Poised PABP-RNA hubs implement signal-dependent mRNA decay in development

Abstract

Signaling pathways drive cell fate transitions largely by changing gene expression. However, the mechanisms for rapid and selective transcriptome rewiring in response to signaling cues remain elusive. Here we use deep learning to deconvolve both the sequence determinants and the trans-acting regulators that trigger extracellular signal-regulated kinase (ERK)-mitogen-activated protein kinase kinase (MEK)-induced decay of the naive pluripotency mRNAs. Timing of decay is coupled to embryo implantation through ERK-MEK phosphorylation of LIN28A, which repositions pLIN28A to the highly A+U-rich 3' untranslated region (3'UTR) termini of naive pluripotency mRNAs. Interestingly, these A+U-rich 3'UTR termini serve as poly(A)-binding protein (PABP)-binding hubs, poised for signal-induced convergence with LIN28A. The multivalency of AUU motifs determines the efficacy of pLIN28A-PABP convergence, which enhances PABP 3'UTR binding, decreases the protection of poly(A) tails and activates mRNA decay to enable progression toward primed pluripotency. Thus, the signal-induced convergence of LIN28A with PABP-RNA hubs drives the rapid selection of naive mRNAs for decay, enabling the transcriptome remodeling that ensures swift developmental progression.

Fig. 1. Machine learning predicts LIN28A-dependent control of developmental switch in mRNA stability.

a, Temporal dynamics of relative mRNA levels (compared with 2iLIF at 0 h) of naive factors (green) and factors associated with primed cell fate (blue) downregulated and upregulated during PS cell naive-to-primed differentiation. FC, fold change. b, Protocol for SLAMseq measurements of mRNA half-life during naive-to-primed pluripotency transition and machine learning framework for classification of genes based on their mRNA stability (Methods). t_1/2, half-life. c, ROC curves and AUROC values of all resulting binary classifiers trained on different feature sets. d, Accuracy and MCC of individual and combined features in predicting the direction of change in mRNA half-life during naive-to-primed conversion. e, Feature importance ranking for the best-performing classifiers: mouse PS cell CLIP datasets (n = 62) and POSTAR2 features (n = 54) (see Methods for details). f, Violin plots depicting mRNA half-life fold change in WT and LIN28A-KO priming PS cells compared with nPS cells of naive mRNA regulon genes (left, n = 23), and genes encoding an extended early pluripotency program (outlined in Extended Data Fig. 1a) (right, n = 931). A two-sided two-sample t-test was performed to assess the significance of differences between the annotated groups. g, Volcano plot displaying gene expression fold changes and their respective statistical score (Benjamini–Hochberg-adjusted P value determined using Fisher’s exact test with a false discovery rate (FDR) control level of α = 0.05), comparing Dox-induced and untreated iLIN28A–GFP PS cells at 6 h (n = 4 biological replicates per condition, two clones). Naive pluripotency markers are labeled in red (marked factors fulfilled P < 2⁻¹⁰) and extended pluripotency genes are labeled in orange. h, Density of fold change in gene expression changes of mRNAs downregulated upon 48 h of transferring nPS cells into primed medium (n = 931). Genes that are downregulated upon 6 h of Dox-induced iLIN28A–GFP overexpression are labeled in red, while the remaining genes are labeled in black. OE, overexpression. i, Flow cytometry quantification of primed (2 days of naive-to-primed transition) PS cells that were treated with Dox (iLIN28A–GFP) or left untreated (LIN28A KO), compared with WT PS cells and immunostained for SSEA1 and SSEA4. The mean of three independent primed experiments is depicted below (error bars, s.d.) and a two-sided two-sample t-test was performed to assess the significance of differences between the annotated groups. Source data

Fig. 2. MEK–ERK-driven LIN28A S200 phosphorylation is required for clearance of the naive regulon.

