Abstract
Recent studies in eukaryotes have demonstrated the translation of alternative open reading frames (altORFs) in addition to annotated protein coding sequences (CDSs). We show that a large number of small proteins could in fact be coded by altORFs. The putative alternative proteins translated from altORFs have orthologs in many species and evolutionary patterns indicate that altORFs are particularly constrained in CDSs that evolve slowly. Thousands of predicted alternative proteins are detected in proteomic datasets by reanalysis with a database containing predicted alternative proteins. Protein domains and co-conservation analyses suggest potential functional cooperation or shared function between small and large proteins encoded in the same genes. This is illustrated with specific examples, including altMID51, a 70 amino acid mitochondrial fission-promoting protein encoded in MiD51/Mief1/SMCR7L, a gene encoding an annotated protein promoting mitochondrial fission. Our results suggest that many coding genes code for more than one protein that are often functionally related.
Introduction
Current protein databases are cornerstones of modern biology but are based on a number of assumptions. In particular, a mature mRNA is predicted to contain a single CDS; yet, ribosomes can select more than one translation initiation site (TIS)1–3 on any single mRNA. Also, minimum size limits are imposed on the length of CDSs, resulting in many RNAs being mistakenly classified as non-coding (ncRNAs)4–11. As a result of these assumptions, the size and complexity of most eukaryotic proteomes have probably been greatly underestimated. In particular, few small proteins (defined as of 100 amino acids or less) are annotated in current databases. The absence of annotation of small proteins is a major bottleneck in the study of their function and to a full understanding of cell biology in health and disease. This is further supported by classical and recent examples of small proteins of functional importance, for instance many critical regulatory molecules such as F0 subunits of the F0F1-ATPsynthase16, the sarcoplasmic reticulum calcium ATPase regulator phospholamban17, and the key regulator of iron homeostasis hepcidin18. This limitation also impedes our understanding of the process of origin of genes de novo, which are thought to contribute to evolutionary innovations. Because these genes generally code for small proteins19–22, they are difficult to detect by proteomics or even impossible to detect if they are not included in proteomics databases.
Functional annotation of ORFs encoding small proteins is particularly challenging since an unknown fraction of small ORFs may occur by chance in the transcriptome, generating a significant level of noise13. However, given that many small proteins have important functions and are ultimately one of the most important sources of functional novelty, it is time to address the challenge of their functional annotations13.
We systematically reanalyzed several eukaryotic transcriptomes to annotate previously unannotated ORFs which we term alternative ORFs (altORFs), and we annotated the corresponding hidden proteome. Here, altORFs are defined as potential protein-coding ORFs in ncRNAs or exterior to, or in different reading frames from annotated CDSs in mRNAs (Figure 1a). For clarity, predicted proteins translated from altORFs are termed alternative proteins and proteins translated from annotated CDSs are termed reference proteins.
Our goal was to provide functional annotations of alternative proteins by (1) analyzing relative patterns of evolutionary conservation between alternative and reference proteins and their corresponding coding sequences;(2) estimating the prevalence of alternative proteins both by bioinformatics analysis and by detection in large experimental datasets; (3) detecting functional signatures in alternative proteins; and (4) predicting and testing functional cooperation between alternative and reference proteins.
Results
Prediction of altORFs and alternative proteins
We predicted a total of 551,380 altORFs compared to 67,765 annotated CDSs in the human transcriptome (Figure 1b, Table 1). Because identical ORFs can be present in different RNA isoforms transcribed from the same genomic locus, the number of unique altORFs and CDSs becomes 183,191 and 51,818, respectively. AltORFs were also predicted in other organisms for comparison (Table 1). By convention, only reference proteins are annotated in current protein databases. As expected, these altORFs are on average small, with a size ranging from 30 to 1480 codons. Accordingly, the median size of human predicted alternative proteins is 45 amino acids compared to 460 for reference proteins (Figure 1c), and 92.96 % of alternative proteins have less than 100 amino acids. Thus, the bulk of the translation products of altORFs would be small proteins. The majority of altORFs either overlap annotated CDSs in a different reading frame (35.98%) or are located in 3’UTRs (40.09%) (Figure 1d). Only about 10% of altORFs are located in repeat sequences (Figure 1-figure supplement 1). To assess whether observed altORFs could be attributable solely to random occurrence, due for instance to the base composition of the transcriptome, we estimated the expected number of altORFs generated in 100 shuffled human transcriptomes. Overall, we observed 62,307 more altORFs than would be expected from random occurrence alone (Figure 1e; p <0.0001). This analysis suggests that a large number are expected by chance alone but that at the same time, a large absolute number could potentially be maintained and be functional. The density of altORFs observed in the CDSs, 3’UTRs and ncRNAs (Figure 1f) was markedly higher than in the shuffled transcriptomes, suggesting that these are maintained at frequency higher than expected by chance, again potentially due to their coding function. In contrast, the density of altORFs observed in 5’UTRs was much lower than in the shuffled transcriptomes, supporting recent claims that negative selection eliminates AUGs (and thus the potential for the evolution of altORFs) in these regions 23,24.
Although the majority of human annotated CDSs do not have a TIS with a Kozak motif (Figure 1g)25, there is a correlation between a Kozak motif and translation efficiency26. We find that 27,539 (15% of 183,191) human altORFs encoding predicted alternative proteins have a Kozak motif, as compared to 19,745 (38% of 51,818) for annotated CDSs encoding reference proteins (Figure 1g). The number of altORFs with Kozak motifs is significantly higher in the human transcriptome compared to shuffled transcriptomes (Figure 1-figure supplement 2), again supporting their potential role as protein coding.
