Charting DENR-dependent translation reinitiation uncovers predictive uORF features and links to circadian timekeeping via Clock

NIH3T3 and HEK293FT cells were cultured under standard conditions (DMEM; 10% FCS, 1% penicillin/streptomycin, all from Invitrogen; 37°C; 5% CO₂).

Lentiviral particle production in HEK293FT cells using envelope pMD2.G and packaging psPAX2 plasmids, and viral transduction of NIH3T3 cells, were performed following published protocols (10), with puromycin selection at 5 μg/ml for 4 days.

For the generation of lentiviral shRNA expression vectors, two different sequences targeting Denr were cloned into pLKO.1puro backbone vector (Addgene no. 10878 (11)): shRNA1: GTACCACAGAAGGTCACGATA, corresponding to clone TRCN0000308443 of the TRC shRNA Library from the Broad Institute; and shRNA2: GTGCCAAGTTAGATGCGGATT, corresponding to clone TRCN0000098826. pLKO.1puro constructs containing Gfp, and Scramble (Addgene no. 1864) shRNAs served as controls.

For the generation of dual luciferase (Firefly/Renilla) reporter plasmids, fragments containing the 5′ UTR and first 5-14 codons of the selected DENR targets/controls were amplified by polymerase chain reaction (PCR) from genomic DNA or cDNA, cloned into the BamHI site of the prLV1 dual luciferase reporter plasmid (12), and validated by sequencing. The following primers were used for PCR: Etaa1, forward (for): aaaggatccGACTTGCAAAGATGGCGCTGCAC, reverse (rev): tttggatccactagtGTCCTTCAGCTGCATTCACATTT; Map2k5, for: aaaggatccGGTTCCGGAGTAACAGCGGTCTAAC, rev: tttggatccactagtGGCCAGCCACAGCATTACAGGTTAA; Lrrc28, for: aaaggatccGCCGCTGAGTGCCGGTCAGCGGGC, rev: tttggatccactagtGATTTCGGATGCCATGACTGAACA; Klhdc8a, for: aaaggatccGTCTCCGACCCTGTAGACACTGCAG, rev: tttggatccactagtATTGGGCACTTCCATGGCAGCCCGG; Clock, for: aaaggatccGGGGAGGAGCGCGGCGGTAGCGGTG, rev: tttggatccactagtAACAATTGAGCTCATTTTACTACAGCTTACGG; Fth1, for: aaaaggatcCAGACGTTCTCGCCCAGAGTCGC, rev: tttggatcccatGGTGGCGGCGGGGCGAGGCGC; Rpl30, for: aaggatcCAGCGGCGGGTGGACAGAGATGCC, rev: tttggatccacgCGTCTTCTTTGCGGCCACCATCTTC; Deaf1, for: aaggatccGGGCGTCAGACTGCCTCTTCGGTC, rev: tttggatccacgcgtGGCAGCGGCCGCCACAGCTGCCGC; Cry1, for: aaaaagatctactagtCACCCGGGCAGCCTCGGGACCG, rev: tttagatctTCGGAACCAGTGCACGGCGTTCAC.

5′ UTR mutants were generated by site-directed mutagenesis of the corresponding plasmids, using the following primers (mutated nucleotides in small case with mutated start/stop codon underlined in the forward sequence): Klhdc8a-Mut, forward (for): GGCAGGGAGCCCGTGCAGCGCGgtaccGAGGCTGAGAGAGGGGACGCGCC, reverse (rev): GGCGCGTCCCCTCTCTCAGCCTCggtacCGCGCTGCACGGGCTCCCTGCC; Map2k5-Mut, for: CTCTTACTCACAAGGACTACAGgctagcTTGTGGGTTCTGTTTTGTCCAC, rev: GTGGACAAAACAGAACCCACAAgctagcCTGTAGTCCTTGTGAGTAAGAG; Etaa1-Mut, for: cgaaggatcCGACTTGCAAAGgtaccGCTGCACGGTGGCGCGCGGGCTC, rev: GAGCCCGCGCGCCACCGTGCAGCggtacCTTTGCAAGTCGgatccttcg; Clock-uORF-Stop-Mut, for: GTTATGGTGTTTACCGTcAGCTGTAGTAAAATGAG, rev: CTCATTTTACTACAGCTgACGGTAAACACCATAAC; Klhdc8a-uORF-Stop-Mut, for: GCCGACCGGCGCCGTGcacCCCGCGGGCCCCGGGGCT, rev: AGCCCCGGGGCCCGCGGGgtgCACGGCGCCGGTCGGC; Etaa1-uORF-Stop-Mut (small case correspond to an insertion that changes uORF frame), for:TGCTCCGCTGTAGCAgcAAATGTGAATGCAGC, rev: GCTGCATTCACATTTgcTGCTACAGCGGAGCA; Clock-uORF1-Mut, for: CCCAAAATCACCAGCAAGAGTTgctagcGTCAGTCACACAGAAGACGGCC, rev: GGCCGTCTTCTGTGTGACTGACgctagcAACTCTTGCTGGTGATTTTGGG; Clock-uORF2-Mut, for: AAAGTGAAAGAGGAGAAGTACAggtacCTACCACAAGACGAAAACATAAT, rev: ATTATGTTTTCGTCTTGTGGTAGgtaccTGTACTTCTCCTCTTTCACTTT; Clock-AnnotCDS-Mut, for: CAAGACGAAAACATAATGTGTTtcGGTGTTTACCGTAAGCTGTAG, rev: CTACAGCTTACGGTAAACACCgaAACACATTATGTTTTCGTCTTG; Clock-AlterCDS-Mut, for: GTTTACCGTAAGCTGTAGTAAAcaGAGCTCAATTGTTactagtg, rev: cactagtAACAATTGAGCTCtgTTTACTACAGCTTACGGTAAAC.

