Evolution of termination codons of proteins and the TAG-TGA paradox

In most eukaryotes and prokaryotes TGA is used at a significantly higher frequency than TAG as termination codon of protein-coding genes. Although this phenomenon has been recognized several years ago, there is no generally accepted explanation for the TAG-TGA paradox. Our analyses of human mutation data revealed that out of the eighteen sense codons that can give rise to a nonsense codon by single base substitution, the CGA codon is exceptional: it gives rise to the TGA stop codon at an order of magnitude higher rate than the other codons. Here we propose that the TAG-TGA paradox is due to methylation and hypermutabilty of CpG dinucleotides. In harmony with this explanation, we show that the coding genomes of organisms with strong CpG methylation have a significant bias for TGA whereas those from organisms that lack CpG methylation use TGA and TAG termination codons with similar probability.

among single cells may provide advantages to the microbial population by enhancing phenotypic diversity and facilitating adaptation to the changing environment and that single cells with high readthrough levels are more adapted to tolerate stress conditions 15,16 .The functional significance of readthrough is most evident in cases where programmed stop codon readthrough is used to generate peroxisomal isoforms of cytosolic enzymes.For example, it has been shown that readthrough of the leaky UGA codon of NAD-dependent lactate dehydrogenase B and NAD-dependent malate dehydrogenase 1 results in C-terminally extended protein variants containing a peroxisomal targeting signal 7,17,18 .
It should be mentioned, however, that a competing hypothesis assumes that stop-codon read-through arises mostly from molecular errors and is largely nonadaptive 19 .As arguments in favor of their error hypothesis and against the adaptive hypothesis the authors point out that read-through rates decrease with gene expression levels, read-through motifs are avoided in highly expressed genes and that read-through regions do not show increased sequence conservation.There are several examples illustrating the deleterious nature of C-terminal extensions resulting from readthrough, questioning the general validity of the "adaptive hypothesis" as an explanation for the codon usage bias of wild type stop codons.It has been shown that some 3'UTR-encoded readthrough peptides mark their resulting products for destruction, mitigating their deleterious effects 20 .It has also been shown that some stop codon read-through mutant proteins are degraded via the ubiquitin-proteasome system 21 .
It must be emphasized, however, that the adaptive and error hypotheses for stop codon readthrough are not necessarily mutually exclusive.Assuming that-on a proteome scale-the selective benefits of readthrough may be more significant than its potential deleterious effects, readthrough could still provide an explanation for the TAG-TGA paradox.
In the present work we have tested the validity of the adaptive hypothesis of the TAG-TGA paradox of human genes by analyzing nonsense mutations accumulated during cancer evolution in different groups of cancer genes that are expected to be selected for or against stop codon readthrough.The rationale of our approach was that in the case of tumor suppressor genes readthrough of inactivating stop codons would counteract the effect of the driver mutation, thus in this case selection would favor TAA and TAG over the leaky TGA codon.Conversely, in the case of oncogenes and tumor essential genes readthrough of stop codons may rescue these pro-oncogenic genes from inactivation by nonsense mutations, therefore selection is expected to favor the leaky TGA over the more efficient TAA and TAG termination codons.
Analysis of mutation data of cancer tissues has revealed that the relative rates of the three nonsense mutations of tumor suppressor genes did not differ significantly from those of oncogenes and tumor essential genes or passenger genes.These observations suggest that differences in termination efficiency are unlikely to explain the TAG-TGA paradox.
Unexpectedly, in the case of all gene groups, the rates of TGA mutations were significantly higher than those expected based on codon frequencies of the genes and the known mutation bias of single base substitutions of tumors and, as a consequence, the fraction of TGA was higher than that of TAG.To get an insight into the source of this excess of TGA, we have surveyed the mutation data of the genes.These analyses revealed that out of the eighteen sense codons that can give rise to a nonsense codon by single base substitution, the CGA codon was exceptional, in as much as it gave rise to a stop codon (the TGA codon) at a significantly higher rate than the other codons, suggesting that the inherent hypermutability of the CpG dinucleotide of CGA codons may underly the TAG-TGA paradox.
An implicit prediction of this hypothesis is that organisms with non-methylated genomes are not expected to favor TGA over TAG as termination codons.In harmony with this prediction, we show that the TAG-TGA distinction parallels the the evolutionary gain and loss of CpG methylation.