Conservation analyses
Next, we compared evolutionary conservation patterns of altORFs and CDSs. A large number of human alternative proteins have homologs in other species. In mammals, the number of homologous alternative proteins is higher than the number of homologous reference proteins (Figure 2a), and 9 are even conserved from human to yeast (Figure 2b), supporting a potential functional role. As phylogenetic distance from human increases, the number and percentage of genes encoding homologous alternative proteins decreases more rapidly than the percentage of genes encoding reference proteins (Figure 2a, 2c). This observation indicates either that altORFs evolve more rapidly than CDSs or that distant homologies are less likely to be detected given the smaller sizes of alternative proteins. Another possibility is that they evolve following the patterns of evolution of genes that evolve de novo, with a rapid birth and death rate, which accelerates their turnover over time20.
If altORFs play a functional role, they would be expected to be under purifying selection. The first and second positions of a codon experience stronger purifying selection than the third27. By definition, CDS regions overlapping altORFs with a shifted reading frame do not contain such third positions because the third codon positions of the CDSs are either the first or the second in the altORFs. We analyzed conservation of third codon positions of CDSs for 100 vertebrate species for the 53,862 altORFs completely nested within the 20,814 CDSs from 14,677 genes (Figure 3). We observed that in regions of the CDS overlapping altORFs, third codon positions were evolving significantly more slowly than third codon positions of random control sequences from the entire CDS for a large number of altORFs (Figure 3), reaching up to 22-fold for conservation at p<0.0001. This is illustrated with three altORFs located within the CDS of NTNG1, RET and VTI1A genes (Figure 4). These three genes encode a protein promoting neurite outgrowth, the proto-oncogene tyrosine-protein kinase receptor Ret and a protein mediating vesicle transport to the cell surface, respectively. Two of these alternative proteins have been detected by ribosome profiling (RET, IP_182668.1) or mass spectrometry (VTI1A, IP_188229.1) (see below, supplementary files 1 and 2).
Evidence of expression of alternative proteins
We provide two lines of evidence indicating that thousands of altORFs are translated into proteins. First, we re-analyzed detected TISs in publicly available ribosome profiling data28,29, and found 26,531 TISs mapping to annotated CDSs and 12,616 mapping to altORFs in these studies (Figure 5a; Supplementary file 1). Although predicted altORFs3’ are more abundant than altORFs5', only a small fraction of TISs detected by ribosomal profiling mapped to altORFs3'. Only a small fraction of TISs detected by ribosomal profiling mapped to altORFs3’ even if those are more abundant than altORF5’ relative to shuffled transcriptomes, likely reflecting a recently-resolved technical issue in the ribosome profiling technique30. New methods to analyze ribosome profiling data are being developed and will likely uncover more translated altORFs9. In agreement with the presence of functional altORFs3’, cap-independent translational sequences were recently discovered in human 3’UTRs31. New methods to analyze ribosome profiling data are being developed and will likely uncover more translated altORFs9. Second, we re-analyzed proteomic data using our composite database containing alternative proteins in addition to annotated reference proteins (Figure 5b, Supplementary file 2). False discovery rate cut-offs were set at 1% for peptide-spectrum match, peptides and proteins. We selected four studies representing different experimental paradigms and proteomic applications: large-scale 32 and targeted 33 protein/protein interactions, post-translational modifications 34, and a combination of bottom-up, shotgun and interactome proteomics 35 (Figure 5b). In the first dataset, we detected 7,530 predicted alternative proteins in the interactome of reference proteins 32, providing a framework to uncover the function of these proteins. In a second proteomic dataset containing about 10,000 reference human proteins 35, a total of 1,658 predicted alternative proteins were detected, representing more than 10% of the detectable proteome. Using a phosphoproteomic large data set34, we detected 1,424 alternative proteins. The biological function of these proteins is supported by the observation that some alternative proteins are specifically phosphorylated in cells stimulated by the epidermal growth factor, and others are specifically phosphorylated during mitosis (Figure 6; Supplementary file 3). We provide examples of spectra validation using synthetic peptides (Figure 6-figure supplement 1–2). A fourth proteomic dataset contained 113 alternative proteins in the epidermal growth factor receptor interactome33 (Figure 5b). A total of 10,362 different alternative proteins were detected in these proteomic data. Overall, by mining the proteomic and ribosomal profiling data, we detected the translation of a total of 22,155 unique alternative proteins. 823 of these alternative proteins were detected by both MS and ribosome profiling (Figure 7), providing a high-confidence collection of nearly one thousand small alternative proteins for further studies.
Functional annotations of alternative proteins
An important goal of this study is to associate potential functions to alternative proteins, which we can do through annotations. Because the sequence similarities and the presence of particular signatures (families, domains, motifs, sites) are a good indicator of a protein’s function, we analyzed the sequence of the predicted alternative proteins in several organisms with InterProScan, an analysis and classification tool for characterizing unknown protein sequences by predicting the presence of combined protein signatures from most main domain databases36 (Figure 8; Figure 8-figure supplement 1). We found 41,511 (23%) human alternative proteins with at least one InterPro signature (Figure 8b). Of these, 37,739 (or 20.6%) are classified as small proteins. Interestingly, the reference proteome has a smaller proportion (840 or 1.6%) of small proteins with at least one InterPro signature, supporting a biological activity for alternative proteins.
