RNA regulons in Hox 5' UTRs confer ribosome specificity to gene regulation.

Emerging evidence suggests that the ribosome has a regulatory function in directing how the genome is translated in time and space. However, how this regulation is encoded in the messenger RNA sequence remains largely unknown. Here we uncover unique RNA regulons embedded in homeobox (Hox) 5' untranslated regions (UTRs) that confer ribosome-mediated control of gene expression. These structured RNA elements, resembling viral internal ribosome entry sites (IRESs), are found in subsets of Hox mRNAs. They facilitate ribosome recruitment and require the ribosomal protein RPL38 for their activity. Despite numerous layers of Hox gene regulation, these IRES elements are essential for converting Hox transcripts into proteins to pattern the mammalian body plan. This specialized mode of IRES-dependent translation is enabled by an additional regulatory element that we term the translation inhibitory element (TIE), which blocks cap-dependent translation of transcripts. Together, these data uncover a new paradigm for ribosome-mediated control of gene expression and organismal development.

In addition to transcription, an important layer of gene expression control may be conferred through a regulatory function of the ribosome[1-6]. For example, RPL38, one of 80 Ribosomal Proteins of the eukaryotic ribosome, helps establish the mammalian body plan by selectively facilitating the translation of subsets of Hox mRNAs[1], genes critically required for formation of the body plan[7]. How ribosome-mediated regulation of gene expression is encoded within mRNA sequence remains an unanswered question.

Eukaryotic cellular mRNAs are capped, and their translation is largely cap-dependent[8,9]. In many viral mRNAs that are not capped, IRESes provide an alternative mechanism for ribosome recruitment to promote translation initiation[8,10]. Interestingly, IRES elements have also been discovered in certain cellular mRNAs, including c-myc, XIAP, Apaf-1 and p53[11-16]. As these cellular mRNAs are capped, the IRES elements act as a “fail-safe” mechanism to promote translation under stress conditions, such as apoptosis or hypoxia, when cap-dependent translation is downregulated[17,18]. Cellular IRES elements[19] have little sequence or structural homology, and their relevance for translational control and gene expression in normal cell physiology, tissue development, or patterning is unclear.

We report below the unexpected identification of IRES elements within Hox mRNAs, particularly those regulated by RPL38. These IRESes are unusual in that they are critically required for accurate gene expression during normal development. We also identify the TIE, an additional regulatory element in these transcripts, which inhibits general cap-dependent translation and thereby facilitates ribosome-directed, spatiotemporally-restricted protein expression to guide organismal development.

RPL38-regulated Hox mRNAs possess IRES elements

To address the nature of specific cis-regulatory elements within transcripts that confer greater regulation by the ribosome, we studied RPL38-mediated control of Hox mRNA translation. We previously showed that while RPL38 is dispensable for general cap-dependent translation, it is required for the translation of a subset of the eleven HoxA mRNAs (Hoxa4, Hoxa5, Hoxa9, Hoxa11) (Fig. 1a)[1]. We employed a bicistronic reporter assay in the mesenchymal stem cell line C3H10T1/2, which expresses Hox transcripts[20], to delineate whether Hox 5′UTRs direct cap-independent translation (Fig. 1a, b). In this reporter system, the first cistron (Renilla luciferase, RLuc) is translated by a cap-dependent mechanism, and the second cistron (Firefly luciferase, FLuc) is translated only if the preceding 5′UTR element can recruit ribosomes by a cap-independent mechanism (Fig. 1a); the latter is considered to reflect IRES-dependent translation initiation[21].

Figure 1

Select HoxA 5′UTRs contain IRES elements regulated by RPL38

a, HoxA genes are located in tandem along a chromosomal locus, and their location directs their anterior expression boundaries along the embryo. Striped boxes, Hox genes translationally regulated by RPL38. IRES activity is the ratio between Fluc and Rluc. b, IRES activity of HoxA 5′UTRs. IRES activity of UTRs was normalized to that of empty vector (pRF) and RPL38-regulated Hox genes are denoted in red. ** p< 0.01 (t-test as compared to pRF). n= 5. c, RPL38 regulates translation of specific HoxA IRES elements. Bicistronic reporters were transfected into control cells (CTRL) or cells where one copy of Rpl38 was deleted (RPL38 KD). Top: Western blot of RPL38 and β-actin with quantification of RPL38 levels below blot. Bottom: IRES activity of specific Hox 5′UTRs in RPL38 KD cells normalized to that of control cells. ** p< 0.01 (t-test). n= 3. d, Deletion analysis of the Hoxa9 5′UTR. IRES activity was normalized to that of the full-length Hoxa9 5′UTR (A9). AS: anti-sense. **p<0.01 (t-test as compared to pRF). n= 3. e, IRES activity of the Hoxa9 IRES (nt 944–1266) in RPL38 KD cells. **p<0.01 (t-test). n= 3. f, Alignment of the Hoxa9 5′UTR in vertebrates. g, The 5′UTR of one of two zebrafish Hoxa9 paralogs (Hoxa9b) has similar IRES activity as the mouse Hoxa9 IRES. IRES activity was normalized to pRF. **p< 0.01 (t-test as compared to pRF). n= 3. All experiments were performed in triplicates. Data represent mean ± s.d.

These experiments revealed that many Hox 5′UTRs possess IRES activity that is as strong as, or even stronger than, the Hepatitis C Virus (HCV) IRES element (Fig. 1b, Supp. Table 2). To ensure that Hox 5′UTRs do not contain putative promoters or splice sites that may produce a monocistronic capped FLuc transcript, we performed additional controls. As expected for a single transcript containing both Rluc and Fluc: (i) the RNA levels of the two reporters are expressed in the same ratio among all bicistronic constructs; (ii) an shRNA against RLuc downregulated both Rluc and Fluc expression; (iii) these transcripts do not contain cryptic splice sites; and (iv) no ribosome read-through of the first reporter is evident (Extended Data Fig. 1). There is good correlation between RPL38 regulation and the presence of IRES elements; all RPL38-regulated mRNAs in the HoxA locus possess IRES elements. These include Hoxa4, which we show by 5′ RACE possesses a longer 5′UTR of 1.1 kb that contains an IRES element.

