In order to study translation and, more specifically, stop codon readthrough in D. melanogaster, we sought to develop a robust ribosome profiling assay for this organism. We initially developed our protocol in S2 cells, a macrophage-like lineage derived from late-stage Drosophila embryos.

In previous studies, ribosome-protected fragments or ‘footprints’ were generated by digesting eukaryotic polysome lysates with RNase I (Ingolia et al., 2009, 2011). In contrast to yeast and mammalian cell lines, we found that Drosophila ribosomes are highly sensitive to RNase I, potentially due to their unusual rRNA sequences and structures (Figure 1—figure supplement 1A; Hancock et al., 1988; Jordan, 1975; Jordan et al., 1976; Pavlakis et al., 1979). By contrast, we found that Drosophila ribosomes tolerate micrococcal nuclease (MNase) over a wide range of concentrations (Figure 1—figure supplement 1B–D). In contrast to RNase I, MNase has a strong 3′ A/T bias. This gives rise to a small amount of positional uncertainty with P-site mapping in MNase datasets, and prevents us from achieving the sort of sub-codon resolution seen in ribosome profiling datasets generated with RNase I.

Nonetheless, replicate experiments established that our measure of translation rate (the ribosome footprint density, defined as the number of ribosome-protected fragments per kilobase of coding region per million aligning reads in the dataset; RPKM), is highly reproducible and insensitive to changes in buffer conditions (Figure 1—figure supplement 1E, Figure 1—figure supplement 2A,B; full data in supplementary table 1 at Dryad: Dunn et al., 2013). Focusing on coding regions that had a minimum of 128 reads, we observed strong correlation between replicates (r² = 0.998; Figure 1—figure supplement 2) and an inter-replicate standard deviation of 1.07-fold, comparable to our protocols in yeast and mammalian cells. Furthermore, our measurements are robust to the number of isoforms per gene, the fraction of sequence-degenerate positions in a gene, gene length, A/T content, and distribution of ribosome density within a gene (Figure 1—figure supplement 3).

In early (0–2 hr) Drosophila embryos, the vast majority of transcripts are maternally supplied and therefore regulated by post transcriptional processes, such as poly- or deadenylation, capping or de-capping, localization, degradation, and control of translation initiation. The early Drosophila embryo has thus been an important system for the study of post-transcriptional and specifically translational regulation (reviewed in Lasko, 2011).

To enable the broad analysis of these processes, we developed a sample harvesting strategy that captures the translational state of early embryos with minimal perturbation. Specifically, we developed a cryolysis protocol in which embryos are collected directly from egg-laying dishes into liquid nitrogen, homogenized while frozen, and thawed in the presence of translation inhibitors to prevent post-lysis translation. Notably, we omit dechorionation and rinsing, steps which could induce cold shock, anoxia, and related translational artifacts.

We collected replicate samples of 0–2 hr embryos, and subjected them to ribosome profiling and RNA-seq of poly(A)-selected mRNA. A subset of ribosomes partition into heavy polysomes (Figure 1A), consistent with reports that a distinct subset of messages is well-translated at this stage (Qin et al., 2007). Ribosome density measurements from replicate embryo collections are correlated nearly as well (r² = 0.984; Figure 1B; supplementary table 1 at Dryad: Dunn et al., 2013) as measurements from technical replicates from a single culture of S2 cells (r² = 0.998; Figure 1—figure supplement 1E). The Drosophila embryo thus provides a system in which experimental noise approaches the precision of our measurements, a fact that will facilitate detection of even small expression differences between wild-type and mutant fly strains.

Figure 1 with 4 supplements see all

Download asset Open asset

Translational control is measured by a gene’s translation efficiency, estimated as the ratio of ribosome footprint density (from ribosome profiling) to mRNA abundance (from mRNA-seq) for each gene. Translation efficiency measurements between replicate embryo collections are highly reproducible (r² = 0.946; Figure 1C) and consistent with prior measurements made by semiquantitative methods (Figure 1—figure supplement 4). The standard deviation of fold-changes between biological replicates is 1.19-fold (Figure 1C, inset), allowing detection of even modest changes in translation efficiency.

Remarkably, we find that the range of translation efficiencies for different messages spans four orders of magnitude, a range comparable to that observed for mRNA abundance of well-counted genes (Figure 1D). Moreover, translation efficiency is uncorrelated with mRNA abundance (r² = 8.29 × 10⁻⁵; Figure 1E) and mRNA abundance predicts only one third of the variance in the rate of protein production as measured by ribosome footprint density (Figure 1F). Translational regulation is therefore a major determinant of gene expression in the early embryo (supplementary table 1 at Dryad: Dunn et al., 2013), and ribosome profiling provides a quantitative and robust means to monitor translational regulation during development.

In addition to measuring gene expression, ribosome profiling maps the physical positions of ribosomes on each transcript, and thus provides a powerful tool to annotate which portions of mRNAs are translated. Consistent with our previous work in mammals (Ingolia et al., 2011) and yeast (Ingolia et al., 2009; Brar et al., 2012), many 5′ UTRs in Drosophila contain substantial footprint density (Figure 2A, Figure 2—figure supplement 1; Supplementary file 1A) covering sequences that appear to be upstream open reading frames (uORFs; example in Figure 2C).

Figure 2 with 3 supplements see all

Download asset Open asset

We attribute this density to translating 80S ribosomes rather than 48S preinitiation complexes for three reasons: first, the length distribution of protected fragments in 5′ UTRs (25–35 nt) is indistinguishable from the length distribution of ribosome-protected fragments in coding regions (Figure 2—figure supplement 2), while the protected footprint of a preinitiation complex is reported to be larger (40–70 nt; Lazarowitz and Robertson, 1977; Pisarev et al., 2008). Second, our measurements of 5′ UTR density are indistinguishable whether we enrich digested monosomes by sedimentation through a sucrose cushion (which collects all heavy particles) or specifically separate them from preinitiation complexes by fractionation of a sucrose gradient (Figure 2B, Figure 2—figure supplement 3). Thus, the dominant signal contributing to our measurement of footprint density in 5′ UTRs is derived from fragments protected by 80S ribosomes. Third, because initiation and termination of translation are slow compared to elongation, initiation and termination events produce peaks of ribosome density (Figure 2—figure supplement 1; Ingolia et al., 2011). Such peaks are frequently visible at the boundaries of predicted uORF sequences (example in Figure 2C), again arguing that reads aligning to 5′ UTRs represent translation events. Given the known roles of uORFs in regulating both the translation and the stability of mRNAs (reviewed in Meijer and Thomas, 2002) we anticipate that our methods will facilitate future analyses of the contributions of uORFs to control of gene expression throughout fly development.

