Pervasive functional translation of non-canonical human open reading frames

Abstract

Ribosome profiling has revealed pervasive but largely uncharacterized translation outside of canonical coding sequences (CDSs). Here, we exploit a systematic CRISPR-based screening strategy to identify hundreds of non-canonical CDSs that are essential for cellular growth and whose disruption elicit specific, robust transcriptomic and phenotypic changes in human cells. Functional characterization of the encoded microproteins reveals distinct cellular localizations, specific protein binding partners, and hundreds that are presented by the HLA system. Interestingly, we find multiple microproteins encoded in upstream open reading frames, which form stable complexes with the main, canonical protein encoded on the same mRNA, thus revealing the diverse use of functional bicistronic operons in mammals. Together, our results point to a family of functional human microproteins that play critical and diverse cellular roles.

Sentence Summary

Systematic interrogation of unannotated open reading frames reveals extensive translation of novel microproteins with important functional roles, expanding our understanding of the proteome.

Main Text

Efforts to bioinformatically discover and annotate protein-coding open reading frames (ORFs) in genomes, termed coding sequences (CDSs), have traditionally relied on rules such as amino acid conservation and homology, translation initiation from an AUG start codon, and minimum length (e.g. 100 amino acids)(1). The widespread adoption of these rules has been based on the assumption that short peptides are unlikely to fold into stable structures to perform functions. However, the generality of these rules has been challenged. For example, the ribosomal protein RPL41 is a 25 amino acid (aa) peptide; Sarcolipin (SLN, 31 aa) and Phospholamban (PLN, 52 aa) bind to and regulate the sarcoplasmic Ca²⁺ transport ATPase SERCA(2, 3). In addition, MYC can be translated from a non-canonical start codon CUG(4), demonstrating that non-AUG initiation can produce functional proteins. Recent studies have added a handful of remarkable examples of short proteins, or “microproteins” (also called “micropeptides” or just peptides), performing diverse functions(5–18), some encoded on transcripts annotated as long non-coding RNAs (lncRNAs). Finally, upstream ORFs (uORFs), located in the 5’ untranslated regions (UTR) of mRNAs, have long been implicated in cis-acting translational control of the main, canonical CDS(19–21), though it has remained unclear whether they can generate stable, functional peptides.

Systematic identification of functional short CDSs remains challenging. Recent ribosome profiling (deep sequencing of ribosome protected fragments) and mass spectrometry (MS) studies have identified thousands of previously unannotated CDSs(22–25) across bacteria, yeast, viruses, and mammalian cells. However, for most cases, the cellular function of these identified CDSs or their peptide products remains unexplored. We reasoned that the advent of CRISPR and its ability to precisely disrupt protein-coding regions(26), when combined with ribosome profiling, provides a unique opportunity to define and characterize empirically the functional protein-coding capacity of a given genome. Here, we applied a diversity of approaches, including ribosome profiling, mass spectrometry, and multiple CRISPR-based techniques, to systematically discover non-canonical CDSs encoded in the human genome and validate their critical roles in diverse cellular pathways.

To annotate potential CDSs comprehensively and accurately, we first investigated genome-wide translation by ribosome profiling across multiple cell types and conditions, including human induced pluripotent stem cells (iPSCs), iPSC-derived cardiomyocytes, human foreskin fibroblasts (HFFs), and HFFs infected with cytomegalovirus(27, 28) (fig. S1A). We leveraged the ORF-RATER algorithm to annotate ORFs(27), incorporating multiple lines of evidence to identify ORFs undergoing active translation. This included consideration of accumulation of ribosome densities at the start and stop codons, three-nucleotide periodicity, and additional experimental results such as data from harringtonine-treated cells, in which ribosomes are stalled at initiation sites(27). In iPSCs and cardiomyocytes, in addition to 9490 annotated CDSs (62% of the identified CDSs), we identified 3455 distinct, non-canonical CDSs (22%, i.e. with no in-frame overlap with previously annotated CDSs) and 2466 variant CDSs of annotated proteins (16%) in our high statistical confidence set (Fig. 1A, see Methods and (27)). Among the distinct CDSs, 818 were CDSs on transcripts lacking prior protein-coding annotations (“new”, i.e. lncRNAs), 2342 were upstream CDSs (i.e. uORFs or “start overlaps”, CDSs that overlap annotated start codons in a different reading frame), and only 13 were downstream CDSs. Similar numbers of CDSs were present in HFFs (fig. S1B), with 75% of the CDSs shared between the two cell types. 96% of the distinct CDSs are less than 100 aa in length, and 36% of the CDSs use non-AUG start codons (Fig. 1B, C, and see fig. S2 for further characterizations).