Extended Data Figure 1

HoxA IRES controls confirming that Fluc activity from bicistronic vector is due to IRES activity

a, qPCR of both Rluc and Fluc from transfected cells shows that Rluc and Fluc are produced at the same ratio. All Rluc and Fluc values are normalized to that of HCV (set to 1). n = 3; individual experiments performed in duplicates. b–c, shRNA against Rluc decreased reporter activity of both Rluc (a) and Fluc (b), confirming that Rluc and Fluc were transcribed on the same mRNA. n= 3 individual experiments performed in triplicates. d, RT-PCR using primers in Rluc and Fluc show that there is no cryptic splice site in the cloned Hox 5′UTR. Primer locations are shown as arrows in the diagram. e, Inserting a strong hairpin (−67 kcal/mol) after the Rluc reporter did not affect Fluc activity, suggesting Fluc activity was not due to ribosome readthrough.

To see if cap-independent translation from the Hox 5′UTRs is controlled by RPL38 in C3H10T1/2 cells, a copy of Rpl38 was disrupted by TALEN nucleases, yielding a 40% reduction in RPL38 protein expression (Fig. 1c, Extended Data Fig. 2a). In these cells, while there is no change in cap-dependent translation, there is a specific decrease in IRES-dependent translation of Hox target mRNAs that are regulated by RPL38 in vivo (Fig. 1c, Extended Data Fig. 2b). Consistent with the fact that RPL38 control of IRES-dependent translation is transcript-specific, Hoxa3 IRES-dependent translation, which is not regulated by RPL38 within the embryo[1], or an HCV IRES are similarly unaffected by RPL38 knockdown (Fig. 1c). These results reveal the unexpected presence of IRES elements within Hox 5′UTRs and a critical function of RPL38 in regulating their IRES-dependent translation.

Extended Data Figure 2

Disruption of RPL38 in C3H10T1/2 by TALEN nucleases

a, Location of TALEN pairs. 2 pairs of TALENs were designed to bind at the end of Exon 2 and the beginning of Exon 3 to make a genomic break close to the ATG. Sequencing of a positive clone shows a deletion of the ATG and most of the intron after it. Coding sequence is highlighted in green.

b RPL38 knock down does not change cap-dependent translation (Rluc) but decreases IRES-dependent translation (Fluc) from specific Hox 5′UTRs. Luciferase activity was normalized to amount of Fluc RNA in the cells as quantified by qPCR. *p< 0.05 (t-test compared to Ctrl). n = 2 individual experiments performed in duplicates.

The Hoxa9 minimal IRES is evolutionarily conserved

To further understand IRES-dependent regulation of Hox mRNAs, we made a series of deletions within the 1.2-kb Hoxa9 5′UTR. We localized the minimum fragment for RPL38-dependent IRES activity to nt 944–1266 (which we term the Hoxa9 IRES element) (Fig. 1d). Although this fragment does not fully recapitulate the IRES activity observed with full-length Hoxa9 5′UTR, it has strong IRES activity on its own and its removal abolishes all IRES activity (Fig. 1d, e). This region of the Hoxa9 5′UTR shows remarkable evolutionary conservation in all vertebrates (from fish to mammals) that is significantly greater than either the remaining 5′UTR or the 3′UTR (Fig. 1f, Extended Data Fig. 3). IRES function is also evolutionarily conserved, as the zebrafish Hoxa9b 5′UTR shows strong IRES activity in murine C3H10T1/2 cells (Fig. 1g). These findings suggest that the Hoxa9 IRES element may have arisen early during vertebrate evolution for post-transcriptional regulation of Hox expression.

Extended Data Figure 3

Alignment of Hoxa9 IRES element between vertebrate species

Nucleotides 944 to 1266 of the mouse Hoxa9 5′UTR were aligned with sequences from other vertebrates and show high sequence homology. Nucleotides are colored based on their homology.

The Hoxa9 IRES forms an RNA structure in vitro that recruits the ribosome

Many viral IRESes possess structures such as conserved helices, asymmetric bulges, and pseudoknots that interact with initiation factors or the ribosome to promote translation initiation[22-24]. We therefore examined whether the Hoxa9 IRES element has structural properties that are functionally important.

Within the full-length Hoxa9 5′ UTR, protections of sub-regions of the 944–1266 domain from selective 2′-hydroxyl acylation analyzed by primer extension (SHAPE)[25] were consistent with the minimal IRES domain forming a specific RNA structure in vitro (Extended Data Fig. 4). Automated modeling indicated that this domain formed a four-way junction with two long hairpin arms (P3 and P4; Fig. 2a, Extended Data Fig. 6) and a ‘right-angle’ asymmetric bulge (between P3b and P3c), but with little interaction with other domains of the 5′UTR (Extended Data Fig. 4). Analogous chemical mapping and functional studies on the 5′UTR from a different Hox mRNA, Hoxa5, also revealed a structured subdomain (nt 77–132) necessary for its full IRES activity (Extended Data Fig. 5a, d, e). Deletion of the subdomain drastically reduces IRES activity (Extended Data Fig. 5b). Structure conservation between different cellular IRES elements has so far been poorly defined[19]. Interestingly, the ‘right angle’ asymmetric bulges observed in Hoxa5 and Hoxa9 (Extended Data Fig. 5c) bear qualitative similarity to asymmetric bulges in, for example, the HCV IRES RNA[26], suggesting at least certain common structural features are present amongst cellular Hox IRES RNAs and viral IRESes.

Extended Data Figure 4

Chemical mapping and secondary structure prediction of full-length Hoxa9 5′UTR

a, Secondary structure prediction of full-length Hoxa9 using ligSHAPE data. The Hoxa9 IRES element (nt 957–1132, shaded in green) is predicted as the same secondary structure shown in Fig. 2a. b, Normalized SHAPE reactivity of Hoxa9 IRES (nt 957–1132 and 944–1266 from CE-based 1-dimensional SHAPE, full-length 1–1266 from MiSeq-based ligSHAPE). c, Normalized SHAPE reactivity of Hoxa9 TIE (nt 1–342 from CE-based 1-dimensional SHAPE, full-length 1–1266 from MiSeq-based ligSHAPE).

