Translational contributions to tissue-specificity in rhythmic and constitutive gene expression

BACKGROUND The daily gene expression oscillations that underlie mammalian circadian rhythms show striking tissue differences and involve post-transcriptional regulation. Both aspects remain poorly understood. We have used ribosome profiling to explore the contribution of translation efficiency to temporal gene expression in kidney, and contrasted our findings with liver data available from the same mice. RESULTS Rhythmic translation of constantly abundant mRNAs affected largely non-overlapping transcript sets with distinct phase clustering in the two organs. Moreover, tissue differences in translation efficiency modulated the timing and amount of protein biosynthesis from rhythmic mRNAs, consistent with organ-specificity in clock output gene repertoires and rhythmicity parameters. Our comprehensive datasets provided insights into translational control beyond temporal regulation. Between tissues, many transcripts showed differences in translation efficiency, which were, however, of markedly smaller scale than mRNA abundance differences. Tissue-specific changes in translation efficiency were associated with specific transcript features and, intriguingly, globally counteracted and compensated transcript abundance variations, leading to higher similarity at the level of protein biosynthesis between both tissues. CONCLUSIONS We show that tissue-specificity in rhythmic gene expression extends to the translatome and contributes to define the identities, the phases and the expression levels of rhythmic protein biosynthesis. Moreover, translational compensation of transcript abundance divergence leads to overall higher similarity at the level of protein production across organs. The unique resources provided through our study will serve to address fundamental questions of post-transcriptional control and differential gene expression in vivo.

6 similar between tissues, with 95% of genes falling into a less than 3-fold range for the kidney/liver 119 TE ratio, as compared to a greater than 100-fold range for the transcript abundance ratio ( Figure   120 2B). This observation was coherent with the considerably broader spread of mRNA abundances 121 vs. TEs across genes within each organ (greater than 500-fold vs. just over 10-fold, respectively; 122 Figure S5A-B) and is in line with a dominant role for the regulation of mRNA levels (i.e., 123 transcription and mRNA decay) in controlling quantitative differences in gene output.

124
Intuitively, we had expected that RNA levels that were widely dissimilar between kidney and liver 125 and subsequently further modulated by organ-specific TEs, would probably give rise to even 126 greater cross-organ divergence at the RPF level. Intriguingly, however, the global correlation 127 between kidney and liver was better for footprint abundances than for transcript abundances 130 dominant protein-coding transcript isoform ("single isoform genes" in the following) that was 131 common to both organs, or whether they gave rise to different (including tissue-specific) mRNA 132 variants ("multiple isoform genes") ( Figure 2D). The observed higher cross-organ concordance of 133 RPFs could have simply had technical reasons, for example if the RPF-seq protocol gave more 134 reproducible results than the RNA-seq protocol. We addressed this caveat by comparing 135 measurement errors (MEs) for RNA and RPF data using a similar approach as in a recent 136 publication [18]. We found that MEs scaled inversely with expression levels, as expected, and 137 showed some variation due to organ ( Figure S6A,B,F,G). Especially in liver and among low 138 expressed transcripts, a tendency towards smaller MEs for RPF than for RNA was indeed visible 139 (differences statistically non-significant). In most other cases, however, measurement errors were 140 (in part significantly) higher for the transcripts' RPF counts than for their RNA counts. Of note, the 141 better cross-organ correlation of RPF vs. RNA levels seen in the full transcript set ( Figure 2C) 142 was also evident within various transcript subsets ( Figure S6C-E, H-L), including such subsets for 143 which RPF MEs were higher than RNA MEs ( Figure S6E, L). It is thus unlikely that technical bias 144 was the reason for the higher RPF correlation. Finally, an analysis that we performed on 145 independent ribosome profiling datasets from rat liver and heart [19] allowed us to confirm the 146 phenomenon of higher concordance of RPF vs. RNA abundance also between these organs 7 ( Figure S7A-C). Taken together, these findings are suggestive of a potentially broader biological 148 phenomenon that consists in the partial compensation of differences in a gene's mRNA 149 expression through counteracting effects exerted through its TE, resulting in the convergence at 150 the level of protein biosynthetic output (footprints, RPF) across tissues.  Table S2). Of these, 533 represented "single isoform genes" with no (or negligible 157 amounts of) expression of tissue-specific mRNA variants. For these genes, we examined whether 158 a higher TE in kidney (N=193) or in liver (N=340) was associated with specific transcript 159 characteristics. Of several features tested, we found that CDS and transcript lengths showed the 160 most significant association with differential TE ( Figure S8A-B). Of note, we had previously seen 161 in liver that shorter coding sequences -i.e., transcripts encoding smaller proteins -are more 162 efficiently translated [10]. Our present analyses suggest that such transcripts are also more prone 163 to tissue-specific regulation at the translational level. Other sequence features showed some bias 164 within the differential TE gene sets as well, although the effects were overall weaker and less 165 consistent. Briefly, the 5′ UTRs of genes with higher TEs in liver were longer and predicted to fold 166 more strongly. By contrast, transcripts with higher kidney TEs were associated with lower 5′ UTR 167 GC content and slightly shorter 3′ UTRs. No association with differential TE was found for the 168 Kozak sequence context score.