Similar to reference proteins, signatures linked to membrane proteins are abundant in the alternative proteome and represent more than 15,000 proteins (Figure 8c-e; Figure 8-supplemental figure 1). With respect to the targeting of proteins to the secretory pathway or to cellular membranes, the main difference between the alternative and the reference proteomes lies in the very low number of proteins with both signal peptides and transmembrane domains. Most of the alternative proteins with a signal peptide do not have a transmembrane segment and are predicted to be secreted (Figure 8c, d), supporting the presence of large numbers of alternative proteins in plasma37. The majority of predicted alternative proteins with transmembrane domains have a single membrane spanning domain but some display up to 27 transmembrane regions, which is still within the range of reference proteins that show a maximum of 33 (Figure 8e).
A total of 585 alternative proteins were assigned 419 different InterPro entries, and 343 of them were tentatively assigned 192 gene ontology terms (Figure 9). 17.1% (100/585) of alternative proteins with an InterPro entry were detected by MS or/and ribosome profiling, compared to 13.7% (22,055/161,110) for alternative proteins without an InterPro entry. Thus, predicted alternative proteins with InterPro entries are more likely to be detected, supporting their functional role (p-value = 0.000035, Fisher’s exact test and chi-square test). The most abundant class of predicted alternative proteins with at least one InterPro entry are C2H2 zinc finger proteins with 110 alternative proteins containing 187 C2H2-type/integrase DNA-binding domains, 91 C2H2 domains and 23 C2H2-like domains (Figure 10a). Seventeen of these (15.4%) were detected in public proteomic and ribosome profiling datasets, a percentage that is similar to reference zinc finger proteins (20.1%) (Figure 2, Table 2). Alternative proteins have between 1 and 23 zinc finger domains (Figure 10b). Zinc fingers mediate protein-DNA, protein-RNA and protein-protein interactions38. The linker sequence separating adjacent finger motifs matches or resembles the consensus TGEK sequence in nearly half the annotated zinc finger proteins39. This linker confers high affinity DNA binding and switches from a flexible to a rigid conformation to stabilize DNA binding. The consensus TGEK linker is present 46 times in 31 alternative zinc finger proteins (Supplementary file 4). These analyses show that a number of alternative proteins can be classified into families and will help deciphering their functions.
Evidence of functional coupling between reference and alternative proteins coded by the same genes
Since one gene codes for both a reference and one or several alternative proteins, we asked whether paired (encoded in the same gene) alternative and reference proteins have functional relationships. There are a few known examples of functional interactions between different proteins encoded in the same gene (Table 3). If there is functional cooperation or shared function, one would expect orthologous alternative-reference protein pairs to be co-conserved40. Our results show a large fraction of co-conserved alternative-reference protein pairs in several species (Figure 11). Detailed results for all species are presented in Table 4.
Another mechanism that could functionally associate alternative and reference proteins from the same transcripts would be that they share protein domains. We compared the functional annotations of the 585 alternative proteins with an InterPro entry with the reference proteins expressed from the same genes. Strikingly, 89 of 110 altORFs coding for zinc finger proteins (Figure 10) are present in transcripts in which the CDS also codes for a zinc finger protein. Overall, 138 alternative/reference protein pairs share at least one InterPro entry and many pairs share more than one entry (Figure 12a). The number of shared entries was much higher than expected by chance (Figure 12b, p<0.0001). The correspondence between InterPro domains of alternative proteins and their corresponding reference proteins coded by the same transcripts also indicates that even when entries are not identical, the InterPro terms are functionally related (Figure 12c; Figure 12-figure supplement 1), overall supporting a potential functional association between reference and predicted alternative proteins. Domain sharing remains significant even when the most frequent domains, zinc fingers, are not considered (Figure 12-figure supplement 2).
Recently, the interactome of each of 131 human zinc finger proteins was determined by affinity purification followed by mass spectrometry 41. This study provides a unique opportunity to test if, in addition to posessing zinc finger domains, some pairs of reference and alternative proteins coded by the same gene also interact. We re-analyzed the MS data using our alternative protein sequence database to detect alternative proteins in this interactome. Five alternative proteins were identified within the interactome of their reference zinc finger proteins. This number was higher than expected by chance (p<10−6) based on 1 million binomial simulations of randomized interactomes. This result strongly supports the hypothesis of functional cooperation between alternative and reference proteins coded by the same genes.
Finally, we integrated the co-conservation and expression analyses to produce a high-confidence list of predicted functional and co-operating alternative proteins and found 3,028 alternative proteins in mammals (H. sapiens to B. taurus), and 51 in vertebrates (H. sapiens to D. rerio) (supplementary file 6). In order to further test for functional cooperation between alternative/reference protein pairs in this list, we focused on alternative proteins detected with at least two peptide spectrum matches. From this subset, we selected altMID51 (IP_294711.1) among the top 3% of alternative proteins detected with the highest number of peptide spectrum matches in proteomics studies, and altDDIT3 (IP_211724.1) among the top 3% of altORFs with the most cumulative reads in translation initiation ribosome profiling studies.