Figure 2

The Hoxa9 IRES forms a stable RNA structure that facilitates recruitment of the ribosome

a, Secondary structure of Hoxa9 5′UTR (nt 957–1132) by SHAPE and mutate-and-map analysis. Nucleotides are colored with SHAPE reactivities. Percentages give bootstrap support values for each helix, including a tentative pseudoknot (pk). Base-pairs verified by mutation and rescue are circled green. Inset: mutations made in helix P3a for functional testing in b. b, Deletions as well as mutation-and-rescue of specific hairpins identified in Hoxa9 IRES structure shown in a affects IRES activity in bicistronic reporter assays. IRES activity was normalized to full-length Hoxa9 5′UTR (A9, set to 1). **p< 0.01 (t-test as compared to A9). n= 3 experiments performed in triplicates. c, Schematic of RNA pull-down experiments. d, e, Representative western blot for RPs (d) and qPCR for rRNA (e) from elutes of two independent RNA pull-down experiments. The 18S and 28S rRNA amounts of all Hoxa9 deletion mutants were normalized to amount of full-length Hoxa9 5′UTR (set to 1) eluted from beads. n= 2. Data represent mean ± s.d.

Extended Data Figure 6

Secondary structure prediction and mutate-and-map (M2) dataset of Hoxa9 IRES element

a–b, Entire M2 dataset and Z-score contact-map of Hoxa9 nt 957–1132 across 177 single mutants probed by 1M7. c, Secondary structure prediction of Hoxa9 nucleotides 957–1132 using M2 data alone. The model contains the same helices as the model from combined SHAPE/ M2 analysis in Fig. 2a, up to register shifts and edge base pairs; the small rearrangements are labeled P3b′, P3c′, P3d, P4b′ and P4c′. d, Secondary structure of Hoxa9 nt 957–1132 using 1-dimensional SHAPE data alone. Nucleotides are colored with SHAPE reactivity. The model contains the same helices as the model from combined SHAPE/M2 analysis in Fig. 2a, up to register shifts and edge base pairs; the small rearrangements are labeled P3b′, P3c′, and P4c. e, Secondary structure prediction and bootstrap support matrix of Hoxa9 nt 944–1266 using 1-dimensional SHAPE data. The model agrees with the model derived from SHAPE analysis of the 957–1132 subdomain in Fig. 2a.

Extended Data Figure 5

Chemical mapping and secondary structure prediction of full length Hoxa5 5′UTR

a, Secondary structure prediction of Hoxa5 using 1-dimensional SHAPE data. Nucleotides are colored with SHAPE reactivities. Percentage labels give bootstrap support values for each helix. The feature highlighted in blue resembles P3 in Hoxa9 and the tip highlighted in pink is deleted in b. b, The deletion of the tip identified in Hoxa5 IRES structure shown in a decreases IRES activity in bicistronic reporter assays. IRES activity was normalized to full length Hoxa5 5′UTR (A5, set to 1). **p< 0.01 (t-test as compared to A5). n= 2 experiments; performed in triplicates. c, Both Hoxa9 and Hoxa5 contain an asymmetric bulge in a region important for IRES activity. d-e, Normalized SHAPE (d) and DMS (e) reactivity of Hoxa5 (CE-based and MiSeq-based).

To gain further insight into structural and functional properties of a Hox IRES, we applied multidimensional chemical mapping to the Hoxa9 minimal IRES subdomain (Fig. 2a, Extended Data Fig. 6). Comprehensive mutate-and-map analysis[27] gave strong independent evidence for the same structural topology seen in the context of the full-length 5′ UTR, revealing exposure of RNA segments upon mutation of their predicted pairing partners (Extended Data Fig. 6a-d). Importantly, compensatory mutation analysis[28] unambiguously confirmed specific base pairs in the P3 and P4 segments (green circles, Fig. 2a, and Extended Data Fig. 7). After this structural characterization, we tested the functional significance of these hairpins towards IRES-dependent translation through bicistronic assays. Deletion of P3, P4, or both hairpins in the context of the full-length Hoxa9 5′UTR gave a significant reduction in IRES activity. Most strikingly, the mutation of nucleotides in the base of P3a reduced IRES activity, while compensatory mutations gave quantitative rescue of IRES activity (Fig. 2b). These results demonstrate the importance of the P3 and P4 hairpins in IRES function and connect the detailed in vitro Hoxa9 IRES structural model to functional activity. Many viral IRES elements contain pseudoknots that may position the ribosome at the initiation codon[29]. A potential pseudoknot was suggested by mutate-and-map analysis (Fig. 2a), but this awaits further analysis. Together, these data confirm a structure defined at base-pair resolution for a Hox IRES element and indicate that a specific RNA structure is important for cellular Hox IRES activity.

Extended Data Figure 7

Mutation/rescue results of Hoxa9 IRES structure (nt 944–1266) probed by 1M7

Electropherograms of mutation/rescue to test base-pairings in P3c (a–b), P3b (c–e), P3a (f–k), P4b (l–o), P4a (p–u) and pk3-4 (v–ai). Perturbation of the chemical mapping reactivities by mutations of one strand and restoration by mutations in the other strand provide strong evidence for the tested pairings in P3c (a–b), P3b (c–e), P3a (f–k), P4b (m–o) and P4a (p–q). Lack of rescue in other tested pairings is consistent with either absence of those pairings or higher-order structure (e.g., base triples) interacting with those pairings.

To date, only a small class of viral IRES elements have been shown to interact with both the large and small ribosome subunits to form a translationally competent 80S ribosome[30,31]. However, a biotinylated full-length Hoxa9 5′UTR, as well as the minimal IRES element contained within nt 944–1266, are able to pull down RPs from both the large and small subunits, including RPL38 (Fig. 2c, d). The full-length 5′UTR also pulls down both 28S and 18S rRNAs (Fig. 2e), suggesting that the 80S ribosome is able to form on the uncapped Hoxa9 5′UTR. Importantly, the full-length Hoxa9 5′UTR lacking the nt 944–1266 IRES element (A9ΔIRES) or an antisense Hoxa9 5′UTR (A9 AS) is no longer able to efficiently recruit RPs or rRNA (Fig. 2d, e), consistent with the fact that these fragments do not possess IRES activity (Fig. 1d). These data support the idea that a structured element from a Hox IRES facilitates 80S ribosome recruitment to the mRNA, either directly or through intermediary RNA-binding proteins.