For iCLIP experiments, we used either NIH3T3 cells without overexpression to immunoprecipitate the endogenous DENR or NIH3T3 cells transiently transfected with Denr cloned into a FLAG-Venus-pLenti6 vector (Addgene 36948). Cells were grown to confluency in a 15-cm dish. One dish was used for overexpression, and six for non-transfected experiments.

iCLIP was performed as described (13,14). RNA–protein complexes were crosslinked with UV-C (254 nm) at 150 mJ/cm² in a Stratalinker 2400. After cell lysis (50 mM Tris–HCl, pH 7.4; 100 mM NaCl, 1% Igepal (Sigma), 0.1% sodium dodecyl sulphate (SDS), 0.5% Na deoxycholate, ethylenediaminetetra acetic acid (EDTA)-free protease inhibitors), RNA was partially digested with 0.2 U RNase I (Thermo Fisher). DENR–RNA complexes were pulled down with anti-DENR antibody (Abcam 108221), pre-immobilized on protein G-coated magnetic beads (Dynal). An IgG immunoprecipitation was used as control. Immuno-complexes were washed twice with High Salt Buffer (50 mM Tris–HCl pH 7.4, 1 M NaCl, 1 mM EDTA, 1% Igepal, 0.1% SDS, 0.5% sodium deoxycholate) and twice with PNK Wash Buffer (20 mM Tris–HCl pH 7.4, 10 mM MgCl₂, 0.2% Tween-20). An infrared adaptor was ligated to RNA 3′ ends, and protein–RNA complexes were visualized in a LiCor CLx Imager (Odyssey), after denaturing PAGE and nitrocellulose membrane transfer. About 60–130 kDa DENR–RNA complexes were excised from the membrane and liberated by proteinase K (Roche) digestion. This size range likely reflects DENR–MCTS1 complexes (26 and 22 kDa, respectively) bound to RNA, rather than DENR–RNA alone, indicating that the DENR–MCTS1 interaction is strong. During library preparation, cDNAs were ligated on their 5′ end to a 5-nt sample barcode to allow demultiplexing. The barcode is preceded by a 4-nt and followed by a 3-nt random sequence that serves as unique molecular identifier (UMI) to mark individual cDNA molecules. cDNA libraries were then PCR amplified (24 cycles) and subjected to high-throughput sequencing (Illumina HiSeq2500 platform). Experiments were done in triplicate for DENR-overexpression, and in duplicate for non-transfected and IgG IP conditions.

iCLIP data were processed and analysed using the iCount platform (https://github.com/tomazc/iCount↗), as described (15). Briefly, the sample barcodes were removed, and the UMI was excised and added to the read ID prior to mapping with Bowtie 0.12.7 to the mouse genome (mm10), allowing two mismatches. Sequences with the same UMI and starting at the same genomic position—likely representing PCR duplicates—were excluded from subsequent analyses. High-confidence crosslink sites were calculated with the iCount function peaks, with options RegionsAsOne and 15-nt flank.

One 15-cm dish of confluent NIH3T3 cells was used per replicate. Ribosome profiling and parallel RNA-seq were performed in triplicate for Denr shRNA1, Denr shRNA2 and scr shRNA, and in duplicate for Gfp shRNA, following the ART-Seq protocol by Epicentre-Illumina (Cat.no. RPHMR12126) with minor modifications as described below. Cells were treated with cycloheximide (100 μg/ml) for 2 min at 37°C to arrest translation elongation, and immediately harvested, followed by lysate preparation. For the generation of ribosome-protected fragments, cell lysates were treated with 5 units RNase I (ART-Seq) per OD260, and monosomes were purified in MicroSpin S-400 columns (GE Healthcare). For both RPF- and RNA-seq, 5 μg of RNA was used for rRNA depletion following the Ribo-Zero Magnetic Kit instructions from Epicentre-Illumina (Cat.no. MRZH11124). Finally, libraries were amplified using 12 PCR cycles, and sequenced (HiSeq2500 platform).

Initial processing, mapping and quantification of mRNA and footprint abundance were performed as previously (16). Briefly, adaptor sequence was trimmed with Cutadapt (17) and trimmed sequences were filtered using a custom python script to keep 26–35 nt and 21–60 nt reads for RPF- and RNA-seq, respectively. Reads were then sequentially aligned to mouse rRNA, tRNA and cDNA (Ensembl mouse release 84) using Bowtie2 (18) and mouse genome using Tophat v2.0.11 (19). Trimmed and filtered sequences from RNA-seq were also directly mapped against the mouse genome with Tophat to estimate expressed transcript isoforms (using Cufflinks v2.2.1 (20)). Transcriptome-mapping reads from the sequential alignment were counted towards their location on 5′ UTR, CDS or 3′ UTR per gene, using only the set of expressed isoforms as estimated above. The location of the putative A-site of the footprint was used for RPF-seq counting (position +15 from the 5′ for reads ≤30 nt and position +16 for reads >30 nt), and the 5′ end was used for RNA-seq reads. For genes with multiple expressed isoforms, reads that did not map unambiguously to a single feature were assigned to one in preference order: CDS, 5′ UTR, 3′ UTR.