AltMiD51 is a 70 amino acid alternative protein conserved in vertebrates42 and co-conserved with its reference protein MiD51 from humans to zebrafish (supplementary file 6). Its coding sequence is present in exon 2 of the MiD51/MIEF1/SMCR7L gene. This exon forms part of the 5’UTR for the canonical mRNA and is annotated as non-coding in current gene databases (Figure 13a). Yet, altMID51 is robustly detected by MS in several cell lines (Supplementary file 2: HEK293, HeLa, HeLa S3, LNCaP, NCI60 and U2OS cells), and we validated some spectra using synthetic peptides (Figure 13-figure supplement 1), and is also detected by ribosome profiling (Supplementary file 1)37,42,43. We confirmed co-expression of altMiD51 and MiD51 from the same transcript (Figure 13b). Importantly, the tripeptide LYR motif predicted with InterProScan and located in the N-terminal domain of altMiD51 (Figure 13a) is a signature of mitochondrial proteins localized in the mitochondrial matrix44. Since MiD51/MIEF1/SMCR7L encodes the mitochondrial protein MiD51, which promotes mitochondrial fission by recruiting cytosolic Drp1, a member of the dynamin family of large GTPases, to mitochondria45, we tested for a possible functional connection between these two proteins expressed from the same mRNA. We first confirmed that MiD51 induces mitochondrial fission (Figure 13-figure supplement 2). Remarkably, we found that altMiD51 also localizes at the mitochondria (Figure 13c; Figure 13-figure supplement 3) and that its overexpression results in mitochondrial fission (Figure 13d). This activity is unlikely to be through perturbation of oxidative phosphorylation since the overexpression of altMiD51 did not change oxygen consumption nor ATP and reactive oxygen species production (Figure 13-figure supplement 4). The decrease in spare respiratory capacity in altMiD51-expressing cells (Figure 13-figure supplement 4a) likely resulted from mitochondrial fission46. The LYR domain is essential for altMiD51-induced mitochondrial fission since a mutant of the LYR domain, altMiD51(LYR→AAA) was unable to convert the mitochondrial morphology from tubular to fragmented (Figure 13d). Drp1(K38A), a dominant negative mutant of Drp1 47, largely prevented the ability of altMiD51 to induce mitochondrial fragmentation (Figure 13d; Figure 13-figure supplement 5a). In a control experiment, co-expression of wild-type Drp1 and altMiD51 proteins resulted in mitochondrial fragmentation (Figure 13-figure supplement 5b). Expression of the different constructs used in these experiments was verified by western blot (Figure 13-figure supplement 6). Drp1 knockdown interfered with altMiD51-induced mitochondrial fragmentation (Figure 14), confirming the proposition that Drp1 mediates altMiD51-induced mitochondrial fragmentation. It remains possible that altMiD51 promotes mitochondrial fission independently of Drp1 and is able to reverse the hyperfusion induced by Drp1 inactivation. However, Drp1 is the key player mediating mitochondrial fission and most likely mediates altMiD51-induced mitochondrial fragmentation, as indicated by our results.
AltDDIT3 is a 34 amino acid alternative protein conserved in vertebrates and co-conserved with its reference protein DDIT3 from human to bovine (supplementary file 6). Its coding sequence overlaps the end of exon 1 and the beginning of exon 2 of the DDIT3/CHOP/GADD153 gene. These exons form part of the 5’UTR for the canonical mRNA (Figure 15a). To determine the cellular localization of altDDIT3 and its possible relationship with DDIT3, confocal microscopy analyses were performed on HeLa cells co-transfected with altDDIT3GFP and DDIT3mCherry. Interestingly, both proteins were mainly localized in the nucleus and partially localized in the cytoplasm (Figure 15b). This distribution for DDIT3 confirms previous studies 48,49. Both proteins seemed to co-localize in these two compartments (Pearson correlation coefficient of 0.92, Figure 15c). We further confirmed the statistical significance of this colocalization by applying Costes’ automatic threshold and Costes’ randomization colocalization analysis and Manders Correlation Coefficient (Figure 15d) 50. This was tested by co-immunoprecipitation. In lysates from cells co-expressing altDDIT3GFP and DDIT3mCherry, DDIT3mCherry was immunoprecipitated with anti-GFP antibodies, confirming an interaction between the small altDDTI3 and the large DDIT3 proteins encoded in the same gene.
Discussion
In light of the increasing evidence from approaches such as ribosome profiling and MS-based proteomics that the one mRNA-one canonical CDS assumption is strongly challenged, our findings provide the first clear functional insight into a new layer of regulation in genome function. While many observed altORFs may be evolutionary accidents with no functional role, at least 9 independent lines of evidence support translation and a functional role for thousands of alternative proteins: (1) overrepresentation of altORFs relative to shuffled sequences; (2) overrepresentation of altORF Kozak sequences; (3) active altORF translation detected via ribosomal profiling; (4) detection of thousand alternative proteins in multiple existing proteomic databases; (5) correlated altORF-CDS conservation, but with overrepresentation of highly conserved and fast-evolving altORFs; (6) underrepresentation of altORFs in repeat sequences; (7) overrepresentation of identical InterPro signatures between alternative and reference proteins encoded in the same mRNAs; (8) several thousand co-conserved paired alternative-reference proteins encoded in the same gene; and (9) presence of clear, striking examples in altMiD51, altDDI3T and 5 alternative proteins interacting with their reference zinc finger proteins. While 5 of these 9 lines of evidence support an unspecified functional altORF role, 4 of them (5, 7, 8 and 9) independently support a specific functional/evolutionary interpretation of their role: that alternative proteins and reference proteins have paired functions. Note that this hypothesis does not require binding, just functional cooperation such as activity on a shared pathway.
Upstream ORFs here labeled altORFs5’ are important translational regulators of canonical CDSs in vertebrates51. Interestingly, the altORF5’ encoding altDDIT3 was characterized as an inhibitory upstream ORF 52, but the corresponding small protein was not sought. The detection of altMiD51 and altDDI3T suggests that a fraction of altORFs5’ may have dual functions as translation regulators and functional proteins.
