1. Computational and Systems Biology
  2. Genes and Chromosomes
Download icon

Tunable protein synthesis by transcript isoforms in human cells

  1. Stephen N Floor Is a corresponding author
  2. Jennifer A Doudna Is a corresponding author
  1. University of California, Berkeley, United States
  2. Howard Hughes Medical Institute, University of California, Berkeley, United States
  3. Lawrence Berkeley National Laboratory, United States
  4. University of California, Berkeley, California, United States
Research Article
Cited
18
Views
6,620
Comments
0
Cite as: eLife 2016;5:e10921 doi: 10.7554/eLife.10921

Abstract

Eukaryotic genes generate multiple RNA transcript isoforms though alternative transcription, splicing, and polyadenylation. However, the relationship between human transcript diversity and protein production is complex as each isoform can be translated differently. We fractionated a polysome profile and reconstructed transcript isoforms from each fraction, which we term Transcript Isoforms in Polysomes sequencing (TrIP-seq). Analysis of these data revealed regulatory features that control ribosome occupancy and translational output of each transcript isoform. We extracted a panel of 5′ and 3′ untranslated regions that control protein production from an unrelated gene in cells over a 100-fold range. Select 5′ untranslated regions exert robust translational control between cell lines, while 3′ untranslated regions can confer cell type-specific expression. These results expose the large dynamic range of transcript-isoform-specific translational control, identify isoform-specific sequences that control protein output in human cells, and demonstrate that transcript isoform diversity must be considered when relating RNA and protein levels.

https://doi.org/10.7554/eLife.10921.001

eLife digest

To produce a protein, a gene’s DNA is first copied to make molecules of messenger RNA (mRNA). The mRNAs pass through a molecular machine known as the ribosome, which translates the genetic code to make a protein. Not all of an mRNA is translated to make a protein; the “untranslated” regions play crucial roles in regulating how much of the protein is produced.

In animals, plants and other eukaryotes, many mRNAs are made up of small pieces that are “spliced” together. During this process, proteins are deposited on the mRNA to mark the splice junctions, which are then cleared when the mRNA is translated. Many different mRNAs can be produced from the same gene by splicing different combinations of RNA pieces. Each of these mRNA “isoforms” can, in principle, contain a unique set of features that control its translation. Hence each mRNA isoform can be translated differently so that different amounts of the corresponding protein product are produced. However, the relationship between the variety of isoforms and the control of translation is complex and not well understood.

To address these questions, Floor and Doudna measured the translation of over 60,000 mRNA isoforms made from almost 14,000 human genes. The experiments show that untranslated regions at the end of the mRNA (known as the 3′ end) strongly influence translation, even if the protein coding regions remain the same. Furthermore, the data showed that mRNAs with more splice junctions are translated better, implying an mRNA has some sort of memory of how many junctions it had even after the protein markers have been cleared.

Next, Floor and Doudna inserted regulatory sequences from differently translated isoforms into an unrelated “reporter” gene. This dramatically changed the amount of protein produced from the reporter gene, in a manner predicted by the earlier experiments. Untranslated regions at the beginning of the mRNAs (known as the 5′ end) controlled the amount of protein produced from the reporter consistently across different types of cells from the body. On the other hand, the 3′ regions can tune the level of protein production in particular types of cells.

Floor and Doudna’s findings demonstrate that differences between mRNA isoforms of a gene can have a big effect on the level of protein production. Changes in the types of mRNA made from a gene are often associated with human diseases, and these findings suggest one reason why. Additionally, the ability to engineer translation of an mRNA using the data is likely to aid the development of mRNA-based therapies.

https://doi.org/10.7554/eLife.10921.002

Introduction

Eukaryotic genes can produce a staggering diversity of messenger RNA products. In humans, there is a median of five annotated transcript isoforms per gene, with more than 80 annotated isoforms in some cases (Ensembl release 75). These transcript isoforms arise from the combined action of alternative transcription initiation, splicing and cleavage, and polyadenylation, which are often cell type specific (Barbosa-Morais et al., 2012; Mele et al., 2015; Merkin et al., 2012; Wang et al., 2008; Xiong et al., 2015; Zheng and Black, 2013). As each of these transcript isoforms may in principle harbor a unique set of translational control elements, gene-level expression, which averages out the translational potential of each individual transcript isoform, may not be an accurate measure of protein levels. Since changes in both translation and splicing are linked to numerous human disorders, it is critical to understand the relationship between the two (Maslon et al., 2014; Piccirillo et al., 2014; Ruggero, 2013; Xiong et al., 2015).

Eukaryotic mRNAs are decorated with diverse sequence features that can control translation, which can vary between transcript isoforms (de Klerk and 'tHoen, 2015; Ingolia et al., 2011; Sterne-Weiler et al., 2013). Classic examples of translational control include upstream ORFs (uORF)-mediated translational control of yeast GCN4 (Hinnebusch, 2005), protein binding such as the iron regulatory protein (Gray and Hentze, 1994), and the action of micro-RNAs (Nottrott et al., 2006; Wilczynska and Bushell, 2015) or DEAD-box proteins such as eIF4A and Ded1 (Chuang et al., 1997; Hinnebusch and Lorsch, 2012; Sen et al., 2015). Alternative 5′ leader sequences, uORFs, and select tandem 3′ untranslated region (UTR) isoforms have been demonstrated to influence protein production (Brar et al., 2012; Hinnebusch, 2005; Ingolia et al., 2011; Mayr and Bartel, 2009; Sandberg et al., 2008; Zhang et al., 2012). Any of these features may in principle be different between transcript isoforms, but the prevalence and dynamic range of isoform-specific translational control across the human genome is currently unknown.

Previous work measuring genome-wide translation in human cells has focused largely on the relationship between gene-level mRNA abundance and protein levels, which is blind to the contribution of transcript isoforms. Ribosome profiling is not well-suited for measuring transcript isoform-specific translation, primarily due to the short ~30 bp length of ribosome-protected fragments (Ingolia, 2014). Prior attempts to characterize isoform-specific translation have measured the effects of 5′ end diversity in yeast (Arribere and Gilbert, 2013) and 3′ end diversity in mouse cells (Spies et al., 2013), or splicing differences between cytoplasmic and aggregate polysomal mRNAs (Maslon et al., 2014; Sterne-Weiler et al., 2013). However, sequencing just the ends of transcripts cannot distinguish between transcript isoforms of the same gene harboring degenerate termini. In addition, aggregating polysome fractions averages lowly- and highly-ribosome-associated messages. Therefore, a different strategy is required to understand how the diversity of the human transcriptome impacts translational output.

Here, we adapt a classic approach of polysome profiling coupled with global gene expression analysis (Arava et al., 2003) to measure transcript-isoform specific translation using deep sequencing, which we term Transcript Isoforms in Polysomes sequencing (TrIP-seq). By using high gradient resolution and sequencing depth, this approach yields polysome profiles for over 60,000 individual transcript isoforms representing almost 14,000 protein coding genes. We observe frequent intron retention on ribosome-associated transcripts, even in high-polysome fractions, identifying a population of retained but not nuclear-detained introns (Boutz et al., 2015). Properties of 3′ untranslated regions predominate over the 5′ leader sequence as the driving force behind differential polysome association for transcript isoforms of the same gene among the transcript features tested. We show that regulatory sequences differentially included in transcript isoforms of the same gene are modular and can trigger differences in the translation of reporters spanning two orders of magnitude. These findings provide a lens through which to ascribe functional consequences to RNA-seq-generated transcriptomes. Moreover, TrIP-seq analysis uncovers regulatory elements that can be utilized to tune translation of synthetic messages robustly in cells.

Results

TrIP-seq measures transcript isoform-specific translation in human cells

We determined the ribosomal association of transcript isoforms by sequencing transcripts cofractionating with different numbers of ribosomes with sufficient depth to determine isoform abundances, as was performed at the gene level in yeast (Arava et al., 2003). We treated HEK 293T cells with cycloheximide to stall translation and fractionated the cytoplasm into ribosome-containing samples including one to eight or more ribosomes (Figures 1A and Figure 1—figure supplement 1A; see Materials and methods for details). We made RNA sequencing libraries from each fraction in biological duplicate and obtained transcript-level abundances using the Cufflinks suite (Figure 1—source data 1 and 2; [Trapnell et al., 2010]). Clustering of the samples recapitulates the gradient order (Figure 1B), indicating the polysome profile was accurately fractionated. Four subgroups emerge from this clustering: the 80S (monosome), low polysomes (two-four ribosomes), high polysomes (five-eight+ ribosomes), and total cytoplasmic RNA. We tested the robustness of the clustering of samples by computing the average Jaccard distance between clusters from data subjected to three different resampling methods, which was ≥0.75, suggesting stable clusters (Figure 1—figure supplement 1G; Materials and methods). This suggests that in cells, transcript isoforms are predominantly poorly- or highly-ribosome associated, causing low polysomes to cluster away from high polysomes, which could be have numerous biological origins including highly abundant short ORFs.

Figure 1 with 3 supplements see all
Transcript Isoforms in Polysomes sequencing (TrIP-seq) measures transcript isoform specific translation.

(A) HEK 293T cells were treated with cycloheximide and the cytoplasmic fraction was extracted and applied to a sucrose gradient, which was further fractionated into individual polysomes that were converted into sequencing libraries. (B) Intersample clustering recapitulates the gradient order of polysomes, indicating the sequenced fractions are faithful to the gradient profile. (C,D) RT-PCR analysis and transcript-level quantification for ACTB and ATF4 (C) and three transcripts of EEF1B2 (D) demonstrating concordance of sequencing and transcript-specific RT-PCR. TPM – transcripts per million. *nonspecific amplicon. (E) Members of transcript classes with more than 100 reads in indicated fractions show that coding genes are represented across the polysome while noncoding genes are preferentially in low polysome fractions. (F) Spearman’s correlation (RS) between gene (left) or isoform (right) read counts from ribosome profiling or TrIP-seq. See also Figure 1—figure supplement 1 and Materials and methods for calculation of TrIP-seq polysome counts.

https://doi.org/10.7554/eLife.10921.003

To confirm the sequencing data, we first analyzed beta-actin (ACTB), which is known to be heavily translated with many ribosomes on each message, and migrates with high polysomes in a gradient (Figure 1C; (Sterne-Weiler et al., 2013; Zhang et al., 2015). However, ACTB only has one transcript isoform, so we examined ATF4, which has one dominant isoform exhibiting low-polysome association in TrIP-seq data and by RT-PCR (Figure 1C), consistent with it being translationally repressed in the absence of cellular stress (Harding et al., 2000). We then analyzed eukaryotic elongation factor 1 beta 2 (EEF1B2), which has three isoforms with the same coding sequence but alternative 5′ leaders as a representative gene that exhibits isoform-specific ribosome association (Figure 1—figure supplement 1B). The isoform-specific polysome profiles and RT-PCR conducted from polysome fractions agree qualitatively, validating the accuracy of the isoform-level quantifications (Figure 1D). Weaker amplification of the longer isoform (EEF1B2-003) in high polysomes may be due to PCR bias towards smaller amplicons (Walsh et al., 1992), and replicate experiments show EEF1B2-003 present in high polysomes (Figure 1—figure supplement 1C). We additionally validated the TrIP-seq data using qRT-PCR from polysome fractions and find that the two measurements agree for both EEF1B2 and SRSF5 (Figure 1—figure supplement 2). Both coding mRNAs and long noncoding RNAs (lncRNAs) associate with ribosomes (Ingolia et al., 2014; van Heesch et al., 2014), so we measured their polysomal abundance. Protein-coding genes and isoforms are found across the polysome, but noncoding and lncRNA genes are predominantly in the low polysome fractions indicating that they are generally weakly ribosome associated or in other large macromolecular complexes (Figure 1E and Figure 1—figure supplement 1D).

To determine whether TrIP-seq is a measure of translation as opposed to cryptic association with large macromolecular complexes or stalled ribosomes (Darnell et al., 2011; Ishimura et al., 2014), we compared TrIP-seq data to ribosome profiling and proteomics datasets. First, we reanalyzed a ribosome profiling dataset from HEK 293T cells and compared the number of ribosome-protected fragments to a weighted sum of the TrIP-seq polysome reads (Materials and methods; nribo ≥2) (Sidrauski et al., 2015). We observe a strong correlation between ribosome profiling and TrIP-seq at the gene level (RS = 0.88; Figure 1F). However, at the isoform level, the correlation decreases (RS = 0.47), which is worse than the correlation between TrIP-seq replicates (RS = 0.90; Figure 1—figure supplement 1E), suggesting the discrepancy is not due to variability in TrIP-seq but instead is likely due to known issues quantifying transcript isoforms in ribosome profiling (Ingolia, 2014). We next computed the translation efficiency (TE) of ribosome profiling and TrIP-seq data by dividing by cytoplasmic RNA levels. TE measured by the two methods correlated with RS = 0.24, perhaps due to variability in culture conditions or in either technique. We compared the data of Sidrauski et al to an unpublished ribosome profiling dataset in 293T cells from the Yoon-Jae Cho lab. Ribosome-protected fragments between experiments correlate with RS = 0.78, while TE has a substantially lower correlation of RS = 0.41, suggesting biological or technical variability disproportionately affects TE. We then compared TrIP-seq data to protein abundances from HEK 293T cells, and find that the two datasets correlate (RS = 0.57, Figure 1—figure supplement 1F, (Geiger et al., 2012), which is better than RNA-seq (RS = 0.45; not shown). The observations that most translational stalling in mammalian cells is transient and translation elongation rates are homogeneous (Ingolia et al., 2011), TrIP-seq appears to underestimate rather than overestimate protein abundance (Figure 1—figure supplement 1F), and TrIP-seq correlates well at the gene-level with ribosome profiling (Figure 1F) suggest that TrIP-seq primarily measures translating ribosome association as opposed RNAs bound to other large complexes.