a, Fold change of DESeq-normalized counts for naive mRNA regulon upon 6 h of LIN28A WT induction in WL and 2iLIF conditions. A two-sided Welch’s t-test was used to compute the P value. b, DESeq-normalized counts of naive marker Klf4, comparing Dox-induced and untreated iLIN28A in 2iL and MEK–Wnt conditions at 6 h. c, Representative immunofluorescence photomicrographs and fluorescence arbitrary units of immunostained naive marker KLF, comparing Dox-induced and untreated iLIN28A in 2iL and MEK–Wnt conditions at 6 h (n > 200. A two-sided Welch’s t-test was used to compute P values. d, Temporal dynamics of LIN28A S200 phosphoproteome measurements as quantified using liquid chromatography–tandem mass spectrometry (LC–MS/MS) during the naive-to-primed pluripotency transition. e, Differential expression of protein-coding genes obtained by comparing 3′-end RNA sequencing data for iLIN28A WT (left) or iLIN28A S200A overexpression (right) cells to KO cells in the FGF2-treated conditions. DESeq2 was used to calculate adjusted P values (FDR control level of α = 0.1) and fold changes. Naive regulon genes are marked in red. f, Violin plot showing the expression change of LIN28A-target (n = 438, twofold downregulated upon 6 h of LIN28A–GFP overexpression as in Fig. 1h; Benjamini–Hochberg-adjusted P value < 0.005) and remaining mRNAs upon induction of LIN28A WT or LIN28A S200A together with MEK–ERK activation. NS, not significant. g, Density plot outlining the interaction effect for LIN28A-target and remaining mRNAs between the additive destabilizing effect of MEK activation and induction of WT or S200A mutant LIN28A in 2iLIF. h, The direct RNA sequencing of total RNA and quantification of poly(A) length using Nanopolish (see Methods for details). The mean poly(A) length for each gene is plotted for three groups of genes: downregulated, control and upregulated (670, 141 and 371 genes, respectively). The circle marks the mean poly(A) tail length in the gene group and the error bars represent the interquartile range of the distribution. A two-sided Mann–Whitney–Wilcoxon test was used to compute P values within each sample, comparing downregulated and control groups. i, Quantifications of alkaline phosphatase-positive WT, iLIN28A WT and iLIN28A S200A PS cell colonies after 24 h in 2iLIF, LFW (LIF, Fgf2, and Wnt), and LFI (LIF, FGF2 and IWP2) with and without Dox induction. Dots represent replicates from three independent experiments. A two-sided Welch’s t-test was used to assess whether LIN28A rescues had different effects on the number of alkaline phosphatase colonies in LFW and LFI. j, Graphical model of MEK-dependent LIN28A phosphorylation and its binding to 3′UTRs of naive pluripotency mRNAs to induce their selective decay. Source data

Fig. 3. Multivalent AUU-rich regions at 3′UTR termini enable selective mRNA decay and pLIN28A binding.

a, Line plots represent summarized trimer importance scores for the classification of 3′UTR sequences into downregulated and upregulated transcripts. Nucleotide-level importance scores were obtained with SHAP and summarized for each trimer along the evaluated groups of 3′UTRs (see Methods for details on analysis and sample size). Each line summarizes the predictive impact of trimers on the selection of mRNAs for decay and trimers with the highest predictive impact are highlighted. b, Line plots represent the mean nucleotide importance scores for the classification of 3′UTR sequences into downregulated and upregulated transcripts, obtained with SHAP. The scores are visualized in a region 100 nt upstream and 20 nt downstream of the terminal PAS in the 3′UTRs. The window of 5 nt was used for smoothing and the shaded areas represent the 95% confidence interval for the mean (computed with bootstrapping; n = 1,000). Inset: the metamotif represents an example sequence motif, identified with TF-MoDISco, by clustering 3′UTR regions of high importance for the downregulated prediction class (see Methods for details). c, Representative Li-Cor imaging of the iCLIP nitrocellulose membrane, representing the protein–RNA complexes (top) and the amount of IPed LIN28A protein in the experiment (middle). The bar plot shows the quantification of RNA intensity comparing the LIN28A iCLIP capture of protein–RNA complexes with or without MEK–ERK inhibition (n = 3). A two-sided two-sample t-test was performed to assess the significance of differences between the annotated groups. d–f, Top: crystal structure of LIN28A (PDB 3TRZ) in complex with GAGG (d), GAU (e) and AUU (f) RNA motifs through its ZnF (d) and CSD domains (e,f). Middle: corresponding metamotif representations of three motif groups that were identified by PEKA as enriched in 3′UTRs in LIN28A CLIP data (Methods). Bottom: line plots showing the distribution of motif group coverage around cross-link sites located in the 3′UTRs (Methods). Source data

Fig. 4. pLIN28A converges with PABP–RNA hubs at 3′UTR termini to trigger selective mRNA decay.