Comparative analysis of the genomes of 12 sequenced Drosophila species has provided a powerful strategy for annotating protein-coding regions in Drosophila genomes (Lin et al., 2007, 2011). Using this approach, 283 transcripts in D. melanogaster were demonstrated to contain clear phylogenetic signatures of amino acid conservation in the region between the annotated and next in-frame stop codons. It was therefore concluded that these regions encode C-terminal protein extensions (hereon called ‘predicted extensions’), produced by stop codon readthrough events (Lin et al., 2007; Jungreis et al., 2011).

In our data, the density of ribosomes on 3′ UTRs is several orders of magnitude lower than in coding regions and 5′ UTRs (Figure 2A, Supplementary file 1A), and many genes show highly efficient termination (example in Figure 3B). However, a subset of transcripts exhibit high footprint density within the predicted extensions. To determine whether the footprint density was consistent with stop codon readthrough (as opposed to alternate explanations, like frameshift), we manually scored each predicted extension whose corresponding structural gene was sufficiently expressed in our embryo sample (158 in total). An extension was scored positively if there existed ribosome density in the extension, ribosome density vanished or unambiguously decreased following the first in-frame stop codon, and positions occupied by ribosomes in the putative extension evenly covered the majority of the extension’s length (see ‘Materials and methods’ for further details). By these criteria, 43 of the 283 transcripts predicted to undergo stop codon readthrough contained ribosome density consistent with a readthrough event (example Figure 3C, full data in supplementary table 2 at Dryad: Dunn et al., 2013), including one example of double readthrough (Figure 3D). We expect that the many of the remaining 240 transcripts also undergo readthrough, either at levels too low to detect at our sequencing depth, or at other developmental stages (discussed further below).

Figure 3 with 1 supplement see all

Download asset Open asset

Surprisingly, we observed that a distinct set of transcripts not predicted to undergo readthrough also exhibits substantial footprint density between the annotated and next in-frame stop codons (Figure 3E). We therefore searched for C-terminal extensions among all transcripts that met the following criteria: (a) a minimum of 128 footprint in the corresponding CDS, (b) a minimum footprint read density of 0.2 RPKM in the extension, (c) a minimum readthrough rate of 0.001, and (d) a lack of methionine codons in the first three codons of the extension, as this latter group could be explained by initiation within the extension rather than readthrough of the upstream stop codon. We additionally excluded extensions whose translation could be explained by alternately spliced transcript isoforms that omit the stop codon. Scoring this group by the same criteria used for the predicted extensions, we identified 307 additional examples of stop codon readthrough (hereon referred to as ‘novel extensions’; see example Figure 3F), including another example of double readthrough (Figure 3G). In addition, we identified several transcripts that contained 3′ UTR footprint density more consistent with ribosomal frameshift (Figure 3—figure supplement 1A,B), or the presence of additional downstream cistrons, RNA structure, or protein binding (Figure 3—figure supplement 1C,D). These were excluded from further analysis.

Because footprint density generally is far lower in 3′ UTRs than in 5′ UTRs or coding regions (Figure 2A), it is possible that various sources of noise (e.g. regions of mRNA protected by RNA structures or by RNA-binding proteins) might contribute more substantially to this density than to the density in coding regions. We therefore asked whether footprints in 3′ UTRs exhibited behaviors specific to footprints protected by 80S ribosomes.

In order to distinguish whether reads mapping to extensions were either protected by ribosomes or derived from alternate sources, we compared the total number of reads aligning to extensions in samples prepared from sucrose cushions, which collect all heavy macromolecular complexes, to those in which we specifically isolated 80S ribosomes on sucrose gradients. Footprint count measurements for each extension are highly correlated between libraries made using these two sample preparation methods, indicating that these footprints are either protected by 80S ribosomes, or by another RNA binding protein that co-sediments with 80S ribosomes (Figure 4A; r² = 0.945).

Figure 4 with 1 supplement see all

Download asset Open asset

Because various ribosome-binding proteins protect nucleotide fragments of distinct lengths, the size distribution of protected mRNA fragments provides a powerful approach for distinguishing 80S footprints from other sources (Ingolia et al., manuscript in preparation). Footprints in C-terminal extensions exhibit a length distribution very similar to footprints in coding regions, while those derived from non-coding sources, such as snoRNAs and tRNAs, show dramatically different length distributions (Figure 4B). Thus, footprints aligning to extensions appear to be protected by 80S ribosomes.

Finally, we sought to determine whether the ribosomes that appear to translate extensions are engaged in active translation, as opposed to some aberrant process of stalling or slippage (e.g., as described in Skabkin et al., 2013). Because terminating ribosomes produce a characteristic peak of ribosome density over annotated stop codons (Figure 2—figure supplement 1; Ingolia et al., 2009), we asked whether the stop codons that terminate the C-terminal extensions also showed this behavior. Indeed, C-terminal extensions exhibit peaks at their stop codons, clearly arguing that footprint density in C-terminal extensions is attributable to actively-translating ribosomes (Figure 4C).

Because this meta-gene analysis represents a group average, we also compiled individual statistics on ribosome release in a manner similar to the RRS score described by Guttman et al. (2013) . Briefly, we tabulated the ratio of the total number of reads aligning within a five codon window immediately downstream of a stop codon to the number of reads aligning to the five codon window immediately upstream of that codon, with the expectation that if ribosomes terminate at a given stop codon, the score for that codon should approach zero. We performed this calculation separately for: (1) stop codons that terminate annotated coding regions, (2) stop codons that terminate C-terminal extensions, and (3) as a negative control, randomly selected codons internal to annotated coding regions. We find that the scores of stop codons that terminate C-terminal extensions fall within the distribution of scores for stop codons that terminate annotated coding regions (Figure 4—figure supplement 1), again arguing that the read density covering putative C-terminal extensions are in fact produced by ribosomes that have undergone stop codon readthrough rather than other processes.

It is possible that the population of ribosomes that read through stop codons is engaged in a pathological translation process that might not produce detectable protein products. We therefore asked whether we could detect translation products by immunoprecipitation (IP) and western blotting. We created reporter constructs for a panel of transcripts including five predicted and 10 novel extensions that exhibited readthrough in both 0–2 hr embryos and S2 cells. In each construct, we fused Venus (a GFP variant) upstream of a portion of each transcript containing the C-terminal 120 codons of the annotated coding sequence and the entire endogenous 3′ UTR. To visualize readthrough, we fused a double FLAG epitope to the C-terminus of the putative C-terminal extension. We transfected these constructs into S2 cells, immunoprecipitated the reporter at the N-terminus using anti-GFP beads, and detected the extensions by western blotting using an anti-FLAG antibody. We detected readthrough products of the correct size for eight of the reporters, arguing that at least this subset of extensions yields C-terminally extended proteins in vivo (Figure 4D).