Diverse human isoform-specific polysome association patterns

To extract global trends in isoform-specific translation, we hierarchically clustered transcript isoform polysome profiles and selected eight clusters that are representative of general trends in the data (Figure 2; Materials and methods). The depth of sequencing (Figure 1—figure supplement 3), augmented by the fractionation strategy, enables detection of 62,703 transcript isoforms in the polysome profile (Figure 2—source data 1). Isoforms in the observed clusters exhibit diverse average patterns across polysomes (Figure 2A,B), from clusters 1 and 2, which contain isoforms primarily in high polysomes, to cluster 3 with isoforms in the middle, to clusters 6 and 7 where isoforms are in low polysomes. Independent clustering of the two biological replicate TrIP-seq datasets shows that the high- and low-polysome clusters (1, 2 and 6) appear more robust when comparing between averaged and individual replicate clusterings (Figure 2—figure supplement 1A). Many clusters have similar total polysome abundance but different distributions, indicating that to obtain accurate measurements of isoform-specific translatability it is crucial to fractionate the polysome profile.

Figure 2 with 1 supplement see all
Clustering of transcript distributions yields eight major clusters with diverse behavior across the polysome profile.

(A) The average relative abundance of all isoforms in each cluster across polysomes is shown. Error is s.d. (B) Hierarchical clustering of 62,703 transcript isoform distributions across the polysome profile and cytoplasmic fraction. Yellow: above isoform average, cyan: below isoform average. (C) Transcript type distribution per cluster from Ensembl-annotated biotypes. Dotted lines mark the abundance of each transcript type in all isoforms that went in to the clustering. See also Figure 1—source data 1 and 2 and Figure 2—figure supplement 1.

https://doi.org/10.7554/eLife.10921.009

Surprisingly, the poorly translated cluster 7, which has ~9,000 transcripts, contains a similar number of annotated retained intron or protein-coding isoforms (Figure 2C). We queried the gene ontology (GO) terms associated with each cluster and found that the poorly translated clusters 6 and 7 are enriched for translation and splicing genes (Figure 2—figure supplement 1B), implicating the alternative splicing-nonsense-mediated decay (AS-NMD) pathway (Jangi et al., 2014; Lareau et al., 2007). However, not all retained intron transcripts are subject to NMD (Boutz et al., 2015; Gohring et al., 2014), and we also observe retained intron transcripts in the best-translated clusters (1 and 2), indicating at least some of the introns may be translated. In sum, clustering of isoform abundance distributions across polysomes immediately provides insight into the diverse patterns of ribosome association by individual isoforms and transcript types in cells, and reveals intron-rich clusters of transcripts associated with polysomes that escape nuclear detention.

3′ UTRs and introns drive isoform-specific polysome association

We reasoned we could extract features regulating the translation of a gene by comparing transcript isoforms of the same gene that are well- or poorly-translated. We selected 24 features to explore that are involved in translational control, such as the length of the coding sequence and untranslated regions, predicted secondary structure, and microRNA binding sites. As their average polysome profiles and composition are broadly similar (Figure 2), we merged clusters 1 and 2 to generate a larger pool of high-polysome isoforms (Figure 3A), and compared these to cluster 6, representing poorly translated isoforms in low polysomes. We then compared transcript isoforms of the same gene that are in high- and low-polysome clusters, or gene-linked isoforms, to extract transcript features that influence translation. The number of gene-linked isoforms per feature per set varies between 569 and 6491. We then measured the effect size and calculated statistical significance for all features between gene-linked isoforms (Figures 3B and Figure 3—figure supplement 1A; Materials and methods; [Cliff, 1993]).

Figure 3 with 1 supplement see all
Effect of transcript features on polysome association.

(A) Meta-transcript distributions for two high polysome clusters (1 and 2) and one low polysome cluster (6). Clusters 1 and 2 were pooled for the analysis in (B). (B) The distance between distributions for 24 different transcript features evaluated for transcripts strongly or weakly associated with polysomes. Distance is the nonparametric effect size, measured as the dimensionless quantity Cliff’s d (see Materials and methods) and error bars are bootstrapped 95% confidence intervals. All differences except 5′ UTR and transcript length are significant at the p = 0.001 level based on two-tailed Mann-Whitney U-tests (Figure 3—figure supplement 1A). See Materials and methods for a description of all features and how they were tabulated. UTR – untranslated region; CDS – coding sequence. (C) Enrichment of either cognate (ATG start codon) or non-cognate (non-ATG) uORFs in high polysome versus low polysome clusters. Density is uORFs per 100 isoforms. See also Figure 3—figure supplement 1.

https://doi.org/10.7554/eLife.10921.012

We find that longer coding sequences and highly abundant versions of gene-linked isoforms are biased towards high polysomes, likely because shorter coding sequences cannot accommodate as many ribosomes (Figure 3B and Figure 3—figure supplement 1A). However, the length of the coding sequence is not the sole determinant of polysome association, since the ribosome density (measured by dividing the weighted sum of TrIP-seq polysome reads by the number of cytoplasmic reads) is also higher in gene-linked isoforms found in high polysomes (Figure 3B). We also find that transcripts from gene-linked isoforms with more exons also tend to be better translated, as was observed using reporter genes (Nott et al., 2003; 2004). This is not due to overall transcript length, as there is no significant difference between transcript length in gene-linked isoforms, reflecting a positive influence of splicing on translation. Highly translated isoforms contain fewer rare codons on average, but contain more stretches of rare codons, perhaps indicating a need for translational pausing (Figure 3B). The observation that exons promote translatability has been shown for select genes, and here we show that this extends across the human genome.

One of the strongest effects seen on polysome association of gene-linked isoforms comes from the length and content of the 3′ UTR (Figure 3B and Figure 3—figure supplement 1A). Specifically, we find that gene-linked 3′ UTRs in low polysomes are considerably longer than those on high polysomes (mean length 1551 nt versus 982 nt). There are numerous regulatory elements contained within 3′ UTRs, including microRNA binding sites, AU-rich elements, and protein-binding sites (Bartel, 2009; Szostak and Gebauer, 2013). The fraction of the 3′ UTR containing AU-rich elements is increased in low polysome gene-linked isoforms, possibly due to translational repression (Brooks and Blackshear, 2013; Moore et al., 2014). Conserved predicted binding sites for miRNAs (Garcia et al., 2011) are also more abundant in poorly translated gene-linked isoforms, but this could be due to a correlation between 3′ UTR length and the number of miRNA binding sites. We, therefore, filtered the miRNAs to those bound to AGO1 or AGO2 in HEK 293T cells (Ender et al., 2008). In this HEK 293T-expressed miRNA set, the difference for both the number of conserved miRNA binding sites and their score increases between gene-linked isoforms, suggesting that miRNA binding may be functionally relevant in this context (Figure 3B and Figure 3—figure supplement 1A). Increased AU-rich elements and miRNA binding should lead to decreased half-life for low polysome isoforms, which is borne out in half-life comparisons (Figure 3B and Figure 3—figure supplement 1A; [(Tani et al., 2012]).

Surprisingly, gene-linked isoforms in high polysomes have increased predicted structure both at the cap and in 75-nucleotide windows in the 5′ leader. There are at least three possibilities that could explain this result. First, RNA structures can both repress or promote translation in a context dependent manner (Xue et al., 2015), so it is possible that some isoforms are being driven to high polysomes by recruitment of transacting factors. Second, recent studies of global RNA secondary structure have consistently observed different structure in cells than in vitro (Rouskin et al., 2014; Spitale et al., 2015). Lastly, HEK 293T cells may translate messages with inhibitory 5′ leaders more efficiently than other cell types. In contrast, we find no significant dependence for the length of the 5′ leader region on polysome association based on this comparison (Figures 3B and Figure 3—figure supplement 1A). However, isoforms containing 5′ leaders over 1000 nucleotides long are poorly ribosome-associated relative to shorter 5′ leaders (Figure 3—figure supplement 1B) and cluster 7, which is associated with few ribosomes, contains longer 5′ leaders (Figure 3—figure supplement 1C). It is likely that more subtle features of the 5′ leader also influence translatability, as in the examples in Figure 1D.

Upstream open reading frames (uORFs) can positively or negatively influence translation (Brar et al., 2012; Calvo et al., 2009; Ferreira et al., 2013; Hinnebusch, 2005). We, therefore, counted the number of experimentally determined uORFs (Wan and Qian, 2014) in each cluster and observe surprisingly complex behavior. Cognate uORFs (those starting with ATG) promote or repress translation, while noncognate uORFs generally promote translation (Arribere and Gilbert, 2013; Brar et al., 2012). Along these lines, we find that cluster 1 (high polysomes) is enriched for cognate and noncognate uORFs while cluster 6 (low polysomes) is enriched for cognate uORFs. However, cluster 2 (high polysomes) is enriched for only noncognate uORFs, indicative of the complex and idiosyncratic behavior of uORFs. Taken together, we find predominant influences for the 3′ UTR and the number of introns in determining polysome occupancy of gene-linked transcript isoforms in human cells, while a diversity of other features can influence translatability to a lesser extent.

Predicting translation changes during preimplantation human development

RNA-seq characterizes the isoform diversity of a sample, but not the functional consequences of this diversity. We reasoned that TrIP-seq data could be used to predict which isoform changes are likely to lead to translation changes in other systems. We chose a human embryonic dataset containing 124 individual cells from seven preimplantation developmental stages and reprocessed these data to directly compare to TrIP-seq data (Materials and methods; [Yan et al., 2013]). We mapped transcripts expressed during each developmental stage onto the TrIP-seq clusters, to attempt to gain insight into global translational properties in human embryos (Figure 4A). In embryos, our analysis predicts a shift towards the extremes of translation, with an increase in the percent of transcript isoforms that are both highly (e.g. clusters 1 and 2) and lowly (e.g. cluster 6) translated isoforms in HEK 293T cells (Figure 4A). Localized translation is widespread in development, and it is possible that the observed increase in poorly translated isoforms reflects a greater need for translational control (Besse and Ephrussi, 2008; Jung et al., 2014). It is also possible that translation in early embryos is differentially regulated than in HEK 293T cells, which is now testable by applying TrIP-seq to other cell types, yielding cell type-specific translational control programs.

Figure 4 with 2 supplements see all
Predicted isoform-specific translational control changes during human embryogenesis.

(A) For each embryonic stage, transcripts are mapped onto all eight TrIP-seq clusters, as in Figure 2. The percentage of transcripts mapping to each cluster per stage is then calculated, and compared to the percentage for TrIP-seq data. hESC – human embryonic stem cell; p0, p10 – passage zero or passage ten. Error is s.d. between single cells at each embryonic stage or between TrIP-seq biological replicates. (B) Expression levels for two transcripts of CSDE1 across embryonic development and polysome fractions demonstrating a switch in translational status. Error is S.D. TPM – transcripts per million. (C) Diagram of two transcript isoforms of CSDE1. Regions different between the two isoforms are in yellow and select shared intronic regions have been shortened for clarity. See also Figure 4—figure supplement 1.

https://doi.org/10.7554/eLife.10921.014

We then collected genes with isoforms that change between early and late embryonic stages. Clustering of abundances for 45,895 isoforms across embryogenesis yielded seven clusters with varying profiles, which we then analyzed by comparing transcripts of the same gene located in different embryonic clusters as before (Figure 4—figure supplement 1A). These isoforms were filtered by those that move between low- and high-polysome clusters in the TrIP-seq data, yielding 366 isoform pairs belonging to 270 genes that are developmentally regulated and exhibit differential translation in HEK 293T cells. For example, the CSDE1 gene expresses one isoform in oocytes, early development and in human ES cells, which is poorly translated (CSDE1-002), but a second appears following fertilization that is well-translated (CSDE1-007; Figure 4B,C). As global zygotic transcription begins in the two-cell stage (Vassena et al., 2011; Yan et al., 2013), it is possible that the alternative CSDE1 isoform is supplied by the spermatid (Fischer et al., 2012; Soumillon et al., 2013) or that it may be precociously transcribed. We also present three other genes (EIF4A2, RNF170, and TBC1D15) that show stage-specific expression in the embryo data and differential polysome association in HEK 293T cells (Figure 4—figure supplement 2). Therefore, widespread changes at the transcriptome level during human embryogenesis may produce concomitant changes in protein production, which can now be predicted using TrIP-seq data.

Tunable translation by isoform-specific regulatory elements

Engineering the translation of a transcript without altering its coding sequence is desirable when introducing exogenous mRNA to cells or patients. We hypothesized that the regulatory features we discovered by comparing highly to lowly translated isoforms (Figure 3B) should be transferrable to an arbitrary reporter. Nine different 5′ leaders and eight different 3′ UTRs derived from gene-linked isoforms (Figure 3) were appended onto Renilla luciferase with a synthetic poly-A60 tail, which were individually in vitro transcribed, capped, and 2′-O-methylated (Figure 5—figure supplement 1A and Materials and methods). We elected to transcribe RNA both because plasmid-encoded transcripts can be heterogeneous and to mimic a scenario where one is delivering RNA to affect cell activity or for therapeutic intervention (Kormann et al., 2011; Warren et al., 2010).

All tested 5′ leader sequences modulate protein production by the luciferase reporter in concordance with the observed polysome association by TrIP-seq, when transfected into HEK 293T cells (Figure 5A,B). The three 5′ leaders from EEF1B2 alter luciferase production in a stepwise manner by roughly a factor of 20. The TrIP-seq profiles from NAE1 are the most distinct among those tested, and the two luciferase constructs differ in output by two orders of magnitude. Surprisingly, the two 5′ leaders from RICTOR differ by only nine nucleotides, yet still exhibit differential protein production, possibly due to altered local RNA secondary structure near the 5′ cap (Figure 5—figure supplement 1B). Both 5′ leaders from SRSF5 contain two uORFs, but they are close (SRSF5-002) or far (SRSF5-005) from the start codon (Figure 5—figure supplement 1A), suggesting reinitiation following uORF translation may be impacting luciferase production (Grant et al., 1994; Hinnebusch, 2005). Not only do these data provide strong evidence that TrIP-seq data can be used to predictably tune the output of heterologous mRNAs using isoform-specific untranslated regions, it additionally validates that the polysome profiles observed are connected to translational output.