a,c Heat maps in blue showing the distribution of iCLIP signal across 200 downregulated 3′UTRs for LIN28A (a) and PABPC4 (c) (Methods). For LIN28A, iCLIPs in 2iLIF-treated (top) and FGF2-treated (bottom) cells are shown. For PABPC4, iCLIPs in FGF2-treated LIN28A-KO cells without (top) and with (bottom) LIN28A overexpression are shown. Heat maps in orange depict the total level of CLIP signal in each 3′UTR, normalized by 3′UTR length, expression and library size, thus reflecting the overall RBP abundance in these regions. b,d, Line plots indicating the mean of expression-normalized cross-link coverage in the region 200 nt upstream and 50 nt downstream of the PAS for iCLIPs of LIN28A WT and LIN28A S200A in the FGF2-treated condition (+MEK) and LIN28A WT in the 2iLIF-treated condition (−MEK) (b) and PABPC4 in LIN28A-KO cells with or without the induction of LIN28A WT in the presence of FGF2 (d). Included 3′UTRs were filtered for minimum length and expression level as described in Methods. Top: heat map showing the mean percentage of nucleotides covered by AUU motif groups (each square is a 10-nt bin) in downregulated (n = 831) and control or upregulated (n = 2,372) gene. Shaded areas indicate the 95% confidence interval. e, Line plot indicating the metaprofile of mean expression-normalized cross-link coverage for LIN28A and PABPC4 across the 5′UTR, CDS and 3′UTR regions of transcripts in downregulated (n = 831) and control or upregulated (n = 2,372) genes. Overlapping cross-links for each transcript region were normalized and divided into 100 bins. The mean normalized cross-link coverage over all transcripts in a group was obtained for each bin (dot) across the transcript region (see Methods for details). The line shows a rolling mean across ten bins. TSS, transcription start site. f, Graphical overview outlining how the convergent binding of pLIN28A together with PABP at AUU-rich terminal regions triggers selective mRNA decay essential for epiblast morphogenesis. Source data

Extended Data Fig. 1. Earliest developmental gene and protein expression changes during naïve-to-primed transition.

(a) Temporal dynamics of relative mRNA levels (compared to 2iLIF) of all genes downregulated during hours of PSC naive-to-primed differentiation (n = 20936). (b) Relative mRNA levels and proteome levels (compared to 2iLIF) indicated by line plots and fluorescence units of naïve markers SOX2, KLF2, TBX3, NR5A2 (n = >200) indicated by boxplots during hours of naïve-to-primed differentiation. Below: temporal dynamics of relative protein levels during naïve-to-primed transition of top-ranking RBPs predicting transcripts with decreased stability, LIN28A and PABPC1. (c) Ranking of all half-life transcripts from UP (half-life higher in nPSCs) towards DOWN (half-life lower in nPSCs) compared with the class probability predicted with the binary model trained using features as presented in Fig. 1d. Class probabilities were smoothed using the running mean (k = 10). (d) Permutation analysis of feature importance impacting mRNA stability comparing naïve and priming PSCs. (e) Left: Western blot analysis of LIN28A and histone H3 in LIN28A WT nPSCs, non-targeted LIN28A clone and independent verified LIN28A KO clones. Right: Representative immunofluorescence photomicrographs of WT nPSCs and LIN28A KO counterpart. (f) Protocol for SLAMseq measurements of mRNA half-life in LIN28A KO vs WT priming PSCs, and Machine learning framework for classifier implementation (See methods). FA stands for Fgf2 + ActA medium during naïve-to-primed conversion. (g) Accuracy and Matthews correlation coefficient (MCC) in predicting the half-life change of test-set mRNAs comparing LIN28A KO vs WT priming PSCs trained with indicated features. (h) Left: Feature importance ranking computed as relative influences, for best performing classifiers of PSC CLIP datasets comparing LIN28A KO to WT priming PSCs. Right: Permutation analysis of feature importance by shuffling a particular feature. (i) Left: Western blot analysis of LIN28A, LIN28A-GFP and GAPDH in wild-type EpiSCs, doxycycline-treated and -untreated iLIN28A-GFP nPSCs. Right: Live cell representative fluorescence micrographs of nPSCs expressing endogenous FBL-mCherry and dox-inducible LIN28A-GFP overexpression in 2iLIF, validating the cytoplasmic accumulation of LIN28A-GFP protein. (j) Fluorescence arbitrary units of immunostained naïve markers SOX2, KLF4, NANOG, NR5A2 comparing 20 hrs doxycycline-induced and untreated iLIN28AGFP PSCs. n = >200. Two-sided two-sample t-test; ***P = < 0.0001. (k) Bar-chart depicting relative levels of miR let-7 in iLIN28A-GFP nPSCs with and without doxycycline treatment (n = 3). Source data

Extended Data Fig. 2. MEK/ERK-driven LIN28A S200 phosphorylation and its role in the clearance of naïve pluripotency regulon.