RPF- and RNA-seq read counts were normalized separately with the upper quartile normalization from edgeR (21) and then transformed to RPKM values as read count per gene kilobase per normalized library size. Genes with <10 reads in 2/3 of samples and mitochondrially encoded genes were removed before normalization.

Translation efficiencies were then calculated per sample as the ratio of RPF RPKM to RNA RPKM of the region of interest (5′ UTR, CDS or uORF), log2 transformed and averaged over all replicates.

Significant changes in TE between knockdown and control cells were assessed with several available algorithms: Xtail (22) with options bins=10000, RiboDiff (23) with options -d 1 -p 1, Riborex (24) with options engine=“DESeq2” and Babel (25) with options nreps=100000.

Briefly, all four methods, with the options implemented in our study for maximal consistency among them, model raw read counts by following a negative binomial distribution and use global normalization methods for sequencing depth that would not account for global changes in TE if existing. RiboDiff and Riborex use DESeq2 to estimate a single mean and dispersion for both RPF- and RNA-seq, and detect differentially translated genes by fitting RPF- and RNA-seq read counts to a single GLM, whereas Xtail uses DESeq2 for mean and dispersion estimation separately for RPF- and RNA-seq and fits two independent GLMs. Xtail then derives probability distributions for fold changes of RNA or RPF, or for RPF-to-RNA ratios, and evaluates statistical significance between conditions from one of the two (the most conservative) pipelines. Babel relies on edgeR for parameter estimation and uses an errors-in-variables regression to test for significant changes in translation within and between conditions.

Differential TE was set at FDR < 0.1 for all four methods and performance for false positive discovery rate was further assessed by randomly label-shuffling knockdown and control datasets, which should not yield differential TE genes.

For uORF detection, a set of transcripts was used for which reads can be unambiguously assigned to the 5′ UTR, i.e. transcripts that are the only protein-coding isoform expressed (N = 4637). In order to include as many transcripts as possible for the annotation, genes with multiple protein-coding isoforms were also considered if (i) one isoform represented >65% of all expressed (as derived from our transcript expression estimation, see ‘Sequencing pre-processing, alignment and quantification’, N = 3211) or (ii) all expressed isoforms had the same CDS start, in which case the transcript with the highest expression was selected (N = 1826). For the selected transcripts, uORFs were annotated and considered as translated with the following criteria: (i) started with AUG, CUG, GUG or UUG, (ii) had an in-frame stop codon within the 5′ UTR or overlapping/within the CDS, (iii) were at least 9 nt long (including stop codon), (iv) had a coverage >25%, (v) showed a frame preference and (vi) the preferred frame was the one corresponding to the uORF 5′ end/start codon. Importantly, we considered only uORFs coding for at least two amino acids (i.e. 9 nt, including the stop codon), as for shorter uORFs frame bias assessment would be highly prone to false positives, and Kozak context scoring would be compromised (for 6 nt uORFs, the position +4 would always be the T of the stop codon, and thus cannot be a G of the strongest Kozak context).

For the modelling, all transcripts containing translated uORFs in Denr knockdown or control cells were included (N = 4755) and the following 5′ UTR features were used as predictors of the fold change in CDS TE across conditions: (i) the number of uORFs, (ii) the uORFs length, using the median length in transcripts with several uORFs, (iii) the proportion of AUG uORFs, (iv) the uORF Kozak context, scored relative to the consensus GccA/Gcc###G, where upper-case letters denote highly conserved bases (scored +3), lower-case letters indicate the most common nucleotides (scored +1) and ### represents the start codon (not scored). The median Kozak score was used when several uORFs were present on a transcript. (v) The CDS Kozak context, scored in the same manner, (vi) the distance of the first uORF start codon from the 5′ cap, (vii) the distance of the last uORF stop codon to the CDS start, (viii) the uORF GC content, using the median for transcripts with several uORFs, (ix) the presence of overlapping uORFs in the transcript (binary predictor). Predictors with skewed distributions (number of uORFs, uORF length, distance from the cap and distance to the CDS start) were log2 transformed. Partial regression plots were generated with the R package ‘visreg’. Absence of multicollinearity between predictors was assessed with the variance inflation factor using R function vif (car package).

Functional enrichment analysis was performed using g:Profiler (26), including Gene Ontology Biological Process and Molecular Function terms and biological pathway information from Reactome. Functional categories of more than 2 and less than 1500 members were used and P-values were corrected by Benjamini–Hochberg FDR. Categories with FDR <0.05 were considered as significant and were used for visualization with Cytoscape v.3.5.1 EnrichmentMap tool (27).