Fig. 1. Ribosome profiling reveals translation of unannotated CDSs.

(A) ORF-RATER analysis of ribosome profiling data: 62% are previously annotated coding sequences, while 16% are variants of canonical coding sequences that share portions of the coding sequence, and 22% are distinct from annotated coding sequences. The naming convention of the identified ORFs is shown on the right. (B) Start-codon usage of the identified CDSs. (C) Cumulative distribution of CDS length. For distinct CDSs, 96% are smaller than 100 amino acids. (D) Example ribosome profiling traces of a lncRNA peptide from LINC00998 and a uORF peptide from ARL5A displaying the hallmarks of translation, including peaks of density around the start codon following harringtonine treatment and three nucleotide periodicities along the coding region. (E) Metagene analysis shows that the signatures of translation, including three-nucleotide periodicity in the expected reading frame, for uORFs and lncRNA CDSs are similar to annotated coding regions. (F) Identification of more than 200 non-canonical CDS peptides from HLA-I peptidomics, cross-validating their existence across the whole abundance range, with a mean Andromeda score of 141 compared to a total mean Andromeda score of 144. See Methods.

Multiple lines of evidence suggested that the non-canonical CDSs are actively translated. The average ribosome density (metagene) of the lncRNA CDSs and of the translated uORFs closely mirrors footprints from that of annotated coding regions with strong 3-nucleotide periodicities, a hallmark of active translation, as exemplified by traces from the lncRNA LINC00998 transcript and a uORF of ARL5A (Fig. 1D, E and see fig. S3). Our analysis also successfully recapitulated well-characterized short ORFs, such as the uORF on ATF4(29), and the recently discovered lncRNA-encoded microproteins MOXI/mitoregulin(11, 12) and NoBody(10). Bona fide lncRNAs such as XIST, HOTAIR, and NEAT1 were not identified to be protein coding (fig. S3E). Moreover, many of the CDSs were differentially translated during iPSC differentiation or viral infection (fig. S3F), providing evidence for translational control in different cell states.

MS-based proteomics in iPSCs and major Human Leukocyte Antigen Class I (HLA-I) peptidomics confirmed the stable expression of hundreds of non-canonical CDS peptides (Fig. 1F, fig. S4–5). HLA-I peptidomics identified 240 non-canonical peptides, suggesting that these peptides enter the HLA-I presentation pathway and contribute to the antigen repertoire and possible immunogenicity (Fig. 1F)(30). HLA class I prediction analysis cross-validated strong binding (Kd ≤ 50 mM) of non-canonical CDS HLA-I peptides to their respective allotype (fig. S6)(30). MS-based proteomics based on tryptic digestion identified far fewer non-canonical peptides, which may be due to challenges in detecting the trypsin-digested products from short, non-canonical CDSs, or possibly due to more rapid turnover of these non-canonical peptides (fig. S7).