Our results raise the question of the evolutionary origins of these altORFs. A first possible mechanism involves the polymorphism of initiation and stop codons during evolution 53,54 For instance, the generation of an early stop codon in the 5’end of a CDS could be followed by the evolution of another translation initiation site downstream, creating a new independent ORF in the 3’UTR of the canonical gene. This mechanism of altORF origin, reminiscent of gene fission, would at the same time produce a new altORF that shares protein domains with the annotated CDS, as we observed for a substantial fraction (24%) of the 585alternative proteins with an InterPro entry. A second mechanism would be de novo origin of ORFs, which would follow the well-established models of gene evolution de novo20,55,56 in which new ORFs are transcribed and translated and have new functions or await the evolution of new functions by mutations. The numerous altORFs with no detectable protein domains may have originated this way from previously non-coding regions or in regions that completely overlap with CDS in other reading frames.
Detection is an important challenge in the study of small proteins. A TIS detected by ribosome profiling does not necessarily imply that the protein is expressed as a stable molecule, and proteomic analyses more readily detect large proteins that generate several peptides after enzymatic digestion. In addition, evolutionary novel genes tend to be lowly expressed, again reducing the probability of detection 20. Here, we used a combination of five search engines and false discovery rate cut-offs were set at 1% for peptide-spectrum match, peptides and proteins, thus increasing the confidence and sensitivity of hits compared to single-search-engine processing57,58. This strategy led to the detection of several thousand alternative proteins. However, ribosome profiling and MS have technical caveats and the comprehensive contribution of small proteins to the proteome will require more efforts, including the development of new tools such as specific antibodies.
In conclusion, our deep annotation of the transcriptome reveals that a large number of small eukaryotic proteins, which may even represent the majority, are still officially unannotated. Our results also suggest that many small and large proteins coded by the same mRNA may cooperate by regulating each other’s function or by functioning in the same pathway, confirming the few examples in the literature of unrelated proteins encoded in the same genes and functionally cooperating59–63. To determine whether or not this functional cooperation is a general feature of small/large protein pairs encoded in the same gene will require much more experimental evidence, but our results strongly support this hypothesis.
Materials and methods
Generation of alternative open reading frames (altORFs) and alternative protein databases
Throughout this manuscript, annotated protein coding sequences and proteins in current databases are labelled annotated coding sequences or CDSs and reference proteins, respectively. For simplicity reasons, predicted alternative protein coding sequences are labelled alternative open reading frames or altORFs. To generate MySQL databases containing the sequences of all predicted alternative proteins translated from reference annotation of different organisms, a computational pipeline of Perl scripts was developed as previously described with some modifications 37. Genome annotations for H. sapiens (release hg38, Assembly: GCF_000001405.26), P. troglodytes (Pan_troglodytes-2.1.4, Assembly: GCF_000001515.6), M. musculus (GRCm38.p2, Assembly: GCF_000001635.22), D. melanogaster (release 6, Assembly: GCA_000705575.1), C. elegans (WBcel235, Assembly: GCF_000002985.6) and S. cerevisiae (Sc_YJM993_v1, Assembly: GCA_000662435.1) were downloaded from the NCBI website (http://www.ncbi.nlm.nih.gov/genome). For B. taurus (release UMD 3.1.86), X. tropicalis (release JGI_4.2) and D. rerio (GRCz10.84), genome annotations were downloaded from Ensembl (http://www.ensembl.org/info/data/ftp/). Each annotated transcript was translated in silico with Transeq64. All ORFs starting with an AUG and ending with a stop codon different from the CDS, with a minimum length of 30 codons (including the stop codon) and identified in a distinct reading frame compared to the annotated CDS were defined as altORFs.
An additional quality control step was performed to remove initially predicted altORFs with a high level of identity with reference proteins. Such altORFs typically start in a different coding frame than the reference protein but through alternative splicing, end with the same amino acid sequence as their associated reference protein. Using BLAST, altORFs overlapping CDSs chromosomal coordinates and showing more than 80% identity and overlap with an annotated CDS were rejected.
AltORF localization was assigned according to the position of the predicted translation initiation site (TIS): altORFs5’, altORFsCDS and altORFs3’ are altORFs with TISs located in 5’UTRs, CDSs and 3’UTRs, respectively. Non-coding RNAs (ncRNAs) have no annotated CDS and all ORFs located within ncRNAs are labelled altORFsnc. The presence of the simplified Kozak sequence (A/GNNATGG) known to be favorable for efficient translation initiation was also assessed for each predicted altORF65.
Identification of TISs
The global aggregates of initiating ribosome profiles data were obtained from the initiating ribosomes tracks in the GWIPS-viz genome browser28 with ribosome profiling data collected from five large scale studies2,9,66-68. Sites were mapped to hg38 using a chain file from the UCSC genome browser (http://hgdownload.soe.ucsc.edu/goldenPath/hg19/liftOver/hg19ToHg38.over.chain.gz) and CrossMap v0.1.6 (http://crossmap.sourceforge.net/). Similar to the methods used in these studies, an altORF is considered as having an active TIS if it is associated with at least 10 reads at one of the 7 nucleotide positions of the sequence NNNAUGN (AUG is the predicted altORF TIS). An additional recent study was also included in our analysis29. Raw sequencing data for ribosome protected fragments in harringtonine treated cells was aligned to the human genome (GRCh38) using bowtie2 (2.2.8). Similar to the method used in this work, altORFs with at least 5 reads overlapping one position in the kozak region were considered as having an experimentally validated TIS.
Generation of shuffled transcriptomes
Each annotated transcript was shuffled using the Fisher-Yates shuffle algorithm. In CDS regions, all codons were shuffled except the initiation and stop codons. For mRNAs, we shuffled the 5’UTRs, CDSs and 3’UTRs independently to control for base composition. Non-coding regions were shuffled at the nucleotide level. The resulting shuffled transcriptome has the following features compared to hg38: same number of transcripts, same transcripts lengths, same nucleotide composition, and same amino-acid composition for the proteins translated from the CDSs. Shuffling was repeated 100 times and the results are presented with average values and standard deviations. The total number of altORFs is 551,380 for hg38, and an average of 489,073 for shuffled hg38. AltORFs and kozak motifs in the 100 shuffled transcriptomes were detected as described above for hg38.