Figure 5 with 1 supplement see all
Regulatory UTR sequences are sufficient to control translation of a heterologous reporter.

HEK 293T cells were transfected with indicated RNAs and Renilla luciferase units are plotted. (A–D) Renilla light units normalized to control UTR mRNA (A,C) and corresponding traces from TrIP-seq data (B,D) for 5′ leaders (A,B) or 3′ UTRs (C,D) exhibit differential protein production over two log-units in cells. Inset graphs show comparisons between mock-transfected cells and the lowest output mRNAs. Error bars are S.E.M. from at least four biological and three technical replicates (twelve total; A) or three biological and three technical replicates (nine total; B). (B,D) Error is S.D. Rluc – Renilla luciferase units; UTR – untranslated region. (E) Luciferase fold change versus the average number of ribosomes on each transcript, computed by averaging the plots in B and D. See also Figure 5—figure supplement 1.

https://doi.org/10.7554/eLife.10921.017

We additionally selected eight 3′ UTRs from four sets of gene-linked isoforms to test their ability to control translation, and find good agreement with the TrIP-seq polysome abundance for two out of four 3′ UTR pairs (Figure 5C,D). In all cases, replacing the 3′ UTRs leads to decreased protein production compared to the short 3′ UTR in the control RNA. The paired 3′ UTRs of NAB1 and RICTOR each differ by two kilobases, and a factor of 35 and 44 in luciferase output, respectively, indicating that 3′ UTRs can strongly modulate protein production (Mayr and Bartel, 2009; Sandberg et al., 2008). However, the two other paired 3′ UTRs, despite also differing by two kilobases each, are not distinguishable at the protein production level in cells. The two tested CCNE2 isoforms differ in both the 5′ and 3′ UTRs, so it is possible that the isoform-level translational control is occurring via the 5′ UTR, and there is a small but not significant difference between the two NDC1 3′ UTRs, suggesting regulation of this 3′ UTR may be subtle. We also find a positive relationship between the average number of ribosomes on each transcript isoform in TrIP-seq and the luciferase fold change (Figure 5E). In sum, we show that elements found to control translation at the isoform level using TrIP-seq can be grafted on to heterologous coding sequences to control translation over a range of two orders of magnitude in human cells.

Translational control of reporters by the 5′ leader is robust between cell types

As different cell types contain different macromolecules, they may also translate messages with the same regulatory features differently. We, therefore, tested the panel of gene-specific untranslated regions fused with luciferase (Figure 5) in four additional cell lines: A549 (lung carcinoma), K-562 (chronic myelogenous leukemia), MCF-7 (breast adenocarcinoma), and Hep G2 (hepatocellular carcinoma). We additionally tested HEK 293T cells at a shorter timepoint of 2 hr, to explore the role of RNA stability in the observed luciferase output.

Remarkably, the luciferase production from all tested 5′ leader reporters was qualitatively similar in all six conditions (Figure 6A). However, the difference between the two RICTOR 5′ leader sequences is decreased. In contrast, the 3′ UTR reporters show considerably more variability (Figure 6B). Calculation of the coefficient of variation of all reporters across all six conditions highlights the greater variability found in 3′ UTR reporters (Figure 6—figure supplement 1A). We additionally compared HEK 293T cells at 2-hr and 8-hr post-transfection to ascertain differences in RNA stability. Most 5′ leader reporters are similar between these two timepoints (Figure 6A), but 3′ UTR reporters show larger changes. Specifically, NAB1-001 and RICTOR-001 generate much more luciferase at the 2-hr timepoint, suggesting the long UTRs of these genes may promote RNA degradation leading to decreased protein at eighteen hours.

Figure 6 with 1 supplement see all
Translational control by transcript 5′ leader sequences is robust across cell types.

(A,B) Five cell lines in six conditions were transfected with Renilla luciferase fused to the UTR indicated and luciferase units normalized to control UTR mRNA. Both 5′ leaders (A) and 3′ UTRs (B) were tested. Error is S.E.M. between three technical and three biological replicates (nine total) per condition. See also Figure 6—figure supplement 1.

https://doi.org/10.7554/eLife.10921.019

Thus, translational control by the 5′ leader sequence is robust to changes in macromolecular composition across the cell lines tested, perhaps because the abundance of most translation initiation factors is high (Kulak et al., 2014), which will buffer against small changes in their expression level. However, factors interacting with 3′ UTRs, specifically miRNAs, can vary considerably between cell types, and this variability may be the cause of the diverse behavior of the 3′ UTR reporters across cell lines. Therefore, at least for the panel tested here, transcript 5′ leaders confer robust control of protein production across cell types, while transcript 3′ UTRs may be better suited to tune the production of a protein to a particular cell type.

Discussion

In this work, we showed that the dynamic range of transcript-isoform-specific translational control spans at least two orders of magnitude in human cells (Figure 5), indicating that it is crucial to take isoform-level effects into account when assessing translation in organisms with extensive alternative transcript processing (Figure 7A). Globally, we find that alternative 3′ UTRs broadly influence translation, with longer isoforms of the same gene associated with lower protein production (Figure 3B). We further show that regulatory regions are sufficient to control the translation of unrelated coding sequences, enabling predictable tuning of translational output of arbitrary genes (Figures 5 and 6B). We demonstrate that translational control conferred by a panel of 5′ leaders is more robust across cell types than 3′ UTRs, suggesting predictable control of protein production. Future work focused on measuring isoform-specific translation in different cell types will yield cell type-specific regulatory sequences, which could then be used to engineer cell type-specific translation of mRNA, as in derivation of pluripotent stem cells (Warren et al., 2010), or mRNA therapeutics (Kormann et al., 2011).

Transcript processing has widespread effects on protein production.

(A) The mapping between gene and protein is complex, and RNA processing has a strong influence on the translatability of individual transcript isoforms and therefore protein production. (B) Regulatory elements can be discovered by comparing transcript isoforms that are differentially translated, and subsequently fused to heterologous genes to control protein output in cells. ORF – open reading frame.

https://doi.org/10.7554/eLife.10921.021

Several observations indicate that TrIP-seq is a faithful measure of translation in cells. First, TrIP-seq and ribosome profiling are highly correlated at the gene level (Figure 1F) and both ribosome profiling and TrIP-seq correlate with mass spectrometric measurements of protein abundance ([Ingolia et al., 2009], Figure 1—figure supplement 1F). Second, protein-coding genes are enriched in high ribosomal fractions (Figure 1E), while lncRNAs are found primarily in the low ribosomal fractions (Figure 1E and Figure 1—figure supplement 1D). Third, a systematic investigation of translational stalling in mouse embryonic stem cells found this phenomenon to primarily be transient and not lead to ribosome accumulation on transcripts (Ingolia et al., 2011), although in specific circumstances it can be more widespread (Darnell et al., 2011; Ishimura et al., 2014Richter and Coller, 2015). Fourth, known highly- and lowly-translated transcript isoforms are enriched in high- and low-polysome fractions in TrIP-seq data (Figure 1C). Fifth, reads derived from stalled polysomes would artificially inflate the apparent translatability, however, TrIP-seq primarily underestimates rather than overestimates protein abundance (Figure 1—figure supplement 1F). Sixth, cryptic 'pseudo-polysomes' induced by miRNA complexes would preferentially enhance polysome association for long 3′ UTR transcript isoforms (Maroney et al., 2006; Nottrott et al., 2006; Olsen and Ambros, 1999; Thermann and Hentze, 2007), which is the opposite of what is observed (Figure 3B). Lastly, transcript isoform changes found to lead to differential ribosome association using TrIP-seq are sufficient to modulate the protein output from a reporter RNA in a manner predicted by the TrIP-seq data (Figure 5), indicating that transcript-specific ribosome association is a correlate of protein output. Therefore, we conclude that TrIP-seq is a measure of transcript-specific translation.

What are the mechanistic foundations of the observed transcript-specific translation? Each of the different untranslated regions tested by reporter assays operate using different mechanisms. For example, the three EEF1B2 transcript isoforms all share the same uORF, but alternative splicing changes the distance between the uORF and the start codon, which can influence the efficiency of downstream initiation (Grant et al., 1994; Hinnebusch, 2005). Globally, we found a large difference between the 3′ UTR lengths of poorly- versus well-translated isoforms by querying the prevalence of features likely to influence translation (Figures 3 and 5). There are at least three possibilities for how 3′ UTRs could influence translation. First, miRNA-mediated regulation targets the 3′ UTR (Mayr and Bartel, 2009; Sandberg et al., 2008). Second, numerous RNA binding proteins target the 3′ UTR through AU-rich elements or other specific binding sites and are known to influence translation, mRNA decay and even protein localization (Berkovits and Mayr, 2015; Szostak and Gebauer, 2013; Zhao et al., 2014). Lastly, it is possible that increasing the length of the 3′ UTR influences the impact of the mRNA closed loop on translation initiation or reinitiation (Amrani et al., 2008; Costello et al., 2015). Targeted work on specific 3′ UTR isoform sets could test the different possibilities for 3′ UTR-mediated translational control. We additionally demonstrated isoform-specific translational control for select 5′ leaders (Figures 1 and 5) as was shown genome-wide in yeast, where 5′ leaders are short enough to directly sequence (Arribere and Gilbert, 2013), and find that globally uORFs can both up- and down-regulate translation (Figure 3). However, we find a surprisingly minor global dependence of other 5′ leader features tested on translation. The mechanisms of translational control of transcript isoforms therefore follow some general trends but are likely highly idiosyncratic; these data empower investigations into the mechanisms underlying transcript-specific translational control for thousands of human genes.

We predicted the consequences of transcript isoforms observed during human embryonic development on protein production (Figure 4). We show that even small changes to transcripts can dramatically affect protein production (Figure 5), so changes at the isoform level (Figure 4) can considerably affect protein abundances even if the gene-level RNA expression remains similar. The application of TrIP-seq to different cell types should yield cell type-specific translation enhancer and repressor elements, contributing to the understanding of cell type-specific translational control, and augmenting our ability to precisely engineer translation in complex systems. If simpler sample preparation is desired, it is likely sufficient to sequence the cytoplasmic and monosome fractions, as well as pooled two-four ribosome and five-eight+ ribosome fractions, as these cluster together (Figure 1B). Accumulation of TrIP-seq data in additional cell types may enable building a holistic model of translational control with isoform resolution in human cells.

We showed that the translation of an arbitrary gene, Renilla luciferase, can be controlled in a manner predicted by TrIP-seq data over two orders of magnitude (Figure 5) and in different cell types (Figure 6). These data can thus be used to select regulatory regions to control translation, without redesign of the coding sequence of the message (Figure 7B). It may prove superior to use 3′ UTRs to design cell type specific translation, as the repertoire of miRNAs and RNA-binding proteins that may affect translation through the 3′ UTR vary between cell types (Figure 6B). Indeed, in Caenorhabditis elegans, 3′ UTRs are sufficient to specify germline-specific expression of the attached ORF (Merritt et al., 2008). Even without engineering of unnatural mRNAs, it may be possible to use antisense oligonucleotides to direct splicing of poorly- or well-translated isoforms to adjust protein expression in situ (Kole et al., 2012), for example to downregulate the oncogenic NAE1 gene [Figure 5A, Xie et al., 2014]), which is currently being targeted by small molecules (Luo et al., 2012; Wu and Yu, 2015). We anticipate the ability to tune the translational output of mRNA will facilitate research and therapeutic uses of designed and endogenous mRNA molecules.

Materials and methods

Sucrose gradient fractionation of polysome profiles

Two independently passaged, biological replicate 15 cm dishes of HEK 293T cells obtained from the University of California, Berkeley cell culture facility were grown to ~70% confluency in DMEM + 10% FBS. The cell line was authenticated by DDC Medical (Fairfield, OH) and were verified to be free of mycoplasma contamination. Cells were actively growing when harvested. The media was aspirated and replaced by PBS + 100 μg/ml cycloheximide and incubated at 37°C for 10 min. We chose cycloheximide because it induces rapid protein synthesis arrest (Han et al., 2014), has been successfully used in ribosome profiling of HEK 293T cells (Sidrauski et al., 2015), and 100 μg/ml is ~100 times higher than the concentration required to inhibit protein synthesis in reticulocytes (Godchaux et al., 1967). Each dish was then placed on ice, media aspirated, and replaced by ice cold PBS + 100 μg/ml cycloheximide. Cells were scraped, pelleted at 16,000× g for 30 s, and re-suspended in three pellet-volumes ice cold hypotonic lysis buffer (10 mM HEPES pH 7.9, 1.5 mM MgCl2, 10 mM KCl, 0.5 mM DTT, 1% Triton X-100 and 100 μg/ml cycloheximide) (Folco et al., 2012). After 10 min, cells were lysed on ice by ten strokes through a 26-gauge needle and nuclei were pelleted at 1,500× g for 5 min. Lysate from ~15 million cells (one dish) was layered on top of triplicate 10–50% (w/v) sucrose gradients (20 mM HEPES:KOH pH 7.6, 100 mM KCl, 5 mM MgCl2, 1 mM DTT and 100 μg/ml cycloheximide) made using a Biocomp Instruments (Canada) gradient master. Gradients were centrifuged for 2 hr at 36,000 RPM in a SW-41 rotor, punctured, and manually peak fractionated using real-time A260 monitoring with a Brandel (Gaithersburg, MD) gradient fractionator and ISCO (Lincoln, NE) UA-6 detector.