To test whether translation of the non-canonical CDSs are important for cell growth and potentially yield functional peptides, we measured the growth phenotypes resulting from CRISPR-mediated ORF knockout in pooled screens(26). We designed a Cas9 ORF single-guide RNA (sgRNA) library to specifically knock out thousands of the non-canonical CDSs identified by ribosome profiling (Fig. 2A, see Methods)(31, 32), targeting 1098 uORFs, 613 lncRNA CDSs, 352 extensions of annotated coding regions, 283 “start overlaps”, and 7 downstream CDSs. We performed pooled Cas9 knockout screens in iPSC and K562 chronic myeloid leukemia cells expressing Cas9 and the sgRNA library, akin to conventional pooled screens for essential proteins(26, 31). We measured sgRNA abundance in the cell populations shortly after library transduction and after 10 additional population doublings by deep sequencing to quantify the fitness defect conferred by each sgRNA. We then calculated a “phenotype” score (γ) and confidence (P-value) for each ORF from the relative enrichment or depletion of sgRNAs targeting a particular ORF (Fig. 2B and Methods). In iPSCs, our screen identified over 500 ORF knockout “hits” that result in statistically significant phenotypes. The hits include 169 genes that are variants of annotated proteins, 78 “start overlap” hits, 230 uORF hits, 91 lncRNA CDS hits, and 2 downstream CDS hits. iPSC and K562 cells had 401 shared hits, suggesting housekeeping or general cellular roles, as well as CDSs that may play cell-specific functions (fig. S8). A fraction of the uORF hits do not have main, canonical CDSs with fitness defects upon knockout, suggesting an independent function of the uORFs or that disruption of the uORFs result in increases in main CDS expression that results in the growth phenotype (fig. S8E). Thus, unannotated CDSs with important functions across multiple cell types are an abundant feature of the genome.

Fig. 2. Genome-scale CRISPR screens to identify functional, non-canonical CDSs.

(A) Schematic of CRISPR library design and screening strategies, either by growth screens or Perturb-Seq. For growth screens, frequencies of cells expressing a given sgRNA are determined by next-generation sequencing, and phenotype scores are quantified with the formula shown. For Perturb-Seq, single-cell transcriptomes and sgRNA identities were obtained by single-cell RNA-Seq. (B) Volcano plot summarizing knockout phenotypes and statistical significance (Mann-Whitney U test) for ORFs targeted in the pooled screen in iPSCs. Each dot represents a targeted ORF, and ORF hits are labeled in purple, with a more negative phenotype score indicating a stronger growth defect. See Methods. (C) Plot of the sgRNA phenotypes and distance from the start codon, across all ORF hits. sgRNAs targeting the genome immediately upstream of the ORF (shown in red) have significantly lower phenotype than sgRNAs targeting within the ORF (shown in blue). Note the axis is increasingly negative (stronger) phenotype. The difference is not due to differences in sgRNA on-target efficiencies, as quantified by the Doench v2 score. (D) The PhyloCSF Score per codon (higher is more conserved across the Euarchontoglires) is generally higher for ORF hits (P = 10⁻²⁰, Kolmogorov–Smirnov test) and ORFs with a stronger phenotype. Note that lack of a growth phenotype does not necessarily imply a low PhyloCSF score.

Several lines of evidence further suggested that our screen reports specifically on the phenotypes of the selected ORFs. First, the phenotypes of control sgRNAs targeted directly upstream of each ORF in the genome (Fig. 2C) were significantly weaker than those of sgRNAs targeting within the ORF (P = 10⁻²⁶, Mann-Whitney test). Second, sgRNA phenotypes are independent of distance to other annotated proteins, splice sites, or transcriptional start sites (fig. S9A). Functionally, ORF hits are on average more phylogenetically conserved, with a higher conservation score than non-hits (PhyloCSF score per codon, P = 10⁻²⁰, Kolmogorov–Smirnov test, Fig. 2D)(33), and have other distinguishing sequence features (e.g. enrichment for Kozak consensus sequence) (fig. S10). Though, interestingly, the non-canonical CDSs on average have lower PhyloCSF scores compared to canonical proteins (fig. S2B). Finally, sgRNAs targeting ORF hits versus non-hits have indistinguishable off-target and on-target scores (fig. S9B)(32). We then performed validation follow-ups with individual sgRNAs, which recapitulated the growth phenotypes from our genome-scale screen (fig. S8D). Sequencing of the targeted genomic regions revealed indels of <50 bp (fig. S9C, D). Together, these analyses independently supported the conclusion that our screen phenotypes result specifically from the disruption of the target ORFs.