169
We also investigated two functional classes of sequence features, miRNA binding sites and 170 upstream open reading frames (uORFs), for association with differential TE. Of note, the 960 "TE 171 different" transcripts were not enriched for any predicted miRNA binding sites, making it unlikely 172 that this class of post-transcriptional regulators is a major player in establishing tissue-specific 173 TEs (data not shown). We had previously observed that in the liver the presence of a translated 174 uORF in the 5′ UTR was strongly predictive of low TE at the main ORF [10]. An analogous 8 relationship was evident in kidney as well ( Figure S5C). To assess whether uORF translation was 176 associated with TE differences across organs, we compared how the identified uORF-containing 177 transcripts (i.e., single isoform genes showing translated uORFs in at least one organ; N=1377) 178 distributed to the differential vs. non-differential TE gene sets. The group of genes with higher TE 179 in liver was significantly enriched for transcripts with translated uORFs (p=6.08e-04; Fisher's 180 exact test) and there was slight depletion among genes with higher TE in kidney (not significant) 181 ( Figure S8C). Only few differential TE genes exhibited uORF translation that was exclusive to one 182 organ, but there was a tendency for kidney-specific translation of uORFs to be associated with 183 higher TE on the CDS in liver, and vice versa ( Figure S8D). For the genes with uORFs translated 184 in both tissues, we expected that cross-organ differences in the strength of uORF usage would 185 negatively correlate with TE differences at the CDS. However, such a trend was only visible for 186 liver differential TE genes ( Figure S8E), and globally, uORF and CDS TEs even showed slightly 187 positive correlation. In summary, these analyses suggested that uORF translation contributed to 188 some extent (and especially for genes that were more efficiently translated in the liver) to cross-189 organ differences in TE; however, the overall impact appeared limited (see Discussion).

190
We next included the "multiple isoform genes" in the analyses and asked whether transcript 191 isoform diversity between the two organs -i.e., the occurrence of tissue-specific mRNA variants 192 generated by alternative transcriptional start sites, splicing, and 3′ processing -had any 193 relationship to differential TE. Briefly, using our RNA-seq data we first compiled an inventory of all 194 annotated, protein-coding transcript isoforms and their estimated relative expression levels per 195 gene and tissue. We then used the Hellinger distance [20] as a measure of dissimilarity in isoform 196 expression levels between kidney and liver. A value of 0 for this metric indicates that a gene has 197 identical isoform distribution in both tissues (i.e., these are essentially the "single isoform genes" 198 described above), while a value of 1 denotes a lack of overlap in expressed isoforms. Globally, 199 the 960 genes with differential TE showed significantly higher Hellinger distances than the 200 remainder of the expressed genes (p=3.74e-04; Kolmogorov-Smirnov-test) ( Figure 2F).