Results and discussion
Codon usage bias and mutation bias of single base substitutions do not explain why TGA is more abundant than TAG as termination codon of human genes.Eighteen of the 61 sense codons can give rise to stop codons through single base substitutions.The different amino acids and their stopogenic sense codons differ markedly in the type of stop codon they can generate by single base substitution (Supplementary Table 1).For example, codons of tyrosine (TAC, TAT) can give rise to both TAA and TAG, the tryptophan codon (TGG) can generate both TGA and TAG, whereas substitutions of codons of cysteine (TGC, TGT) and arginine (AGA, CGA) can lead only to the formation of TGA (Supplementary Table 1).Accordingly, the amino acid composition and codon usage of proteins have a marked influence on the relative probability of the formation of the three stop codons.Taking into account the frequency of stopogenic codons in the human proteome (assuming that there was no difference in the probability of the substitution classes), the fractions of termination codons are expected to be fTAA = 0.3073; fTAG = 0.4448; fTGA = 0.2478.
Thus, based on just the frequency of stopogenic codons in the human proteome, single base substitutions leading to TAG would far exceed those leading to TGA, just the opposite of what we find in the case of natural termination codons of the human proteome.As shown by Ho and Hurst 4 , in the case of the human proteome, fTAA = 0.2692; fTAG = 0.2096; fTGA = 0.5212.
Differences in the probability of the six substitution classes have a major impact on the relative probability of the three nonsense mutations (Supplementary Table 1).It is noteworthy, however, that although mutation bias may affect the relative proportion of TAA vs. TGA and TAG, the relative probability of the TAG -TGA mutations are insensitive to such differences, as they have the same composition.
In summary, the amino acid composition, codon usage bias or mutation bias do not provide an explanation as to why TGA is much more abundant than TAG as termination codons of human protein-coding genes.
Nonsense mutation spectra of tumor suppressor genes, oncogenes, tumor essential genes and passenger genes are similar, suggesting that differences in termination efficiency have no major impact on the choice of nonsense mutations.To test whether the differences in the efficiency of TAG and TGA as signals for translation termination play a role in the differences in the abundance of TAG and TGA, we have analyzed nonsense mutations accumulated during cancer evolution in cancer genes that are expected to be most sensitive to truncation by nonsense mutations.In the case of tumor suppressor genes (positively selected for nonsense mutations) readthrough of inactivating stop codons would dampen the effect of the driver mutation, therefore selection should favor TAG over the leaky TGA codon.Conversely, in the case of oncogenes and tumor essential genes that are known to be negatively selected against nonsense mutations 22 , readthrough of stop codons may rescue these pro-oncogenic genes from inactivation therefore selection should favor the leaky TGA codon over TAG.In these studies we have used the lists of oncogenes (OGs) and tumor suppressor genes (TSGs) defined by Vogelstein et al. 23 .The list of tumor essential genes (TEGs) consisted of genes identified by Bányai et al. 22 .As a control group, we have used the list of selectively neutral passenger genes (PGs) characterized by Bányai et al. and Bányai et al. 22,24 .
We have retrieved the mutation data of these genes from the COSMIC database and identified all single base substitutions resulting in nonsense mutations.The primary data of nonsense mutations for the four sets of proteins are found in Supplementary file 1.This dataset contains information on the positions and identities of the wild type amino acids and wild type codons, the single base substitutions, the resulting stop codon and the number of times the given nonsense mutation has been observed.The observed spectra of the three types of nonsense mutations were determined by including the recurrence number of the given mutation in the calculations (Supplementary file 2).
Since amino acid composition, codon frequency and mutation bias has a major impact on the spectrum of nonsense mutations (see section "Codon usage bias and mutation bias of single base substitutions do not explain why TGA is more abundant than TAG as termination codon of human genes"), for each protein-coding gene we have calculated the fTAA, fTAG and fTGA values expected on the basis of their stopogenic codon frequency and the mutation bias characteristic of tumor cells (Supplementary file 3).
Comparison of the ratios of observed and expected fTAA, fTAG and fTGA values revealed no significant differences among tumor suppressor genes, oncogenes and tumor essential genes (Fig. 1, Supplementary file 4), suggesting that the differences in termination efficiency of the three stop codons has no major impact on the pattern of nonsense mutations observed in cancer tissues.