To survey function of the non-canonical CDSs at scale, we combined CRISPR screening with single-cell RNA sequencing (Perturb-seq) (34, 35). Disruptions of the various non-canonical CDSs resulted in broad and diverse changes in RNA-seq profiles across a variety of critical pathways, suggesting that the candidate CDSs play diverse cellular roles (fig. S11). As an example, disruption of the CDS on LINC00998 resulted in differentially expressed genes related to glycosylation (P-value < 10⁻¹⁰), suggesting a function at the Golgi or ER (see below). Importantly, the transcriptional phenotype also allowed us to functionally profile CDSs that are not essential for robust growth (fig. S11C). Furthermore, we found that CRISPR-targeted transcripts did not show detectable changes in abundance that might result from processes such as nonsense-mediated decay, indicating that the phenotypes we observed were not due to decreasing the abundance of the entire transcript (fig. S11D). Thus, similar to screens for essential protein-coding genes(26, 31), our screen for non-canonical CDSs required for robust cell growth underestimated the true number of functional CDSs in the genome, further underscoring the pervasiveness of functional, unannotated CDSs in the genome affecting a wide range of cellular activities.

We next explored the functional role of the peptides encoded by the non-canonical CDSs identified from our screen, first focusing on lncRNA CDSs. For seven lncRNAs, we ectopically expressed the transcript encoding for the peptide and found in all cases knockout-induced growth defect was partially or completely rescued. This rescue was abrogated by the removal of the initiating start codon (Δstart codon) (Fig. 3A), suggesting an essential role of the peptide itself in cell growth. To further interrogate the specific functions of the non-canonical microproteins, we adopted a split-fluorescent protein (FP) approach using mNeonGreen (mNG), in which we fused each peptide with a minimally disruptive 16 aa tag (mNG11). Co-expression of the tagged peptide with the remainder of the mNG protein (mNG1–10) results in a fluorescence signal upon complementation (36, 37), creating both a fluorescent reporter to detect stable expression and cellular localization, and a handle for co-immunoprecipitation (co-IP) and MS to define interaction partners(36) (fig. S12A). We probed the functions of six essential lncRNA CDSs and found five to form specific complexes that were consistent with their subcellular localization. For example, the 62 aa peptide encoded by lncRNA RP11_469A15.2 specifically localized to the mitochondria. The peptide has a predicted transmembrane domain and co-immunoprecipitates with the cytochrome c oxidase (COX) complex and the mitochondrial Prohibitin complex (Fig. 3B). Moreover, the 70 aa peptide encoded by RP11–84A1.3 localizes to the plasma membrane and interacts with various cell surface proteins (Fig. 3C). Thirdly, the 59 aa peptide encoded by lncRNA LINC00998, which contains two predicted transmembrane domains, localizes specifically to both the ER and Golgi, and co-immunoprecipitates with lysosomal and vesicular transport proteins (Fig. 3D). Finally, the 55 aa peptide encoded on TOPORS-AS1 and the 124 aa peptide on RP11–132A1.4 also form functional complexes consistent with their cellular localization (fig. S12C, D and fig. S13). Consistent with prior studies(5–18), these examples demonstrated that lncRNAs can encode uncharacterized proteins, and highlight the need to fully extend the annotation of lncRNAs and the proteome.

Fig. 3. Short lncRNA CDSs encode functional microproteins.