Sequencing library construction and deep sequencing

RNA was extracted from pooled technical triplicate sucrose gradient fractions by ethanol precipitation followed by acid phenol:chloroform extraction. Direct phenol:chloroform extraction was precluded by phase inversion in high sucrose fractions. RNA was then DNase treated, acid phenol:chloroform extracted, and ethanol precipitated. Cytoplasmic RNA was TRIzol extracted (Life Technologies, Grand Island, NY) and ethanol precipitated. Total RNA integrity was verified using a BioAnalyzer (Agilent, Santa Clara, CA). Ribosomal RNA was then depleted using Ribo-Zero (Illumina, San Diego, CA) and biological duplicate sequencing libraries were generated using the TruSeq RNA Sample Prep v2 kit (Illumina) without the poly-A selection steps. An equal mass (100 ng) of rRNA-depleted RNA was used as input to each individual library preparation. Libraries were verified using a BioAnalyzer and quantified using a Qubit (Life Technologies) prior to pooling for sequencing. Library insert sizes were typically ~150 ± ~25 bp. Pooled libraries were 75-bp paired-end sequenced on an Illumina HiSeq 2500 and runs of the same library in different lanes or flowcells were merged.

Sequencing data processing and transcriptome alignment

Adapters were trimmed using Cutadapt v1.5 (Martin, 2011) followed by subtractive alignments against the repeatmasker (RMSK) database (retrieved from UCSC on 2/11/2015) and abundant sequences from the Illumina iGenomes project (e.g. ribosomal RNA and the mitochondrial chromosome) using Bowtie2 v2.2.4 (Langmead and Salzberg, 2012). Unaligned reads were then aligned to the Ensembl release 75 transcriptome using Tophat v2.0.13 with parameters "-r 5 --mate-std-dev 50 -g 100 --report-secondary-alignments" (Trapnell et al., 2009). Mapping percentages for each pipeline stage are presented in Figure 1—figure supplement 3. Transcript isoform level abundances were calculated using Cuffquant v2.2.1 (Roberts et al., 2011; Trapnell et al., 2010), normalized between samples with Cuffnorm v2.2.1, and transcripts per million (TPM) (Wagner et al., 2012) values were calculated according to:

TPMi=RPKMi 106gall genes RPKMg

Biological duplicate datasets were processed independently. Quantified abundances are in Figure 1—source data 1 and Figure 1—source data 2. TrIP-seq plots are available for all Ensembl GRCh37 isoforms at http://meru.qb3.berkeley.edu/tripseq.

Hierarchical clustering of isoform distributions across the polysome profile

Cuffnorm counts for each replicate were subjected to a variance stabilizing transformation (VST) using the DESeq2 R package to correct for heteroscedasticity (Love et al., 2014); the VST approaches log2 for large counts but compresses low counts to suppress Poisson noise. Variance stabilized counts were averaged between replicates, filtered such that the mean across all nine samples was greater than one (roughly translating to 100 reads), and then mean subtracted to generate relative expression values. Inter-row distance was computed using Spearman’s rank correlation, and hierarchical clustering was performed using the fastcluster R package (Müllner, 2013) with Ward’s agglomeration method. The resulting dendrogram was split at increasing heights into subtrees until clusters contained similar overall trends, generating the large clusters presented in Figure 2. The samples were clustered similarly except with Euclidean distance and complete agglomeration. The R packages magrittr, dendextend Galili, 2015, and ggplot2 Wickham, 2009 were used to generate figures shown and are available through CRAN. Transcript isoforms in each cluster are in Figure 2—source data 1.

Clustering of TrIP-seq samples

The intersample clustering in Figure 1B was performed by hierarchically clustering the VST-transformed isoform-level counts as above between samples. The statistical significance of the resulting clustering was analyzed by measuring the Jaccard distance between clusterings of these data after subjecting the data to subsampling by bootstrap, jittering (adding random noise to each point), or replacing random points by noise using the R package fpc. The mean Jaccard distances of 100 such subsamplings as well as the average are presented in Figure Figure 1—figure supplement 1G S1G; a Jaccard distance ≥.75 is considered a stable cluster by the R package fpc.

Cloning of untranslated regions and in vitro transcription

Isoform-specific 5′ and 3′ untranslated regions were amplified from anchored oligo-dT primed cDNA libraries from HEK 293T cells and Gibson cloned (Gibson et al., 2009) into a vector based on pUC57 containing Renilla luciferase and a synthetic polyA60 tail (pA60; Fukaya and Tomari, 2011). Gibson cloning was performed such that untranslated regions were precisely cloned next to the ATG but contained two guanosines as the 5′-most nucleotides for T7 transcription for 5′ leaders, or the stop codon used by the isoform, and the six nucleotides CTGCAG at the 3′ end of the 3′ UTR immediately preceding the polyA60 tail for 3′ UTRs. The differences between cloned isoforms are shown in Figure 5—figure supplement 1, which was generated using IGV (Thorvaldsdottir et al., 2013). The entirety of all untranslated regions was verified using dideoxy sequencing. Transcription templates were generated by PCR using Phusion polymerase (NEB, Ipswich, MA), size verified by gel electrophoresis, and gel purified. Transcription was performed using T7 polymerase with 1 μg template in a buffer containing 7.5 mM each NTP, 1 μg pyrophosphatase (Roche, Pleasanton, CA), 30 mM DTT, 35 mM MgCl2, 2 mM spermidine, 0.01% Triton X-100 and 30 mM Tris pH 8.1 for four hours at 37°C followed by DNase treatment with RQ1 RNase-free DNase (Promega, Madison, WI) for 30 min. Transcription products were purified by ethanol precipitation and a Zymo (Irvine, CA) Clean & Concentrator column, followed by simultaneous capping using Vaccinia capping enzyme (NEB) and 2′-O-methylation (NEB). Capped products were purified using a Zymo Clean & Concentrator column, followed by size verification on glyoxylated samples (Ambion, Foster City, CA) using an agarose gel, and full-length 3′ UTR-containing RNAs were gel purified from an agarose gel (Zymo).

Transfections and luciferase assays

The concentration of RNAs containing 5′ untranslated regions was determined using A260, which was normalized by the intensity of the full-length product on an agarose gel, and then by the molar ratio of the construct to empty pA60. Molar-adjusted amounts of each RNA relative to 100–200 ng of pA60 were transfected into three technical triplicate wells of ~50% confluent cells in a 96-well plate using the TransIT-mRNA reagent (Mirus, Madison, WI). The concentration of RNAs containing 3′ UTRs was determined using a Qubit RNA HS assay (Life Technologies), normalized for the molar ratio of the construct to empty pA60, and molar-adjusted amounts of RNA relative to 7 ng of pA60 were transfected using TransIT-mRNA. A reduced amount of the 3′ UTR RNAs was used due to decreased yield of the longest 3′ UTRs. HEK 293T cells were grown in DMEM + 10% FBS, Hep G2 cells were grown in EMEM + 10% FBS, MCF7 cells were grown in DMEM:F12 + 10% FBS, A549 cells were grown in F12-K media + 10% FBS, and K-562 cells were grown in RPMI media + 10% FBS. Cells were harvested after ~18 hr (Figure 5) or 2 hr (Figure 6) and Renilla luminescence was measured (Promega).

Tabulation of isoform features

The Ensembl release 75 annotation set from the Illumina iGenomes project was first decomposed into 5′ leader, start codon, CDS, 3′ UTR, and whole-transcript regions. Length, GC-content, and number of exons were computed directly. Cytoplasmic expression and the median expression from the 80S through the eight+ ribosome fraction were extracted from TrIP-seq data. Transcript halflife data were derived from HeLa cell measurements (Tani et al., 2012). The structure was computed using RNALfold from the ViennaRNA package (Lorenz et al., 2011) in a 75-nt window. Codon usage statistics were downloaded from http://www.kazusa.or.jp/codon/ on 11/20/2014 and the minimum codon frequency is the average of codon usage across a five-codon window. The fraction of AU-elements is calculated as the percentage of the 3′ UTR that is of repeating A or U nucleotides for more than 5nt in a row. TargetScan 6.2 scores (Garcia et al., 2011) were downloaded from http://targetscan.org and parsed for the properties indicated, and an identical comparison was performed after filtering the miRNA list by those expressed in HEK 293T cells (Ender et al., 2008). All feature tabulation was performed using custom Python programs, which are available through GitHub at https://github.com/stephenfloor/tripseq-analysis.

Effect size measurement between isoform feature distributions

Isoforms belonging to the same gene present in different clusters were compiled, yielding gene-linked isoforms. For example, if the gene A has isoform 001 in cluster one and isoforms 002 and 003 in cluster two, features of isoform 001 are added to the cluster one set and features of 002 and 003 are added to the cluster two set and 001–002 as well as 001–003 would be gene-linked isoforms. Features in each set were then compared for statistical significance using the Mann-Whitney U test and visualized as empirical cumulative distribution functions (Figure 3—figure supplement 1A). The effect size was then computed between distributions using Cliff’s d, which is a nonparametric, dimensionless measure of the distance between distributions (Cliff, 1993). Cliff’s d is a measure of the number of times that a point xi in one distribution is greater than all points xj in the second distribution, or

d=#(xi>xj)#(xi<xj)mn

where # denotes the number of times, the two distributions are of sizes n and m, and xi and xj are items of the two sets. Cliff’s d is also related to the Mann-Whitney U statistic, by

d=2Umn1

Cliff’s d and the boostrap confidence intervals shown in Figure 3 were computed using the R package orddom, which is available through CRAN.

Ribosome profiling comparison and TrIP-seq polysome count calculation

Ribosome profiling was performed in HEK 293T cells (Sidrauski et al., 2015). These data were downloaded from the NCBI (GEO: GSE65778) and reprocessed by subtractive Bowtie 2 alignment to rRNA and RMSK sequences, and then mapped onto the human transcriptome using Tophat, as for the TrIP-seq data. Aligned reads were quantified at the gene and isoform level using Cuffquant onto the Ensembl release 75 transcriptome and normalized with Cuffnorm as above, to facilitate direct comparison to the TrIP-seq data. The number of reads derived from polysomes for TrIP-seq data was computed as the sum of read counts for fractions containing two to eight+ ribosomes multiplied by the number of ribosomes in each fraction:

TrIPseq polysome counts=i=17ini+24n8

where ni is the number of reads in the ith polysome fraction averaged for both biological replicates. The factor of 24 is applied to the last fraction based on a ceiling of ~40 ribosomes per transcript which then leads to an average of 24 ribosomes in the eighth peak (40 + 8 / 2 = 24). The number of ribosomes in the final peak is not directly measured by the sucrose gradients used Figure 1—figure supplement 1A).

Reprocessing of embryo sequencing data

RNA sequencing datasets of human preimplantation embryos (Yan et al., 2013) were downloaded from the NCBI (GEO: GSE36552). Reads were converted to FASTQ using fastq-dump (NCBI) and then processed as for TrIP-seq data by adapter trimming with Cutadapt, subtractive alignment to the RMSK and abundant sequences, and transcriptome alignment with Tophat to Ensembl release 75. Transcript abundances were calculated using Cuffquant and normalized with Cuffnorm. Data from individual cells were processed independently and averaged for each stage to generate the data shown in Figure 4 and Figure 4—figure supplement 1. The data were reprocessed for consistency to facilitate direct abundance comparisons.

RT-PCR

Cytoplasmic lysate was fractionated using a sucrose gradient as for the RNA-seq libraries and RNA was extracted from each fraction. Libraries of cDNA were then synthesized from these fractions using random primers (Applied Biosystems) from an equal amount of RNA as measured using a Qubit (Life Technologies; RNA HS assay). Template RNA was removed using RNase H treatment (NEB) and cDNAs were purified over an oligo cleanup column (Zymo). PCR was then performed for 25 (ACTB) or 30 (EEF1B2, ATF4) cycles with gene-specific primers (below) using Taq Titanium (Clontech) using an equal amount of cDNA input into each PCR, as measured using a Qubit (ssDNA assay). Reactions were then run on a 1% agarose gel and stained with SYBR-Gold (Life Technologies). Gels are representative of three biological replicates.

Primers used for RT-PCR are:

ACTB-forward AGAGCTACGAGCTGCCTGAC

ACTB-reverse AGCACTGTGTTGGCGTACAG

EEF1B2-forward TTCCCGTCATCTTCGGGAGCCGT

EEF1B2-reverse CTTTTCAGGTCTCCGAAACCCATGG

ATF4-forward GGCTCTGCAGCGGCAACCCC

ATF4-reverse CGACTGGTCGAAGGGGGACA

Quantitative RT-PCR

RNA was extracted from polysome fractions as for TrIP-seq. qRT-PCR was performed using the SuperScript III Platinum SYBR Green One-Step kit (Life Technologies) with 1ng input RNA in a 20 ul reaction volume. CT values were converted to fold changes over the cytoplasmic abundance and plotted.

Primers used for qRT-PCR are:

EEF1B2-001-fwd AGCCGTGGAGCTCTCGGATA

EEF1B2-001-rev AGCTCTTGTCCGCCAGGTAA

EEF1B2-003-fwd TCCAAACGCAACGAAAGGTCC

EEF1B2-003-rev CGCCGGACCGAAGGTTAAAG

EEF1B2-201-fwd AGCCGTGGAGCGTGGG

EEF1B2-201-rev TCGGCTGTATCCGAGAGCTG

SRSF5-002-fwd GACCCCGTCCGGTAGGAAGTACTAGCC

SRSF5-002-rev CAATATCTCTTATCCGTCCATATCCC

SRSF5-005-fwd GGTGAGTGGCTCACTTTGAGGGCAAG

SRSF5-005-rev CGAATCAACTGCGCTCATTAGACGC

Gene Ontology analysis

The DAVID server was used to calculate gene ontology (GO) terms for biological processes (BP) associated with the individual clusters (Jiao et al., 2012). Transcripts associated with each cluster were input into DAVID and GO BP terms with a Benjamini-corrected p-value of <0.05 were tabulated.

Accession numbers

Raw sequencing reads for all samples are available through the NCBI via the GEO Accession ID GSE69352.