202
By comparing the genes for which all expressed variants affected exclusively one single feature, 203 or for which this particular feature was not affected at all, it became apparent that transcript 204 9 diversity in the 5′ UTR was particularly strongly associated with differential TE ( Figure 2G). By 205 contrast, variation in the CDS showed significantly less association with cross-organ differences 206 in translation efficiency ( Figure S9A-B). Although the low number of available transcripts bearing 207 exclusively 3′ UTR differences precluded a rigorous interpretation, 3′ UTR variation did not appear 208 to be associated with differential TE either ( Figure S9C). Altogether, we thus concluded that TE 209 differences between tissues may, at least in part, have their origin in tissue-specific transcript 210 variants, especially through alternative 5′ UTRs.

211
Finally, we were interested in whether cross-organ differences in translation efficiency affected 212 specific pathways. For the 640 "TE different" genes that showed increased TE in liver ( Figure 2D), 213 gene ontology (GO) analyses revealed significant enrichment for categories related to 214 transcription (Supplementary Table S2). Conceivably, tissue-specific translational control of 215 transcriptional regulators may thus impact also on the organs' transcriptomes. The 320 "TE 216 different" genes that were translated better in kidney did not show any significant enrichment.

218
We next turned to the analysis of factor time across the datasets. We annotated rhythmic events 219 in kidney with the same methodology as previously for liver, including a 1.5-fold cut-off on peak-220 to-trough amplitudes [10]. A list of the detected RNA and RPF rhythms and genome-wide gene 221 expression plots are provided in Supplementary Table S3 and Supplementary Dataset S1, 222 respectively. Our analyses yielded 1338 and 977 genes that cycled at the RNA abundance and 223 footprint level, respectively, with an overlap of 542 genes ( Figure 3A). As discussed later, this 224 relatively modest overlap (542 genes corresponds to 41% and 55% of all "RNA rhythmic" and 225 "footprints rhythmic" cases, respectively) likely underestimates the full extent of shared rhythmicity 226 and only contains the most robustly oscillating gene expression events, which we further explored 227 in the following.

228
Interestingly, the analysis of rhythmicity parameters across the 542 genes revealed that the timing 229 of their RPF peaks relative to their RNA peaks had a significantly different and broader 230 distribution than the corresponding set from liver (p<1.0e-04; permutation test) ( Figure 3B). This 231 observation suggested that the phase of protein biosynthesis rhythms was subject to marked 232 10 translational modulation in kidney. In liver, by contrast, RPF peaks were more tightly gated by 233 RNA abundance peaks. Surprisingly, maximal translation tended to precede maximal RNA 234 abundance in kidney ( Figure S10A), as globally the mean RPF peak phase was advanced (-0.123 235 hours) and also RPF rhythms were enriched for phase advances (282) vs. delays (260), albeit 236 neither reaching statistical significance (p=0.16, Wilcoxon rank sum test).

237
The above analyses used different rhythmic gene sets for kidney than for liver, potentially 238 compromising comparability. The observed differences in the RPF-RNA phase relationships could 239 thus have simply arisen from transcript-specific rather than from tissue-specific differences in the 240 timing of translation. We thus analysed the group of 178 genes whose RNA and RPF profiles

242
Again, the distribution of RPF-RNA offsets was significantly broader in kidney than in liver ( Figure   243 3D; p=0.007, permutation test) with an RPF peak phase advance in kidney (mean -0.143 h) and a 244 phase delay in liver (mean 0.036 h) (Figure S10B-C). We next calculated the gene-wise RPF-245 RNA peak phase difference in kidney relative to that in liver. More genes showed their RPF 246 maxima earlier (96) than later (82) in kidney vs. liver, with a mean advance of -0.178 h (Figure 247 S10D), but again without passing statistical significance (p=0.152, Wilcoxon rank sum test).