Identification of paralogs/orthologs in alternative proteomes
Both alternative and reference proteomes were investigated. Pairwise ortholog and paralog relationships between the human proteomes and the proteomes from other species, were calculated using an InParanoid-like approach69, as described below. The following BLAST procedure was used. Comparisons using our datasets of altORFs/CDS protein sequences in multiple FASTA formats from Saccharomyces cerevisiae, Caenorhabditis elegans, Drosophila melanogaster, Danio rerio, Xenopus tropicalis Bos taurus, Mus musculus, Pan troglodytes, Homo sapiens were performed between each pair of species (human against the other species), involving four whole proteome runs per species pair: pairwise comparisons (organism A vs organism B, organism B vs organism A), plus two self-self runs(organism A vs organism A, organism B vs organism B). BLAST homology inference was accepted when the length of the aligned region between the query and the match sequence equalled or exceeded 50% of the length of the sequence, and when the bitscore reached a minimum of 4070. Orthologs were detected by finding the mutually best scoring pairwise hits (reciprocal best hits) between datasets A-B and B-A. The selfself runs were used to identify paralogy relationships as described69.
Co-conservation analyses
For each orthologous alternative protein pair A-B between two species, we evaluated the presence and the orthology of their corresponding reference proteins A’-B’ in the same species. In addition, the corresponding altORFs and CDSs had to be present in the same gene.
In order to develop a null model to assess co-conservation of alternative proteins and their reference pairs, we needed to establish a probability that any given orthologous alternative protein would by chance occur encoded on the same transcript as its paired, orthologous reference protein. Although altORFs might in theory shift among CDSs (and indeed, a few examples have been observed), transposition events are expected to be relatively rare; we thus used the probability that the orthologous alternative protein is paired with any orthologous CDS for our null model. Because this probability is by definition higher than the probability that the altORF occurs on the paired CDS, it is a conservative estimate of co-conservation. We took two approaches to estimating this percentage, and then used whichever was higher for each species pair, yielding an even more conservative estimate. First, we assessed the percentage of orthologous reference proteins under the null supposition that each orthologous alternative protein had an equal probability of being paired with any reference protein, orthologous or not. Second, we assessed the percentage of non-orthologous alternative proteins that were paired with orthologous reference proteins. This would account for factors such as longer CDSs having a higher probability of being orthologous and having a larger number of paired altORFs. For example, between humans and mice, we found that 22,304 of 51,819 reference proteins (43%) were orthologs. Of the 157,261 non-orthologous alternative proteins, 106,987 (68%) were paired with an orthologous reference protein. Because 68% is greater than 43%, we used 68% as the probability for use in our null model. Subsequently, our model strongly indicates co-conservation (Fig. 11; p<10−6 based on 1 million binomial simulations; highest observed random percentage =69%, much lower than the observed 96% co-conservation).
Analysis of third codon position (wobble) conservation
Basewise conservation scores for the alignment of 100 vertebrate genomes including H. sapiens were obtained from UCSC genome browser (http://hgdownload.soe.ucsc.edu/goldenPath/hg38/phyloP100way/). Conservation PhyloP scores relative to each nucleotide position within codons were extracted using a custom Perl script and the Bio-BigFile module version 1.07. The PhyloP conservation score for the wobble nucleotide of each codon within the CDS was extracted. For the 53,862 altORFs completely nested inside 20,814 CDSs, the average PhyloP score for wobble nucleotides within the altORF region was compared to the average score for the complete CDS. To generate controls, random regions in CDSs with a similar length distribution as altORFs were selected and PhyloP scores for wobble nucleotides were extracted. We compared the differences between altORF and CDS PhyloP scores (altORF PhyloP – CDS PhyloP) to those generated based on random regions. We identified expected quantiles of the differences (“DQ” column in the table), and compared these to the observed differences. Because there was greater conservation of wobble nucleotide PhyloP scores within altORFs regions located farther from the center of their respective genes (r = 0.08, p < 0.0001), observed differences were adjusted using an 8-knot cubic basis spline of percent distance from center. These observed differences were also adjusted for site-specific signals as detected in the controls.
Human alternative protein classification and in silico functional annotation
Repeat and transposable element annotation
RepeatMasker, a popular software to scan DNA sequences for identifying and classifying repetitive elements, was used to investigate the extent of altORFs derived from transposable elements71. Version 3-3-0 was run with default settings.
Alternative protein analysis using InterProScan
InterProScan combines 15 different databases, most of which use Hidden Markov models for signature identification72. Interpro merges the redundant predictions into a single entry and provides a common annotation. A recent local version of InterProScan 5.1453.0 was run using default parameters to scan for known protein domains in alternative proteins. Gene ontology (GO) and pathway annotations were also reported if available with -goterm and -pa options. Only protein signatures with an E-value ≤ 10−3 were considered.
We classified the reported InterPro hits as belonging to one or several of three clusters; (1) alternative proteins with InterPro entries; (2) alternative proteins with signal peptides (SP) and/or transmembrane domains (TM) predicted by at least two of the three SignalP, PHOBIUS, TMHMM tools and (3) alternative proteins with other signatures. The GO terms assigned to alternative proteins with InterPro entries were grouped and categorised into 13 classes within the three ontologies (cellular component, biological process, molecular function) using the CateGOrizer tool73.