(a-b) Representative immunofluorescence photomicrographs (a) and fluorescence arbitrary units of immunostained naïve marker NANOG (b), comparing 6 hrs doxycyclin-induced and untreated iLIN28A in naïve and MEK/WNT conditions. The size of all samples was =>200 and Welch’s t-test was used to assess whether KLF4 measurements upon LIN28A dox-induction and MEK activation significantly differ in their mean. ***P = < 0.0001 (c) Schematic visualisation of AlphaFold predicted LIN28A structure entailing C-terminal intrinsically disorder region, where S200 is located. Below, schematic of LIN28A knockout cells with doxycycline-inducible LIN28A-WT or LIN28A-S200A. (d) Fractionation western blot analysis of LIN28A and naïve pluripotency factor KLF4 in LIN28A knockout cells with doxycycline-inducible LIN28A-WT or LIN28A-S200A and the same cells left untreated as LIN28A KO. N and C indicate nucleus and cytoplasm, respectively. (e) Number of differentially regulated genes when comparing 3’end RNA sequencing data for LIN28A-WT (grey) or LIN28A-S200A overexpression (black) to KO, in the FGF2 treated condition. (f) Boxplots display the length of 3’UTRs for genes in DOWN, Control, and UP group. For each gene, representative 3’UTR was annotated based on the most abundant mRNA isoform in cells expressing LIN28A-WT cultured in 2iL medium (Methods). Welch’s t-test was performed to assess whether the groups of genes significantly differ in mean length of 3’UTRs. (g) Gene Ontology annotation of DOWN and UP genes. Bar graph representing fold change of GO annotation of biological processes. (h) Normalised Quantseq counts comparing 6 hrs doxycyclin-induced and untreated iLIN28A in naïve and FGF/MEK/WNT conditions for indicated transcripts, without or with MEK activation, respectively (n = >3). (i) Immunofluorescence quantification of KLF4 abundance in LIN28A KO, iLIN28A-WT and iLIN28A-S200A in Wnt/LIF media with or without MEK/ERK activation. The size of all samples was =>200. Source data

Extended Data Fig. 3. Cellular surface markers and alkaline phosphatase stainings dependent on MEK/ERK-driven LIN28A S200 phosphorylation.

(a) Quantifications of alkaline phosphatase positive wild-type, iLIN28A-WT and iLIN28A-S200A PSC colonies after 24 h in 2iLIF, LFW (Lif/Fgf2/Wnt), and LFI (Lif/FGF2/IWP2) with and without dox induction. Dots represent replicates from three independent experiments, Two-sided t-test; ***P = < 0.0001. (b) Representative images of alkaline phosphatase stainings as quantified in Fig. 2i. (c) Quantification of flow cytometry analyses of PSCs that were treated with doxycycline (iLIN28A-S200A or iLIN28A-WT), or left untreated (LIN28A KO), and immunostained for naïve pluripotency marker CD117 and primed pluripotency marker SSEA-4 grown in indicated media for 24 h. n = 3. Two-sided t test, ***p < 0.001, **<0.01. Source data

Extended Data Fig. 4. Cis-acting determinants of mRNA decay.

(a) Schematic visualisation of the deep learning model for prediction of pLIN28A-dependent transcript destabilisation from 3’UTR nucleotide sequences (see Methods for details). (b) ROC curve benchmark shows high efficacy and robustness of the model in classifying the transcripts into DOWN or UP groups based on 3’UTR nucleotide sequences. The area under the ROC curve (auROC) is indicated. (c) Heatmap indicates summarised importance scores for DOWN prediction class for each of the four nucleotides, quantifying their contributions to model predictions across the full 3’UTR sequence. For each 3’UTR, importance scores at each nucleotide position were normalised by 3’UTR length, binned into 100 equal-sized bins, and summed within each bin. Finally, the mean score in each bin was calculated across all evaluated 3’UTRs and visualised (see Methods for details). (d) Metaprofile shows mean nucleotide composition 100nt upstream and 20nt downstream of PAS for DOWN (top) and UP/Control (bottom) transcripts (see Methods for details). The line plots indicate the increased A/U nucleotide content in DOWN transcripts around PAS, compared to UP/Control. (e) Top 4 sequence motifs identified with TF-MoDISco, by clustering 3’UTR regions of high-importance for the DOWN prediction class (see Methods for details). (f) Control experiment for Fig. 3c and representative Li-Cor imaging of the iCLIP nitrocellulose membrane, representing the protein-RNA complexes (above) and the amount of IPed phosphomutant LIN28A-S200A protein in the experiment (below). Source data