The observed rates of single base substitutions leading to TGA are significantly higher than those expected based on codon usage and mutation bias.Comparison of the observed and expected spectra of nonsense substitutions in cancer tissues revealed that the observed fTGA values are 1.689fold higher than those expected based on codon usage and mutation bias of single base substitutions in cancer (Supplementary file 4, Fig. 1).To get an insight into the possible source of this excess of TGA, we have surveyed the spectrum of nonsense mutations observed in cancer (Supplementary file 1).Our analyses have revealed that the eighteen stopogenic sense codons differed markedly in their probability to give rise to a stop codon (Fig. 2).In all four groups of genes the CGA codon gave rise to a stop codon (TGA) at a significantly higher rate than the other codons (Supplementary file 5).The exceptionally high rate of the CGA to TGA mutation is even more evident when the recurrence of nonsense mutations is also taken into account (Fig. 3, Supplementary file 6).Our studies on nonsense mutations in the human germline (Supplementary file 7) have shown that the relative probability of the use of stopogenic codons is very similar to that observed in the case of somatic mutations in cancer, indicating that germline and somatic cells share the hypermutability of CGA codons.In all four groups of genes (TSGs, OGs, TEGs and PGs) the CGA codon gave rise to a stop codon (TGA) at a significantly higher rate than the other stopogenic codons (Supplementary file 8, Supplementary Fig. 1).

Hypermutability of CpG dinucleotide of the CGA codon is responsible for the excess of CGA → TGA mutations.
It should be noted that the CGA to TGA mutation involves the mutation of a cytosine of a CpG dinucleotide that is known to have inherent hypermutability in methylated genomes.In mammalian genomes CpG dinucleotides are methylated at position 5 of cytosine and spontaneous deamination of methylated cytosine yields thymine, thus leading to the dinucleotide TpG.Thanks to spontaneous deamination, CpG dinucleotides of human genes are known to undergo germ-line transition to TpG at frequencies six to seven times the base mutation rate 25 .
Since CGA is the only stopogenic codon that gives rise to a stop codon through mutation of a CpG dinucleotide, we suggest that the methylation-based hypermutability of CGA codons and the increased rate of TGA formation may provide a clue for understanding the TAG-TGA paradox.
Nevertheless, the fTGA/fTAG ratios of nonsense mutations observed in cancer (fTGA/fTAG ~ 1.236, Supplementary file 2, Fig. 1,) are lower than the values for natural termination codons of human protein-coding genes (fTGA/fTAG ~ 2.49) calculated from the data of Ho and Hurst 4 (Table 1).It should be pointed out, however, that there is a major difference between somatic nonsense substitutions arising de novo in cancer tissues and natural termination codons of protein-coding genes: the former appear during the lifespan of the organism, whereas termination codons are the results of the long evolutionary history of protein-coding genes.The ordinates show fN*, the fraction of stopogenic codons that experienced a nonsense mutation at least once (corrected for codons that can give rise to two nonsense codons), multiplied by the number of times the mutation was observed.The abscissa lists the stopogenic codons.In each panel the horizontal lines indicate the average values for the eighteen stopogenic codons.According to the currently accepted evolutionary models, protein-coding genes evolve from proto-genes that emerge de novo when non-genic sequences become transcribed and translated 26,27 .Phylostratigraphy studies suggest that de novo evolution of genes has occurred continuously throughout evolutionary time and should therefore be considered as a general mechanism for the emergence of new gene functions 28 .
The proto-genes that emerge de novo utilize randomly occurring, pre-existing start and stop codons, therefore they are likely to encode very short polypeptides, but during proto-gene to gene transition the open reading frames lengthen significantly 26,29 .As part of the latter process the original termination codons of the nascent proto-genes may be replaced by new ones in a complex interplay of nonsense to sense, sense to nonsense substitutions and frame-shift mutations affecting the C-terminal regions (Fig. 4).Genes may recruit pre-existing hidden stop codons as termination codons through mutations that cause shifts of reading frame (indels, changes in splicing) or nonsense to sense substitutions (Fig. 4a,b), but new termination codons may also be gained through de novo nonsense mutations affecting the translated region (Fig. 4c).