NIH3T3 cells were first lentivirally transduced with the dual luciferase reporters containing the 5′ UTR construct of interest and then transduced with Denr-shRNA/scramble-shRNA. After puromycin selection, cells were lysed in passive lysis buffer and luciferase activities were quantified using DualGlo luciferase assay system and a GloMax 96 Microplate luminometer (all from Promega). Firefly/Renilla luciferase (FL/RL) activities were first normalized to that of the empty vector in scramble shRNA-transduced cells. For DENR target experiments, FL/RL signals were also normalized to the scramble shRNA-transduced, non-mutated construct. For Clock mutational experiments, FL/RL signals were secondly normalized to the scramble shRNA-transduced Clock-WT construct.

mRNAs were produced using in vitro transcription at the ‘ivt mRNA production and formulation platform’ in Zurich (http://www.cancer.uzh.ch/en/Research/mRNA-Platform.html↗). The 5′ end consisted of a CleanCapTM (Trilink) followed by the WT or mutated Klhdc8a 5′ UTR and by a codon optimized firefly luciferase open reading frame. The 3′ end consisted of a tandem repeat of the mouse β-globin 3′ UTR and a poly-A tail (28). Uridine residues were fully substituted by 1-methyl pseudouridine residues (this modification can prevent immunostimulation, avoiding production of type I interferon and consecutive inhibition of translation in vivo (28)).

Animal procedures were approved by the Veterinary Office and its research ethics review committee of the University of Zurich (Kanton Zurich, Health Direction, Veterinary Office, Zollstrasse 20; 8090 Zurich; licence number ZH215/17). Animals were purchased from Envigo (Netherlands). Four- to eight-week-old mice were injected intravenously with 1 μg of mRNA formulated with TransIT transfection reagent following the manufacturer’s suggestions (Mirus). The procedure for one mouse was as follows: the nucleic acid was diluted in 38 μl of Opti-MEM (Gibco) before the addition of 0.72 μl of the ‘mRNA Boost Reagent’. After mixing, 1.12 μl of the TransIT reagent was added. The solution was thoroughly homogenated and injected immediately: 40 μl per mouse. At different time points after mRNA injection, in vivo bioluminescence imaging was performed on an IVIS Lumina instrument (PerkinElmer). Immediately before each measurement, D-luciferin (Synchem) dissolved in phosphate-buffered saline (15 mg/ml stock) and sterile filtered was injected (150 μg/g intraperitoneally). Emitted photons from live animals were quantified 10-20 min post luciferin injections, with an exposure time of 3 min. Regions of interest (ROI) were quantified for average radiance (photons s-1 cm-2 sr-1) (IVIS Living Image 3.2).

Total protein extracts from NIH3T3 cells were prepared according to the NUN procedure (29). Antibodies used were rabbit anti-CLOCK 1:3000 (ab3517, abcam), mouse anti-U2AF65 1:5000 (Sigma), rabbit anti-DENR 1:4000 (ab108221, abcam) and anti-rabbit-HRP or anti-mouse-HRP 1:10000 (Promega). Blots were visualized with the Fusion Fx7 system and quantified with ImageJ 1.48k.

We hypothesized that if DENR is required for efficient protein production from the main CDS of mRNAs that undergo reinitiation after uORF translation, we would be able to identify DENR-regulated transcripts by comparing ribosome occupancies transcriptome-wide in DENR loss-of-function versus control cells by ribosome profiling. Conceivably, DENR target mRNAs should exhibit decreased translation efficiencies (TEs) on their CDS when DENR is absent (Figure 1A). We therefore performed shRNA-mediated knockdown of Denr in mouse NIH3T3 fibroblasts, a cell line in which we had previously observed that Denr depletion led to a shortening of the period of circadian oscillations (9). Cells were lentivirally transduced with two different Denr-targeting shRNAs (shRNA1, shRNA2; in triplicate) or control RNAs (scramble shRNA, in triplicate; Gfp shRNA in duplicate), grown to confluency and harvested in the presence of cycloheximide (CHX) to stabilize elongating ribosomes. After preparation and sequencing of ribosome-protected mRNA fragment (RPF-seq) and total RNA libraries (RNA-seq) as previously described (9), we obtained ribosome footprints that were predominantly 29–30 nt in length (Supplementary Figure S1A), which preferentially mapped to protein coding sequences (Supplementary Figure S1B) and that showed the characteristic triplet periodicity expected from high-quality footprint data (Supplementary Figure S1C). Across replicates, CDS-mapping normalised reads were highly correlated (R² > 0.97, Supplementary Figure S1D), indicating good reproducibility. The RNA-seq and RPF-seq data confirmed the efficiency of Denr knockdown (77–90% reduction in Denr transcript levels), whereas the abundance of Mcts1 mRNA, encoding the DENR heterodimerization partner, or of Eif2d, encoding a structurally and functionally related homologue of the DENR-MCTS1 dimer (4), remained unchanged (Figure 1B,C and Supplementary Figure S1E).

A reduction in Denr levels that alters reinitiation efficiencies should engender differences in ribosome occupancies on the CDS and, possibly, on uORFs. First, we compared CDS footprint levels for all expressed genes (N = 10 516, Figure 1D, grey), and then selected the subset of transcripts that also displayed footprint coverage within the 5′ UTR (N = 5148, Figure 1D, blue). Of note, the quantification of footprints from the full 5′ UTR can serve as a useful proxy for uORF translation (30,31), given that uORFs are considered the major source of 5′ UTR-mapping reads. This group of transcripts, thus presumably enriched for uORF-containing genes, showed an overall similar distribution as all genes (Figure 1D). We concluded that the effect of DENR deficiency on this gene set did not markedly differ from the effect seen genome-wide, consistent with previous reports indicating that acute Denr loss did not lead to major global effects on translational output (6). Next, we calculated the ratio of 5′ UTR-mapping to CDS-mapping footprints per gene, and determined how this measure of ‘relative 5′ UTR translation’ was affected in Denr-depleted versus control cells. Our analysis revealed a striking increase in relative translation on 5′ UTRs in Denr-deficient cells, indicative of a redistribution of translating ribosomes within transcripts (Figure 1E and Supplementary Figure S1F). This effect was absent from the RNA-seq data (Supplementary Figure S1G). Conceivably, these changes on the translational landscape may reflect complex alterations in initiation efficiency and uORF usage, and are in line with a role of DENR in translation reinitiation (6).