While we did not seek to detect C-terminally extended proteins generated by endogenous transcripts (e.g., through mass spectrometry), we do believe our reporter constructs to be at least as faithful as those used in earlier literature, as we included substantially more nucleotide context (120 codons upstream of stop plus the entire endogenous 3′ UTR) than other groups screening through candidate genes to find readthrough signals (2–8 codons upstream and 3–15 codons downstream of the stop codon; Fearon et al., 1994; Harrell et al., 2002; Namy et al., 2002, 2003).

The appearance of stop codon readthrough, both in ribosome profiling data and in IP-westerns, could result from several other processes, such as selenocysteine insertion, genomic mutation of stop codons to sense codons, or the editing of stop codons in mRNAs. We consider each of these in turn.

UGA stop codons may be decoded by specialized translation machinery as the unconventional amino acid selenocysteine if the 3′ UTR contains a selenocysteine insertion (SECIS) element. However, UGA stop codons represent only 25% of the readthrough events we report, and none of these are annotated as selenoproteins in either FlyBase (Marygold et al., 2013) or SelenoDB (Castellano et al., 2008). Furthermore, we were unable to detect SECIS elements in any of their 3′ UTRs using SeciSearch 2.19 (Kryukov et al., 2003). Thus, at most, even unannotated selenocysteine insertion events could only account for a small fraction of the readthrough events we report.

We also exclude the possibility that readthrough might result from genomic polymorphisms or RNA editing at the stop codon. Because both types of events would be represented in our data as mismatches between read alignments and the reference transcript sequence, we counted the total number of matching and mismatching reads covering each nucleotide position in each stop codon in our mRNA-seq and ribosome footprint datasets. For each dataset, we calculated a global average proportion of mismatching reads, and used the binomial test to identify stop codon nucleotides whose individual proportion of mismatches significantly deviated from the corresponding global average.

Together, the mRNA and footprint datasets identified a total of 10 nucleotide positions whose mismatch rates significantly exceeded the average (Figure 4E,F). Three positions contained genomic polymorphisms that changed one stop codon to another stop codon (Figure 4E,F, green). Two (Figure 4E,F, red) contained genomic polymorphisms that converted the stop codon to a sense codon. These two transcripts were therefore excluded from further study. The remaining five positions contained a variety of mismatches each occurring at low frequency. These observations are inconsistent with the presence of a genomic polymorphism at those positions, which should cause a 50% or 100% frequency of a single mismatch, depending on whether the polymorphism is hetero- or homozygous (Figure 4E,F, black).

An alternate explanation for a low but elevated proportion of mismatches is RNA editing, the conversion of one nucleotide to another in an mRNA. In Drosophila, the only mechanism known to edit mRNA is the deamination of adenine to inosine, which is converted to guanine by reverse transcriptase (Ramaswami et al., 2013). A-to-I editing thus appears in sequencing data as a preference for A-to-G transitions among mismatches. Of the five mismatching positions we could not ascribe to genomic polymorphisms, four contain thymine or guanine rather than adenine residues in the reference sequence, and therefore cannot be edited by this pathway. We therefore attribute these mismatches to sequencing error. The majority of mismatches at the single remaining position are transversions from adenine to thymine, similarly arguing that these mismatches are more likely due to sequencing error than to A-to-I editing.

Formally, it is possible that a minor fraction of transcripts are edited, but that this fraction, even if small as measured in the RNA-seq or total footprint data, might account for all of the stop codon readthrough we observe. Analysis of the ribosome footprint data allows us to explore this possibility directly. Specifically, were this the case, the sequences of all the footprints deriving from ribosomes that have undergone readthrough—namely, those whose A-sites have already translated the stop codon—should contain evidence of editing (Figure 4G, top). We therefore separately analyzed the footprints deriving from this specific pool of ribosomes. Our dataset provided sufficient coverage to test 419 of 450 such positions (93% of the total). Of these, only four stop codon positions exhibited significantly elevated levels of mismatch (Figure 4G, bottom). All of these were identified in the mRNA and total footprint datasets above as having genomic polymorphisms (Figure 4E–F). Thus, our most stringent dataset contains no positive evidence of RNA editing.

Further, this dataset contains positive evidence against RNA editing. Under the null hypothesis that A-to-I editing drives readthrough, one would expect nearly all footprints (for our purposes, conservatively assuming 90%) in the A-site footprint dataset to contain an edited base. Under this assumption, we used a binomial test to estimate the probability of observing the proportion of A-to-G mismatches in the A-site footprint dataset at each adenine residue sufficiently covered by reads (217 positions, representing roughly 50% of A positions in all readthrough events reported). In this analysis, all positions contained significantly fewer A-to-G mismatches than expected under the hypothesis of A-to-I editing, (Bonferonni-corrected p<<0.05 for all transcripts), indicating that A-to-I editing plays no part in any of the readthrough events we could test.

Because we detected far more readthrough events in Drosophila than were predicted from phylogenetic data, we collected yeast datasets and examined them for empirical evidence of readthrough. Importantly, because the [PSI⁺] form of the yeast eRF3 homologue is known to promote readthrough, we limited our analysis to data collected from [psi⁻] strains.

In contrast to MNase (which exhibits a 3′ A/T cutting bias, yielding positional uncertainty of the ribosomal P-site, see ‘Materials and methods’), RNAse I shows little cutting bias. Therefore, libraries prepared with RNase I (e.g., yeast and mammalian libraries) offer superior spatial resolution along mRNAs. In such libraries, the locations of ribosome-protected footprint fragments in coding regions exhibit a characteristic three-nucleotide periodicity or phasing from which reading frames can be deduced (Ingolia et al., 2009, 2011). We therefore tabulated the phasing of ribosome-protected footprint fragments in all annotated coding regions, putative C-terminal extensions, and the 40 codon windows downstream of the putative extensions as an approximation of the portion of the 3′ UTR distal to the putative extension (hereafter called ‘distal 3′ UTRs’). To control for cloning biases caused by skewed nucleotide frequencies at each phase, we tabulated the phasing of randomly-fragmented mRNA fragments that were cloned using the same protocol and aligned to the same regions. Non-random phasing consistent with translation is apparent in both the coding regions and the putative extensions, but not the distal 3′ UTR (p=3.98 × 10⁻²⁶, Χ² test, footprints vs mRNA fragments in extension, dof = 2; Figure 5A). Importantly, the major component of phasing in the putative extensions occurs in the same reading frame as that of coding regions, indicating that readthrough (as opposed to, e.g., frameshift) is a major contributor to protected fragment density in 3′ UTRs in yeast. Having found global evidence for readthrough, we manually scored a subset of yeast genes to identify individual examples of readthrough, using the same filtering and scoring criteria we used in the Drosophila datasets. We found 30 clear examples of readthrough in yeast (examples in Figure 5B,C; full results in supplementary table 3 at Dryad: Dunn et al., 2013), demonstrating that readthrough is not unique to Drosophila.