Each unique alternative protein with InterPro entries and its corresponding reference protein (encoded in the same transcript) were retrieved from our InterProscan output. Alternative and reference proteins without any InterPro entries were ignored. The overlap in InterPro entries between alternative and reference proteins was estimated as follows. We went through the list of alternative/reference protein pairs and counted the overlap in the number of entries between the alternative and reference proteins as 100*intersection/union. All reference proteins and the corresponding alternative proteins were combined together in each comparison so that all domains of all isoforms for a given reference protein were considered in each comparison. The random distribution of the number of alternative/reference protein pairs that share at least one InterPro entry was computed by shuffling the alternative/reference protein pairs and calculating how many share at least one InterPro entry. This procedure was repeated 1,000 times. Finally, we compared the number and identity of shared InterPro entries in a two dimensional matrix to illustrate which Interpro entries are shared. In many instances, including for zinc-finger coding genes, InterPro entries in alternative/reference protein pairs tend to be related when they are not identical.
Mass Spectrometry identification parameters
Wrapper Perl scripts were developed for the use of SearchGUI v2.0.1174 and PeptideShaker v1.1.057 on the Université de Sherbrooke’s 39,168 core high-performance Mammouth Parallèle 2 computing cluster (http://www.calculquebec.ca/en/resources/compute-servers/mammouth-parallele-ii). SearchGUI was configured to run the following proteomics identification search engines: X!Tandem75, MS-GF+76, MyriMatch77, Comet78, and OMSSA79. SearchGUI parameters were set as follow: maximum precursor charge, 5; maximum number of PTM per peptide, 5; X!Tandem minimal fragment m/z, 140; removal of initiator methionine for Comet, 1. A full list of parameters used for SearchGUI and PeptideShaker is available in Supplementary file 2, sheet 1. For PXD000953 dataset35, precursor and fragment tolerance were set 0.006 Da and 0.1 Da respectively, with carbamidomethylation of C as a fixed modification and Nter-Acetylation and methionine oxidation as variable modifications. For PXD00078833 and PXD00061234 datasets, precursor and fragment tolerance were set to 4.5 ppm and 0.1 Da respectively with carbamidomethylation of cysteine as a fixed modification and Nter-Acetylation, methionine oxidation and phosphorylation of serine, threonine and tyrosine as variable modifications. For PXD002815 dataset32, precursor and fragment tolerance were set to 4.5 ppm and 0.1 Da respectively with carbamidomethylation of cysteine as a fixed modification and Nter-Acetylation and methionine oxidation as variable modifications.Datasets were searched using a target-decoy approach against a composite database composed of a target database [Uniprot canonical and isoform reference proteome (16 January 2015) for a total of 89,861 sequences + custom alternative proteome resulting from the in silico translation of all human altORFs (available to download at https://www.roucoulab.com/p/downloads)], and their reverse protein sequences from the target database used as decoys. False discovery rate cut-offs were set at 1% for PSM, peptides and proteins. Only alternative proteins identified with at least one unique and specific peptide, and with at least one confident PSM in the PeptideShaker Hierarchical Report were considered valid57.
Peptides matching proteins in a protein sequence database for common contaminants were rejected80.
For spectral validation (Figure 13-figure supplement 1; Supplementary Figures 1-4), synthetic peptides were purchased from the peptide synthesis service at the Université de Sherbrooke. Peptides were solubilized in 10% acetonitrile, 1% formic acid and directly injected into a Q-Exactive mass spectrometer (Thermo Scientific) via an electro spray ionization source (Thermo Scientific). Spectra were acquired using Xcalibur 2.2 at 70000 resolution with an AGC target of 3e6 and HCD collision energy of 25. Peaks were assigned manually by comparing monoisotopic m/z theoretical fragments and experimental (PeptideShaker) spectra.
In order to test if the interaction between alternative zinc-finger/reference zinc-finger protein pairs (encoded in the same gene) may have occurred by chance only, all interactions between alternative proteins and reference proteins were randomized with an in-house randomisation script. The number of interactions with reference proteins for each altProt was kept identical as the number of observed interactions. The results indicate that interactions between alternative zinc-finger/reference zinc-finger protein pairs did not occur by chance (p<10−6) based on 1 million binomial simulations; highest observed random interactions between alternative zinc-finger proteins and their reference proteins = 3 (39 times out of 1 million simulations), compared to detected interactions=5.
Code availability
Computer codes are available upon request with no restrictions.
Data availability
Most Data are available in Supplementary information. Alternative protein databases for different species can be accessed at https://www.roucoulab.com/p/downloads with no restrictions.
Cloning and antibodies
Human Flag-tagged altMiD51(WT) and altMiD51(LYR→AAA), and HA-tagged DrP1(K38A) were cloned into pcDNA3.1 (Invitrogen) using a Gibson assembly kit (New England Biolabs, E26115). The cDNA corresponding to human MiD51/MIEF1/SMCR7L transcript variant 1 (NM_019008) was also cloned into pcDNA3.1 by Gibson assembly. In this construct, altMiD51 and MiD51 were tagged with Flag and HA tags, respectively. MiD51GFP and altMiD51GFP were also cloned into pcDNA3.1 by Gibson assembly. For MiD51GFP, a LAP tag32 was inserted between MiD51 and GFP. gBlocks were purchased from IDT. Human altDDIT3mCherry was cloned into pcDNA3.1 by Gibson assembly using coding sequence from transcript variant 1 (NM_001195053) and mCherry coding sequence from pLenti-myc-GLUT4-mCherry (Addgene plasmid # 64049). Human DDIT3GFP was also cloned into pcDNA3.1 by Gibson assembly using CCDS8943 sequence. gBlocks were purchased from IDT. For immunofluorescence, primary antibodies were diluted as follow: anti-Flag (Sigma, F1804) 1/1000, anti-TOM20 (Abcam, ab186734) 1/500. For western blots, primary antibodies were diluted as follow: anti-Flag (Sigma, F1804) 1/1000, anti-HA (BioLegend, 901515) 1/500, anti-actin (Sigma, A5441) 1/10000, anti-Drp1 (BD Transduction Laboratories, 611112) 1/500, anti-GFP (Santa Cruz Biotechnology, sc-9996) 1/10000, anti-mCherry (Abcam, ab125096) 1/2000.