248
Conceivably, we lost statistical power and introduced error in the above analyses by restricting 249 the phase comparisons merely to the peaks of the rhythmic curve fits. We thus sought a method 250 that would take into account phase differences between RPF and RNA profiles over all data 251 points. To this end, we used cross-correlation to quantify the similarity between the RPF and RNA 252 time series as a function of sliding one series on the time axis relative to the other. When the time 253 series were not shifted against each other at all (RPF-RNA lag = 0 h), the RPF-RNA cross-254 correlation values were overall highest, as expected, and they were significantly higher in liver, in 255 line with stronger gating of RPF rhythms relative to RNA oscillations in this organ ( Figure 3E).

256
Importantly, when cross-correlation of RNA was calculated with earlier RPF time points (negative 257 RPF-RNA lags; see in particular lags of -2 h to -8 h in Figure 3E), kidneys scored significantly 258 higher than livers. Sliding the series in the other direction, however, rather led to overall better 259 correlations in the liver (see lags of +4 to +8 in Figure 3E; liver-kidney difference non-significant).

260
11 Taken together, these analyses underscored that there was asymmetry in the data with RPF 261 rhythms preceding RNA rhythms specifically in the kidney.

262
We confirmed kidney-specific translational phase advances by visual inspection of individual gene 263 expression profiles. Figure 3F shows the profiles for the genes Hlf, Nampt, Slc5a6, Tardbp, 264 Dnajb4, Cgn and Etnk2, which all show an RPF phase advance of up to several hours relative to 265 RNA. Cross-correlation analysis for the individual genes confirmed kidney-specific, phase-266 advanced translation as well ( Figure 3G).

267
At first sight, translation that is phase-advanced to mRNA abundance is counterintuitive.

278
Rhythmicity detection algorithms are sensitive to false-negatives i.e., to classify gene expression 279 profiles as "non-rhythmic" (for example because they fail imposed thresholds on amplitude or 280 FDR) although the underlying temporal patterns may still be more similar to, and more likely to 281 be, rhythmic than invariable. Of note, the lack of canonical methods to reliably determine true 282 absence of rhythms is a common problem in the field (see recent review by [6] for discussion).

283
Venn diagrams that simply overlap rhythmic gene sets hence need to be interpreted with caution.

284
For these reasons, the extent of "RNA only" and of "footprints only" oscillations in Figure 3A is 285 likely not reported reliably and subject to overestimation. The heatmaps of the corresponding 286 RNA and RPF profiles support this notion as well ( Figure S12B, D). 12 In order to identify the true-positive "translation only" cycling transcripts with higher reliability, we 288 implemented the same methodology as in our previous study [10]. Briefly, we used the analytical 289 framework Babel [22] to preselect all transcripts whose translation efficiency changed significantly 290 over the day (and/or whose TEs deviated significantly from the global transcript population).

291
Rhythmicity analyses were then performed on this gene subset and yielded 92 cases with the 292 sought-after temporal profiles of rhythmic translation on non-rhythmic mRNAs ( Figure 4A).

293
Comparison with the 142 genes of the analogous set from liver revealed near-perfect tissue 294 specificity of translationally driven oscillations. Only two genes, Abcd4 and Lypla2, were shared 295 between the organs; they were both among the least compelling cases of "translation only 296 rhythms" that our method had identified, as judged by visual inspection ( Figure 4B).

297
Interestingly, not only the identity of rhythmically translated genes, but also the time-of-day at 298 which the majority of rhythmic translation events occurred, was highly tissue-specific. The phase 299 histograms thus showed striking differences in the peak time distributions between the organs

318
In summary, we concluded that temporal changes in TE were strikingly tissue-specific and overall 319 relatively rare in kidney. Specifically for transcripts encoding RPs and other components of the 320 translation machinery, which are the most prominent group of TE rhythmic genes in liver, it has 321 been suggested that feeding-dependent mTOR-signaling underlies the translational upsurge at 322 the light-dark transition via a mechanism involving the 5′-terminal oligopyrimidine (5′-TOP) motifs 323 that these transcripts carry [11,12]. Interestingly, the TE comparison between both tissues