Figure 5 with 1 supplement see all

Download asset Open asset

Because readthrough has been observed in two mammalian genes (Geller and Rich, 1980; Yamaguchi et al., 2012), we collected data from primary human foreskin fibroblasts and sought evidence of readthrough in humans. We identified 42 readthrough events in the human data (Figure 5D,E; full results in supplementary table 4 at Dryad: Dunn et al., 2013). These events are not explained by selenocysteine insertion, and, as in Drosophila, read lengths mapping to extensions in the yeast and human datasets are similar to those mapping to coding regions in these organisms (Figure 5—figure supplement 1). Thus, readthrough appears to be prevalent in all three organisms.

To estimate how many of the novel extensions we detected might be translated at a biologically significant level, we estimated a threshold for biological significance as the fifth percentile of readthrough rates for the phylogenetically conserved extensions that were translated in the D. melanogaster embryo, a rate of 1.2%. Out of all the extensions for which we could measure readthrough rates (i.e., those sufficiently long not to be covered by stop codon peaks, see ‘Materials and methods’), 61.8% of the novel extensions in Drosophila, 94.7% of the extensions in human foreskin fibroblasts, and 40.0% of the extensions in yeast exceeded this threshold, arguing that readthrough might be important in all three organisms (Figure 5F).

Because 307 of the 350 readthrough events we discovered were not predicted phylogenetically, we sought to determine whether any of them showed signs of protein-coding conservation through the Drosophila phylogeny. To this end, we used PhyloCSF, which reports a log-likelihood ratio reflecting the relative probabilities of observing a given alignment of orthologous nucleotide sequences under models of protein-coding or non-coding evolution (Lin et al., 2011). By this metric, only 14 of the 307 novel extensions score positively (Figure 6A), and their distribution of PhyloCSF scores was not markedly different from the global distribution (Figure 6—figure supplement 1A), indicating a lack of phylogenetic evidence for amino acid conservation.

Figure 6 with 1 supplement see all

Download asset Open asset

The lack of detectable phylogenetic evidence of amino acid conservation among the novel extensions suggests two models: either (1) the novel extensions, on average, are selectively neutral, and occur only because they do not incur too great a fitness disadvantage, or (2) the novel extensions are under selection, but originated after the divergence of D. melanogaster from its closest sequenced relatives, making conservation in this group undetectable by cross-species tools such as PhyloCSF. To distinguish these possibilities, we used two tests to detect signs of selection for protein coding specifically within D. melanogaster.

To determine whether the nucleotide sequences of novel extensions show signs of selection for protein coding potential, we implemented a Z-curve classifier, a machine-learning technique that separates coding regions from non-coding regions based upon phased differences in nucleotide k-mer frequency (Gao and Zhang, 2004). We trained the classifier to distinguish annotated coding regions from distal 3′ UTRs (see ‘Materials and methods’ for details). Consistent with a long history of protein-coding selection, extensions predicted by phylogenetic conservation showed a nucleotide character indistinguishable from annotated coding regions (Figure 6B). By contrast, novel extensions exhibit a nucleotide character intermediate between coding regions and distal 3′ UTRs (p=1.02 × 10⁻²³, Mann-Whitney U test, distal 3′ UTR vs novel extensions), which is consistent with an evolutionary trajectory towards coding-like character from a 3′ UTR. This effect is not due to specific nucleotide signals found in distal 3′ UTRs (p=3.81 × 10⁻²², Figure 6—figure supplement 1B), and was robust across Z-curve classifiers trained on different windows drawn from distal 3′ UTRs (see ‘Materials and methods’).

To obtain more direct evidence for or against protein-coding selection, we analyzed SNP data from 50 individuals of D. melanogaster from the Drosophila Population Genomics Project (http://www.dpgp.org). We determined the proportion of SNPs that would be synonymous if translated in-frame in coding regions, predicted extensions, novel extensions, and distal 3′ UTRs. Novel extensions show a modest but significant preference for synonymous SNPs above the background level of distal 3′ UTRs (Figure 6C; p=1.42 × 10⁻⁵, one-sided Fisher’s exact test), but below that of the predicted extensions (p=8.42 × 10⁻⁹, one-sided Fisher’s exact test). This pattern suggests that a subset of the novel extensions is undergoing selection for protein coding, and that the contribution from this subset to the average SNP preference outweighs the contributions from other subsets of extensions that are selectively neutral or undergoing diversifying selection. Together, these results favor the hypothesis that at least a fraction of the novel extensions are of recent evolutionary origin and have come under selection within the melanogaster lineage.

In order to determine whether C-terminal extensions might be functional, we sought evidence for biological regulation of readthrough rates. We therefore queried our S2 cell and embryo datasets for evidence of differential regulation of readthrough in all genes that were sufficiently expressed in both datasets and contain only one, unique annotated coding region across all transcripts.

For each gene meeting these criteria, we tabulated the number of ribosome-protected footprints in the corresponding coding region and extension in each tissue type, and calculated a p value for the observed distribution of counts using Fisher’s exact test. Controlling the false discovery rate at 5%, we found nine of 182 testable transcripts to significantly change between samples, indicating that all nine should be true positives (Table 1; full data in supplementary table 2 at Dryad: Dunn et al., 2013). Thus, readthrough is differentially regulated between Drosophila cell types.

In principle, readthrough could be regulated by: (1) changes in the expression or activities of global factors (e.g., eukaryotic release factors, charged tRNA abundance etc), (2) by gene- or transcript-specific elements, like mRNA structures, or (3) by a combination of both. In the first scenario, readthrough rates for all transcripts should increase or decrease monotonically in one cell or tissue type compared to another. In the latter two scenarios, readthrough rates should increase for some transcripts, but decrease for others. We identified four significant increases and five significant decreases in readthrough rate in embryos compared to S2 cells, indicating that readthrough is at least in part regulated on a transcript-by-transcript basis. The distribution of fold-changes in readthrough rate spans several orders of magnitude, indicating that transcripts that are robustly read through in one cell type are not necessarily read through in another (Table 1). This result implies that extensions function in specific cellular or developmental contexts, consistent with earlier reports that readthrough of specific genes is regulated in metazoans (Robinson and Cooley, 1997; Yamaguchi et al., 2012).