TIE inhibits cap-dependent translation of IRES-containing mRNAs

Because all eukaryotic mRNAs are capped and translated by a cap-dependent mechanism[8], the function of an IRES-element in normal translational control remains largely unknown. Surprisingly, removing the IRES element from a Hoxa9 5′UTR monocistronic reporter construct dramatically decreases translational efficiency by 70-fold, despite the fact that the mRNA remains capped (Fig. 4a). We therefore hypothesized that the Hoxa9 5′UTR may contain additional element(s) that intrinsically inhibit cap-dependent translation. The Hoxa9 5′UTR contains 14 putative upstream open reading frames (uORFs), which are generally thought to inhibit the translation of the main open reading frame[32]. However, mutating all 14 ATGs to TTGs in the HoxA9 5′UTR (A9ΔIRESΔuORF) (Extended Data Fig. 8a) does not affect cap-dependent translation of Hoxa9. We therefore placed various regions of the Hoxa9 5′UTR upstream of a monocistronic reporter to assay for effects on cap-dependent translation. Strikingly, nt 1–342 of the Hoxa9 5′UTR strongly inhibit cap-dependent translation (Fig. 4b) despite this region not being particularly GC-rich (55%). Other similarly sized pieces of the 5′UTR (nt 343–643), or the very long remaining portion of the 5′UTR (nt 343–1266), do not inhibit cap-dependent translation and are likely translated by a combination of cap and IRES-mediated translation. To our knowledge, the 1–342 fragment of the Hoxa9 5′UTR is a unique example of a 5′UTR regulon that is sufficient to strongly inhibit cap-dependent translation under physiological conditions. We therefore termed this 5′UTR mRNA element the Translation Inhibitor Element (TIE). Interestingly, SHAPE analysis of the full-length Hoxa9 5′UTR lacks evidence for structural interactions between the TIE and the minimal IRES (Extended Data Fig. 4a), suggesting they function as independent modules. These findings reveal that the TIE enables a more specialized mode of translation initiation directed by the IRES.

Figure 4

Identification of the TIE that blocks cap-dependent translation

a, Deletions within the Hoxa9 5′UTR in a monocistronic reporter assay. Fluc activity was normalized to amount of luciferase mRNA determined by qPCR. All values are normalized to Hoxa9 full-length 5′UTR (A9), set to 1. n= 4. b, A TIE that inhibits cap-dependent translation was identified in the Hoxa9 5′UTR (nt 1–342). n= 3. c-e, Deletion of the minimal IRES elements within monocistronic reporters of HoxA 5′UTRs and identification of TIEs close to the mRNA cap in Hoxa3 (c), Hoxa4 (d) and Hoxa5 (e). All values are normalized to the full-length constructs (FL). n= 3. f, HoxA TIEs were placed upstream of the β-globin 5′UTR (Hbb) in a monocistronic reporter to assay inhibition of cap-dependent translation. n= 3. All experiments were performed in duplicates. Data represent mean ± s.d. **p< 0.01 (t-test).

Extended Data Figure 8

Putative uORFs within the 5′UTRs of Hoxa9 and Hoxa5 do not inhibit cap-dependent translation and Hoxa9Δ targeting strategy

uORFs are marked by black circles on the diagram of monocistronic reporter for the (a) Hoxa9 and (b) Hoxa5 5′UTR. All the ATGs in the 5′UTR were mutated to TTG in A9ΔuORF construct and GTG in A5ΔuORF. The IRES element (944-1266) was removed in A9 IRES construct. The IRES element was removed from the A9ΔuORF construct in A9ΔIRESΔuORF. n= 3 individual experiments in duplicates. Data represent mean ± s.d. c, Diagrams of the Hoxa9 locus and the targeting vector. Boxes represent exons, grey boxes represent UTRs and black boxes represent the coding sequence. Nucleotides 944-1145 were replaced by a floxed Neo cassette in the targeting vector. Locations of Southern blot probes, restriction enzymes used for Southern analysis and expected sizes are marked on the diagrams. d, Southern blot analysis of targeted cells using both the 5′ and 3′ probes showing that both arms integrated correctly into the Hoxa9 locus.

A unique 5′UTR topology allows for specialized translation

We next asked whether the unique 5′UTR topology of both a TIE close to the 5′ end of the mRNA and an IRES element is a more generalizable mechanism of translation regulation. Our findings reveal that all the minimal IRES domains of HoxA 5′UTRs lie within the 100–300 nt closest to the AUG, with the exception of Hoxa5 (Fig. 3). Importantly, deletions of these minimal IRES domains in monocistronic reporter constructs reveal that they are critically required for normal translation (Fig. 4c-e). In addition, within these HoxA 5′UTRs, a TIE is also present near the mRNA cap (Fig. 4c-e). Moreover, although the Hoxa9, Hoxa3, Hoxa11, and Hoxa4 TIEs do not share obvious sequence similarity, they all dramatically block translation from the β-globin (Hbb) 5′UTR (Fig. 4f), which initiates translation exclusively by a cap-dependent mechanism[33]. Hoxa5 is the only 5′UTR that does not appear to possess a TIE and a putative uORF also does not decrease cap-dependent translation (Extended Data Fig. 8b). However, Hoxa5 has the shortest 5′UTR (237 nt) and almost the entire UTR is required for IRES activity, leaving little space for an additional TIE (Fig. 3c). Importantly, the Hoxa5 transcript can be produced as a bicistronic mRNA with the neighboring Hoxa6 gene[34] and therefore may not need a TIE to inhibit cap-dependent translation in this context. These findings reveal a unique topology to Hox 5′UTRs that suppresses a more generic mode of cap-dependent translational control and facilitates specialized, ribosome-mediated control of protein expression.

Figure 3

Identification of minimal IRES domains in all IRES-containing HoxA 5′UTRs

Deletion analysis of HoxA 5′UTRs performed in bicistronic constructs maps the IRES element in Hoxa3 (a), Hoxa4 (b), Hoxa5 (c) and Hoxa11 (d). IRES activity was normalized to full-length 5′UTRs (FL). ** p< 0.01 (t-test compared to pRF). n= 3 experiments performed in triplicates. Data represent mean ± s.d.