Cell culture, immunofluorescence, knockdown and western blots
HeLa cell (ATCC CCL-2) cultures, transfections, immunofluorescence, confocal analyses and western blots were carried out as previously described. Mitochondrial morphology was analyzed as previously described82. A minimum of 100 cells were counted (n=3 or 300 cells for each experimental condition). Three independent experiments were performed. For Drp1 knockdown, 25,000 HeLa cells in 24-well plates were transfected with 25 nM Drp1 SMARTpool: siGENOME siRNA (Dharmacon, M-012092-01-0005) or ON-TARGET plus Non-targeting pool siRNAs (Dharmacon, D-001810-10-05) with DharmaFECT 1 transfection reagent (Dharmacon, T-2001-02) according to the manufacturer’s protocol. After 24h, cells were transfected with pcDNA3.1 or altMiD51, incubated for 24h, and processed for immunofluorescence or western blot. Colocalization analyses were performed using the JACoP plugin (Just Another Co-localization Plugin) 50 implemented in Image J software.
Mitochondrial localization, parameters and ROS production
Trypan blue quenching experiment was performed as previously described83. A flux analyzer (XF96 Extracellular Flux Analyzer; Seahorse Bioscience, Agilent technologies) was used to determine the mitochondrial function in HeLa cells overexpressing AltMiD51Flag. Cells were plated in a XF96 plate (Seahorse Biosciences) at 1×104 cells per well in Dulbecco’s modified Eagle’s medium supplemented with 10% FBS with antibiotics. After 24 hours, cells were transfected for 24 hours with an empty vector (pcDNA3.1) or with the same vector expressing AltMiD51Flag with GeneCellin tranfection reagent according to the manufacturer’s instructions. Cells were equilibrated in XF assay media supplemented with 25 mM glucose and 1 mM pyruvate and were incubated at 37°C in a CO2-free incubator for 1h. Baseline oxygen consumption rates (OCRs) of the cells were recorded with a mix/wait/measure times of 3/0/3 min respectively. Following these measurements, oligomycin (1 μM), FCCP (0.5 μM), and antimycin A/rotenone (1 μM) were sequentially injected, with oxygen consumption rate measurements recorded after each injection. Data were normalized to total protein in each well. For normalization, cells were lysed in the 96-well XF plates using 15 μl/well of RIPA lysis buffer (1% Triton X-100, 1% NaDeoxycholate, 0.1% SDS, 1mM EDTA, 50 mM Tris-HCl pH7.5). Protein concentration was measured using the BCA protein assay reagent (Pierce, Waltham, MA, USA).
Reactive oxygen species (ROS) levels were measured using Cellular ROS/Superoxide Detection Assay Kit (Abcam #139476). HeLa cells were seeded onto 96-well black/clear bottom plates at a density of 6,000 cells per well with 4 replicates for each condition. After 24 hours, cells were transfected for 24 hours with an empty vector (pcDNA3.1) or with the same vector expressing AltMiD51Flag with GeneCellin according to the manufacturer’s instruction. Cells were untreated or incubated with the ROS inhibitor (N-acetyl-L-cysteine) at 10mM for 1 hour. Following this, the cells were washed twice with the wash solution and then labeled for 1 hour with the Oxidative Stress Detection Reagent (green) diluted 1:1000 in the wash solution with or without the positive control ROS Inducer Pyocyanin at 100μM. Fluorescence was monitored in real time. ROS accumulation rate was measured between 1 to 3 hours following induction. After the assay, total cellular protein content was measured using BCA protein assay reagent (Pierce, Waltham, MA, USA) after lysis with RIPA buffer. Data were normalised for initial fluorescence and protein concentration.
ATP synthesis was measured as previously described84 in cells transfected for 24 hours with an empty vector (pcDNA3.1) or with the same vector expressing AltMiD51Flag.
Acknowledgements
This research was supported by CIHR grants MOP-137056 and MOP-136962 to X.R; MOP-299432 and MOP-324265 to C.L; a Université de Sherbrooke institutional research grant made possible through a generous donation by Merck Sharp & Dohme to X.R; a FRQNT team grant 2015-PR-181807 to C.L. and X.R; Canada Research Chairs in Functional Proteomics and Discovery of New Proteins to X.R, in Evolutionary Cell and Systems Biology to C.L and in Computational and Biological Complexity to A.O; A.A.C is supported by a CIHR New Investigator Salary Award; M.S.S is a recipient of a Fonds de Recherche du Québec – Santé Research Scholar Junior 1 Career Award; V.D is supported in part by fellowships from Région Nord-Pas de Calais and PROTEO; A.A.C, D.J.H, M.S.S and X.R are members of the Fonds de Recherche du Québec Santé-supported Centre de Recherche du Centre Hospitalier Universitaire de Sherbrooke. We thank the staff from the Centre for Computational Science at the Université de Sherbrooke, Compute Canada and Compute Québec for access to the Mammouth supercomputer.