Because we observe such a large magnitude of regulation, we believe the 350 readthrough events we report here to represent a small subset of a larger group that occur throughout the lifetime of an individual fly. We therefore expect many of the extensions that were phylogenetically predicted but not observed in our samples are in fact translated at other developmental stages in Drosophila. Finally, because transcripts with significant p values are statistically more highly counted in their extensions than those without significant p values (p=2.4 × 10⁻³, Mann-Whitney U test), we surmise that our ability to detect regulation was limited by sequencing depth and that the true number of transcripts whose readthrough rates are regulated in tissue- or condition-specific manners is in fact larger than we report.

Many peptide sequences—such as signal sequences, degrons, and phosphorylation sites—affect the localization, stability, or activity of proteins. Because these sequences are frequently short and/or degenerate, a high proportion of even random peptide sequences confer function (Kaiser et al., 1987; Kaiser and Botstein, 1990). Thus, a C-terminal extension produced by termination failure could purely by chance alter the function or behavior of its host protein, and thus come under selection. Indeed, Freitag et al. (2012) reported two readthrough events in fungi that append peroxisomal localization signals (PTS1) to the C-termini of glyceraldehyde-3-phosphate dehydrogenase (GAPDH) and 3-phosphoglycerate kinase, enabling these typically cytosolic enzymes to function in peroxisomal metabolism. We therefore searched our full set of C-terminal extensions for short peptide signals that direct peroxisome localization, nuclear localization (NLS), prenylation, or ER retention, or that resemble transmembrane domains (see ‘Materials and methods’). PTS1 signals were detected in one extension. 10 proteins not annotated as nuclear in FlyBase contain predicted NLSes in their extensions. Eight extensions contain predicted transmembrane domains and one contains a C-terminal prenylation signal. No extension contained an ER retention signal (Table 2).

To determine whether any of the putative nuclear localization signals (NLSes) function in vivo, we constitutively fused C-terminal extensions containing putative NLSes to the C-terminus of a GFP-mCherry-GST reporter, which is excluded from the nucleus (Figure 7, left column; Chan et al., 2007). When expressed in S2 cells, three of four NLSes relocalized the cytosolic reporter to the nucleus at levels above background (Figure 7, columns 3–5), arguing that these extensions can regulate the localization of their endogenous host proteins. Given the large number of short peptide signals (e.g., phosphorylation motifs, degradation motifs, ubiquitination sequences, etc) that have been discovered, and the limited number of reporters we tested here, we likely underestimate the number of extensions that confer function. Nonetheless, our results clearly establish that C-terminal extensions can alter protein function.

Figure 7

Download asset Open asset

Wild-type (y w) flies were cultured according to standard procedures. S2 cells were cultured in Schneider’s (Gibco by Life Technologies, Carlsbad, California) media supplemented with 10% heat-inactivated FBS (UCSF cell culture facility, San Francisco, California) and antibiotics (UCSF cell culture facility). S2 cells were transfected using Effectene reagent (Qiagen, the Netherlands) following the manufacturer’s instructions. For stable transfectants, the plasmid of interest was co-transfected at a 10:1 molar excess with pCoPuro. Stable integrants were selected and maintained in Schneider’s media supplemented as above, but additionally containing 10 μg/ml puromycin.

Concentrations of total RNA in lysates were determined using the RiboGreen kit (Molecular Probes by Life Technologies). For each sample, 35–100 μg total RNA was diluted 2:1 in digestion buffer (50 mM Tris pH 7.5, 5 mM MgCl2, 0.5% Triton x-100, 1 mM DTT, 20 U/ml SuperaseIn, 20 μg/ml emetine, 15 mM CaCl₂, and 3 U micrococcal nuclease [Roche Applied Science, Indianapolis, Indiana] per μg of total RNA in the sample), to bring the final concentration of NaCl to 100 mM and CaCl₂ to 5 mM. Samples were digested for 40 min at 25°C in a Thermomixer (Eppendorf, Hamburg, Germany). Digestions were quenched by adding EGTA to a final concentration of 6.25 mM and placing the reactions on ice. 1 U MNase is defined as previously (Oh et al., 2011) as an increase of 0.005 A260 per min, measured in a Spectramax M2 plate reader (Molecular Devices, Sunnyvale, California) using 10 μg/ml salmon sperm DNA (Sigma-Aldrich) with 5 mM Ca+ and 20 mM Tris, pH 8.0 in a 0.1 ml reaction at 25°C.

10–50% sucrose gradients were prepared in polysome gradient buffer (250 mM NaCl, 15 mM MgCl₂, 20 U/ml SuperaseIn, 20 μg/ml emetine) using a GradientMaster (Biocomp Instruments, Fredericton, New Brunswick, Canada) in polyclear centrifuge tubes (Seton Scientific, Petaluma, California). Up to 200 µl of samples was applied to the top of each gradient. Gradients were resolved by spinning for 3 hr at 35 krpm at 4°C in an SW-41 rotor (Beckmann Coulter, Brea, California), and fractionated using the GradientMaster. When appropriate, monosome fractions were collected, flash-frozen in liquid nitrogen, and stored at −80°C.

Up to 0.5 ml of digested sample was layered atop 1.0 ml of a solution of 34% sucrose in polysome gradient buffer. Monosomes were sedimented by spinning for 4 hr at 70 krpm at 4°C in a TLA-110 rotor (Beckmann Coulter). Pellets were resuspended in 600 µl 10 mM Tris, pH 7.0 and stored at −20°C.

Lysates were prepared and footprinted as above. Unless otherwise indicated, monosomes were enriched by sedimentation through 34% sucrose cushions and resuspended in 600 µl 10 mM Tris, pH 7.0. Resuspended monosomes were extracted once with 700 µl 65°C acid phenol and 40 µl 10% SDS, followed by 650 µl acid phenol and a final extraction with chloroform. RNA was precipitated for at least 2 hr at −30°C, resuspended in 10 mM Tris, pH 7.0, and quantitated on a NanoDrop spectrophotometer (Thermo Scientific, Asheville, North Carolina). 5–35 μg RNA was dephosphorylated for 1 hr at 37°C using T4 polynucleotide kinase (New England Biolabs, Ipswich, Massachusetts) in a 50 µl reaction and resolved on a 15% TBE-urea gel (Invitrogen by Life Technologies). A gel slab spanning 28–34 nt (as measured by oligoribonucleotide size standards in a neighboring lane; see Supplementary file 2) was excised from the gel, eluted, and precipitated. Samples were then carried through all steps of library generation (see below).

