sites in bacteria (Lionnet and Singer, 2012) and eukaryotes (Wu et al., 2016) . The origin of frameshifting bursts is difficult to pinpoint. It is tempting to speculate that there is a unique structure the RNA takes that enhances frameshifting, particularly given the distinct pause times in our best-fit model. According to our observations, multi-RNA translation factories appear to be more susceptible to frameshift bursts. Such factories have been previously observed (Morisaki et al., 2016; Pichon et al., 2016) . Local concentration of translation factors is higher within factories, so translation kinetics are enhanced (Pichon et al., All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint 2016). However, the crowding of machinery at these factory sites may cause them to be more prone to errors and even more so with increased interactions with other FSSencoding RNA, the FSS being a proverbial wrench within translational factory machinery. Such low-fidelity factories may benefit viral robustness, where mutations are desirable. Consistent with this idea, HIV-1 RNAs prefer a dimeric state which may promote translation factory formation (Barajas et al., 2018) . In this study, we observed frameshifting in the context of factories in three different circumstances. First, in puromycin experiments, we observed frameshifting sites separate into two frameshifting sites upon translation shut-down (Movie S7). Second, we observed on two occasions non-translating RNA interact with a frameshifting site, after which the frameshifting signal at the site markedly increased (Figs. 4E and S11). Third, when we co-transfected FSS oligo RNA into cells (Fig. 3) , frameshifting increased in proportion to the amount we co-transfected. Collectively, these data suggest a causal ordering, whereby interactions between FSS-encoding sequences lead to enhanced frameshifting. According to our best-fit model, pauses always occur at the frameshift sequence, with longer pauses associated with frameshifting RNA compared to non-frameshifting RNA. Pausing in and of itself is therefore only a weak predictor of frameshifting, as others have shown (Kontos et al., 2001; Tu et al., 1992) . Nevertheless, the longer pauses associated with frameshifting suggest the frameshift sequence can adopt more than one The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint an elevated ribosomal occupancy as far back as ~400 codons from the FSS, a length that coincides almost perfectly with the length of the FSS-upstream sequence in HIV-1 gagpol. This is not unprecedented, as ribosome profiling experiments have also found evidence for relatively high ribosomal densities as far back as the start codon of the gag protein (Napthine et al., 2017) . These data suggest the strength of the HIV-1 pause may have evolved to on occasion produce the longest possible traffic jams. As clearance of these jams takes time, frameshifting can persist on RNA for hours despite a global shutdown of translation initiation. In effect, the jam acts like a battery that continually fuels the Beyond the imaging of HIV-1 translational frameshifting, the multi-frame tag can now be used in a variety of other contexts. For example, it can immediately be used to investigate frameshifting dynamics along other viral RNA sequences, as well as frameshifting thought to occur along endogenous human mRNA (Cardno et al., 2015) . Likewise, the multi-frame tag can be used to examine other non-canonical translation processes involving more than one open reading frame, including start-codon selection, leaky scanning, ribosomal shunting, and general translation fidelity. In fact, in an accompanying paper, a similar multi-frame tag was used to examine upstream and downstream open reading frame selection (Boersma et al.) . Like us, the authors also saw a high degree of heterogeneity between translating RNAs, with bursts of translation All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint initiation in multiple open reading frames similar to the bursts of frameshifting we observed at the HIV-1 frameshift sequence. Translational heterogeneity may therefore be far more common than originally appreciated, particularly when it comes to non-canonical translation. We therefore believe multi-color single molecule imaging of both canonical and non-canonical translation will become a powerful new tool for dissecting complex RNA regulatory dynamics in a variety of important contexts. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint Dulude, D., Baril, M., and Brakier-Gingras, L. (2002) . Characterization of the frameshift stimulatory signal controlling a programmed -1 ribosomal frameshift in the human immunodeficiency virus type 1. Nucleic Acids Res. 30, 5094-5102. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. is the translation initiation rate, is the translation elongation rate, is the rate at which RNA switch to the frameshifting state, is the rate at which RNA switch to the non-frameshifting state, is the pause rate at the FSS in the non-frameshifting state, * is the pause rate at the FSS in the frameshifting All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint longer ribosomal traffic jams. These allow frameshifting RNA to continue to be translated for up to an hour longer than non-frameshifting RNA upon global translation initiation shutdown. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint RNA (0, 1, and 4 mg) encoding the frameshift sequence were co-transfected into cells with the +FSS multi-frame reporter or the control -FSS reporter. The bar graph above shows the Gaussian fit intensity (normalized to non-translating RNA, i.e. 0 TS sites). (B) The distributions of RNA signals from the four experiments normalized to their mean value. In all +FSS experiments, the fraction of bright RNA (defined here as having an intensity greater than or equal to 1.6 times the mean) remained constant. The -FSS experiment had fewer bright RNA, suggesting the FSS sequence facilitates incorporation into multi-RNA sites. For completeness, an additional experiment was performed to generate the ribosome run-off curve from FLAG epitopes in the HA multi-frame tag. Because we can only image two epitopes at the same time (since RNA is imaged in a third color), we examined FLAG and Sun epitopes (Sun epitopes were required to distinguish frameshifting and nonframeshifting sites). The FLAG run-off from this experiment is on the right (green triangles). This curve is compared to the run-off curve of HA in HA only sites (orange squares) from Figure 4B . Both experiments represent non-frameshifting sites (no Sun signal detected). Error bars represent S.E.M. A single frameshifting translation site encoding the HA multi-frame tag was tracked after harringtonine addition. On the top, a montage of image trims shows the detected RNA-, HA-, and Sun shown signals through time. Below, the positions of the detected signals within the site are plotted through time. On the right, the normalized total intensity of the HA Fab signal (marking all ribosomes) and the SunTag scFv signal (marking frameshifting All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint ribosomes) is plotted through time. At about the fourth timepoint, another non-translating RNA interacts with the frameshifting RNA, leading to a burst of frameshifting signal (Sun). Gray arrows signify a burst of frameshifting, coinciding with multi-RNA interactions. and frameshifting sites (gray), plotted with data in Fig. 4A (black triangles and gray circles). On the right, simulated run-off (solid lines) from the start site (AUG) of all ribosomes in non-frameshifting (black) and frameshifting sites (gray), plotted with data in The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint channels separated to highlight two frameshifted RNA that split from a single translation site. On the right, the position of the spot is plotted through time. Intensity distribution of non-frameshifted HA signals (black) versus frameshifted HA signals (gray) produced from the HA multi-frame tag. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint Movie S1. Dynamics of cell presented in Figure 1B Max projection of a 13 z-stack movie showing an example U-2 OS cell expressing the +FSS multi-frame tag and bead-loaded with probes: MCP-Halo JF646 (Ch1 red, RNA), FLAG Fab Cy3 (Ch2 green, 0 ORF), and scFv-sfGFP (Ch3 blue, -1 ORF). Images were acquired every 4 seconds for a total of 76 seconds. Circled regions of interest correspond to species identified in Figure 1B and C. Scale bar, 10 µm. Single plane movies showing example U-2 OS cells expressing either +FSS multi-frame tag (left), -FSS control multi-frame tag (middle), or -FSS (+1 nt) tag with XXLb1 shifted into the 0 frame (right). All cells were bead-loaded with probes: MCP-Halo JF646 (Ch1 red, RNA), FLAG Fab Cy3 (Ch2 green, AlexX), and scFv-sfGFP (Ch3 blue, XXLb1). Images were acquired every 7 seconds for a total of 42 seconds. Scale bar, 10 µm. Figure 1D Max projection of a 13 z-stack movie showing an example U-2 OS cell expressing the +FSS multi-frame tag and bead-loaded with probes: MCP-Halo JF646 (Ch1 red, RNA), FLAG Fab Cy3 (Ch2 green, 0 ORF), and scFv-sfGFP (Ch3 blue, -1 ORF). Images were acquired every 7 seconds for a total of 112 seconds prior to puromycin and 280 seconds following puromycin. Scale bar, 10 µm. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint Max projection of a 13 z-stack movie showing an example U-2 OS cell expressing the +FSS multi-frame tag and bead-loaded with probes: MCP-Halo JF646 (Ch1 red, RNA), FLAG Fab Cy3 (Ch2 green, 0 ORF), and scFv-sfGFP (Ch3 blue, -1 ORF). Images were acquired every 7 seconds for a total of 63 seconds. Circled regions of interest correspond to species identified in Fig. S2 . Scale bar, 10 µm. Max projection of a 13 z-stack movie showing an example U-2 OS cell expressing the +FSS reverse multi-frame tag and bead-loaded with probes: MCP-Halo JF646 (Ch1 red, RNA), FLAG Fab Cy3 (Ch2 green, -1 ORF), and scFv-sfGFP (Ch3 blue, 0 ORF). Images were acquired every 7 seconds for a total of 63 seconds. Circled regions of interest correspond to RNA with no translation (numbered '1'), 0 and -1 frame translation site (numbered '2'), and 0 frame only translation site (numbered '3'). Scale bar, 10 µm. Max projection of a 13 z-stack movie showing an example U-2 OS cell expressing the -FSS control tag and bead-loaded with probes: MCP-Halo JF646 (Ch1 red, RNA), FLAG Fab Cy3 (Ch2 green, 0 ORF), and scFv-sfGFP (Ch3 blue, -1 ORF). Images were acquired every 7 seconds for a total of 63 seconds. Circled regions of interest correspond to RNA with no translation (numbered '1'), and 0 frame only translation site (numbered '2'-'8'). Scale bar, 10 µm. Figure 1D All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint Max projection of a 13 z-stack movie showing an example U-2 OS cell expressing the + FSS multi-frame tag and bead-loaded with probes: MCP-Halo JF646 (Ch1 red, RNA), FLAG Fab Cy3 (Ch2 green, 0 ORF), and scFv-sfGFP (Ch3 blue, -1 ORF). Images were acquired every 7 seconds for a total of 217 (of 280) seconds total post puromycin addition. This movie shows the same cell from Movie S3, but now zoomed in to accentuate a frameshifting site splitting (red arrows). Scale bar, 10 µm. Max projection of a 13 z-stack movie showing an example U-2 OS cell expressing the + FSS 2x multi-frame tag and bead-loaded with probes: MCP-Halo JF646 (Ch1 red, RNA), FLAG Fab Cy3 (Ch2 green, 0 ORF), and scFv-sfGFP (Ch3 blue, -1 ORF). Images were acquired every 10 seconds for a total of 100 seconds pre-FRAP and 790 seconds post-FRAP. Control spot (far left red arrow) and two frameshifting sites (center arrows) were tracked through time to measure their fluorescence (0 frame translation) after photobleaching. Scale bar, 10 µm. Max projection of a 13 z-stack movie showing an example U-2 OS cell expressing the HA multi-frame tag and bead-loaded with probes: MCP-Halo JF646 (Ch1 red, RNA), HA Fab Cy3 (Ch2 green, all ORFs), and scFv-sfGFP (Ch3 blue, -1 ORF). Images were acquired post harringtonine every 120 seconds for a total of 3480 seconds. Scale bar, 10 µm. Movie S10. Dynamics of tracked frameshifting sites presented in Figure S13 All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. Movie S12. Sample simulation of best-fit bursty model of frameshifted and nonframeshifted translation sites during harringtonine run-off All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint A best-fit bursty model simulation of multiple non-frameshifting (top four labeled 'nonFS Spots') and frameshifting (bottom four labeled 'FS Spots') translation sites encoding the HA multi-frame tag. In the simulation, harringtonine is added after 1000 s and the simulation continues for an additional 2000 s. Individual ribosomes (gray dots) are shown moving along the reporters (gray lines), with green, blue, and cyan squares marking the positions of FLAG, SunTag, and HA epitopes, respectively. As the ribosomes move along the repeat epitopes, the translated nascent chains appear as green (FLAG-tagged nascent chains), blue (Sun-tagged nascent chains), and cyan (HA-tagged nascent chains) dots that increase in size as more epitopes are translated. When the RNA is in the non-frameshifting state, the FSS is labeled 'FSS(OF)' in green; when the RNA is in the frameshifting state, the FSS is labeled 'FSS(-1F)' in blue. The simulated time is shown to the right of each reporter. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint Requests for further information and resources should be directed to and will be fulfilled by the corresponding authors, Brian Munsky (brian.munsky@colostate.edu) and Tim Stasevich (tim.stasevich@colostate.edu). Key plasmids will be deposited on Addgene. The HIV-1 frameshift sequence (FSS) followed by either the multi-frame (MF) tag or the reverse multi-frame (revMF) tag were synthesized by GeneArt® gene synthesis service (Thermo Fisher Scientific). The gene fragments were flanked by NotI and NheI, and fused to the upstream of the beta-actin zipcode and 24x MS2 stem loops in the 3' UTR of plasmid pUB_smFLAG_ActB_MS2 (Plasmid #81083, addgene) to obtain the FSS-MF and FSS-revMF, respectively. To double