338
We first investigated transcript and footprint RPKMs as averages over timepoints to assess the 339 cumulative daily production of clock RNAs and proteins. Most core clock genes showed a 340 considerable degree of organ-specificity in their expression levels that was readily appreciable in 341 the footprint vs. transcript abundance representation with both organs overlaid in a single graph 342 ( Figure 5A). Two tissue differences caught our particular attention. First, the balance between the 343 transcriptional activators Rora/Rorc and repressors Nr1d1/Nr1d2 differed markedly between 14 organs and was skewed towards repression in kidney (i.e., higher Nr1d1/2 and lower Rora/c 345 RPKMs in kidney, Figure 5A-B). These transcriptional regulators bind to shared sequence 346 elements on DNA and form the "interconnecting limb" within the rhythm-generating clock circuitry.

347
In addition, they also control an output branch of the oscillator [1, 2]. It is hence conceivable that   periods than liver clocks [5], as would be predicted from the increased PER biosynthesis that our 361 analyses revealed. As a more general concept, we deem it conceivable that the modulation of 362 biosynthesis levels for individual clock proteins may be a more general mechanism to engender 363 distinct differences in clock parameters across cell types.

366
indicating that translation efficiencies partially counteracted RNA expression differences. Only in 367 four cases (Clock, Arntl, Nr1d2, Rora) TEs exacerbated transcript abundance differences and led 368 to higher tissue differences at the RPF level. Interestingly, this observation could also be made in 369 the time-resolved data. As a measure of similarity between rhythmic profiles, we used the 370 Euclidean distances calculated between the four rhythmic traces of each individual gene (i.e.,

371
RNA and RPF in kidney and liver; Figure S13). Hierarchical clustering of the similarities for the 372 ensemble of the 12 main core clock genes showed that RPF rhythms from the two organs 373 15 grouped together ( Figure 5C). Clock protein biosynthesis rhythms between organs were thus 374 more similar than RNA and RPF rhythms within organs. By contrast, the 178 common rhythmic 375 genes identified in Figure 3C -serving as a control set for this analysis -revealed within-organ 376 clustering ( Figure 5D). These findings underscored that translational compensation was occurring 377 within the core clock, where it led to more similar rhythms in clock protein biosynthesis than would 378 have been predicted from the rhythmic RNA abundance profiles. This phenomenon was, 379 however, not a general feature of all rhythmic gene expression.

380
The transcriptome-wide analyses described further above had shown only weak signs of 381 association between cross-organ differences in TE and in uORF usage ( Figure S8C-E). However, 382 we knew from our previous work in liver that at least five core clock transcripts (Nr1d1, Nr1d2, 383 Cry1, Clock, Arntl) contained translated, potentially regulatory, AUG-initiated uORFs [10]. We 384 therefore examined whether for any of these concrete cases there was evidence for a connection 385 between uORF translation and cross-organ TE differences. The read distribution along the 386 transcripts ( Figure S14A) and the marked frame preference of RPF reads ( Figure S14B) 387 confirmed that the footprints mapping to our annotated uORFs likely reflected active translation.

388
However, only in one case, Nr1d2, there was a distinct anticorrelation between uORF usage and 389 TE differences on the CDS ( Figure 5E). Nr1d2 contains two translated uORFs in the 5′ UTR 390 ( Figure 5F), whose decreased usage in kidney was accompanied with higher TE on the CDS in 391 this organ ( Figure 5E). For Nr1d2, differential uORF usage could thus represent a plausible 392 mechanism that contributes to regulating organ-specific gene expression output at the 393 translational level, keeping NR1D2 biosynthesis low in liver and high in kidney.

395
Given that the functionally relevant output of most gene expression is the protein, quantitative and  i.e. its tissue-and its time of day-dependence, is among the first of its kind and, together with the 401 16 associated datasets and resources, will likely be of wide interest and utility to researchers working 402 in the chronobiology and gene expression fields.