For each sample, 375 µl of undigested polysome lysate was diluted into 3 vol Trizol LS (Invitrogen) and total RNA was extracted following the manufacturer’s instructions. 20–50 μg Poly(A)+ RNA was selected on oligo-dT25 DynaBeads (Invitrogen) per manufacturer’s instructions, and fragmented at 95°C in fragmentation buffer (2 mM EDTA, 100 mM NaCO₃/NaHCO₃, pH 9.2) to a mean size of roughly 100 nt. Fragmented RNA was precipitated, dephosphorylated for 1 hr at 37°C with T4 polynucleotide kinase (New England Biolabs), and resolved on a 15% TBE-urea gel. A gel slab corresponding to 55–65 nt was excised from the gel, eluted, and precipitated. Samples were then carried through all steps of library generation (see below).

We performed two sequential rounds of subtractive hybridization on each sample. To 5 µl cDNA the following were added: 1 µl 20× SSC, 3 µl nuclease-free water, and 1 µl of a 60 μM mixture of the biotinylated oligonucleotides oJGD132, oJGD133, oJGD134, oJGD135, oJGD136, oJGD161, oJGD162, oJGD163, and oJGD164 (sequences in Supplementary file 2) mixed in a ratio of 25.5:1:13:17:4:6:2:11:21. Samples were denatured for 90 s at 95°C and annealed for 20 min at 25°C. MyOne Streptavidin C1 DynaBeads (Invitrogen) were prepared as follows: for each sample, 45 µl of beads were aliquoted into a microcentrifuge tube and washed three times in 50 µl 2 × binding buffer (10 mM Tris, pH 7.5, 1 mM EDTA, 2 M NaCl), and resuspended in 22.5 µl 2 × binding buffer. 10 µl equilibrated beads were added to 10 µl hybridized sample. The mixture was incubated at 20 min in a room temperature Thermomixer with shaking at 850 rpm. Beads were then separated on a magnetic manifold (Invitrogen) and the supernatant recovered to a microcentrifuge tube.

For the second round of subtraction, 1 µl 60 μM biotinylated oligo mix and 1 µl 20X SSC were added to the supernatant from the first subtraction, and the denaturation and annealing repeated. 10 µl of equilibrated beads were pelleted on a magnetic manifold. The buffer was removed, and the beads resuspended in the mixture from the second hybridization. Samples were then incubated at 20 min in a room temperature Thermomixer with shaking at 850 rpm. The supernatant was recovered on a magnetic manifold, transferred to a microcentrifuge tube, precipitated, and resuspended in 15 µl 10 mM Tris, pH 8.0.

RNA concentrations were measured using the Small RNA Series II Bioanalyzer assay (Agilent Technologies, Santa Clara, California). 10–15 picomoles of RNAs were ligated to 1 μg 3′ miRNA cloning linker 1 (Integrated DNA Technologies, Coralvaille, Iowa) for 2 hr 30 min at 25°C in ligase buffer (1× T4 RNA ligase 2 buffer [New England Biolabs], 40% PEG-100 [Sigma-Aldrich], 5% DMSO, T4 RNA ligase 2 K227Q, truncated [a kind gift from Calvin Jan]) in a 20 µl reaction. Ligated fragments were precipitated for at least 2 hr at −30°C, purified on a 10% TBE-urea gel, eluted, and precipitated. Ligation products were then reverse-transcribed using SuperScript III (Invitrogen) in a 16.7 µl reaction using using the primer o225-link1 (see Supplementary file 2). RNA template was hydrolyzed by addition of 1/10 vol 1 M NaOH and incubation at 95°C for 20 min cDNAs were purified on a 10% TBE-urea gel (Invitrogen), eluted, precipitated, and resuspended in 5 µl 10 mM Tris pH 7.0. cDNAs from footprint samples were subjected to two rounds of subtractive hybridization as described above.

Subtracted samples were circularized using CircLigase (Epicentre, Madison, Wisconsin), following manufacturer’s instructions in a 20 µl reaction. An additional microliter of CircLigase was then added, and the circularization repeated a second time. Circularized libraries were amplified by 6–12 cycles of PCR using oNTI231 and any of four indexing primers oCJ30–33 (Supplementary file 2) using Phusion polymerase (Finnzymes by ThermoScientific) in a 17 µl reaction. Amplification products were size-selected on 8% TBE gels (Invitrogen), eluted, precipitated, and resuspended in 10 µl 10 mM Tris, pH 8.0. Samples were then quantitated using the Bioanalyzer High Sensitivity DNA assay (Agilent Technologies), diluted to 2 nM, multiplexed as needed, and subjected to 50–57 cycles of single-end sequencing on an Illumina HiSeq sequencer (Illumina, San Diego, CA) using version 3 clustering and sequencing kits with a 6-cycle index read (Illumina).

For all Drosophila experiments we used revision 5.43 of the FlyBase genome annotation and the corresponding genome assembly (Marygold et al., 2013). Reads were demultiplexed and cleaned of 3′ cloning adapters using in-house scripts. Reads shorter than 25 nt were discarded. Remaining reads were aligned using Bowtie version 0.12.7 (Langmead et al., 2009) sequentially to Bowtie indices composed of the following sequences: (a) D. melanogaster rRNAs (GenBank accession #M21017 [Tautz et al., 1988] and from FlyBase), (b) D. melanogaster tRNAs, snoRNAs, and snRNAs (from FlyBase), (c) cloning oligos, (d) the S288C yeast genome version R64-1-1 (downloaded on 6 June 2011 from http://downloads.yeastgenome.org/sequence/S288C_reference/genome_releases/), (e) Wolbachia (GenBank accession #AE017196), (f) D. melanogaster chromosome arms, and (g) splice junctions (from FlyBase and, in the case of embryos—figure supplemented with junctions discovered in the pooled embryo mRNA datasets using HMMSplicer 0.95 [Dimon et al., 2010]). For all quantitative analyses, we counted only uniquely-mapped reads.