The Hoxa9 IRES is essential for protein expression in vivo

To determine whether IRES-dependent translation is critically required for the conversion of the Hoxa9 transcript into protein in vivo, we generated the first targeted mouse knockout of a cellular IRES by removing nt 944–1194 of the Hoxa9 5′UTR, leaving the Kozak sequence and the rest of the 5′UTR intact (Fig. 5a, Extended Data Fig. 8c-d). A caveat to the previously generated HOXA9 targeted knockout is the retention of a Neomycin (Neo) cassette[35,36], which can have indirect effects on the expression of neighboring Hox genes[37]. Indeed, our current genetic studies reveal that the presence of the Neo cassette within the Hoxa9 targeting locus is sufficient to produce phenotypes previously attributed to Hoxa9 loss of function (Extended Data Fig. 9). Therefore, in our Hoxa9ΔΔ mouse, we were careful to remove the Neo cassette from the targeted locus by Cre-mediated excision.

Figure 5

The Hoxa9 IRES is required for axial skeleton patterning but not for Hoxa9 mRNA expression

a, Scheme of Hoxa9 IRES targeted deletion in the mouse. LoxP sites are depicted by yellow triangles. The region deleted is highlighted in orange. b, Representative in situ hybridization of Hoxa9 mRNA expression at E11.5. n = 3 embryos of each genotype. Arrows indicate the anterior boundaries of expression in the neural tube and somites. c, qPCR shows little change in Hoxa9 transcript levels in neural tube and somites of Hoxa9Δ and Hoxa9ΔΔ embryos. Hoxa9Δ p = 0.06; Hoxa9ΔΔ p = 0.03, t-test as compared to Hoxa9+/+. n ≥ 4 embryos of each genotype. d, Representative skeletons of Hoxa9+/+, Hoxa9Δ and Hoxa9ΔΔ mice. Arrows point to the missing rib(s) on T13, normally present in WT mice (arrowheads) revealing a posterior homeotic transformation to L1. n = 5 mice of each genotype.

Extended Data Figure 9

The presence of Neo cassette in the Hoxa9 locus is linked to the presence of an L1 →T13 homeotic transformation

a, Diagram of the Hoxa9 locus (top) and axial skeleton phenotype (bottom) in different Hoxa9 mouse mutants. The original Hoxa9 was made by replacing the homeodomain with a Neo cassette. Vertebra with homeotic transformation is colored red. b, Representative skeletons of Hoxa9+/+, Hoxa9 and Hoxa9. Arrows point to the additional rib(s) on L1, revealing a homeotic transformation to T13. These results show that it is the presence of Neo in the targeting locus, which may affect the expression of neighboring Hox gene, that is sufficient to cause the L1 → T13 phenotype. When the Neo cassette is removed from the targeting locus by crossing the Hoxa9 mouse with a CMV Cre line, the L1 → T13 phenotype is no longer present. n = 3 embryos of each genotype.

Hoxa9ΔΔ embryos do not show any major change in Hoxa9 transcript boundaries or expression levels (Fig. 5b-c). Strikingly, however, these mice show a fully penetrant homeotic transformation of the thirteenth thoracic vertebra (T13) to the first lumbar vertebra (L1), where a full set of ribs normally present on T13 is missing (Fig. 5d). Although we cannot at present ascertain whether the Hoxa9 IRES deletion completely phenocopies a Hoxa9 knock-out mouse, this dramatic homeotic transformation is precisely at the anterior boundary of Hoxa9 expression at the thoracic to lumbar transition[35]. This transformation is also evident in ~50% of Hoxa9Δ mice, although small rudimentary ribs can still form on T13. Consistently, there is a dramatic decrease in HOXA9 protein levels in the neural tube and somites of Hoxa9ΔΔ embryos from the thoracic to sacral level (Fig. 6a) revealing a block in HOXA9 protein production. Furthermore, in Hoxa9ΔΔ embryos there is a dramatic decrease in ribosome association of the Hoxa9 transcript. In fact, the majority of the Hoxa9 transcript in Hoxa9ΔΔ embryos, but not control transcripts such as β-actin, accumulates in pre-polysomal fractions, reflecting an accumulation of mRNA not bound by translationally active ribosomal subunits (Fig. 6b, Extended Data Fig. 10). These findings reveal that the Hoxa9 IRES element is crucial in recruiting the ribosome to this transcript to pattern the mammalian body plan in vivo.

Figure 6

The Hoxa9 IRES is critically required for HOXA9 translation in vivo

a, Representative immunostaining of HOXA9 (green) in cross-section of E11.5 Hoxa9 (top) and Hoxa9ΔΔ (bottom) embryos from the thoracic to sacral levels. DAPI staining is in blue. NT: neural tube, som: somite. n = 3 embryos of each genotype. b, Somites and neural tubes of E11.5 Hoxa9+/+ and Hoxa9ΔΔ embryos were microdissected and fractionated on a sucrose gradient. Middle: qPCR of Hoxa9 mRNA from each fraction. Early pre-polysome fractions are highlighted. Right: quantification of Hoxa9 mRNA in fractions. Fractions 1–8 are pre-polysome and 9–16 polysome fractions. **p< 0.01 (t-test compared to Hoxa9+/+). n = 3 embryos of each genotype. c, Model for how ribosome-mediated regulation of gene expression is encoded within mRNA sequence. Stress-responsive cellular IRES elements are only active when cap-dependent translation is down-regulated through a decrease in eIF4F activity. HoxA mRNAs possess a TIE, which normally inhibits cap-dependent translation. The TIE enables ribosome-mediated control of HoxA IRES expression during embryonic development.

Extended Data Figure 10

Sucrose gradient fractionation shows no difference in β-actin association with polysomes in Hoxa9+/+ and Hoxa9ΔΔ embryos

a, Overlay of A260 trace during fractionation showing no difference in polysome profiles between E11.5 Hoxa9+/+ and Hoxa9ΔΔ embryos. b, qPCR from each fraction reveals no difference in B-actin mRNA accumulation between Hoxa9+/+ and Hoxa9ΔΔ embryos. c, Quantification of B-actin mRNA in fractions. Fractions 1-8 are pre-polysomes and 9-16 are polysome fractions. n = 3 embryos of each genotype.