References

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
    Alternative 3' UTRs act as scaffolds to regulate membrane protein localization
    1. BD Berkovits
    2. C Mayr
    (2015)
    Nature, 522, 10.1038/nature14321.
  8. 8
  9. 9
  10. 10
  11. 11
  12. 12
  13. 13
  14. 14
  15. 15
  16. 16
  17. 17
  18. 18
  19. 19
    Tuning gene expression with synthetic upstream open reading frames
    1. JP Ferreira
    2. KW Overton
    3. CL Wang
    (2013)
    Proceedings of the National Academy of Sciences of the United States of America 110:11284–11289.
    https://doi.org/10.1073/pnas.1305590110
  20. 20
  21. 21
  22. 22
  23. 23
  24. 24
  25. 25
  26. 26
  27. 27
  28. 28
  29. 29
  30. 30
    Iron regulatory protein prevents binding of the 43S translation pre-initiation complex to ferritin and eALAS mRNAs
    1. NK Gray
    2. MW Hentze
    (1994)
    The EMBO Journal 13:3882–3891.
  31. 31
  32. 32
  33. 33
  34. 34
    The mechanism of eukaryotic translation initiation: new insights and challenges
    1. AG Hinnebusch
    2. JR Lorsch
    (2012)
    Cold Spring Harbor Perspectives in Biology, 4, 10.1101/cshperspect.a011544.
  35. 35
  36. 36
  37. 37
  38. 38
  39. 39
  40. 40
  41. 41
  42. 42
  43. 43
  44. 44
  45. 45
  46. 46
  47. 47
  48. 48
  49. 49
  50. 50
  51. 51
  52. 52
  53. 53
  54. 54
  55. 55
  56. 56
  57. 57
    The translational landscape of the splicing factor SRSF1 and its role in mitosis
    1. MM Maslon
    2. SR Heras
    3. N Bellora
    4. E Eyras
    5. JF Cáceres
    (2014)
    eLife, 3, 10.7554/eLife.02028.
  58. 58
  59. 59
  60. 60
  61. 61
  62. 62
  63. 63
    fastcluster : fast hierarchical, agglomerative clustering routines for r and python
    1. D Müllner
    (2013)
    Journal of Statistical Software, 53, 10.18637/jss.v053.i09.
  64. 64
  65. 65
  66. 66
  67. 67
  68. 68
  69. 69
  70. 70
  71. 71
  72. 72
  73. 73
  74. 74
  75. 75
  76. 76
  77. 77
  78. 78
  79. 79
    The small molecule ISRIB reverses the effects of eIF2α phosphorylation on translation and stress granule assembly
    1. C Sidrauski
    2. AM McGeachy
    3. NT Ingolia
    4. P Walter
    (2015)
    eLife, 4, 10.7554/eLife.05033.
  80. 80
  81. 81
  82. 82
  83. 83
  84. 84
  85. 85
  86. 86
  87. 87
  88. 88
  89. 89
  90. 90
  91. 91
  92. 92
  93. 93
  94. 94
  95. 95
  96. 96
  97. 97
  98. 98
  99. 99
  100. 100
  101. 101
  102. 102
  103. 103
  104. 104
  105. 105
  106. 106
  107. 107
  108. 108
  109. 109
  110. 110
  111. 111

Decision letter

  1. Rachel Green
    Reviewing Editor; Johns Hopkins School of Medicine, United States

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

[Editors’ note: this article was originally rejected after discussions between the reviewers, but the authors were invited to resubmit after an appeal against the decision.]

Thank you for submitting your work entitled "Tunable protein synthesis by transcript isoforms in human cells" for consideration by eLife. Your article has been reviewed by two experts in the field, and the evaluation has been overseen by a Reviewing Editor and James Manley as the Senior Editor. Our decision has been reached after extensive consultation between the Reviewing Editor and the two reviewers. Based on these discussions and the individual reviews below, we regret to inform you that your work cannot be considered further for publication in eLife.

Both reviewers felt that the data were overall of very high quality and the general approach an interesting (albeit not novel) one to get at isoform specific information on translational efficiency. In this regard, the TrIP-Seq approach as presented is basically identical to previous experimental designs (Sterne-Weiler et al. 2013; Maslon et al. 2014; Jingyi Hou et al. 2015). Nevertheless, the degree of novelty of the approach was not the key factor in the decision. Further, the reviewers were impressed with the reporter data (in the original cell type HEK293) that provided strong evidence for differential translation of isoforms, as suggested and then followed up from the initial observations from the global approach. The reporter data led to an interesting model where 5' UTR leader sequences contribute generally to overall translational efficiency. The authors further proposed a model wherein 3' UTRs would be generally more important in specifying cell-specific regulation, based on reporter analysis in several different cell types. However, the reviewers felt that these data did not provide sufficient breadth to justify the broad conclusions that were made and felt that further global analysis in different cell types would have provided this support. Overall, the reviewers concluded that this high quality analysis failed to sufficiently support the broad claims made concerning isoform-specific translation patterns but rather represented the beginnings of such an analysis. In the absence of broader exploration of this topic in multiple cell types (and a comparative analysis), the study falls short of its claims and thus is not appropriate for publication in eLife.

Reviewer #1:

In this paper Floor and Doudna apply RNA-Seq to identify the presence of previously annotated isoforms in polyribosome fractions isolated from HEK 293T cells. Then the authors classify the identified isoforms according to their distribution along the sucrose gradient and select 7 genes for confirmation of their results. The work is well done and the experimental confirmation goes beyond previous works, however the "TrIP-Seq" approach is basically identical to previous experimental designs (Sterne-Weiler et al., 2013; Maslon et al., 2014; Jingyi Hou et al., 2015). Taking that in consideration I think that the novelty of the manuscript it is not sufficient to merit its publication in eLife, but rather in a more specialized journal.

Major comments and concerns:

1) In general the claims of the manuscript seem to go beyond what it is actually demonstrated. All the presented data is based in an experiment performed in HEK 293T cells. And the confirmation in other cancer cell lines is restricted to a handful of UTRs. To produce "an atlas of isoform-specific translation patterns" as the authors claim it would be necessary to apply their method in different cell lines and conditions where translation regulation is expected. However that is clearly out of the scope of this first manuscript. To avoid misleading the readers, I would advise the authors to town down their claims and Discussion.

2) In paragraph three, subheading “TrIP-seq measures transcript isoform-specific translation in human cells”, the authors compare their approach with other measures of polyribosome association and translation. It is clear that their method measures polyribosome abundance, however the presented comparisons do not demonstrate that they measure translation. For example in Figure 1F, neither TrIP-Seq nor ribosome profiling counts are corrected by mRNA abundance. Thus the observed correlation could be simple due to mRNA abundance (and not the translation efficiency). I would not expect their method to obtain a perfect correlation with ribosome protection or protein abundance, however it could be expected to be significantly greater than the one measured by mRNA abundance alone.

3) The authors make special emphasis in the distinction between high and lowly- ribosome associated transcripts (e.g. paragraph one, subheading “TrIP-seq measures transcript isoform-specific translation in human cells” and paragraph two subheading “3′ UTRs and introns drive isoform-specific polysome association”). The authors claim that isoforms "with more exons tend to be better translated" and that that is not caused by transcript length. However in Figure 3 and supplement, it can be observed that the length of the CDS could explain part of those differences. If that were the case, it would lead to the trivial observation that genes encoding longer proteins can potentially accommodate more ribosomes in their coding regions. That will cause those isoforms to be present in higher polyribosome fractions (independent of their translational potential). This effect would also potentially affect the clustering shown on Figure 2, which would be affected both by the CDS length and the translation efficiency (polyribosome association).

4) It is not clearly defined how the authors define biological replicates. If the authors refer to different batches of cells grown independently and subjected to different sucrose fractionationations they should specify that in the Methods section.

Reviewer #2: There is a very interesting finding that seems very likely correct. However, the manuscript as presented has many issues that should be rigorously addressed, as noted below:

1) Abstract: There is no value in stating that the relationship of RNA and protein levels is 'controversial'; that's not really scientific. Indeed, this is the case either or both because there is too little data, data are wrong, or the right type of data have not been collected. I think this manuscript makes a great case for the last of these. I.e., prior work (arguably) has been too crude to ferret out or isolate different factors responsible. Here, having the gene the same but changing the 5' or 3' UTR is a much better controlled experiment to sort out these effects. I know that wasn't the authors' intention – they wanted to discover these effects – but serendipity has given them a great controlled experiment as well. My guess is that this is why they're able to get so much more information.

2) Abstract: A suggestion (strongly) not to use the term rheostat. Analogies can be helpful, distracting, or misleading. Here at a minimum the rheostat is not a physical analogy. Presumably the effects arise from molecular interactions – and the results lead to the neat model that 3'UTRs are cell type specific because different cells have different proteins expressed (providing additional evidence in support of the 3' UTR as hosting a complex panoply of proteins; the constancy of 5'UTR effects suggests that these arise from something common to all cells –perhaps the ribosome and translation initiation machinery.

3) Introduction: The first paragraph is considerably under-documented. How well established are 'diverse translational control elements'? There are certainly a number of specific cases, but is their number and breadth clear? And the three examples given are maybe not the most clear or general. While one can't have a review here, an opening perspective that is both broader and contains important details would provide the scholarly context for this work.

4) Introduction: The authors point out that there is not a technique to look at isoform specific translation. They then take nearly a page to note how 'ribo-seq' cannot do this. They then reveal that sucrose (polysome) gradients, followed by deep sequencing, can do this. But of course polysome fractionation is a very old technique. And it is not until the Results (paragraph one, subheading “TrIP-seq measures transcript isoform-specific translation in human cells”) that we learn that Arava et al. did polysome fractions for all yeast genes – at that time using microarrays. And microarrays could be designed to look at isoforms, or looked at another way, the technical extension of this work from the existing technologies is a modest step, but not presented in the appropriate historical context. While this is common, it is not needed. The biology and importance of the question stands for itself. So the authors’ main advance is not a technique – but rather the identification of a (very) important question and unmet need, and the expert execution of addressing this question. (As a related aside, if one goes back to the original Ignolia et al. Science paper introducing ribo-seq, there the method has held – and been improved – and has been very important, but each of the scientific conclusions drawn was either made previously (by Arava et al.) or later shown to be incorrect. That paper still deserved the wide audience it got, as does this one, but let's work hard to get the intellectual and scholarly history right, whether or not it involves prior work that I was associated with.)

5) Subsection “TrIP-seq measures transcript isoform-specific translation in human cells” and Figure 1: It is not clear that had Figure 1B not recapitulated the gradient order that it would have meant that there is something amiss in the experiment.

While the clustering 'broke' between poly 4 and 5, this is not a statistically tested or rigorous result, and there are many possible origins. So the conclusion (from this observation at least) that the data can be considered in terms of low and high polysomes is not supported. One variable to consider is that short genes (which are highly abundant) will by necessity cluster in the low region. Thus, that statement “This suggests that in cells, transcript isoforms are predominantly poorly- or highly-ribosome associated, causing low polysomes to cluster away from high polysomes” is an over-interpretation of the data that are presented.

Technical: Was the clustering done on all reads or on a per-gene basis?

Technical: How is TPM (transcripts per million) determined? What controls for the validity of this measure are there?

Technical: What controls for rapid freezing of translation upon cycloheximide addition were performed? The extent of reads near the 5' end of the ORF should be a measure of this and should be noted and documented.

Figure 1E: While it is true that noncoding are preferentially in low polysome fractions, they ARE in those fractions, and the difference between low and high is only 2x. Is there reason to be concerned by this? (The statement “Protein coding genes and isoforms are found across the polysome, but noncoding and lncRNA genes are predominantly in the low polysome fractions indicating that they are generally weakly ribosome associated or in other large macromolecular complexes (Figure 1E and Figure 1—figure supplement 1D)” seems overstated given the 2x differential.)

Figure 1F: The description is inadequate – what are the points?

6) “Agreement between the isoform-specific polysome profiles and RT-PCR conducted from polysome fractions verified the accuracy of the isoform-level quantifications (Figure 1D)”. Figure 1D does not show even qualitative agreement between RT-PCR and TrIP-seq results. This gel and the one in the supplement should be quantitatively compared to the trip-seq data – quantitative PCR is probably needed here.

7) “Weaker amplification of the longer isoform (EEF1B2-003) in high polysomes may be due to PCR bias towards smaller amplicons (Walsh et al., 1992), and replicate experiments show EEF1B2-003 present in high polysomes (Figure 1—figure supplement 1C)”. The comment indicates less than robust data agreement, and a potential reason for this. That suggests that additional experiments (like shorter PCR reads) should be done to determine if the expected agreement is indeed observed.

8) Subheading “TrIP-seq measures transcript isoform-specific translation in human cells”. Where the trip-seq replicates completely independent biological replicates? If not, this conclusion cannot be drawn.

9) “Therefore, TrIP-seq and ribosome profiling agree at the gene level but diverge at the isoform-level, likely due to difficulty in mapping the very short profiling reads at the isoform level, or the influence of untranslated regions, which are invisible to profiling”. As I understand it there are two models for the differences between full ribosome profiles and that for the isoforms – noise from fewer reads or 'real' effects. This isn't so clear from the wording here, and seems like it needs to be addressed here – since one is an artifact and the other is biology.

10) “We then compared TrIP-seq data to copy number estimates from HeLa cells, and find that the two datasets correlate well (RS = 0.55, Figure 1—figure supplement 1F, (Kulak et al., 2014)”. Rs – 0.55 is viewed as good, but similar correlation coefficients are viewed as bad in other contexts. Maybe instead of saying correlate 'well' we need to better understand the factors that affect this correlation. Figure 1—figure supplement 1F shows a spread in Hela copy number (protein #) that looks much broader than the spread in trip-seq polysome counts. Visually I don't think a neutral reader would say that trip-seq data account well for the protein levels.

Importantly, the rest of this paragraph is logically unclear – statements are made but the underlying evidence for those statements seems an opinion rather than a reasoned series of considerations. The final statement is strong: 'we conclude…' – but the basis for that conclusion does not appear strong or minimally is not clearly laid out. And these are the crucial data for that paper.