Alignments were assigned to genomic coordinates as follows. Randomly-fragmented poly(A)+ mRNA alignments were counted along the entire length of the alignment. Each genomic position covered by a single RNA fragment was incremented 1/l, where l corresponds to the length of the alignment. Ribosome-protected footprint alignments were mapped to their estimated P-sites as follows: 12 nt were pruned from each end of the alignment, leaving a fragment n nt long (where n = l − 2 × 12). Each genomic position covered by a nucleotide remaining in the pruned alignment was then incremented by 1/n. Thus, the P-site of each 25 mer was assigned to one unique position, while the P-site of each 26-mer was spread over two positions, each incremented by 0.5 reads, and so on. Alignment statistics are given in Supplementary file 1B.

mRNA abundance and ribosome density for each genomic feature were measured in reads per kilobase of feature length per million reads aligning to chromosomes or splice junctions in the dataset (RPKM), a unit which corrects for both feature length and sequencing depth. Unless otherwise indicated, the RPKM values we report for mRNA abundance reflect the total number of RNA fragments aligning to all countable exonic positions for a given locus. For ribosome density, we report the total number of ribosome-protected footprint fragments aligning to all countable positions of a coding region (CDS) for a given locus. We calculate translation efficiency as the ratio of footprint RPKM in the CDS to the RNA fragment RPKM across the entire locus. When comparing mRNA fragment or footprint density between samples, we restricted our analyses to genes that had at least 128 summed counts between replicates as determined in Figure 1—figure supplement 2. When comparing translation efficiencies between samples, we required at least 128 exonic counts of mRNA for each gene.

Translation efficiencies for these regions were calculated as the ratio of footprint counts to mRNA counts in each region, for all regions with at least 128 mRNA counts. We excluded all positions that could be labeled as two or more of 5′ UTR, CDS, or 3′ UTR depending upon transcript isoform. To remove variability or bleedthrough introduced by start and stop codon peaks, we additionally excluded the following genomic positions from consideration: 9 nucleotides preceding each start codon, 15 nucleotides following each start codon, the 15 nucleotides preceding each stop codon, and the 15 nucleotides following each stop codon.

For each analyses (Figure 2—figure supplement 1, Figure 4B), we identified regions of interest (ROIs) germane to the analysis. In Figure 2—figure supplement 1, these included roughly 3000 ROIs each for the left and right panels, each of which met the following criteria: (a) all transcripts deriving from that gene had one annotated start codon (left panel) or stop codon (right panel), (b) all transcripts deriving from that locus covered identical genomic positions over the region of interest (ROI) shown, (c) all positions within the ROI were non-degenerate (see ‘Materials and methods’), and (d) at least 10 reads were present in the coding subregion of the ROI. For coding regions in Figure 4C, we kept the same criteria as above but required only 0.5 reads in the coding subregion of each ROI, yielding roughly 7401 ROI for that set. For C-terminal extensions, we required only that the extension be long enough to cover the interval shown, and have 0.5 reads in the coding subregion, allowing us to include 123 of the 350 extensions.

For each ROI, we then generated a ‘coverage vector’ tallying ribosome density at each nucleotide position. We then normalized each coverage vector to the mean number of footprint reads covering the annotated coding region in the ROI, excluding a 3-codon buffer flanking the start or stop codon to avoid bleedthrough from initiation or termination peaks. We then plotted the median value across all normalized coverage vectors at each position.

To improve our sensitivity in detection, we re-aligned our footprint and mRNA datasets to a Bowtie database of spliced transcript models, allowing three mismatches (where we previously only allowed two). For the first, second, and third nucleotide position in each unique, annotated stop codon, we counted the number of matching and mismatching nucleotides in each read alignment covering that position. We ignored mismatches that occurred in the first position of the read alignment, because they frequently arise from non-templated nucleotide addition by reverse transcriptase. Considering the first, second, and third positions of each stop codon separately, we calculated a global average mismatch frequency for each. We then searched for individual stop codon positions that far exceeded the corresponding global average using a binomial test, controlling the false discovery rate at 5% following the procedure of Benjamini and Hochberg (Benjamini and Hochberg, 1995). We performed this analysis separately upon each of three datasets: total mRNA, total footprints, and the subset of footprints whose P-sites had passed the nucleotide position in question, following the P-site assignment rules described above.

4–48 ml of transiently or stably transfected cells were harvested 48 or 72 hr post-transfection by centrifuging for 2 min at 1600 rpm in a tabletop centrifuge. All cell pellets were rinsed once in PBS, and flash-frozen in a bath of dry ice in ethanol. Cell pellets were thawed and lysed for at least 15 min on ice in 0.5–1.5 ml lysis buffer (150 mM NaCl, 50 mM Tris pH 7.5, 1% Triton x-100, 1 mM EDTA and 1× complete protease inhibitor cocktail [Roche Applied Science]), depending upon the pellet size. Lysates were clarified by spinning 10 min at 20,000 × g in a microcentrifuge and supernatants collected. GFP reporters were immunoprecipitated on 10 μl of anti-GFP beads (Chromotek, Planegg-Martinsried, Germany) equilibrated in IP wash buffer (150 mM NaCl, 50 mM Tris pH 7.5, 1 mM EDTA, 0.05% Triton x-100). The bound fraction was washed three times in IP wash buffer, and finally eluted by boiling for at least 5 min in NuPage sample loading buffer (Invitrogen). Supernatants were collected and transferred to new tubes, and stored at −20°C.

For western blotting, samples were resolved on 4–12% NuPage gels (Invitrogen) in MOPS buffer. Gel lanes were loaded such that the amounts of uncleaved GFP reporter in each lane were loaded as equally as possible. GFP was detected using a mouse anti-GFP antibody (Roche Applied Sciences), and visualized using IR800 anti-mouse antibodies on a LI-COR Odyssey system (LI-COR, Lincoln, Nebraska). FLAG was similarly detected on a separate gel, instead using the M2 Mouse anti-FLAG antibody (Sigma-Aldrich).

PhyloCSF analysis was performed on all C-terminal extensions at least five codons long, exclusive of the stop codons. Multiple species alignments were obtained from the Drosophila 12-way multispecies alignment as downloaded from the UCSC genome browser, and stitched together over regions of interest using the Phast utility maf_parse (Hubisz et al., 2011). PhyloCSF was then used to evaluate the extension on the empirical codon model ‘12flies’. Columns in which the D. melanogaster sequence contained gaps were ignored. Alignments that contained no sequence besides that from D. melanogaster were not evaluated.

We calculated the 189-variable Z-curve as previously described (Gao and Zhang, 2004). We empirically determined that the classifier became error-prone if trained on sequences 81 nt or shorter in length. Our training set consisted of 81-nucleotide windows drawn from coding regions (the positive set, 14,507 windows) or from portions of distal 3′ UTRs that did not overlap annotated coding regions or 5′ UTRs (the negative set, 8151 windows). To assay the stability of the classifier’s behavior and control for overfitting, we trained the classifier with fourfold cross-validation training on 2200 windows from the CDS set and 2200 windows from the distal 3′ UTR set, yielding an average misclassification error of 6.9–7.3% with each iteration. We repeated this analysis (and cross-validation) several times selecting different 81-nucleotide windows from each CDS and distal 3′ UTR, obtaining similar levels of error. The classifier was then trained on the entire training set, and used to evaluate randomly chosen 81 nt windows from observed C-terminal protein extensions that were 81 nt or greater in length. These included 26 extensions predicted by Jungreis et al. and 83 novel extensions.