Discussion

Our studies reveal how expression information encoded within the mRNA template is sufficient to direct ribosome-mediated control of protein production. In particular, our findings suggest that IRES elements facilitate important, ribosome-mediated control of gene expression in development and are therefore key regulatory elements within the mammalian genome (Fig. 6c). Cellular IRES elements have been controversial because of their weak IRES activity as compared to viral IRES elements under non-stress conditions. Their relevance to normal gene regulation is also poorly understood as all eukaryotic mRNAs are capped and undergo cap-dependent translation[21,38]. Importantly, in our studies, the requirement for Hox IRES elements is explained by our identification of the TIE within these transcripts, which strongly inhibits general, cap-dependent translation (Fig. 6c). We speculate that such IRES elements may be more widespread in the genome and suggest a two-state mechanism for gene regulation, controlled by transcription and an additional layer of cis-acting RNA regulons within 5′UTRs that direct protein production. Both IRES and TIE elements may be differentially regulated, providing a versatile toolbox for controlling the ultimate expression of transcripts. Interestingly, while Hox IRES elements are highly conserved, and may already be present in Drosophila orthologs[39], the TIE lacks conservation and is absent even in vertebrate species such as fish and amphibians (Fig. 1f). These species often lay their eggs in harsh environments[40] and may have originally employed IRES elements as a fail-safe mechanism to ensure proper expression of key transcripts under stress conditions. Therefore, the acquisition of a TIE later in evolution may have enabled an additional, independent level of regulatory control to IRES-dependent translation and gene expression during normal development. A mechanism that regulates whether a pre-existing transcript is converted into a protein offers a fast and dynamic means to provide greater variation to gene expression.

It will be interesting to determine if additional RPs may promote specialized translation through control of unique subsets of IRES-containing mRNAs, either directly or through RNA-binding proteins. For example, RPS25 is required for IRES-dependent translation of certain viral IRES mRNAs[41]. Moreover, rRNA modifications both at the level of pseudouridylation, and RPL13a-dependent methylation also appear to regulate the translation of certain cellular IRES containing mRNAs[14,16,42,43].We therefore speculate that similar to the complex and highly regulated system of transcriptional control, where specific DNA sequences and histone marks regulate gene expression, cis-acting RNA regulons in conjunction with more specialized ribosome activity provide newfound regulatory control to gene expression critical for mammalian development.

Methods

Plasmids

Hox 5′UTRs were amplified from mouse E10.5–12.5 cDNA and cloned into the bicistronic pRF vector[14] using EcoRI and NcoI sites. For monocistronic assays, Hox 5′UTRs are cloned into pGL3 control vector (Promega) using HindIII and NcoI sites. 5′UTRs of Hoxa3 and Hoxa10 were made by gene synthesis (IDT). pRF-HCV was a kind gift from D. Ruggero (UCSF). Zebrafish Hoxa9b 5′UTR was amplified from zebrafish 24 hpf cDNA. Sequences were based off the Ensembl database[44]. See Supp. Table 1 for a list of PCR primers used.

To ensure no ribosome readthrough of the first cistron in bicistronic assays, a hairpin with a linker (GAATTCAGATCTGGTACCGAGCTCCCCGGGCTGCAGGATATCCTGCAGCC CGGGGACCTCGGTACCAGATCTAGGATCCGGATCCGGCGACACTCCACCA TGAATCACTCCCCTGTGAGGAACTACTGTCTTCACGCAGAAAGCGTCTAGC CATGG) was synthesized and ligated into the pRF vector at the EcoRI and NcoI sites. Hox 5′UTRs and HCV IRES were cloned into this new vector pRF-hpC in the NcoI site. Clones were sequenced to ensure directionality.

5′ RACE

RNA was purified from dissected E11.5 neural tubes and somites using TRIzol (Invitrogen). The reverse transcription template was produced using ExactSTART Eukaryotic mRNA 5′-& 3′-RACE Kit (Epicentre) and reverse transcribed using Superscript III (Invitrogen) with a primer in the coding region of Hoxa4. The 5′ UTR fragment was amplified using Phusion High-Fidelity DNA polymerase (Thermo Scientific) with a gene specific reverse primer and a forward primer at the 5′ linker. PCR products were gel-extracted and sub-cloned into TOPO TA cloning vector (Invitrogen) for sequencing.

TALEN

2 sets of TALEN pairs were designed against the second and third exons of Rpl38. TALENs were generated towards the following sequences: F1 5′-GCGGACGCCAGAGCCGA, R1 5′-CCCTCGTCCCGCACTCA, F2 5′-CATCTTCTGTCCCTC, R2 5′-TGATCTCCTCAATTTTC. Intermediate arrays were produced for each TALEN-pair and assembled into pC-GoldyTALEN[45]. The 2 sets of TALEN pairs were transfected together into C3H10T1/2 cells. Cells were trypsinized into single cell suspension and plated for clonal expansion 2 days after transfection. Positive clones were identified by PCR followed by restriction digest. Rpl38 disruption was confirmed by Sanger sequencing and Western blotting.

Luciferase assay

C3H10T1/2 cells were transfected in 24-well plates with various pRF constructs using Lipofectamine 2000 (Invitrogen) as per manufacturer instructions. Cells were harvested 24h post transfection and assayed using Dual luciferase kit (Promega). IRES activity is expressed as a ratio between Fluc and Rluc. To verify the integrity of the bicistronic construct, an shRNA plasmid against Rluc or scrambled plasmid[16] was cotransfected with the pRF constructs. Cells were harvested 72h post transfection for luciferase assays. For monocistronic luciferase assays, various pGL3 constructs were transfected as above into 12 well plates. Cells were harvested 24 h later. 50% of the cells were used for qPCR analysis with Fluc primers to quantify total transcript abundance, the rest for luciferase assays. Fluc activity was normalized to total luciferase mRNA expression levels.