403
We have addressed several, rather fundamental questions that go beyond the chronobiological influence coming from RNA stability as well), whereas differences in translation rate have more of 419 a modulatory role. It is intriguing that this modulation is overall characterised by directionality, with 420 TE differences between tissues globally counteracting some of the mRNA abundance differences.  infer that an overarching theme of the identified associations is a connection to 5′ UTRs, which is 438 in also in line with the notion that initiation is rate-limiting for most translation events. We thus 439 observed associations of cross-organ TE differences with 5′ UTR length, with uORF usage, with 440 GC content and folding potential, as well as with transcript isoform diversity that affected the 5′ 441 UTR. We would like to point out that comprehensive uORF annotations remain a bioinformatics 442 challenge that is far from resolved. We have therefore restricted our analyses to AUG-initiated 443 ORFs, inevitably leading to a bias towards false-negatives in uORF annotation. As we will learn 444 how to annotate uORFs more comprehensively and more precisely in the future, it may be worth 445 revisiting the relationship between differential TE and uORF translation in our datasets in order to 446 evaluate whether a clearer role for these regulatory sequence elements will emerge.

447
Our study has led to novel insights into rhythmic gene expression. The extent to which rhythmicity 448 is generated by the temporal regulation of translation has been the subject of speculation ever

565
Within each bin, the measurement error was calculated separately for RNA-and RPF-seq and for 566 liver and kidney, using the two replicates (log of normalised CDS counts) to estimate the error, 567 and the 12 timepoint samples to estimate its variability. For the analyses using a filtered gene set 568 ( Figure S6F-G), genes that showed a mean expression ratio (either between organs or between 569 RNA-and RPF-seq) greater than 2 for all timepoints were excluded (9236 genes used in  To test for differential translation efficiency (TE) between liver and kidney we used the Wilcoxon-573 signed rank paired test, using all 24 samples (12 timepoints; 2 replicates/timepoint) as replicates; 574 resulting p-values were FDR-corrected. A gene was defined as having differential TE when FDR 575 < 0.01 and the inter-organ difference in TE was at least 1.5-fold ( Figure 2E).

576
Analysis of transcript usage diversity across organs: For each gene g, P(g) = (p 1 ,...,p n ) is the 577 vector of the relative expression proportions of its n protein-coding transcripts, as estimated from 578 our RNA-seq analysis (see Sequencing data processing, alignment and quantification). To 579 quantify the dissimilarity in relative transcript isoform expression between liver L and kidney K, the 580 Hellinger distance H is defined as:

582
In order to detect the transcript features that were associated with tissue specificity in TE, we 583 selected genes whose transcript diversity between both organs originated from, or was excluded 584 from, 5′ UTR, CDS, or 3′ UTR, based on feature annotation information for the detected protein-585 coding transcripts ( Figure 2G and Figure S9).

586
Study of transcript characteristics: For single-isoform genes, we investigated whether a particular 587 transcript characteristic (length, GC content, Kozak context, structure) could be predictive of  To assess the impact of differential uORF usage on TE differences across organs, uORFs were 597 identified as in our previous study [10]. Briefly, genes expressing a single protein-coding isoform 598 in both organs were used for this analysis (N=5815). We selected uORFs with an AUG start 599 23 codon and a length of at least 18 nt to the first in-frame stop codon and considered them as 600 translated if the reads showed significant frame bias towards the reading frame of the uORF start 601 codon and if coverage was >10%. uORF translation efficiency was calculated from the ratio of 602 RPF-seq to RNA-seq reads whose predicted A-sites mapped to the annotated uORF regions. If 603 several uORFs partially of completely overlapped on a given 5′ UTR, a composite uORF was 604 considered for read counting. uORFs overlapping with the CDS in the same frame were not 605 considered. When they overlapped in different frames, only reads mapping to the 5′ UTR-specific 606 uORF sequence (but not the overlapping sequence) was considered for quantifications.   Figure 5C or all rhythmic genes in Figure 5D), using the "average" clustering

787
Inset: Frame analysis of CDS reads showed preference for the annotated reading frame (frame 1, 788 the same frame as the start codon) in RPF but not in RNA reads. Violin plots extend to the range 789 of the data (N=3694 genes for liver, N=4602 genes for kidney). A separate analysis of the higher 790 level of stop codon footprints in kidney, that also led to the differences in 3′ UTR reads in B, can 791 be found in Figure S2A    circular arrangement was larger for liver than kidney, suggesting a higher contribution of hepatic 799 rhythmic genes to overall variability. Figure S4 shows the scree plot for the ten first components.