11) Subheading “An atlas of human isoform-specific translation patterns”, Figure 2; clustering. Clustering is statistically ambiguous, ultimately, and these data are pretty 'flat' – i.e., there aren't many very strong features to cluster by. But maybe more centrally since clustering is a somewhat arbitrary (though often useful) transformation – e.g., there is no natural or set number of clusters for a dataset- it is not clear that one what's to be making conclusions – rather than just organizing data – from clusters. The authors conclude that it's 'crucial to fractionate the polysome profile.' I think that's strongly expected a priori, but I don't see how that was strongly tested or established from the cluster analysis – other than to note that there are some differences in profiles.

12) Paragraph two, subheading “An atlas of human isoform-specific translation patterns”. The conclusions in the paragraph seem too strong from the data. The authors above said that they don't know why ncRNAs are in the polysomes – which is reasonable. But here they make conclusions about decay and pioneering rounds – these arguments are likely not germane and do not seem to be supported by the data or logic but rather seem to be 'fits' to existing ideas and models.

13) “Surprisingly, the poorly translated cluster 7, which has ~9,000 individual transcripts…”. It is not clear to me that seeing ~10K sequences with introns means that they are being translated. That's one model. The other, stated earlier in the manuscript, is that they are in other large complexes.

14) “In sum, clustering of isoform abundance distributions across polysomes immediately provides insight into the diverse patterns of ribosome association by individual isoforms and transcript types in cells, and reveals intron-rich clusters of transcripts associated with polysomes that escape nuclear detention”. For similar reasons, I don't see how this conclusion is reached, relative to other possibilities.

15) Subheading “3′ UTRs and introns drive isoform-specific polysome association”. Here the authors are using the clusters to explore and identify possible trends. This seems like a powerful exploratory use of the clusters.

16) “We find that longer coding sequences and highly abundant versions of gene-linked isoforms are biased towards high polysomes (Figure 3B and Figure 3—figure supplement 1A)”. This can be because short genes can't hold as many ribosomes – or they reach the end sooner and thus fall off sooner.

17) Figure 3, subheading “3′ UTRs and introns drive isoform-specific polysome association”. A number of interesting points are raised and discussed. The bottom line at the end of this section seems vaguer than one would like – e.g., what is really established vs what might be and what might have other explanations. More basically, is the best way to analyze these data from pre-clustered groups or, since the top and bottom fractions are compared, to use a numerical ratio of those – so that one analyzes a continuous rather than discretized variable (as this distorts error).

18) Subheading “Predicting translation changes during preimplantation human development”. What does 'reprocessed' mean? It's not clear from what's stated what this dataset is? Presumably it is not polysome data – so presumably it's seeing whether gene isoforms change in abundance (lit data) that have isoform specific differences (this data).

19) Paragraph two, subheading “Predicting translation changes during preimplantation human development”. Seems like general conclusion about use of trip-seq from one example – and even that example doesn't seem so striking, and given that there are 1000's of data, the chance of a small number having seemingly statistically significant effects seems high. This might be considered as some form of multiple hypothesis testing.

20) It seems like the best independent test of the data and models is predicting new effects, and that's what this section does. The key statement seems to be “Not only do these data provide strong evidence that the output of heterologous reporters can be predictably tuned by untranslated regions, it additionally validates that the polysome profiles observed are connected to translational output”, stating that the results have predictive value. But the comparison is complex – i.e., showing polysome profiles. There are standard ways to convert polysome occupancy into protein production rate. While imperfect, that's the basis for earlier stuff in this and other work. So what seems clearest and least biased would be to make quantitative predictions for several – preferably 10-20, variants, and compare the output to those quantititave prior predictions. Figure 6 (below) suggests that this should be doable for 5' leaders but not 3' UTRs.

21) Discusssion. Where does the value of two orders of magnitude come from?

While no large translational regulation was found in prior work with 3000 highly expressed genes, are there not examples of translational regulation of genes in humans? Or in related species with conserved genes?

Paragraph two, Discussion. As noted above, the statement that trip-seq provides a 'faithful measure' may not be well supported by the analyses presented.

It seems that by far the strongest support for the manuscripts conclusions is the 'engineering', though that can be improved as noted above.

22) An interesting test would be to demonstrate/test isoform-specific effects of a miRNA.

https://doi.org/10.7554/eLife.10921.029

Author response

[Editors’ note: the author responses to the first round of peer review follow.]

Thank you for reviewing our paper on transcript isoform specific translation in human cells. We were pleased that the reviewers thought that the work is "well done", presents a "very interesting finding" and that our paper deserves as wide of an audience as the pioneering work of Ingolia, 2009 (Science). The reviewers were also "impressed with the reporter data", where we show that isoform-specific sequences derived from TrIP-seq data can be used to control translation of unrelated genes over 100-fold.

We believe that the work conducted in HEK 293T cells represents an advance in our understanding of how the diverse human transcriptome relates to protein translation in cells that warrants a broad audience. No prior study has observed the extent of transcript isoform specific translational control uncovered by our work in any system. We are grateful for the reviewers’ comments, which have helped us to clarify the significance of the work.

New additional data or analyses to include (one experiment and eight analyses):

a) Author response Figure 1: polysome association of noncoding transcripts and associated ribosome profiling protected fragments;

b) Author response Figure 2: retained intron transcripts found in polysomes harbor ribosome profiling protected fragments, suggesting they are not simply associated with other large macromolecular complexes;

c) Correlation between translation efficiency as measured by TrIP-seq and ribosome profiling (regarding Figure 1);

d) Quantitative RT-PCR across polysome profiles to compare to the TrIP-seq data (regarding Figure 1);

e) Bootstrap analysis of inter-fraction clustering (regarding Figure 1);

f) A correlation between TrIP-seq data and HEK 293T proteomics (regarding Figure 1);

g) An analysis of the coding potential of retained intron sequences (regarding Figure 2);

h) Ribosome density analyses to normalize for CDS-length dependent effects in analyzing transcript isoform specific features (regarding Figure 3);

i) Quantitative comparisons between the predicted protein output and luciferase measured (regarding Figure 5).

Reviewer #1:

In this paper Floor and Doudna apply RNA-Seq to identify the presence of previously annotated isoforms in polyribosome fractions isolated from HEK 293T cells. Then the authors classify the identified isoforms according to their distribution along the sucrose gradient and select 7 genes for confirmation of their results. The work is well done and the experimental confirmation goes beyond previous works, however the "TrIP-Seq" approach is basically identical to previous experimental designs (Sterne-Weiler et al., 2013; Maslon et al., 2014; Jingyi Hou et al., 2015). Taking that in consideration I think that the novelty of the manuscript it is not sufficient to merit its publication in eLife, but rather in a more specialized journal.

Polysome profiles have indeed been used previously to assess gene expression levels. However, no prior study has observed the extent of transcript isoform specific translational control uncovered by our work in any system. Prior work has focused on gene-level analyses (Maslon, Hou), isolated splicing events (Sterne-Weiler), or transcript termini (Arribere, Spies), without reconstructing transcript isoforms and without analyzing effects within the polysome profile itself. Our work clearly shows that transcript isoforms of the same gene can be associated with different numbers of ribosomes (Figure 2), which is driven by diverse sequence features of isoforms (Figure 3), and that isoform-specific sequences that alter ribosome association by TrIP-seq can be used to control translation of orthogonal genes in a manner predicted by the sequencing data with a dynamic range of at least 100-fold (Figure 5). Therefore, the data from HEK 293T cells present fundamental insights into transcript isoform specific translation in humans that remain unappreciated. These data and conclusions are new relative to previous work and are of interest to researchers in many fields including translation, splicing, transcription, and cancer biology, where the transcriptome is frequently altered. Furthermore, since we show that isoform-specific sequences can control translation of orthogonal genes in a predictable manner, researchers in mRNA therapeutics or synthetic biology can mine these data for sequences that confer graded translation responses in cells.

Major comments and concerns:

1) In general the claims of the manuscript seem to go beyond what it is actually demonstrated. All the presented data is based in an experiment performed in HEK 293T cells. And the confirmation in other cancer cell lines is restricted to a handful of UTRs. To produce "an atlas of isoform-specific translation patterns" as the authors claim it would be necessary to apply their method in different cell lines and conditions where translation regulation is expected. However that is clearly out of the scope of this first manuscript. To avoid misleading the readers, I would advise the authors to town down their claims and Discussion.

We have revised the manuscript to better match our claims and conclusions with the data presented. Although it would be fascinating to apply TrIP-seq to other systems where differential translational control is expected, we agree with the reviewer that this is beyond the scope of this paper. We will instead highlight the rich set of isoform-specific translational control features observed in 293T cells, and suggest the potential for some of these features to exist in other cell types, rather than concluding that there are broad similarities between cell types. We note that while this manuscript does not describe translation in different biological conditions, it still presents a rich resource of isoform-specific translation patterns in human cells.

2) In paragraph three, subheading “TrIP-seq measures transcript isoform-specific translation in human cells”, the authors compare their approach with other measures of polyribosome association and translation. It is clear that their method measures polyribosome abundance, however the presented comparisons do not demonstrate that they measure translation. For example in Figure 1F, neither TrIP-Seq nor ribosome profiling counts are corrected by mRNA abundance. Thus the observed correlation could be simple due to mRNA abundance (and not the translation efficiency). I would not expect their method to obtain a perfect correlation with ribosome protection or protein abundance, however it could be expected to be significantly greater than the one measured by mRNA abundance alone.

We appreciate that mRNA abundance could influence this correlation. However, the abundance of mRNA within a polysome or protected ribosome fragments is generally a measure of translation, and is what both TrIP-seq and ribosome profiling seek to measure. Thus, the most direct method to test for agreement between the two methods is through comparing polysome/protected fragment abundances directly without additional confounding variability by dividing by RNA-seq measurements.

To measure the correlation between TrIP-seq and ribosome profiling independent of mRNA abundance, we computed the ribosome profiling translation efficiency (TE) using standard methods. We then divided the weighted sum of polysome abundance of the TrIP-seq data as described in the manuscript by the cytoplasmic abundance, yielding a TrIP-seq TE. The correlation between the ribosome profiling and TrIP-seq TE measurements is RS = 0.24, which we have added to the manuscript in paragraph three, subheading “TrIP-seq measures transcript isoform-specific translation in human cells”. Possible sources of this weak correlation include different biological conditions (e.g. cell confluence) and sequencing formats, especially since Sidrauski et al used a linker-ligation based prep while this work uses random primed reverse transcription. We compared the data of Sidrauski et al. to an unpublished ribosome profiling dataset in 293T cells from the Cho lab and find that ribosome protected fragments correlate with RS = 0.78, while TE has a substantially lower correlation of RS = 0.41, suggesting biological or technical variability disproportionately affects TE. Therefore, TrIP-seq polysome counts correlate with ribosome profiling protected fragments similarly to other profiling experiments, while TE appears to be exquisitely sensitive to inter-experiment variability.

To address the relationship between TrIP-seq versus RNA-seq and protein abundance, we compared the weighted TrIP-seq polysome counts to 293T protein abundances, obtaining a correlation of 0.57 (Figure 1—figure supplement 1F), while RNA-levels alone exhibit a weaker correlation of 0.45 (not shown), indicating TrIP-seq is superior to RNA-seq when predicting protein abundance. We thank the reviewer for motivating these analyses.

3) The authors make special emphasis in the distinction between high and lowly- ribosome associated transcripts (e.g. paragraph one, subheading “TrIP-seq measures transcript isoform-specific translation in human cells” and paragraph two subheading “3’ UTRs and introns drive isoform-specific polysome association”). The authors claim that isoforms "with more exons tend to be better translated" and that that is not caused by transcript length. However in Figure 3 and supplement, it can be observed that the length of the CDS could explain part of those differences. If that were the case, it would lead to the trivial observation that genes encoding longer proteins can potentially accommodate more ribosomes in their coding regions. That will cause those isoforms to be present in higher polyribosome fractions (independent of their translational potential). This effect would also potentially affect the clustering shown on Figure 2, which would be affected both by the CDS length and the translation efficiency (polyribosome association).

We thank the reviewer for noting that the CDS length strongly drives polyribosome association, and now explicitly discuss this in the text. We have analyzed the ribosome density across transcripts to explore ribosome association of transcripts independent of the coding sequence length. We find that gene-linked isoforms in low polysome clusters generally have lower ribosome density than those in high polysome clusters (Figure 3). Therefore, the clustering in Figure 2 is not solely due to CDS length. Furthermore, the correlation between CDS length and 3’ UTR length is positive 0.3 (not shown), suggesting that shorter CDSs on average have shorter 3’ UTRs, while Figure 3 shows short CDSs are associated with few polysomes but short 3’ UTRs are associated with many polysomes. We also note that all the reporters tested in Figure 5 and Figure 6 were selected from gene-linked isoforms bearing nearly identical CDS lengths to mitigate this effect (Figure 5—figure supplement 1A). As these reporters show considerable isoform-specific translation, there is clearly signal in the data beyond CDS-length effects.

4) It is not clearly defined how the authors define biological replicates. If the authors refer to different batches of cells grown independently and subjected to different sucrose fractionationations they should specify that in the Methods section.

We regret that this was not clear and have updated the Methods section. Biological replicates refer to separate dishes of cells that were grown, lysed, fractionated and used to prepare sequencing libraries in completely independent experiments.

Reviewer #2:

There is a very interesting finding that seems very likely correct. However, the manuscript as presented has many issues that should be rigorously addressed, as noted below:

1) Abstract: There is no value in stating that the relationship of RNA and protein levels is 'controversial'; that's not really scientific. Indeed, this is the case either or both because there is too little data, data are wrong, or the right type of data have not been collected. I think this manuscript makes a great case for the last of these. I.e., prior work (arguably) has been too crude to ferret out or isolate different factors responsible. Here, having the gene the same but changing the 5' or 3' UTR is a much better controlled experiment to sort out these effects. I know that wasn't the authors' intention – they wanted to discover these effects – but serendipity has given them a great controlled experiment as well. My guess is that this is why they're able to get so much more information.