the MF tag, the MF tag region was digested out from FSS-MF with XbaI and AgeI, and then ligated into FSS-MF flanked with NheI and AgeI. The open reading frame encoding human XXLb1/AlexX (Abramowitz et al., 2004; Aydin et al., 2009 ) was amplified from U-2 OS cells cDNA with the primers: 5'-GTT GTC ATA TGG GCG TGC GCA ACT -3'; 5'-GAT GTA GCT AGC CTA GAA GCA GCA GGC GGT G -3'. The amplified XXLb1/AlexX was flanked with NsiI and NheI, and then inserted into the C-terminal region of the FSS-MF, FSS-2xMF and FSS-revMF to obtain FSS-MF-AlexX (i.e. the +FSS multi-frame tag), FSS-2xMF-AlexX (i.e. the +FSS 2x multi-frame tag), and FSS-revMF-AlexX (i.e. the +FSS reverse multi-frame tag), respectively. To All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint Fab generation and dye-conjugation Fab generation was done using the Pierce mouse IgG1 preparation kit (Thermo Fisher Scientific) per the manufacturer's instructions. Briefly, beads conjugated with ficin were incubated in 25 mM cysteine to digest FLAG (Wako, 012-22384 Anti DYKDDDDK mouse IgG2b monoclonal) or HA (Sigma-Aldrich, H3663 HA-7 IgG1 mouse monoclonal) antibodies to generate Fab. Fab were separated from the digested Fc region using a NAb Protein A column (Thermo Scientific, product # 1860592). Fab were concentrated to ~1 mg/ml and conjugated to either Cy3 or Alexa Fluor 488 (A488). Cy3 Nhydroxysuccinimide ester (Invitrogen) or A488 tetrafluorophenyl ester (Invitrogen) was suspended in DMSO and stored at -20°C. 100 µg of Fab were mixed with 10 µL of 1M NaHCO3, to a final volume of 100 µL. 2.66 µl of Cy3 (or 5 l of A488) was added to this 100 µL mixture and incubated for 2 hours at room temperature with end-over-end rotation. The dye conjugated Fab were eluted from a PBS equilibrated PD-mini G-25 desalting column (GE Healthcare) to remove unconjugated dye. Dye conjugated Fabs then were concentrated in an Ultrafree 0.5 filter (10k-cut off; Millipore) to 1 mg/ml. This conjugation and concentration process was repeated on occasion to ensure a degree of labeling close to one. The ratio of Fab:dye, , was determined using the absorbance at 280 and 550 nm or 495 nm, the extinction coefficient of IgG at 280 nm, Ɛ , the extinction coefficient of the dye, Ɛ , provided by the manufacturer, and the dye correction factor at 280 nm, , provided by the manufacturer. The degree of labeling, , was calculated with the following formula: All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint sfGFP and 33 µg/ml of purified MCP-HaloTag protein were mixed with PBS to a final volume of 4 µl. Second, in cell culture hood, DMEM was aspirated from the MatTek chamber and the 4 µl mix was pipetted on top of cells. Third, ~106 µm glass beads (Sigma Aldrich) were evenly sprinkled over the cells. Fourth, the chamber was tapped carefully ~10 times on the cell culture hood bench top. Fifth, DMEM was immediately added back to the cells. Sixth, cells were returned to the incubator for at least an hour to recover from the loading procedure. In most experiments, we also bead-loaded DNA, which we added to the initial 4 µl mix (so that DNA had a final concentration close to 1 mg/ml). On occasion, DNA was transiently transfected ~2 hours before bead-loading. Around one hour before experiments began, bead-loaded cells were washed with phenol-red-free complete DMEM to remove glass beads, and 200 nM of JF646-HaloTag ligand (a cell permeable fluorogenic ligand (Grimm et al., 2015) ) was added to label MCP-HaloTag protein. After 30 mins of incubation, cells were washed three times using phenol-red-free complete DMEM to remove any unconjugated fluorophores. Cells were then immediately imaged for experiments. To track single molecule translation sites, a custom-built widefield fluorescence microscope based on a highly inclined illumination scheme (Tokunaga et al., 2008 ) was used (Morisaki et al., 2016 . Briefly, the excitation beams, 488, 561 and 637 nm solidstate lasers (Vortran), were coupled and focused at the back focal plane of the objective lens (60X, NA 1.49 oil immersion objective, Olympus). The emission signals were split by an imaging grade, ultra-flat dichroic mirror (T660lpxr, Chroma) and detected using two All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. Readout time for the cameras from the combination of our imaging size, readout mode, and the vertical shift speed was 23.36 msec, resulting in an imaging rate of 13 Hz (70 msec per image). The excitation laser lines were digitally synched to ensure they only illuminated cells when the camera was exposing in order to avoid excessive photobleaching. To capture the entire volume of the cytoplasm of U-2 OS cells, 13 zstacks with step size of 500 nm (6 µm in total) were imaged using the