829
G Cumulative distribution of the kidney-to-liver TE ratio for genes whose transcript diversity 830 originated exclusively from the 5′ UTR (identical CDS and 3' UTR, light blue, N=216; these genes 831 show more TE differences across organs), and genes whose transcripts had identical 5′ UTR 832 (and divergent CDS and/or 3′ UTR, purple, N=314; these genes show less TE differences across 833 organs). The vertical dashed grey line marks the 1.5-fold difference used to define differential TE 834 (as in (E)). These results suggested that tissue specificity in TE was partially achieved by 835 expressing transcript isoforms that differed in their 5′ UTRs (note the significant shift towards 836 smaller TE differences for genes with identical 5' UTRs). See also Figure S9.

99
F and G Same as A and B, but with a filtered gene set in which specifically those 100 genes that showed very different expression levels/high variability between organs or 101 between datasets (RPF-seq, RNA-seq) were removed (see Methods). The reason to 102 also analyse such a filtered set was that we wished to be sure that genes that were 103 widely different in their gene expression level were not distorting the analyses (e.g.

104
specifically causing extreme measurement errors under a condition where 105 expression was very low). Moreover, because the binning into the groups was based 106 on expression level across all sets (calculated as the fourth root of the product of 107 liver RNA-seq, liver RPF-seq, kidney RNA-seq and kidney RPF-seq), the highly 108 variable genes made binning inaccurate. This filtered set thus contained genes with 109 overall better comparability across datasets; of note, the distribution of ME

156
D Organ-specific uORF usage and its association with differential TE. The group of 157 genes with uORFs specifically translated in liver was enriched for transcripts better 158 translated in in kidney, and vice versa, consistent with a role of tissue-specific uORF 159 usage in setting TE differences. However, due to the low number of differential TE 160 genes exhibiting uORF translation that was exclusive to one organ for this analysis, 161 the enrichments and depletions did not reach statistical significance.

162
E Scatterplot of upstream ORF vs. CDS TE differences across organs for genes 163 containing translated uORFs in both organs and detected as differential TE with 164 higher TE in kidney (yellow) or liver (green), or not showing differential TE (grey). An 165 anticorrelation between uORF usage and CDS TE was only observed for genes with 166 differential and higher TE in liver. 167 8 Figure S9. Relationship between transcript diversity and differential TE.

168
A Cumulative distribution of the absolute kidney-to-liver TE ratio for genes whose 169 transcript diversity is present or absent only in the indicated feature. The vertical 170 dotted grey line marks the 1.5-fold difference used to define differential TE. In this

174
Collectively, these results showed that transcript diversity that originated only within 175 the CDS (red), or that was excluded from the 5′ UTR (purple), or that was present 176 only within the 3′ UTR (dark green), all showed smaller TE differences across 177 organs, thus pointing towards variability within the 5′ UTR as a contributor to tissue-178 specific TE.

181
UTRs are identical), there is a significant shift to more similar TEs in both organs.

182
This is consistent with the specific association of 5′ UTR diversity with differential TE 183 that is shown in Fig. 2G.

235
UTRs, corresponding to the same length as the full 5′ UTR, is depicted (exception 236 Nr1d1, for which the 3′ UTR is so short that it is shown full length).

237
B Read distribution to the three translation frames showed a frame bias of footprint 238 reads for most predicted uORFs that was in a similar range as the frame bias on the 239 CDS. This frame preference is indicative of active translation on the uORFs.