Translational remodelling that affects a sizeable proportion of the transcriptome is most consistent with a scenario in which DENR interacts with ribosomes at mRNAs with little, or no, sequence specificity. To address this question, we examined the direct interactions of DENR with RNAs in vivo by performing individual nucleotide resolution crosslinking and immunoprecipitation (iCLIP) (15) of DENR (Supplementary Figure S2A). Over 85% of iCLIP reads mapped to ribosomal RNA, in particular to 18S rRNA (Supplementary Figure S2B and C), with the five most prominent binding sites on helices 24-26a (Supplementary Figure S2D and E). This outcome is in line with structural data that places the DENR–MCTS1 complex in proximity to the mRNA channel within the small ribosomal subunit and that identifies helix 24 as a potential interface of MCTS1 with 18S rRNA (32,33). Moreover, helix 26 has previously been reported to make contacts with sequences upstream of restarting AUG codons on viral RNAs (34). In contrast, only 10.5% of iCLIP reads mapped to mRNAs, which was not sufficient to identify specific mRNA-binding sites. In summary, while we cannot exclude transient direct interactions of DENR–MCTS1 with mRNAs, our iCLIP data are consistent with structural studies and support a model in which the complex binds close to the mRNA channel of initiating ribosomes at a location that could facilitate reinitiation when uORFs are present, but travels indiscriminately along mRNAs.

From the shift in the distribution of 5′ UTR versus CDS footprints that were affecting a significant proportion of the transcriptome in Denr-deficient cells (Figure 1E), we wished to narrow down to the most-DENR-sensitive transcripts that were potentially direct and functionally relevant targets. We reasoned that changes in CDS translation efficiency (TE; defined as the ratio of CDS-mapping RPF-seq to RNA-seq reads) would represent an appropriate measure to identify such targets, as TE differences would alter the amount of protein produced from the transcript. We therefore calculated transcriptome-wide TEs and, given that there is currently no established consensus on the most suitable analytical tool for differential translation analysis, we then applied several available algorithms to detect significant TE changes. Briefly, the four methods we employed, Xtail (22), RiboDiff (23), Riborex (24) and Babel (25), are all conceived for the quantification of translational changes that deviate from possible variations already occurring at the level of mRNA abundances. They differ in parameter estimation and modelling approaches, as well as in sensitivity, false discovery rates and computational speed. When applied using settings that maximize comparability of results (see ‘Materials and Methods’ section), we identified 240, 78, 41 and 30 genes that had a significant change in TE between Denr-deficient and control cells with Xtail, RiboDiff, Riborex and Babel, respectively (FDR < 0.1 for all four algorithms). Importantly, differential TE genes detected by Riborex, RiboDiff and Xtail completely overlapped, while only about half of the genes detected by Babel were also found by any of the other methods (Supplementary Figure S3A). Moreover, most genes detected by Xtail, RiboDiff and Riborex showed decreased TE in Denr-depleted cells, as expected for the knockdown of a reinitiation factor, and they spanned the full range of expression levels (Figure 2A and Supplementary Figure S3B,C). By contrast, Babel was biased towards low abundance transcripts and identified similar numbers of up- and down-regulated TE genes (Supplementary Figure S3B and C). We concluded that Xtail likely represented the most sensitive method to identify DENR targets, and further estimated the false discovery rate of the algorithm by randomly shuffling the dataset labels, which eliminated nearly all differential TE genes (Supplementary Figure S3D). Please note that, for simplicity, we shall further use the term ‘DENR targets’ for the 240 differential TE genes identified by Xtail, although in the absence of information on direct DENR–mRNA interactions, the term ‘most DENR-sensitive transcripts’ would be more precise.

Among the 240 differential TE genes identified by Xtail (see Supplementary File S1) were several for which the presence of regulatory uORFs has been described previously, such as Atf5 (35), Rnaseh1 (36,37), Jak2 and Map2k6 (38). In order to investigate whether the TE decrease among DENR-sensitive transcripts was associated with uORF translation, we annotated uORFs transcriptome-wide. Briefly, we included AUG and non-canonical, near-AUG (i.e., CUG, GUG, UUG) initiation codons located in 5′ UTRs, and assessed the active translation of the corresponding ORFs by their footprint coverage, frame preference and triplet codon periodicity (Supplementary Figure S4A). Using this approach, we detected translated uORFs in nearly half of all expressed genes (N = 4755 genes, Figure 2B), consistent with previous predictions (8). For the majority of genes, the uORFs were fully contained in the 5′ UTRs (51.8%, N = 2464 genes), but genes with uORFs overlapping the CDS (19.5%, N = 926 genes), or that had both overlapping and non-overlapping uORFs (28.7%, N = 1365 genes) were frequent, too (Supplementary Figure S4B). Notably, DENR targets were highly enriched for genes with translated uORFs (72.7%, P = 5.73e-4, Fisher test, Figure 2B) that were non-overlapping uORFs (Supplementary Figure S4B). The depletion of overlapping uORFs (Supplementary Figure S4B)—indicating that they are insensitive to DENR regulation—is consistent with the idea that after terminating on an overlapping uORF, a ribosome would presumably not be able to backtrack to reinitiate on an upstream CDS start codon. Moreover, relative 5′ UTR translation in Denr-depleted versus. control cells calculated over the full 5′ UTR—as a proxy of uORF translation—was significantly higher for DENR targets than for the global set of uORF-containing genes (Supplementary Figure S4C and D). In summary, these observations were fully consistent with an involvement of DENR in reinitiation and in efficient CDS translation of a subset of uORF-containing mRNAs.