Extended Data Fig. 5. Motifs enriched at crosslink sites of both phosphorylated and unphosphorylated LIN28A and their occurrence in 3’UTRs of naïve and DOWN mRNAs.

(a) Heatmap shows hierarchically clustered rankings of 5mers in the 3’UTRs significantly enriched by PEKA in any of the LIN28A iCLIP samples (Methods). (b) Boxplots display the average expression levels for genes in DOWN (n = 1183), Control (n = 2705) and UP (n = 988) groups. Expression for a given gene is assessed by summing up TPM values of its transcripts, and then taking a mean of this value across replicates. The plots show expression levels for LIN28A KO cells (top), with induced LIN28A-S200A (middle) and LIN28A-WT (bottom) expression, without (left; 2iLIF) and with (right, Fgf2/CHIR/LIF) MEK induction. Boxplot whiskers indicate a range of values within the 5 to 95 percentile. Two-sided Welch’s t-test was performed to assess whether groups of genes significantly differ in average expression level and only significant comparisons are indicated. (c) Heatmap (top) shows the mean percentage of nucleotides covered by the AUU-motif group in DOWN (n = 786) and Control+UP (n = 2383) genes (Methods). The mean coverage is computed across evaluated genes in 20nt bins spanning the region of 500 nts before the 3’UTR termini. Line plot (bottom) indicates the mean of expression-normalised crosslink coverage for indicated iCLIPs. Shaded areas indicate 95% confidence interval. (d) Barplot (top) shows the log2 fold-change in expression-normalised crosslink coverage between iCLIPs of LIN28A-WT with and without MEK induction in genes of naïve regulon (n = 16, Methods). Fold-changes are shown in the region of 500 nts upstream of the 3’UTR termini. Error-bars represent 95% confidence interval, computed with bootstrapping (n = 1000). Line plot (bottom) shows the mean percentage of nucleotides in each bin. The shaded areas represent 95% confidence intervals, computed with bootstrapping (n = 1000). (e) Heatmaps show the mean percentage of nucleotides covered by AUU-, WGG- and GAU-motif groups in DOWN (n = 831) and Control/UP (n = 2372) (Methods). Mean coverage is calculated across evaluated genes in 10nt bins in a region 200nts upstream and 50nts downstream of PAS. (f) Boxplots show the count for WGG, GAU, AUU and AAA-motif trimers in a region 100nts upstream of the polyA signal. Each boxplot shows values for three groups of genes DOWN (n = 1002), Control (n = 2126) and UP (n = 778). Only 3’UTRs that contained the canonical PAS motif ‘AAUAAA’ were included in the quantification. Two-sided Mann-Whitney-Wilcoxon test was performed to assess significance. Source data

Extended Data Fig. 6. Heatmaps of iCLIP crosslinking profiles across all DOWN mRNAs.

(a, b) Heatmaps in blue show the iCLIP crosslinking profiles across all 1116 DOWN 3’UTRs for PABPC1 (a) and LIN28A (b). For LIN28A, iCLIPs in 2iLIF- (left) and FGF2-treated cells (right) are show and for PABPC1, iCLIPs in 2iLIF WT PSC cells are shown. For visualisation, iCLIP signal in each 3’UTR was smoothed and min-max normalised (Methods). PolyA signal is annotated with a green dot for the respective mRNA. Source data

Extended Data Fig. 7. AUU-rich mediate a convergence of pLIN28A with PABP at terminal 3’UTR regions.