piezoelectric stage such that the z-position changed every 2 images (one image for Cy3 and one for All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint sfGFP/A488 + JF646). The position of the filter wheel was changed during the camera readout time. This resulted in a total cellular imaging rate of 0.5 Hz (2 sec per volume for 3-colors). Note that all colors described in the text and that are shown in the figures are based on the color of the excitation laser: RNA in red (JF646) and protein in green (Cy3) or blue (sfGFP/A488). Images were first pre-processed using either Fiji (Schindelin et al., 2012) or a customwritten batch processing Mathematica code (Wolfram Research) to make 2D maximum intensity projections from 3D images. Pre-processed images were then analyzed with a custom-written Mathematica code to detect and track particles. Specifically, particles were emphasized with a band-pass filter so the positions could be detected using the built-in Mathmatica routine ComponentMeasurements "IntensityCentroid". Detected particles were linked through time by allowing a maximum displacement of 5 pixels between consecutive frames. Particle tracks lasting at least 5-10 frames were selected and their precise coordinates were determined by fitting (using the built-in Mathematica routine NonlinearModelFit) the original 2D maximum intenstiy projected images to a 2D Gaussians of the following form: where is the background intensity, the particle peak intensity, ( , ) the spread of the particle, and ( 0 , 0 ) the particle location. The offset between the two cameras was All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint registered using the built-in Mathematica routine FindGeometricTransform to find the transform function that best aligned the fitted positions of 100 nm diameter Tetraspeck beads evenly spread out across the image field-of-view. We did not register the images, but only the fitted positions in order to avoid introducing any distortion into images. This is why a slight offset can be observed between the red and the green/blue particles even though they are within a diffraction limited spot, according to our registration. For visuallization and some quantification, average intensity image trims were created by averaging images of all detected particles of a given species (each centered by their intensity centroid). To compute the average number of ribosomes in a translation site (see Fig. S4 ), the average image trims of translation sites and single mature proteins were fitted to Eq. (2). The average intensities were calculated by integrating the fitted Gaussian. The ratio of the average intensity of translation sites to that of single mature proteins was then calculated to give the average number of ribosomes in a translation site. The average intensity of single RNAs were also calculated using the same procedure ( Fig. 3C , S5, and S6). Fits were susceptible to noise, so we also used an alternative strategy to determine the average intensity of translation sites and RNA that was robust to noise (in all Figs aside from Fig. 3C , S4, S5, and S6). Specifically, the intensity of centered images of RNA or translation sites was calculated to be the average intensity within a centered radial spot of four pixels in diameter minus the average background intensity from a centered ring with an outer diameter of twelve pixels and an inner diameter of eight pixels. Mathematica source code is available upon request. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https: //doi.org/10.1101/478040 doi: bioRxiv preprint After RNA particles were identified and tracked using the custom Mathematica code described above, an average centered image of the first five frames from each track was created for RNA (JF646), 0 ORF (FLAG-Cy3 in the +FSS multi-frame tag), and -1 ORF (scFv-sfGFP in the +FSS multi-frame tag). The trims were then hand checked to remove any trims with artifacts, e.g., smears or non-diffraction limited spots. Next, a custom Mathematica code was used to detect particles in the 0 ORF or -1 ORF trim channels, sorting the spots into RNA only, 0 frame translation sites (0 TS only), 0 and -1 TS, and -1 only TS. For all cases, RNA always had to be present. Finally, frameshifting translation sites (the 0 and -1 TS or the -1 only TS) were validated by eye, to further remove artifacts. For example, RNA that briefly colocalized with a mature protein puncta were removed at this stage. After all sites were validated, the total count of each type of species was used to determine the percentage of non-translating RNA (no TS), 0 frame translation sites (0 TS), 0 and -1 TS, and -1 only TS. To confirm active translation elongation, puromycin (Sigma Aldrich) was used