We downloaded SNP data from the Drosophila Population Genomics Project from Ensembl.org (release 67; Flicek et al., 2013) and counted the proportion of SNPs that, if translated in frame, would cause synonymous substitutions in coding regions, extensions, and distal 3′ UTRs.

To test C-terminal extensions for differential readthrough rates, we examined all extensions which met the following criteria: (1) all annotated isoforms covering the extension contain exactly the same CDS, (2) the CDS had at least 128 total footprint reads in each of the S2 cell and embryo samples, and (3) the C-terminal extension had been scored as positive for readthrough in either the S2 and/or the embryo sample. For those extensions, we tabulated the footprint reads that aligned to the CDS and putative extension, masking out regions normally covered by start and stop codon peaks as described (see section ‘Readthrough rates’, above). For each extension, this tabulation yielded a 2 × 2 contingency table of reads aligning to the CDS and extension in the S2 cell and 0–2 hr embryo datasets. We evaluated the statistical significance of asymmetry in the contingency tables using Fisher’s exact test, and controlled the false discovery rate at 5% using the procedure of Benjamini and Hochberg (Benjamini and Hochberg, 1995).

For human cells, we collected data from uninfected human foreskin fibroblasts and processed it as previously described (Stern-Ginossar et al., 2012). Yeast samples were collected from [psi⁻] W303 and processed as previously described (Ingolia et al., 2009), with the exception that a 3′ linker ligation strategy was used instead of poly(A) tailing for fragment capture. For phasing of yeast footprints, we counted only 28-mers, which have previously been shown to be the best-phased footprint population in that organism (Ingolia et al., 2009).

C-terminal extensions 20 amino acids or longer were scanned for transmembrane domains using TmHmm (Krogh et al., 2001) using default settings. Nuclear localization signals were predicted in extensions 20 amino acids or longer using the cNLS mapper (Kosugi et al., 2009), with a score cutoff of 7.0. Peroxisome targeting signals were predicted for all extensions 12 amino acids or longer using PTS1 Predictor (Neuberger et al., 2003) with the signal type set to ‘metazoan’. Prenylation signals were predicted for all extensions 12 amino acids or longer using PrePS (Maurer-Stroh and Eisenhaber, 2005). In addition, we searched for ER retention signals using the consensus [KH]DEL*.

We searched 3′ UTRs (including the predicted extension and entire distal 3′ UTR) for selenocysteine insertion elements using SeciSearch 2.19 (Kryukov et al., 2003) with parameters set as follows: e1 = 05, e2 = −22, Y_filter = True, O_filter = True, B_filter = True, S_filter = True. We searched each 3′ UTR using every available SECIS Pattern (pat_c, pat_Sep20, pat_dm, pat_g, and pat_s), and considered a 3′ UTR receiving a COVE score above the recommended threshold of 15 in any of the pattern searches to contain a SECIS element. Additionally, we excluded any extensions that were annotated as selenoprotein annotations in SelenoDB (Castellano et al., 2008; for Drosophila, yeast, and human data) or FlyBase (Marygold et al., 2013; for Drosophila), For transcripts with no annotated or short 3′ UTRs, we extended the 3′ UTR in uninterrupted genome coordinates until it was 1000 nucleotides in length, an in Jungreis et al. (2011).

S2 cells stably transfected with the reporter of interest were maintained at a density of 1.6–12 million cells/ml. Nuclei were visualized by staining with 1 μg/ml Hoechst 34580 (Invitrogen) for at least 5 min. Live cells were imaged in culture media on an inverted spinning disk confocal Nikon Ti microscope (Nikon Instruments, Melville, NY) in glass-bottom culture dishes (MatTek, Ashland, MA). Images were contrast-adjusted and prepared for presentation in Adobe Photoshop (Adobe Systems, San Jose, CA).

Both the raw data (as FastQ files) and processed data (wiggle files) are available in NCBI’s Gene Expression Omnibus (Edgar et al., 2002) under GEO series accession number GSE49197 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE49197). Supplementary tables 1–4 are available at Dryad (Dunn et al., 2013; http://dx.doi.org/10.5061/dryad.6nr73):

Supplementary table 1: gene expression measurements in 0–2 hr embryos and S2 cells. Source data for Figures 1 and 2, as well as their supplements.

Supplementary table 2: readthrough statistics for Drosophila melanogaster. Source data for Figures 3, 4 and 6, as well as their supplements, and annotations of readthrough events in Drosophila melanogaster.

Supplementary table 3: readthrough statistics for Saccharomyces cerevisiae. Source data for Figure 5 and annotations of readthrough events in [psi^-] W303 yeast.

Supplementary table 4: readthrough statistics for human foreskin fibroblasts. Source data for Figure 5 and annotations of readthrough events in human foreskin fibroblasts.

Supplementary file 1 provides alignment statistics and Supplementary file 2 contains the oligonucleotides used in this study.

We wrote custom scripts in Python 2.7, using the following open-source libraries: Numpy 1.6.0 (http://numpy.scipy.org), Scipy 0.11.0rc2 (http://www.scipy.org), Biopython 1.59 (Cock et al., 2009), PySam, and HTSeq 0.5.1p2 (http://www-huber.embl.de/users/anders/HTSeq/doc/overview.html). Plots and genome browser snapshots were generated using Matplotlib 1.0.1 (Hunter, 2007).

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

We thank John Atkins for pointing out the importance of translation readthrough in Drosophila; Irwin Jungreis and Manolis Kellis for critical comments on the manuscript; Noam Ginossar for supplying the human foreskin fibroblast ribosome profiling data; Jeffrey Farrell, Tony Shermoen, and Hansong Ma for useful conversation and help with handling flies; Onn Brandman, Luke Gilbert, Noam Ginossar, Calvin Jan, Gene-wei Li, and Eugene Oh for advice on laboratory and analytical methods; and Clement Chu, Jessica Lund, and Silvi Rouskin for help with sequencing. We also thank Jessica Walter, the UCSF Cell Propulsion Lab, DeLaine Larsen, and the Nikon Imaging Center at UCSF for help with imaging.

© 2013, Dunn et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.