(A) Rescue of lncRNA CDS knockout growth phenotypes by the ectopic expression of the transcript encoding the peptide, as well as controls where the initiating start codon is removed (Δstart codon). Error bars represent standard deviation of triplicates. P < 0.05 for all comparisons between knockout (KO) and KO + rescue. (B-D) Microscopy images and volcano plots of the co-IP MS of three example lncRNA-encoded microproteins tagged with mNG11, expressed ectopically (in the native transcript context) in a HEK293T cell line expressing mNG1–10. Green is mNG, red is the indicated organelle localization, and blue is Hoechst 33342, which stains for the nucleus. Scale bar dimensions are labeled. Significant interactors are shown in the top, right corner of the volcano plots. Thick threshold line is 1% FDR (false discovery rate), and the thin threshold line is 5% FDR. The bait (the tagged peptide) is labeled in blue. The interactors are colored according to their functional groups. (E) lncRNA-encoded microproteins are uncharacterized proteins that may play important regulatory roles in cells.

We next explored the functional effects of uORF translation, which is complicated by the fact that phenotypes can in principle be mediated by the peptide product (24, 38–41) or the impact of uORF translation on expression of the main, canonical CDS(20), or both. To distinguish between these possibilities, we first separately tagged the uORF and the main CDS, and used Western blot to confirm the independent expression of uORF peptides from the canonical protein (fig. S14). Furthermore, we established that ectopic expression of a transcript encoding only the uORF peptide was able to at least partially rescue the growth phenotype caused by disruption of the endogenous uORF. Critically, in all cases this rescue is dependent on the initiating start codon in the ectopically expressed message, demonstrating that the rescue is due to production of the expressed peptide (Fig. 4A). Consistent with this, in all cases we tested, deleting the start codon for the uORFs only minimally increased (around 20% - 60%) the expression of the main CDS, suggesting that the growth defect seen is mediated by the peptide and not increased expression of the canonical protein (fig. S14E, F). Taken together, these findings established that uORFs could function through the peptide they produce independently of any cis regulatory effects.

Fig. 4. Bicistronic mRNAs can encode uORF peptides that function in trans.

(A) Rescue of uORF knockout growth phenotypes by the ectopic expression of a transcript encoding the uORF peptide alone, as well as a controls where the initiating start codon is removed (Δstart codon). Error bars represent standard deviation of triplicates. P < 0.05 for all comparisons between KO and KO + rescue. (B) Summary of co-IP MS interactions, showing five uORF peptides that interact with their downstream-encoded protein (shown in red). Other significant interactors are shown in blue. (C-E) Examples of uORF peptides tagged with mNG11, expressed alone ectopically (in the native transcript context) in a HEK293T cell line expressing mNG1–10. Volcano plot of co-IP MS reveals significant interactors with uORF peptides. Threshold line is 1% FDR. The bait (the tagged peptide) is labeled in blue. For microscopy in C and D, the main, canonical protein tagged with mCherry (red) is co-expressed. For E, the mNG11-tagged MIEF1 uORF peptide (green) localizes to the mitochondria (red). (F) Volcano plot of co-IP MS from endogenously mNG11-tagged HAUS6 uORF. For microscopy, the mNG11-tagged uORF is expressed alone ectopically (green), and the canonical HAUS6 tagged with mCherry (red) is co-expressed. (G) Percent change for each cell cycle state for HAUS6 knockout (KO) and HAUS6 uORF KO, compared to control cells. (H) Transcriptome response of the MIEF1 uORF KO compared with the main CDS KO from Perturb-Seq. (I) Quantification of mitochondria morphology upon MIEF1 uORF peptide overexpression and knockout, as well as rescue of knockout phenotype. Representative microscopy images of the different mitochondria morphologies are shown to the right. (J) Possible model of uORF peptide functions and regulatory roles in cells.