to release nascent chains from elongating ribosomes, leading to a rapid loss of nascent chain signal at translation sites. Bead-loaded cells with visible translation sites were imaged at a rate of one volume every seven seconds. After acquiring 16 pre-images, cells were treated with a final concentration of 100 µg/ml puromycin and continuously imaged for an additional 100 time points. As a control, the same imaging conditions were performed except that the cells were treated with vehicle (H2O). In this case, nascent chain signals All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint did not disappear (data not shown). Both frameshifting and non-frameshifting translation sites were monitored through time using the tracking code described above. Oligo FSS RNA co-transfection An uncapped RNA oligo containing the FSS from HIV-1 was synthesized from IDTDNA with the following RNA sequence: 5'-CUG GCC UUC CCU UGU GGG AAG GCC AGA UCU UCC CUA AAA AA -3'. Co-transfection of FSS-MF-AlexX and the RNA oligo was carried out via bead-loading, as describe above. Briefly, 1 µg plasmid FSS-MF-AlexX construct DNA, 100 µg/ml of fluorescently labeled Fab, 250 µg/ml of purified (GCN4) scFv-sfGFP or 1 µg of (GCN4) scFv-sfGFP plasmid, 1 or 4 µg of RNA oligo, and 33 µg/ml of purified MCP-HaloTag protein were mixed with PBS to a final volume of ~4 µl of PBS. To measure the average elongation rate, harringtonine (Cayman Chemical) was used to block translation initiation and iduce the run-off of all actively elongating ribosomes, leading to a gradual loss of nascent chain signal at translation sites. Beadloaded cells with visible translation sites were imaged as described above except that cell volumes were acquired every 60 or 120 seconds. After acquiring 5 pre-images, cells were treated with a final concentration of 3 µg/ml Harringtonine and continuously imaged for 30 more time points. As a control, the same imaging conditions were performed except that cells were treated with vehicle (DMSO). In this case, nascent chain signals did not decay (data not shown). The intensities of translation sites measured, as described above. The All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint intensity of all translation sites in each frame (and in all cells) were then totaled to produce the run-off curve. Run-off curves were normalized to the mean of the total intensity of the first four time points after treatment of harringtonine began. These curves were then fit to a linear regression to estimate run-off times (see Fig. S7 ). The linear portion of the runoff decay begins when the normalized run-off intensity reaches a fraction 0 , where where 1 is the length of untagged portion of the open reading frame and 2 is the length of the tagged portion (i.e. the length of the repeated epitopes), as described (Morisaki and Stasevich, 2018). The linear portion of the decay was then interpolated to background levels to estimate the run-off time . The elongation rate was calculated as follows: Fluorescence recovery after photobleaching (FRAP) To confirm elongation rates of frameshifting translation sites, FRAP experiments were performed. Bead-loaded cells with visible translation sites were imaged once every 10 seconds, with an intentional photobleach at frame 10. The photobleach was performed with a 405 nm laser focused to a spot roughly 5 m in diameter and operating at a minimal power such that the nascent chain signal did not completely vanish. This allowed us to track the translation site continously throughout the experiment. Following the intentional photobleach, the fluorescence recovery of translation sites within the bleach zone were monitored for an additional 80 time points. These translation sites were tracked and their intensities quantified, as described above. To correct for unintentional photobleaching, All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint the intensity of a control translation site was also tracked and quantified. The loss of signal from the control translation site was fit to a single exponential decay and this decay was divided out from the FRAP recovery curves of the intentionally photobleached translation sites. The FRAP recovery curve can be thought of as the inverse of the harringtonine runoff curve (Morisaki and Stasevich, 2018) . In this way, the FRAP recovery curve was fit to determine the average translation elongation rate (Fig. S8 ). The frameshift models describe the stochastic dynamics of nascent translation with single-codon resolution. Both the constitutive and bursting models are formulated to track an arbitrary number of individual ribosomes that can perform three possible reactions: (i) A new ribosome can initiate translation at the start codon