(a) Library-normalised crosslink profiles of ~5 kb part of 3’UTR of naïve pluripotency factor Tfcp2l1 for LIN28A-WT (in 2iL and FGF2 treated cells), LIN28A-S200A (FGF2 treated cells) iCLIPs and PABPC1/4 iCLIPs (LIN28A KO with and without LIN28A overexpression). Auxilliary tracks below show motif-based binding sites of LIN28A that correspond to WGG-, GAU- and AUU-motif groups (Methods), to AAA trimer, and to canonical ‘AAUAAA’ PAS. (b) Line plot shows metaprofiles of expression-normalised crosslink coverage for LIN28A iCLIPs around the centers of LIN28A peaks in 3’UTRs, merged together from all samples (Methods). Peaks were stratified into those that overlap with peaks of PABPC1/4 (bottom) and those that do not (top). Metaprofile shows that the crosslinking signal of phosphorylated LIN28A-WT increases around peaks that overlap with PABPC1/4 binding and decreases around peaks that do not overlap with PABPC1/4 peaks. Shaded areas indicate a 95% confidence interval. (c) Line plot shows metaprofiles of expression-normalised crosslink coverage for LIN28A iCLIPs around the centers of PABPC1 peaks (above) and PABPC4 peaks (below) located within the final 200nts of 3’UTRs (right) or the rest of the 3’UTRs (left). Metaprofile shows a prominent increase in crosslink signal of pLIN28A around PABP peaks located in the last 200nt and thus indicates phosphorylation-induced repositioning of LIN28A to the 3′-termini. Shaded areas indicate a 95% confidence interval. More information in Methods. (d, e) Library-normalised crosslink profiles of naïve pluripotency factors Zfp281 (d) and Esrrb (e) 3’UTR for LIN28A-WT (in 2iL and FGF2 treated cells), LIN28A-S200A (FGF2 treated cells) iCLIPs and PABPC1/4 iCLIPs (LIN28A KO with and without LIN28A overexpression). Auxilliary tracks below show motif-based binding sites of LIN28A that correspond to WGG-, GAU- and AUU-motif groups (Methods), to AAA trimer, and to PAS. Source data

Extended Data Fig. 8. Convergence of pLIN28A into the poised PABP-RNA hubs selectively enhances the PABP binding only in the DOWN mRNAs.

(a-b) Line plots indicate the mean of expression-normalised crosslink coverage in the region 500 nts upstream of 3’UTR termini for iCLIPs of PABPC1 (a) and PABPC4 (b) in the FGF2-treated condition (+MEK) with (blue) or without (orange) LIN28A induction. Solid line shows the crosslink coverage in DOWN (n = 786), dashed line shows coverage in Control+UP (n = 2383) genes. Shaded areas indicate a 95% confidence interval. (c) Line plot indicates the location of enriched PABP/LIN28A crosslink coverage upstream of the polyA signal on DOWN (n = 831) transcripts. Shaded areas indicate a 95% confidence interval. (d) Barplots show the mean percentage of significantly enriched (p < 0.05) U-rich and A-rich 5-mers at PABPC1 (left) and PABPC4 (right) binding sites in 3’UTRs. The mean number of k-mers of two replicates indicated with dots and the errorbar. Motif enrichment was analysed with PEKA (see Methods). (e) Quantification of expression-normalised crosslink signal in the region 300nts upstream of 3’UTR termini for PABPC1 with and without MEK induction in the presence of absence of LIN28A–WT expression. Two-sided Mann-Whitney-Wilcoxon test was performed to assess significance. (f) Quantification of expression-normalised crosslink signal in the region 300nts upstream of 3’UTR termini for PABPC1 with and without MEK induction in the presence of absence of LIN28A–WT overexpression. (g) Quantification of expression-normalised crosslink signal in the region 300nts upstream of 3’UTR termini for LIN28A with and without MEK induction. (e-g) DOWN and Control+UP feature 586 and 2125 genes, respectively. The length of boxplot whiskers is limited to a maximum 1.5-times the interquartile range. Two-sided Mann-Whitney-Wilcoxon test was performed to assess significance. Only comparisons that pass the 10e-10 statistical threshold are annotated. (h) Boxplots show log2 fold-change of binding in indicated iCLIP experiments, quantified in the region of 300nts upstream of 3’UTR terminus (yellow) or 300nts downstream of stop codon (cyan). Each boxplot shows coverage values for three groups of genes (150, 284, 46 of DOWN, Control and UP, respectively), passing filters for 3’UTR length, expression and crosslink signal (Methods). Two-sided Mann-Whitney-Wilcoxon test was performed to assess significance. Source data