We next analysed whether specific functional pathways were over-represented among the 240 differential TE genes. GO term analysis revealed enrichment of categories associated with translation and mRNA stability, as well as cellular functions related to protein phosphorylation and catabolism, kinase binding and cell cycle (Figure 2C and Supplementary Figure S4E), in line with previous observations in Drosophila S2 cells (6) and indicating evolutionary conservation of DENR functions.

Among the DENR-regulated transcripts were several relevant to cancer, such as Klhdc8a, Etaa1, Map2k5, Slc20a1 or Vegfd (39–43). Individual inspection confirmed their reduced CDS translation efficiency (Figure 3A), as well as a shift of ribosome occupancy from the CDS to the 5′ UTR in Denr-depleted cells (Figure 3B). This signature was characteristic of many differential TE genes (see Supplementary Figure S5A and B), and correlated with the presence of translated uORFs (Supplementary Figure S5C). Among the strongly DENR-regulated cases, we also noted Atf5 (Supplementary File S1), whose translation (similar to that of Atf4) is known to respond to cellular stress signals via two uORFs, the first of which is reinitiation-permissive whereas the second overlaps the main CDS (35,44). In Denr-depleted cells, translation of the first uORF (on Atf5 and, to lesser extent, on Atf4) remained unchanged compared to control cells, while translation of the second uORF and the CDS were decreased (Supplementary Figure S5D and E). These findings validate that our approach indeed delivered reinitiation-prone transcripts.

In order to confirm the DENR- and uORF-dependence of CDS translation for selected presumed targets, we cloned the 5′ UTR and immediate start codon context of several transcripts of interest in frame and upstream of Firefly luciferase (FL) in a lentiviral vector that also expresses Renilla luciferase (RL, for internal normalization) from the same bidirectional promoter (Figure 3C, upper panel). Denr depletion led to a 18-44% reduction in FL/RL signals for five different control reporters that contained either no cloned 5′ UTR or 5′ UTRs of genes that were not among the differential TE genes (of these, Cry1 and Fth1 harbour translated uORFs as well, whereas Deaf1 and Rpl30 do not) (Figure 3C, lower panel). By contrast, we observed a stronger reduction in FL/RL signals—of 49–90%—for four 5′ UTR reporters of presumed DENR target genes (Etaa1, Klhdc8a, Map2k5 and Lrrc28), indicating that DENR was indeed involved in their regulation. Mutating uORF start codons increased reporter signal and abolished most DENR-dependence (Figure 3C, bottom), indicating that uORF translation was necessary for DENR dependance. Furthermore, in line with the view of DENR as a reinitiation factor, rather than a leaky scanning factor, the presence of a functional stop codon was required for DENR regulation (Supplementary Figure S5F). uORF-dependent regulation of Klhdc8a translation was particularly striking, with a >4-fold translation upregulation in WT cells with a uORF-less construct (Figure 3C, bottom). We next wished to investigate whether uORF-mediated translational regulation could also be recapitulated in vivo and selected Klhdc8a for this experiment, as this mRNA contained the strongest, reinitiation-dependent uORF in the cellular assay. We injected mice intravenously with in vitro-synthesized mRNAs comprising the Klhdc8a 5′ UTR (WT or uORF start codon-mutated) fused to Firefly luciferase CDS and carrying a synthetic polyA tail. We then monitored expression by bioluminescence live imaging in mice. Bioluminescence signal was primarily detected in the livers of the animals, and was 2–4.3 times higher in the uORF-less compared to the WT constructs (Supplementary Figure S6A and B), demonstrating the strong inhibitory activity of the Klhdc8a uORF also in vivo.

Taken together, we concluded that DENR was involved in the regulation of protein biosynthetic output, and that a set of >200 transcripts containing uORFs was particularly sensitive to loss of this reinitiation factor.

uORFs can operate through a bewildering diversity of mechanisms that act on CDS translation (by affecting start site selection/leaky scanning/re-initiation/ribosome shunting), transcript abundance (by eliciting nonsense-mediated mRNA decay) or in some cases through uORF-encoded bioactive peptides (reviewed in (3)). However, rules to predict specific mechanisms from specific uORF features have not yet been established. Our datasets could potentially enable the identification of signatures that render certain uORF-containing transcripts DENR-dependent. Previous studies have suggested that, at least in the context of synthetic reporters in Drosophila S2 (6) and human HeLa (7) cells, very short uORFs (coding for 1–2 amino acids) with favourable Kozak contexts were most strongly dependent on DENR for efficient reinitiation at a downstream reporter CDS.