To explore the functions of uORF-encoded microproteins, we examined their localization and protein binding partners by tagging the uORF peptides with mNG11. Out of the 10 uORF peptides further tested by co-IP MS, we failed to detect statistically significant interaction partners for three of the tagged peptides. Two peptides, encoded by the uORFs of TBPL1 and ARL5A, localize generally to the cytoplasm, while the main CDS proteins exhibit different cellular localization patterns. Consistent with our observed cellular localizations, these two uORF peptides immunoprecipitate specifically proteins with function independent of the main CDS protein (Fig. 4B, C and fig. S12). Thus, these uORF peptides and their main CDS protein have independent functions.

Interestingly, we found that five of the ten uORF peptides co-localized and formed a stable physical complex with the downstream-encoded, canonical protein on their shared mRNA. These include MIEF1, DDIT3, FBXO9, HMGA2, and HAUS6 (Fig. 4B, D, E and fig. S12). In all the cases, we expressed the tagged peptides in their native transcript context but without the downstream CDS, eliminating the possibility of stop codon read-through. We further confirmed this interaction by co-IP of the canonical protein and immunoblotting for the uORF peptide (fig. S12F), as well as with endogenously tagged clonal lines (fig. S15, Fig. 4F). This physical interaction between the proteins encoded by the uORF and the canonical CDS on the same transcript is particularly interesting(39, 42, 43), because it implies an additional layer of regulation beyond the propensity of uORFs to modulate translation of downstream CDSs.

We further explored the function of uORF-expressed microproteins by HAUS6 and MIEF1. In both cases, disrupting the uORF led to minimal increase in the expression of the main CDS protein, and the ectopic expression of a peptide-encoding transcript rescued the knockout-induced growth phenotype (Fig. 4A and fig. S14). mNG11-tagged HAUS6 uORF expressed from its endogenous locus efficiently pulled down key components of the HAUS6 complex, localized to the centrosome, and knockout of the uORF caused cells to arrest in the G1 stage, consistent with the role of HAUS6 microtubule attachment to the kinetochore and central spindle formation (Fig. 4F, G, fig. S12, fig. S15). Similarly, the MIEF1 uORF peptide localized to the mitochondria, consistent with the localization of the MIEF1 protein (Fig. 4E), which regulates mitochondrial fission and fusion (44). The MIEF1 uORF peptide knockout induced differential expression of mitochondrial fusion and fission genes, with a distinct transcriptional signature from that seen in the knockout of the MIEF1 protein (Fig. 4H). We observed that overexpression of the MIEF1 uORF peptide alone induced a fragmented mitochondrial phenotype (increased fission), whereas a clonal knockout of the MIEF1 uORF (with the sequence disrupted but nonetheless preserving an upstream ORF, see fig. S15) resulted in a tubular and more elongated mitochondrial phenotype (increased fusion). Importantly, this knockout morphology could be rescued by the exogenous expression of the MIEF1 uORF peptide (Fig. 4I). Together, our results indicated a possible role of the uORF-encoded peptide in regulating the downstream-encoded protein thereby challenging the monocistronic assumption in mammalian genomes. We speculated that this type of genomic architecture may be general, opening the doors to the investigation of the cooperative and regulatory nature of bicistronic human mRNAs. Indeed, a number of stress-regulated alternate translation initiation factors can modulate translation initiation site choice and uORF usage suggesting that regulation of bicistronic expression could play important roles in both normal biology and diseases states(21, 45).

In summary, we described a strategy that combines ribosome profiling, MS-based proteomics, microscopy and CRISPR-based genetic screens to discover and characterize widespread translation of functional microproteins and define the protein-coding potential of complex genomes. We identified a subset of lncRNAs that can encode stable, functional proteins, suggesting that they may be mis-annotated RNAs or potentially have dual roles at the RNA and protein levels. Furthermore, we provided examples of uORFs encoding functional peptides, highlighting the diverse cellular roles that uORFs may play beyond translational control. Intriguing, we also identified uORF-encoded peptides binding to the downstream-encoded protein on the same mRNA. Thus, our data highlighted an unappreciated complexity to the functional mammalian proteome, as well as the full spectrum of antigens presented by the HLA system.