We agree that prior work, which has largely focused on translation at the gene level, has not collected the right type of data and/or performed the right type of analyses to uncover isoform-specific translational control. We have removed the statement regarding “controversy” and reframed the Introduction to focus on limitations of prior datasets as advised.

2) Abstract: A suggestion (strongly) not to use the term rheostat. Analogies can be helpful, distracting, or misleading. Here at a minimum the rheostat is not a physical analogy. Presumably the effects arise from molecular interactions – and the results lead to the neat model that 3'UTRs are cell type specific because different cells have different proteins expressed (providing additional evidence in support of the 3' UTR as hosting a complex panoply of proteins; the constancy of 5'UTR effects suggests that these arise from something common to all cells –perhaps the ribosome and translation initiation machinery.

We have removed the term rheostat. The Abstract now reads: “These results expose the large dynamic range of transcript-isoform-specific translational control, and identify isoform-specific sequences that control protein output in human cells”.

3) Introduction: The first paragraph is considerably under-documented. How well established are 'diverse translational control elements'? There are certainly a number of specific cases, but is their number and breadth clear? And the three examples given are maybe not the most clear or general. While one can't have a review here, an opening perspective that is both broader and contains important details would provide the scholarly context for this work.

We have added additional citations to the first paragraph of the manuscript documenting classic examples of translational control elements, such as uORF-mediated control of GCN4, the binding of the iron regulatory protein to transcript 5’ leaders, and translational regulation by micro RNAs. We think that the number and breadth of translational control elements are not clear, which is part of the motivation for this study.

4) Introduction: The authors point out that there is not a technique to look at isoform specific translation. They then take nearly a page to note how 'ribo-seq' cannot do this. They then reveal that sucrose (polysome) gradients, followed by deep sequencing, can do this. But of course polysome fractionation is a very old technique. And it is not until the Results (paragraph one, subheading “TrIP-seq measures transcript isoform-specific translation in human cells”) that we learn that Arava et al. did polysome fractions for all yeast genes – at that time using microarrays. And microarrays could be designed to look at isoforms, or looked at another way, the technical extension of this work from the existing technologies is a modest step, but not presented in the appropriate historical context. While this is common, it is not needed. The biology and importance of the question stands for itself. So the authors’ main advance is not a technique – but rather the identification of a (very) important question and unmet need, and the expert execution of addressing this question. (As a related aside, if one goes back to the original Ignolia et al. Science paper introducing ribo-seq, there the method has held – and been improved – and has been very important, but each of the scientific conclusions drawn was either made previously (by Arava et al.) or later shown to be incorrect. That paper still deserved the wide audience it got, as does this one, but let's work hard to get the intellectual and scholarly history right, whether or not it involves prior work that I was associated with.)

We appreciate the view that this study addresses an important question that is interesting to a broad audience. Indeed, the motivation behind this study is not to develop a new technique but to determine the fundamental biology discussed here – how do transcript-specific sequences contribute to translational control? We apologize that this was not presented in the appropriate historical context and have edit the Introduction and Discussion to focus more on the scientific history and context for this work. We tried to strike a balance between selling the work and presenting the raw science and thank the reviewer for shifting us back to science.

5) Subsection “TrIP-seq measures transcript isoform-specific translation in human cells” and Figure 1:

It is not clear that had Figure 1B not recapitulated the gradient order that it would have meant that there is something amiss in the experiment.

We regret that the motivation behind this analysis was not clear. Since the sucrose gradient was fractionated manually, it is possible that technical errors could lead to tube swaps or mis-fractionated peaks. This analysis is meant to be a test of the relationship between the samples with regard to their origin in the gradient, and their clustering into the gradient order demonstrates that the fractions were faithfully collected. We have amended the sentence to read “Clustering of the samples recapitulates the gradient order (Figure 1B), indicating the polysome profile was accurately fractionated”.

While the clustering 'broke' between poly 4 and 5, this is not a statistically tested or rigorous result, and there are many possible origins. So the conclusion (from this observation at least) that the data can be considered in terms of low and high polysomes is not supported. One variable to consider is that short genes (which are highly abundant) will by necessity cluster in the low region. Thus, that statement “This suggests that in cells, transcript isoforms are predominantly poorly- or highly-ribosome associated, causing low polysomes to cluster away from high polysomes” is an over-interpretation of the data that are presented.

We have performed statistical resampling of the cluster boundaries in Figure 1B to test the significance of the “break” in the clustering. We measured the Jaccard distance between clusterings of these data that were subsampled by bootstrap, subjected to jitter (addition of random noise to each point) or had random points replaced by noise using the R package fpc. The average Jaccard distance for each condition as well as the average Jaccard distance across the three tests is presented in Figure 1—figure supplement 1G. A commonly used criterion to define robustly separated clusters is a Jaccard distance of 0.75, showing that on average the clustering in Figure 1B is robust.

Although short ORFs might cause the observed partitioning, this does not change the conclusion that “transcript isoforms are predominantly poorly – or highly –ribosome associated”; instead it suggests a biological basis for this conclusion. We have amended the sentence to read: “This suggests that in cells, transcript isoforms are predominantly poorly- or highly-ribosome associated, causing low polysomes to cluster away from high polysomes, which could be have numerous biological origins including highly abundant short ORFs”.

Technical: Was the clustering done on all reads or on a per-gene basis?

We have added a new section to the Methods to describe the inter-sample clustering “Clustering of TrIP-seq samples”.

Technical: How is TPM (transcripts per million) determined? What controls for the validity of this measure are there?

We apologize for leaving out a description of the unit transcripts per million (TPM). TPM is a unit of gene expression that has superior properties to the more-common reads-per-kilobase-per-million-mapped-reads (RPKM) by removing the dependence on sequencing library size. See e.g. PMID 22872506 “Measurement of mRNA abundance using RNA-seq data: RPKM measure is inconsistent among samples” by Wagner, Kin and Lynch. In our work, TPM is a transformation of RPKM as in Wagner et al:

TPMi=RPKMi 106all genesRPKMg

where RPKM was calculated using the Cufflinks suite. A description of this has been added to the Methods section. Using TPM improves comparability between studies but as it is linearly related to RPKM all the same controls for validity of RPKM, which is widely used, apply to TPM.

Technical: What controls for rapid freezing of translation upon cycloheximide addition were performed? The extent of reads near the 5' end of the ORF should be a measure of this and should be noted and documented.

We appreciate the concern for the acute stalling of translation by cycloheximide. Translational shutoff in mammalian cells by cycloheximide is known to occur rapidly (e.g. within 20 minutes and the fastest inhibitor tested in Han, K et al. Nature Methods 2014) and the concentration used in these experiments (~350 μM; 100 μg ml-1) is ~100 times higher than that required to inhibit protein synthesis in reticulocytes (Godchaux, W. et al., 1967). We cannot use reads near the 5’ end of the ORF as a control for this since reads in TrIP-seq are derived from RNA-seq and not ribosome protected fragments, but similar conditions acutely shut off translation in ribosome profiling protocols in HEK 293T cells where you can perform this analysis (Sidrauski et al., 2014). We have added this to the Methods section of the manuscript.

Figure 1E: While it is true that noncoding are preferentially in low polysome fractions, they ARE in those fractions, and the difference between low and high is only 2x. Is there reason to be concerned by this? (The statement “Protein coding genes and isoforms are found across the polysome, but noncoding and lncRNA genes are predominantly in the low polysome fractions indicating that they are generally weakly ribosome associated or in other large macromolecular complexes (Figure 1E and Figure 1—figure supplement 1D)” seems overstated given the 2x differential.)

Indeed, we were surprised by the number of noncoding genes present in high polysome fractions. We have analyzed noncoding genes in high polysomes but did not include it in this manuscript version because we are still determining what it means. For example, clustering of transcripts derived from annotated noncoding genes reveals a similar heatmap to that for coding genes, with some transcripts strongly polysome associated (Author response image 1A). Analysis of the gene types represented in high polysomes shows a two-fold enrichment for antisense transcripts (Author response image 1B). We validated numerous individual examples of polysome-associated antisense transcripts by comparing to ribosome profiling datasets from HEK 293T cells from the Cho or Ingolia labs, for example the gene pair MCM3AP and MCM3AP-AS1 (Author response image 1C). Therefore, one explanation for the abundance of noncoding genes in high polysome fractions is that a fraction of genes annotated as noncoding are translated. If this is interesting to the reviewers or editors we can include this analysis and more as a main text or supplemental figure.

Author response image 1
Polysome association of noncoding genes.

(A) Clustering of noncoding transcript isoforms as in Figure 2A. (B) The fraction of transcripts per noncoding gene type in high polysome clusters red and blue (gray) versus all annotated noncoding gene derived transcripts (black). (C) The gene MCM3AP and its antisense gene MCM3AP-AS1 are polysome associated and exhibit ribosome profiling reads from two datasets, one of which is stranded (red reads: sense strand; blue reads: antisense strand).

https://doi.org/10.7554/eLife.10921.023

Figure 1F: The description is inadequate – what are the points?

We apologize that this was not clear and have revised the figure legend to read: “(F) Spearman’s correlation (RS) between gene (left) or isoform (right) read counts from ribosome profiling or TrIP-seq. See also Figure 1—figure supplement 1 and Methods for calculation of TrIP-seq polysome counts”. A description of the calculation of TrIP-seq polysome counts is in the Methods section now titled “Ribosome profiling comparison and TrIP-seq polysome count calculation”.

6) “Agreement between the isoform-specific polysome profiles and RT-PCR conducted from polysome fractions verified the accuracy of the isoform-level quantifications (Figure 1D)”. Figure 1D does not show even qualitative agreement between RT-PCR and TrIP-seq results. This gel and the one in the supplement should be quantitatively compared to the trip-seq data – quantitative PCR is probably needed here.

See next response.

7) “Weaker amplification of the longer isoform (EEF1B2-003) in high polysomes may be due to PCR bias towards smaller amplicons (Walsh et al., 1992), and replicate experiments show EEF1B2-003 present in high polysomes (Figure 1—figure supplement 1C)”. The comment indicates less than robust data agreement, and a potential reason for this. That suggests that additional experiments (like shorter PCR reads) should be done to determine if the expected agreement is indeed observed.

We appreciate the concern for rigorous validation of the deep sequencing data. It was not possible to make quantitative comparisons with these gels because it is not quantitative PCR as mentioned. Qualitatively, isoform 003 is primarily in low polysomes and generally decreases towards high polysomes, while isoforms 201 and 001 both qualitatively increase from low to high polysomes, both by RT-PCR and TrIP-seq, so there is qualitative agreement between the validation and sequencing data.

We have performed qRT-PCR with isoform-specific primers from polysome fractions and present these new data for EEF1B2 and SRSF5 in Figure 1—figure supplement 2. Again, qualitative agreement is observed but the relationship may be less quantitative than desired, perhaps due to the difficulty in accurately reconstructing transcript isoforms from short read sequencing data or different RNA normalizations between the two experiments. We believe the strongest validation of TrIP-seq is that it can be used to control the translation of reporters as predicted by the sequencing data as in Figure 5, and future work could use long read sequencing to directly measure transcript isoforms without the need for in silico reconstruction.

8) Subheading “TrIP-seq measures transcript isoform-specific translation in human cells”. Where the trip-seq replicates completely independent biological replicates? If not, this conclusion cannot be drawn.

We apologize for the confusion shared by both reviewers on this issue and have modified the Methods section to reflect that the two TrIP-seq datasets are independent biological replicates.

9) “Therefore, TrIP-seq and ribosome profiling agree at the gene level but diverge at the isoform-level, likely due to difficulty in mapping the very short profiling reads at the isoform level, or the influence of untranslated regions, which are invisible to profiling”. As I understand it there are two models for the differences between full ribosome profiles and that for the isoforms – noise from fewer reads or 'real' effects. This isn't so clear from the wording here, and seems like it needs to be addressed here – since one is an artifact and the other is biology.

We appreciate the request for clarification and have reworded this sentence to read: “However, at the isoform level the correlation is worse (RS = 0.47), which is worse than the correlation between TrIP-seq replicates (RS = 0.90; Figure 1—figure supplement 1E), suggesting the discrepancy is not due to variability in isoform quantification in TrIP-seq but instead is likely due to known issues quantifying transcript isoforms in ribosome profiling (Ingolia 2014)”. The citation here is to a review by Ingolia wherein he specifically discusses the issues associated with quantifying transcript isoforms using ribosome profiling: “transcript variants with different 5ʹ and 3ʹ UTRs affect translation, and ribosome profiling cannot deconvolve these transcripts when multiple variants are present simultaneously”.

10) “We then compared TrIP-seq data to copy number estimates from HeLa cells, and find that the two datasets correlate well (RS = 0.55, Figure 1—figure supplement 1F, (Kulak et al., 2014). “:RS – 0.55 is viewed as good, but similar correlation coefficients are viewed as bad in other contexts. Maybe instead of saying correlate 'well' we need to better understand the factors that affect this correlation. Figure 1—figure supplement 1F shows a spread in Hela copy number (protein #) that looks much broader than the spread in trip-seq polysome counts. Visually I don't think a neutral reader would say that trip-seq data account well for the protein levels.

We have now correlated TrIP-seq data against HEK 293T proteomics data and present the results in Figure 1—figure supplement 1F, which also shows a positive correlation. We agree there is still considerable spread in this plot, but note that TrIP-seq is a better predictor of protein levels than RNA-seq (RS, TrIPseq = 0.57, RS, RNAseq = 0.45). We have revised this sentence to read: “We then compared TrIP-seq data to protein copy number estimates from HEK 293T cells, and find that the two datasets correlate”.

Importantly, the rest of this paragraph is logically unclear – statements are made but the underlying evidence for those statements seems an opinion rather than a reasoned series of considerations. The final statement is strong: 'we conclude…' – but the basis for that conclusion does not appear strong or minimally is not clearly laid out. And these are the crucial data for that paper.