We used multiple regression analysis to determine whether DENR-dependent translation events were associated with specific transcript features. We modelled the CDS TE change observed upon Denr knockdown using nine relevant transcript features, namely the number of uORFs, uORF start codon identity (AUG versus near-cognate) and context, the length, GC content and location of the uORF within the 5′ UTR, the presence of overlapping uORFs and CDS start codon context (Figure 4A). The model was applied to all translated uORF-containing genes (N = 4741) and yielded several features with significant predictive value for DENR-sensitivity of TE (Figure 4B–J; adjusted R² = 0.057, F-statistic = 33.21 on 4731 DF, P < 2.2e-16). Three of the identified features reflected the expected positive relationship between efficient uORF translation and a greater need for reinitiation to subsequently translate the CDS as well. These features were: the number of uORFs present on a transcript (Figure 4B, b = −0.019, P < 2e-16), the presence of AUG, rather than non-canonical uORF start codons (Figure 4C, b = −0.05, P < 1.15e-9), and the strength of Kozak contexts around uORF start codons (Figure 4D, b = −0.0026, P = 0.013). Relatedly, DENR targets were enriched for AUG uORFs as compared to the ensemble of translated uORF-containing transcripts (P = 0.0044, Fisher test, Supplementary Figure S7A), and genes harboring exclusively AUG uORFs showed the strongest effect on CDS TE in Denr-deficient cells (Supplementary Figure S7B). By contrast, the Kozak context around the CDS start codon bore no predictive value (Figure 4E, b = −0.0006, P = 0.38), suggesting no DENR involvement in its recognition. Moreover, the presence of CDS-overlapping uORFs was not associated with DENR-dependent TE change (Figure 4F, b = 0.005, P = 0.44), as expected for a reinitiation factor, given that overlapping uORFs would be reinitiation-incompetent unless the transcripts contained a downstream alternative CDS start codon.

Two noteworthy outcomes from the model were that uORF length lacked a relationship (Figure 4G, b = 0.0006, P = 0.81), whereas the uORF–CDS intercistronic distance had predictive value for DENR-sensitivity of CDS TE (Figure 4H, b = −0.0101, P = 7.89e-7). Briefly, already early studies on a model transcript had suggested that the location of a uORF with respect to the main ORF could dictate the efficiency of reinitiation, as longer spacing would generally allow the reacquisition of initiation factors to the scanning ribosome (45). Our model provides an unbiased and transcriptome-wide validation of this notion for DENR-regulated transcripts, where larger distances from the stop codon of the 3′-most uORF to the start codon of the main ORF were associated with downregulation of TE (Figure 4H); uORFs located further upstream from the CDS were thus more likely to reinitiate in a DENR-dependent fashion. Importantly, this effect of uORF–CDS distance was not caused by longer 5′ UTRs per se, given that the distance from the 5′ cap to the first uORF was not associated with DENR regulation (Figure 4I, b = −0.0018, P = 0.18). Of note, the uORF–CDS distance was only weakly correlated with 5′ UTR length (Spearman rho = 0.16), whereas the correlation of the 5′ cap-first uORF distance with 5′ UTR length was more prominent (Spearman rho = 0.27, Supplementary Figure S7C and D).

No linear relationship was found between uORF length and CDS TE (Figure 4G, b = 0.0006, P = 0.81), in agreement with previous results in human cells using synthetic uORFs encoding >2 amino acids (9 nt) (7). Of note, a length dependence for extremely short uORFs of 1–2 amino acids, as reported by Schleich et al. (7), would go unnoticed in our analyses, as we excluded them due to the uncertainties associated with their annotation (see ‘Materials and Methods’ section).

Finally, the uORF GC content showed a significant relationship with TE fold change (Figure 4J, b = 0.058, P = 0.005), suggesting that potential weaker secondary structures (low GC) favoured reinitiation involving DENR.

For the above model, we used the CDS TE fold change as response, which has the benefit of accounting for DENR regulation in a continuous manner. Of note, however, also a logistic regression model based on a binary response (i.e. DENR target as defined in Figure 2 versus non-target) recapitulated the main findings (Supplementary Figure S7E), further supporting our predictions for the strongest cases of DENR translational control. In summary, although the model explained only a relatively modest proportion of the total variance (see ‘Discussion’ section), our results indicated that DENR was important to enable efficient protein synthesis when the transcripts contained uORFs initiating with AUG in strong Kozak contexts (i.e. uORFs expected to be highly translated), and that the distance between the most downstream uORF and the CDS start was an important parameter for efficient reinitiation.

We recently found that uORF translation was pervasive within circadian core clock transcripts, and that the depletion of Denr in NIH3T3 cells engendered robust circadian period shortening by 1–1.5 h (9). We therefore sought to uncover which core clock transcripts were DENR-regulated, and to map the relevant uORFs. Conceivably, these would represent likely contributors to the observed circadian phenotype.

The expression of several core clock genes was altered in Denr-depleted versus control cells. In most cases (Per1/2/3, Cry1/2, Nr1d1, Rorα), the observed changes occurred already at the level of mRNA abundance and subsequently led to comparable changes at the footprint level as well (Figure 5A and Supplementary Figure S8A). None of these transcripts (nor Bmal1, whose lower RNA abundance was partially compensated at the footprint level) was associated with marked TE reduction (Figure 5B and Supplementary Figure S8B), and we thus considered them unlikely candidates for DENR targets. For Clock, however, the reduction in expression was mostly translational, resulting in a ∼20% decrease in protein biosynthesis, as quantified from Clock RPFs (Figure 5A,B and Supplementary Figure S8A,B). In our initial analysis of differential TE (Figure 2A), Clock had, however, not passed the significance threshold (Supplementary File S1). The reason was that the applied algorithm, Xtail, evaluates statistical significance of TE differences through two separate tests—one regarding differences between the RNA and the RPF fold-change distributions, and a second one evaluating differences in TE distributions directly—and then derives the final P-value from the most conservative of the two tests (22). While the additional test might prevent false positives, it can discard differential TE genes close to the statistical threshold. Importantly, for one of the tests, the TE fold change for Clock was significant (P = 0.025 in test 1; P = 0.19 for test 2; Supplementary File S1); Clock thus may represent a DENR target whose detection is at the borderline of statistical significance.