at rate kini. (ii) An existing ribosome can elongate at rate ke to incorporate one amino acid into the nascent peptide chain. The rate ke is sequence-specific with each codon's rate scaled by the genomic copy of the corresponding codon (Nakamura et al., 1999) . When the ribosome completes the final amino acid, translation is terminated and that ribosome is eliminated. (iii) A ribosome at the frameshift sequence can shift from the 0 to the -1 frame. In addition to these three reactions, ribosomes at the frameshift sequence can pause for an average time of 1/ * or 1/ for the shifted and non-shifted states. For the bursting model, each RNA is assumed to switch back and forth between non-frameshifting and frameshifting states with rates and . With these mechanisms and parameters, the models can be analyzed using either simplified approximations or detailed simulations. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint Several approximate features can be derived directly from the bursting model parameters (see Table 1 ). The average burst time or time spent in a frameshifting state is: The average time spent in a non-frameshifting state is: The fraction of RNA in a frameshifting state is: The number of ribosomes that initiate during the frameshifting state is: The number of ribosomes that initiate during the non-frameshifting state is: The time for a ribosome to clear the FSS in the frameshifting state is: where the second term accounts for ribosome pileup in the 9-codon ribosomal exclusion (Ingolia et al., 2009) , , region upstream from the FSS. Similarly, the time for a ribosome to clear the FSS in the non-frameshifting state is: All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint One can approximate = 1 or 0 if initiation is faster ( > ) or slower ( > ) than the FSS pause. The number of ribosomes to pass the FSS in the frameshifting state The number of ribosomes to pass the FSS in the non-frameshifting state is: The average number of ribosomes to accumulate upstream of the FSS in the frameshifting state is: * = * − * and the number to clear in a non-frameshifting state is: In equations (13) and (14), positive values of correspond to accumulation of ribosomes in the traffic jam, and negative values correspond to clearance of ribosomes. The average elongation time for a ribosome in the frameshifting state is: and the number to clear in a non-frameshifting state is: where is the gene length in codons. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint Simulations were started at t = -10,000 seconds to approximate steady state at t = 0 using the Gillespie algorithm (Gillespie, 1976) . Ribosome densities were found by collecting position statistics for multiple simulations. Simulated ribosome numbers and positions and multi-frame tag probe locations were combined to estimate translation site intensities. Harringtonine assays were simulated by preventing the initiation reaction at the time of treatment. Parameter estimation was performed using genetic algorithms and a multiple-objective cost function that considers the frameshifting efficiency, the number of ribosomes per RNA and the Harringtonine assays. A detailed description of the computational methods and codes is given in the 'Computational Details' section below. Both the constitutive model and the bursting model consist of three general ribosomal reaction types: initiation ( 0 ) , elongation ( or ̂) and frameshifting ( or ̂) as depicted in Eq. 17, Ribosome Initiation. Ribosomes are assumed to bind to the RNA with rate . To account for the fact that ribosomes are large biomolecules that occupy around 20 to 30 All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint nuclear bases on the RNA (Ingolia et. al, 2009) , initiation is blocked by any downstream ribosome within = 9 codons (in either frame). Therefore, initiation rate is set at: Ribosome Elongation. Each ribosome moves along the RNA codon by codon in the 5' to 3' direction. The elongation rate for each ℎ codon, ̂ is assumed to be: where denotes the codon usage frequency in the human genome obtained from (Nakamura et. al, 2000) , and ̅ represents the average codon usage frequency in the human genome. The fit parameter specifies the average elongation rate. Ribosomal termination is assumed to be equivalent to elongation of the final codon. or the ribosome can continue in the -1 frame at rate: All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. For the bursting model, the decision to continue in the 0 or -1 frame depends upon the frameshifting state with the rate given by: Relating model dynamics to experimental fluorescence intensity. To relate ribosome occupancy to experimental translation spots, we proscribe an assumed fluorescence to each ribosome based upon its position. This intensity is proportional to the number of peptide probes upstream from the ribosome location and in the appropriate frame or frames. Ribosomes in the 0 frame include all upstream probes in the 0 frame. Ribosomes in the -1 frame include probes in the 0 frame between the start codon and the FSS (if any) and probes in the -1 frame between the FSS and the current ribosome position. The total intensity vector for the ℎ color is given by where ( ) denotes the position (i.e., frame and codon location) of the ℎ ribosome, and ( ) is the corresponding intensity of the ℎ color. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint Fraction of translation spots. Experimental data was measured between 2 to 6 hours post bead-loading for a period of 120 seconds. To reproduce these experimental data, the model was solved using 100 trajectories of 120 seconds starting at steady state. Spots were classified as 0 frame only ( 0 ) , -1 frame only ( −1 ) and both frames ( ) . Following the experimental observation that frameshifting spots commonly consist of more than one RNA, were equally added to the 0 and −1 . The percentage of nonframeshifting translation spots was calculated as follows: where 0 is the sum of 0 and is the total number of translating spots. The percentage of frameshifting translation spots was calculated as follows: * = ( −1 + ( 1 2 ⁄ ) ) • 100, where −1 is the sum of −1 . Number of ribosomes per RNA. Experimental observations allowed to quantify the number of ribosomes in frameshifting and non-frameshifting translation spots ( Figure 2 in the main text). To relate model simulations and data, the number of ribosomes per RNA was estimated by directly by adding the number of ribosomes in each RNA at steady state. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint Harringtonine assays. Harringtonine inhibits translation by binding to the ribosomal 60S sub-unit, which blocks new initiation events. Experimental data showed that Harringtonine causes the intensity in translating spots to drop to a basal intensity value after a run-off time (Figure 4 in main text). To mimic the effects of Harringtonine in our model we modified the initiation rate as: < and codons {1, … , } unoccupied, where is the time of application of Harringtonine. After Harringtonine application, spots simulated from the original construct were classified as 0 frame only ('0F') or both 0 frame and -1 frames ('BF'). After classification, average spot intensities were quantified as: where (.) is experimental background expression obtained at the end of the experimental time of the run-off assays. Similarly, spots simulated for the extended construct with upstream HA tags and downstream -1 tags were classified as HA in non-shifted spots ('HA') or HA in shifted spots ('HA*'). All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint ( ) = 〈 ( )〉 + , (29) * ( ) = 〈 * ( )〉 + * . Parameter estimation. For each model, we sought to find a single parameter set that reproduces all experimental data. Given the different sources of experimental data we estimated the parameter values using a multi-objective optimization strategy to simultaneously compare the Harringtonine run-off data, the fractions of shifted and nonshifted spots and the ribosome occupancy, as follows: where experimental data are denoted as ̃ and simulations results as ̂. Weights ( , and ) were used to balance constraints by the different experiments and time points. represents the number of experimental points in the Harringtonine assays ( = 30), is the number of studied frames ( = 2), is the number of different gene constructs ( = 2), and is the number of different data types ( = 3). The weight used on the Harringtonine data set was defined as = 1 ( • • ) ⁄ . The weight in the fraction of spots per frame was defined as = 1 ( + * ) ⁄ . And the weight for the number of ribosomes was defined as: = 1 ( + * ) ⁄ . All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint Parameter searches. Parameter searches consisted of optimization routines based on genetic algorithms (GA) using a population of 100 individuals and running the search for 20 generations. The constitutive model has a total of four parameters ( , , and * ), whereas the bursting model has two additional parameters ( , ). The optimization strategy consisted in to fit four free parameters ( , , and * ) in the models, parameters ( , ) were independently determined from data in Fig. 5B ,C). The parameter set that best reproduces the data was selected as: = argmax − log 10 ( ) . Optimized parameter values are given in full detail in All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. . https://doi.org/10.1101/478040 doi: bioRxiv preprint