RNA pull down

Template DNA was made by PCR from pRF-HoxA9 using a 5′ primer containing the T7 promoter and a 3′ primer containing a linker sequence T2 (Supp. Table 1). Biotinylated RNA were in vitro transcribed with Megascript T7 kit (Ambion) using a final concentration of 6.8 mM each of ATP, CTP and GTP, 6.5 mM UTP and 0.3 mM Biotin-16-UTP (Roche). Synthesized RNAs were treated with TURBO DNase (Ambion) and purified by phenol/chloroform extraction. Unincorporated nucleotides were removed by running the RNA through a G50 column (GE Healthcare). The RNA was analyzed by electrophoresis for quality and Nanodrop for quantity.

A total of four 15-cm dishes of subconfluent C3H10T1/2 cells were harvested by trypsinization. Cell pellets were washed and lysed on ice with 1.5 ml lysis buffer (10 mM Tris pH7.5, 10 mM MgCl2, 150 mM NaCl, 0.1% TritonX, 0.25% NP40, 10 mM DTT, RNaseOUT, Halt Protease inhibitor) for 30 min. Lysate was centrifuged at 14000 rpm for 10 min, 4°C. Supernatant was transferred to new tube and cleared with 400 μl avidin-agarose beads (Thermo) for 50 min, 4°C. Cleared lysate was transferred to new tube and quantified by spectrometry.

Folding buffer was added to 100 pmol of biotinylated RNA of each species to a final concentration of 100 mM Tris pH8, 100 mM NaCl, 10mM MgCl2. RNA was heated to 95 °C for 1.5 min, ice for 2 min and 37 °C for 20 min. 100 pmol of folded RNA and 1 mg cleared lysate were incubated together with 100 μg/ml heparin and 100 μg/ml yeast tRNA in 2 ml lysis buffer for 1 h at RT.

Streptavidin M280 Dynabeads (Invitrogen) were washed according to manufacturer’s instructions. 100 μl beads were incubated with each RNA/ lysate sample for 1.5 h at 4 °C. Beads were then washed with lysis buffer with increasing concentrations of salt up to 600 mM NaCl. One-tenth of beads was kept for RNA and the rest was used for protein analysis. RNA was eluted with 50 μl RNA denaturing buffer (95% formamide, 10 mM EDTA) at 65 °C for 5 min. Eluted RNA was purified by phenol/ chloroform extraction. RNA was converted to cDNA and analyzed by qPCR using primers HoxA9 5UTR F and HoxA9 342 R (Supp. Table 1). Proteins were eluted with 40 μl protein elution buffer (5 mM biotin, 3x SDS sample buffer) at 95 °C for 5 min. Proteins were analyzed by Western blotting.

Western blotting

Western blots were performed with antibodies against RPS6 (Cell Signaling #2217) and RPL10a (Santa Cruz sc-100827). Polyclonal RPL38 antibodies were generated against C-EKAEKLKQSLPPGLAVKDLK.

RNA structural probing and data analysis

Double-stranded DNA templates were prepared by PCR assembly of DNA oligomers with maximum length of 60 nt ordered from IDT (Integrated DNA Technologies). DNA templates for mutate-and-map (M2) and mutation/rescue were designed as previously described using NA_Thermo (https://github.com/DasLab/NA_thermo) and HiTRACE[27,28]. DNA template and RNA transcript preparation and quality check were carried out as previously described[27,28].

All chemical mapping were carried out in 96-well format as described previously[27,28,46]. Briefly, prior to chemical modification, RNA was heated up and cooled to remove secondary structure heterogeneity, and folded for 20 min at 37 °C in 10 mM MgCl2, 50 mM Na-HEPES, pH 8.0. RNA was modified by adding freshly made 5 mg/mL 1-methyl-7-nitroisatoic anhydride (1M7)[47] or anhydrous DMSO as control. Modification reactions were incubated at room temperature and then quenched with 0.5 M Na-MES, pH 6.0. Quenches also included poly(dT) magnetic beads (Ambion) and FAM-labeled Tail2-A20 primer for reverse transcription. Samples were separated and purified using magnetic stands, washed with 70% ethanol twice, and air-dried. Beads were resuspended in ddH2O and reverse transcription mix, then incubated at 55 °C for 30 min. RNAs were degraded by base and heat, then cooled and neutralized with acid quench. Fluorescent labeled cDNA was recovered by magnetic bead separation, rinsed twice with 70% ethanol and air-dried. The beads were resuspended in Hi-Di formamide containing ROX-350 ladder (Applied Biosystems), then loaded on capillary electrophoresis sequencer (ABI3100).

MiSeq-based ligSHAPE (ligation-based SHAPE) on long RNAs was carried out, using a protocol analogous to Mod-seq[48]. RNA was folded similarly and probed by SHAPE as above, then recovered by ethanol precipitation. Modified RNA was folded again and fragmented by 5 mM TbCl3 for 5 min (or no Tb3+ as control)[49], quenched by 0.5 M NaEDTA pH, 8.0, then purified by RNA cleanup and concentration columns (Zymo Research). Fragmented RNA was prepared through 3′ end repair, T4 RNA-ligase-based universal linker ligation to the RNA 3′ end, reverse transcription to cDNA using sequencing primers, and circLigase-based Illumina adapter ligation to the cDNA 3′ end as described[50].

The HiTRACE 2.0 package was used to analyze CE data (available as MATLAB toolbox at https://github.com/hitrace[51] and web server at http://hitrace.org[52]). Electrophoretic traces were aligned and baseline subtracted using linear and non-linear alignment routines as previously described[53]. Band intensities were fitted to Gaussian peaks, then background subtracted, signal attenuation corrected and normalized. MiSeq results were analyzed using MAPseeker package[50,54]. The SHAPE reactivity profile was generated by projection of pair-ended reads. Briefly, a correction factor of 0.25 was used for full-length read ligation bias correction, accounting for the empirically observed low circLigase ligation efficiency for full-length cDNAs ending in CC[55]. Reads shorter than 8 nt were discarded due to ambiguity in their positional alignment in the full-length RNA. Modification fractions were determined by the number of cDNAs in which reverse transcription stopped at a given nucleotide divided by the number of cDNAs in which reverse transcription stopped or proceeded beyond that nucleotide[55]. Traces were background subtracted (SHAPE minus nomod, and SHAPE-Tb minus Tb) and normalized by boxplot[56] for Hoxa5, or using a scanning window of 350 nt for Hoxa9. Errors were calculated between SHAPE/nomod and SHAPE-Tb/Tb measurements.