We annotated two translated AUG-initiated uORFs in Clock (Figure 5C and Supplementary Figure S8C–E). The second uORF overlapped with the main CDS in a different frame—a configuration that would render the two translation events (uORF-2 and annotated CDS) mutually exclusive. However, downstream of the uORF-2 stop codon and 10 amino acids into the annotated Clock CDS, we noted an additional AUG in the Clock coding frame. We thus hypothesized that, when uORF-2 was translated, initiation at this alternative AUG could still allow for the biosynthesis of a CLOCK protein variant that was N-terminally truncated by nine amino acids (Figure 5C). In Denr-deficient cells, translation was increased on both uORFs (Figure 5C–D) and, in particular, uORF-2 showed ∼23% higher translation rate upon Denr depletion (Figure 5E), suggesting that DENR-mediated reinitiation after uORF-2 was crucial for the translation of the Clock CDS.

Using GWIPS-viz, an online genome browser for the visualization of ribosome profiling data from published studies (46), we looked for further evidence of uORF-2 and alternative CDS translation across mammalian datasets (of note, both annotated and alternative methionine residues are conserved across mammals, see Supplementary Figure S9). Interestingly, initiating ribosomes can be detected on both uORF-2 and alternative CDS start codons (but not the annotated Clock start codon) across human and mouse datasets, and elongating ribosome profiling studies showed good footprint coverage on uORF-2 and the characteristic accumulation of ribosomes at the start and stop codons (Supplementary Figure S9). Together, the published and our new data thus suggested abundant translation of uORF-2 and of Clock CDS from the alternative AUG, indicating that reinitiation has an important role in ensuring CLOCK protein biosynthesis.

In order to investigate uORF usage, effect, and DENR-dependence for CLOCK biosynthesis from the annotated and alternative CDS, we cloned the Clock 5′ UTR and first 14 codons (containing both the annotated and alternative AUGs) in frame and upstream of Firefly luciferase in our dual luciferase vector (Figure 5F, upper panel). We then systematically mutated the uORF-1, uORF-2 and CDS annotated/alternative start codons, or combinations thereof (Supplementary Figure S8F), and compared the effects on luciferase output in Denr-depleted versus control cells. Of note—and in contrast to Figure 3C—we also adjusted for the effect of Denr depletion on the control reporter (whose levels were set to 1) to display the isolated effects of the Clock variants.

Denr depletion resulted in an almost 30% reduction in expression for the Clock WT construct (Figure 5F, lower panel), in line with the decrease in relative translation quantified for the endogenous transcript (Figure 5D) and with reduced levels of endogenous CLOCK protein (Supplementary Figure S10). Surprisingly, mutation of the annotated CDS initiation codon did not substantially affect reporter levels and DENR dependence when compared to Clock WT. However, mutation of the alternative AUG decreased luciferase signal by ∼95% and abolished DENR dependence (Figure 5F, lower panel), suggesting that most CLOCK biosynthesis in NIH3T3 cells involves the alternative AUG. Mutation of uORF-1 led to a 19% increase in luciferase signal and fully retained DENR dependence. By contrast, uORF-2 mutation led to a ∼50% increase in protein synthesis and eliminated DENR regulation. These findings indicated that uORF-2 translation acts to significantly reduce CLOCK biosynthesis and to render it DENR-dependent. This model was further supported by the combinatorial mutations. Mutation of both uORF-1+2 start codons thus resulted in an additive, >90% increase in protein synthesis, and in a loss of DENR-dependence. In line with minor levels of synthesis of the long CLOCK isoform, mutation of the annotated CDS start codon in combination with uORF mutations led to similar results as mutation of the uORFs alone. Furthermore, mutation of the alternative AUG in combination with uORF mutations led to strongly decreased luciferase signals, consistent with the predominant synthesis of the short CLOCK isoform. In particular, the case where both the alternative CDS and uORF-2 (or uORF-1+2) start codons were mutated—a scenario that would allow unhindered CLOCK synthesis from the annotated AUG—resulted in a >80% reduction in CLOCK protein production (Figure 5F, lower panel). Taken together, we concluded that Clock required DENR for efficient protein synthesis from a CDS that is 9 codons shorter than the annotated one, and that uORF-2 plays a key role in coordinating the complex interactions that occur at the translational level on the Clock transcript (Figure 5G). Moreover, we propose that the CLOCK N-terminus is likely wrongly annotated not only in NIH3T3 cells, but based on our analysis of published ribosome profiling data, also in other cell types and in tissues (Supplementary Figure S9).

High-throughput sequencing data was deposited to GEO, under accession number: GSE124793↗ (SuperSeries composed of ribosome profiling data: GSE116221↗ and iCLIP data: GSE124790↗). Code for ribosome profiling pre-processsing, aligning and quantification, and for uORF annotation, is available online at http://doi.org/10.5281/zenodo.521199↗.