We thank the reviewer for helping us to improve the clarity of this paragraph. We based our reasoning on a series of observations from our data and the literature: cryptic large complexes would systematically overestimate protein abundance, but in general HEK 293T protein abundances are lower than TrIP-seq abundances (Figure 1—figure supplement 1F) and there is a high gene-level correlation with ribosome profiling (Figure 1F), which is an accepted measurement of translation due to e.g. demonstrated ribosome-derived read length patterns (Ingolia 2014 Cell Reports). There are examples of systematic stalling of elongating ribosomes, which would also decouple ribosome occupancy and translation (citations in the manuscript), but we reasoned here based on the literature that this is a minor effect because of e.g. no detectable long-term ribosome stalls in mouse ES cells (Ingolia 2011). Previous work has related polysome profiles to translation (e.g. Arava et al.) and, in addition to the observations laid out here and in the Discussion, translating ribosome association is the most parsimonious explanation for why an mRNA is associated with a polysome peak. We have toned down the final statement of this section and reserve it for the Discussion, after testing the ability of these data to predictably control translation in Figure 5.

11) Subheading “An atlas of human isoform-specific translation patterns”, Figure 2; clustering. Clustering is statistically ambiguous, ultimately, and these data are pretty 'flat' – i.e., there aren't many very strong features to cluster by. But maybe more centrally since clustering is a somewhat arbitrary (though often useful) transformation – e.g., there is no natural or set number of clusters for a dataset- it is not clear that one what's to be making conclusions – rather than just organizing data – from clusters. The authors conclude that it's 'crucial to fractionate the polysome profile.' I think that's strongly expected a priori, but I don't see how that was strongly tested or established from the cluster analysis – other than to note that there are some differences in profiles.

Our goal with this analysis was to find a set of transcripts that are poorly and highly translated and then extract features of gene-linked isoforms that may be associated with differential translation. To do this, one must categorize transcripts somehow and we elected to use clustering. We provide evidence that the specific clusters referenced most in the paper (e.g. clusters 1, 2 and 6) are present in each biological replicate by measuring the Jaccard Index between biological replicate clusters and pooled replicate clusters (Figure 2—figure supplement 1A).

The sentence referenced regarding: “fractionating the polysome profile” is “many clusters have similar total polysome abundance but different distributions, indicating that to obtain accurate measurements of isoform-specific translatability it is crucial to fractionate the polysome profile”. By this we mean that comparing total cytoplasmic RNA to total polysomal RNA as was done in Sterne-Weiler et al., Maslon et al., and other earlier studies is blind to the diverse transcript isoform specific translation patterns highlighted by the clustering analysis. Later we show that regions from isoforms which are associated with many or few polysomes are sufficient to control protein production by Renilla luciferase by 100-fold, so the strong test of this conclusion comes later.

12) Paragraph two, subheading “An atlas of human isoform-specific translation patterns”. The conclusions in the paragraph seem too strong from the data. The authors above said that they don't know why ncRNAs are in the polysomes – which is reasonable. But here they make conclusions about decay and pioneering rounds – these arguments are likely not germane and do not seem to be supported by the data or logic but rather seem to be 'fits' to existing ideas and models.

We have removed this paragraph and Figure 2C.

13) “Surprisingly, the poorly translated cluster 7, which has ~9,000 individual transcripts…”. It is not clear to me that seeing ~10K sequences with introns means that they are being translated. That's one model. The other, stated earlier in the manuscript, is that they are in other large complexes.

We were curious about this point as well, and inspected a variety of retained intron transcripts in other ribosome profiling datasets to see if ribosome protected fragments could be found on these transcripts. In numerous examples we find low levels of ribosome profiling fragments derived from retained intron regions, confirming that at least some of the ~10k retained intron transcripts are indeed engaged with ribosomes (Author response image 2A). Computing the FLOSS score of these retained intron derived fragments shows they are indistinguishable from coding sequence fragments (Author response image 2B). We are happy to provide further examples of retained intron transcripts with ribosome protected fragments and corresponding analysis as a main text or supplemental figure if the reviewers or editors are interested.

Author response image 2
Retained intron transcripts found in polysomes are also observed by ribosome profiling.

(A) An example of a retained intron transcript on polysomes where the retained intron region is detected through the eight+ ribosome fraction. (bottom) Ribosome profiling reads are detected in the retained intron region in two separate 293T ribosome profiling datasets. (B) Shown are fragment length distributions for ribosome profiling reads derived from either retained intro regions (red) or coding sequence regions (black). The FLOSS score (Ingolia et al., 2014) between retained intron and CDS fragment length distributions is 0.023, indicating near perfect agreement. The FLOSS score is the integrated point wise difference between the two distributions. Ribosome profiling data from the Yoon-Jae Cho lab and from Sidrauski et al. (2015).

https://doi.org/10.7554/eLife.10921.024

14) “In sum, clustering of isoform abundance distributions across polysomes immediately provides insight into the diverse patterns of ribosome association by individual isoforms and transcript types in cells, and reveals intron-rich clusters of transcripts associated with polysomes that escape nuclear detention”. For similar reasons, I don't see how this conclusion is reached, relative to other possibilities.

We apologize that the reasoning behind this statement was not clear. The retained intron transcripts observed in these data escape the nucleus since the input material to the sucrose gradient was from the cytoplasmic fraction of cells. If desired, we can include explicit demonstrations of the difference between the abundance of retained intron regions between the nuclear fraction and polysome distribution of retained intron isoforms, as can be seen on the left and right sides of Author response image 1C, “nuc” row, which shows sequencing data from the nuclear fraction of the same cells used for the TrIP-seq experiment.

15) Subheading “3′ UTRs and introns drive isoform-specific polysome association”. Here the authors are using the clusters to explore and identify possible trends. This seems like a powerful exploratory use of the clusters.

We appreciate the comment; this analysis is the reason the data were clustered in Figure 2 and was the major motivation of the entire project!

16) “We find that longer coding sequences and highly abundant versions of gene-linked isoforms are biased towards high polysomes (Figure 3B and Figure 3—figure supplement 1A)”. This can be because short genes can't hold as many ribosomes – or they reach the end sooner and thus fall off sooner.

We agree, and have added wording to this effect.

17) Figure 3, subheading “3′ UTRs and introns drive isoform-specific polysome association”. A number of interesting points are raised and discussed. The bottom line at the end of this section seems vaguer than one would like – e.g., what is really established vs what might be and what might have other explanations. More basically, is the best way to analyze these data from pre-clustered groups or, since the top and bottom fractions are compared, to use a numerical ratio of those – so that one analyzes a continuous rather than discretized variable (as this distorts error).

We chose to use the clustered data rather than numerical comparisons across all transcripts because 1) the clustering allowed us to extract isoforms that were especially lowly or highly polysome associated (which should then have the largest differences), and 2) difficulty in converting the TrIP-seq polysome to a precise number of ribosomes associated with each transcript since the last polysome peak is degenerate. There are multiple ways to analyze these data and other types of analysis would also be interesting in the future.

18) Subheading “Predicting translation changes during preimplantation human development”. What does 'reprocessed' mean? It's not clear from what's stated what this dataset is? Presumably it is not polysome data – so presumably it's seeing whether gene isoforms change in abundance (lit data) that have isoform specific differences (this data).

We apologize for the vague use of “reprocessed”. We meant that since the original data were quantified using a different set of analysis software and onto a different set of reference transcripts the two datasets could not be directly compared as is. The latter interpretation is correct; the purpose of this figure was to demonstrate that further information can be distilled from existing, precious sequencing data (like that collected from human embryos) when interpreted given the unique insight into the translatability of each transcript isoform that TrIP-seq provides. We have added a reference to the Methods section on this line and the full reprocessing is described in the Methods section “Reprocessing of embryo sequencing data”.

19) Paragraph two, subheading “Predicting translation changes during preimplantation human development”. Seems like general conclusion about use of trip-seq from one example – and even that example doesn't seem so striking, and given that there are 1000's of data, the chance of a small number having seemingly statistically significant effects seems high. This might be considered as some form of multiple hypothesis testing.

We have included three additional examples of transcript isoforms that change across embryonic development and are also differentially associated with polysomes in the new Figure 4—figure supplement 2.

20) It seems like the best independent test of the data and models is predicting new effects, and that's what this section does. The key statement seems to be “Not only do these data provide strong evidence that the output of heterologous reporters can be predictably tuned by untranslated regions, it additionally validates that the polysome profiles observed are connected to translational output”, stating that the results have predictive value. But the comparison is complex – i.e., showing polysome profiles. There are standard ways to convert polysome occupancy into protein production rate. While imperfect, that's the basis for earlier stuff in this and other work. So what seems clearest and least biased would be to make quantitative predictions for several – preferably 10-20, variants, and compare the output to those quantititave prior predictions. Figure 6 (below) suggests that this should be doable for 5' leaders but not 3' UTRs.

We have performed quantitative comparisons between the predicted protein output and observed luciferase levels and present these in Figure 5E. There is a positive relationship between the average number of ribosomes in TrIP-seq measurements (Figure 5B, D) and the luciferase fold change across all reporters (Figure 5 A,C). It is challenging to convert TrIP-seq abundances to quantitative protein synthesis rates due to the poor gradient resolution of mRNAs associated with many ribosomes in the poly8+ peak. This is one possible source for the variability in 5E. Testing additional variants as well as the mechanistic foundations of the differential translation observed is the subject of other current work and we agree it will be fruitful to conduct this comparison.

21) Discusssion. Where does the value of two orders of magnitude come from?

We measured luciferase production modulated by untranslated regions derived from different gene-linked isoforms and found one example where two isoforms affect protein production by ~100-fold, and three other examples that affect protein production by ~10-fold. We apologize that this was not clear and have added a reference to Figure 5 with this statement.

While no large translational regulation was found in prior work with 3000 highly expressed genes, are there not examples of translational regulation of genes in humans? Or in related species with conserved genes?

We have removed this statement and added references regarding translational control at the gene level to the Introduction.

Paragraph two, Discussion. As noted above, the statement that trip-seq provides a 'faithful measure' may not be well supported by the analyses presented.

It seems that by far the strongest support for the manuscripts conclusions is the 'engineering', though that can be improved as noted above.

We agree that some of the strongest support that the TrIP-seq data is a measure of translation is derived from the reporter experiments. Given this, the discussion earlier in point 10, the discussion presented in paragraph three, Discussion, and the fact that the most parsimonious explanation for a mRNA being in a polysome profile is that it is associated with a ribosome, we stand by the statement that TrIP-seq by-and-large is a faithful measure of isoform-specific translation in cells. There are certainly exceptions to this conclusion; the Discussion is speaking to the majority of the signal in TrIP-seq.

22) An interesting test would be to demonstrate/test isoform-specific effects of a miRNA.

This would be a fascinating experiment and we will consider the best way to perform it. This is one of many follow-up experiments that could be performed given the observations here regarding isoform-specific translation, like altering the levels of a miRNA, sequence-specific RNA binding protein, DEAD-box protein, or ribosomal protein. Experiments like this are uniquely possible with measurements of translation with transcript isoform resolution provided by TrIP-seq, and we are excited to pursue experiments in this vein.

https://doi.org/10.7554/eLife.10921.030

Article and author information

Author details

  1. Stephen N Floor

    1. Department of Molecular and Cell Biology, University of California, Berkeley, Berkeley, United States
    2. Howard Hughes Medical Institute, University of California, Berkeley, Berkeley, United States
    Contribution
    SNF, Conceptualization, Methodology, Investigation, Software, Writing - Original Draft, Funding Acquisition; Writing - Review & Editing; Supervision, Conception and design, Acquisition of data, Analysis and interpretation of data, Drafting or revising the article
    For correspondence
    stephen.floor@gmail.com
    Competing interests
    The authors declare that no competing interests exist.
  2. Jennifer A Doudna

    1. Department of Molecular and Cell Biology, University of California, Berkeley, Berkeley, United States
    2. Howard Hughes Medical Institute, University of California, Berkeley, Berkeley, United States
    3. Innovative Genomics Initiative, University of California, Berkeley, Berkeley, United States
    4. Physical Biosciences Division, Lawrence Berkeley National Laboratory, Berkeley, United States
    5. Department of Chemistry, University of California, Berkeley, Berkeley, California, United States
    Contribution
    JAD, Writing - Original Draft, Funding Acquisition, Supervision; Writing - Review & Editing, Drafting or revising the article
    For correspondence
    doudna@berkeley.edu
    Competing interests
    The authors declare that no competing interests exist.

Funding

Howard Hughes Medical Institute

  • Stephen N Floor
  • Jennifer A Doudna

Helen Hay Whitney Foundation

  • Stephen N Floor

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

Acknowledgements

We thank Y Bai, P Kranzusch, A S Y Lee, E. Montabana, M O’Connell, D Rio and A Tambe for critical reading of the manuscript, N Ingolia, S Iwasaki, L Lareau and members of the Doudna lab for helpful advice, and A Fischer, K Condon and M Chung for technical assistance. We thank S Oh and Yoon-Jae Cho for sharing data prior to publication. This work used the Vincent J. Coates Genomics Sequencing Laboratory at UC Berkeley, supported by NIH S10 Instrumentation Grants S10RR029668 and S10RR027303.

Reviewing Editor

  1. Rachel Green, Reviewing Editor, Johns Hopkins School of Medicine, United States

Publication history

  1. Received: August 17, 2015
  2. Accepted: January 5, 2016
  3. Accepted Manuscript published: January 6, 2016 (version 1)
  4. Version of Record published: February 10, 2016 (version 2)

Copyright

© 2016, Floor 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.

Metrics

  • 6,620
    Page views
  • 1,620
    Downloads
  • 18
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Comments

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Cell Biology
    2. Developmental Biology and Stem Cells
    Hirohito Shimizu et al.
    Research Article