Data-driven secondary structure models were obtained using the Fold executable of the RNAstructure package[57]. Pseudo-energy parameters of RNAstructure version 5.4 were used by default. To obtain 2D-data-guided secondary structure models, Z score matrices for mutate-and-map datasets were calculated and used as base-pair-wise pseudo-energies with a slope and intercept of 1.0 kcal/mol and 0 kcal/mol as previously described[27]. The ShapeKnots executable of RNAstructure version 5.6[58] enables modeling of pseudoknots but with different parameters for utilizing SHAPE data in the modeling; this newer method was used for pseudoknot prediction of Hoxa9 957–1132 when using 1D and 2D SHAPE data combined. Extension of ShapeKnots to accept 2D data is being incorporated into the next version of RNAstructure. Helix-wise confidence values were calculated via bootstrapping as described previously[27,28]. All predictions were modeled in context with full-length sequences. Secondary structures were visualized by VARNA[59].

Mice

A 2.8-kb 5′arm and a 4.9-kb 3′arm of Hoxa9 were cloned into PGKneolox2DTA.2 (P. Soriano, Addgene plasmid 13449). This results in a replacement of nt 945–1194 of Hoxa9 5′UTR with a LoxP-flanked Neo cassette in the same transcriptional direction as Hoxa9 (Extended Data Fig. 8c). The plasmid was linearized with XhoI and electroporated into JM8NA3.N1 ESCs for homologous recombination. The correctly targeted ESC clones were identified by long range PCR and confirmed by Southern blotting. The clones were injected into C57BL/6 blastocysts. Resulting chimeras were crossed with C57BL/6 mice to generate F1 (Hoxa9). Hoxa9 were crossed with CMV-Cre mice (Jackson Laboratory) to precisely remove the LoxP-flanked Neo cassette within the 5′UTR to generate HoxA9 mice. All mice used for experiments were between 2 and 6 months old. All animal studies were performed in accordance to with Stanford University Animal Care and Use guidelines.

Sucrose gradient fractionation

E11.5 neural tubes and somites were dissected and trypsinized as previously described[1]. The neural tube and somites of each embryo were lysed on ice in 200μl of lysis buffer (20 mM Tris pH 7.5, 200 mM NaCl, 15 mM MgCl2, 1 mM DTT, 8% Glycerol, 1% Triton X-100, 100 μg/ml cycloheximide, 100 μg/ml heparin, 200U/ml SUPERase In (Ambion)). After lysis, nuclei and membrane debris was removed by centrifugation (1,300 g, 5 min at 4 °C, and then 14,000 rpm, 5 min, at 4°C). The supernatant was layered onto a linear sucrose gradient (10–40% sucrose[w/v], 20 mM Tris, pH 7.5, 100 mM NaCl, 15 mM MgCl2, 100 μg/ml cycloheximide) and centrifuge in a SW41Ti rotor (Beckman) for 2.5 h at 40,000 rpm at 4°C. Fractions were collected by Density Gradient Fraction System (Brandel).

qPCR analysis

For cells, RNA was isolated using NucleoSpin RNA kit (Clontech) and DNase-treated on column. For embryos, E11.5 neural tubes and somites were dissected and trypsinized as previously described[1]. RNA was isolated using Trizol (Invitrogen) and treated with TURBO DNA-free kit (Ambion). For sucrose gradient fractions, RNA was purified by acid phenol/chloroform followed by isopropanol precipitation. 0.5 μg of RNA was converted to cDNA using iScript supermix (BioRad). cDNA was diluted 2-fold and 1 μl used to run a SYBR green detection qPCR assay (SsoAdvanced SYBR Green supermix and CFX384, BioRad). Data was analyzed and converted to relative RNA quantity using CFX manager (BioRad). For sucrose gradient fractions, amount of RNA from individual fractions was expressed as a fraction of the total RNA collected from all fractions. Primers were used at 250 nM per reaction. B-actin and Hoxa9 qPCR primers were previously described[1]. Fluc F 5′-CATCACGTACGCGGAATACTT, Fluc R 5′-AAGAGATACGCCCTGGTTC. 18S F 5′ ACATCCAAGGAAGGCAGCAG, 18S R 5′ CATTCCAATTACAGGGCCTC. 28S F 5′ GGGGAGAGGGTGTAAATCTC, 28S R 5′ TCCTTATCCCGAAGTTACGG.

Whole mount in situ hybridization

Whole mount in situ hybridization was carried out using digoxygenin-labeled antisense RNA probes as previously described[1]. Hoxa9 probe was amplified from mouse embryo cDNA. Primers sequences are described in Supp. Table 1.

Skeletal staining

Alcian blue and alizarin red staining of cartilage and bone were performed on neonates as previously described[1].

Immunofluorescence

E11.5 embryos were dissected, fixed and cryosectioned as previously described[1]. WT and Hoxa9ΔΔ littermates sections were placed onto the same slide for direct comparison and identical processing. Sections were equilibrated to RT and rehydrated in blocking buffer (1% heat-inactivated goat serum, 0.1% Triton X-100 in PBS) for 1 h at RT. Sections were incubated with anti-HOXA9 antibody (Abcam ab140631) at 1:1000 in blocking buffer overnight at 4 °C, washed three times, 30 min each and incubated with Alexa 488 anti-rabbit secondary antibodies (Invitrogen) for 1 h at RT. Slides were washed with blocking buffer three times, 30 min each and incubated with 1 ug/ml DAPI (Cell Signaling) for 15 min at RT. They were washed again in blocking buffer for 20 min and mounted with VectaShield Mounting Media (Vector Laboratories).

Statistical analysis

No normalization, blinding or exclusion was used in any experiments. In all cases, multiple independent experiments were performed on different days to verify the reproducibility of experimental findings. For mice experiments, embryos from multiple litters were used to avoid litter-specific bias. Each variable was analyzed using the two-tailed unpaired Student’s t test. For all analyses, a p value of less than 0.05 was considered significant. Results are given as means ± SD.