Our analysis of the hidden stop codons of human protein-coding genes has revealed that the TGA codon is the most abundant (Supplementary file 9, Table 1).Interestingly, the TGA/TAG ratio of out-of-frame stop codons (3.631) is significantly higher than in the case of nonsense mutations (~ 1.236) and even exceeds the ratio of wild type, in-frame termination codons (2.487, Table 1).It must be emphasized that the preponderance of hidden, pre-existing TGA codons and the increased rate of CGA → TGA nonsense mutations have common roots: hypermutability of CpG.It has long been known that-thanks to its hypermutability-the CpG dinucleotide is strongly underrepresented in methylated genomes and that the CpG deficit correlates with increased abundance of TpG dinucleotides [30][31][32] .In fact, CpG hypermutability plays a major, general role in determining the amino acid composition and codon usage of proteins.It has been shown that there is a universal trend of amino acid gain-loss that is caused by CpG hypermutability.CpG hypermutability increases the frequency of amino acids coded by codons with TpG dinucleotides but decreases the frequency of amino acids coded by codons with CpG dinucleotides 33 .CpG hypermutability is also one of the determining factors of codon preferences among synonymous codons, in as much as in the case of codon families synonymous codons containing CpG tend to be the rarest, whereas those containing TpG are the most frequent 34,35 .
The question, however, remains: why is the TAG-TGA difference more pronounced in the case of pre-existing hidden stop codons and wild type termination codons of human protein-coding genes than in the case of de novo somatic nonsense mutations that arose during the evolution of cancer (Table 1).The most plausible explanation is that it stems from the difference in the time-scale of the events.As mentioned above (Section "Hypermutability of CpG dinucleotide of the CGA codon is responsible for the excess of CGA→TGA mutations"), there is a major difference between somatic nonsense substitutions arising de novo in cancer tissues and hidden and functional termination codons: the latter are the results of the long evolutionary history of the organisms, whereas the former are recent, they appear during their lifespan.The higher levels of TpG dinucleotides and hidden TGA codons in genomes accumulated as a result of long evolutionary periods of CpG hypermutability that have also depleted CpG dinucleotides, whereas de novo somatic mutations affect genomes in which CpG dinucleotides are significantly underrepresented.

TAG-TGA usage of nascent protein-coding genes.
There is an increasing body of evidence that de novo gene origination is an important mechanism for the formation of new genes and that recently evolved de novo genes may provide an explanation for some species-or lineage-specific trait 28 .As discussed above, proto-genes may recruit pre-existing hidden stop codons as termination codons, but during the proto-gene to gene transition new termination codons may be gained both through de novo nonsense mutations and through acquisition of hidden stop codons (Fig. 4).However, the relative contribution of pre-existing and de novo nonsense codons to termination codons of protein-coding genes is unclear.
To get an insight into the origin of termination codons of new, nascent protein-coding genes we have analyzed the stop codon usage of de novo genes defined in recent studies.Ruiz-Orera et al. have identified several translated de novo protein-coding genes unique to mouse 36 ; in similar studies on the human genome, Xie et al., An et al. and Sandmann et al. have identified several hominoid-specific de novo protein-coding genes [37][38][39] .In many cases the transcripts of these genes were originally annotated as long non-coding RNAs (lncRNAs) because they lacked conserved open reading frames.The encoded proteins are usually much shorter than average and show no selective constraints, suggesting that their features are determined primarily by the chance occurrence of pre-existing start and stop codons in lncRNAs 29 .For example, the median protein size of the de novo genes identified by Ruiz-Orera et al. was 44 versus the 412 amino acids of conserved ORFs 36 .
Our analysis of the transcripts of de novo murine and human genes with proteomic evidence for the translation of the ORFs has shown that these young "proto-genes" display pronounced preference of TGA over TAG (TGA/TAG > 2, Supplementary file 10, Table 1).In the case of de novo human protein-coding genes, the TGA/ TAG ratio (2.4-3.2) was higher than the value for nonsense mutations (1.24), but lower than the value for hidden stop codons (3.63), suggesting that both pre-existing and de novo nonsense codons contributed to their termination codons.It is noteworthy that the TGA/TAG ratio (3.21) of the youngest de novo microproteins translated from short open reading frames 39 is quite close to the value for hidden stop codons (3.63), suggesting that these microproteins use primarily pre-existing stop codons as termination codons.Despite these age-dependent minor differences, the TGA/TAG ratios of young and old human protein-coding genes are remarkably similar.
The similarity of termination codon usage bias of young and old proteins also holds for eukaryotes with less significant CpG methylation and CpG hypermutability (see section "CpG hypermutability and the TAG-TGA paradox of termination codons").Our analyses of de novo genes of Saccharomyces cerevisae, Caenorhabditis elegans and Drosophila melanogaster [40][41][42][43][44][45] has revealed that the termination codon usage of de novo genes is similar to that of their proteome (Supplementary file 10; Table 1).

CpG hypermutability and the TAG-TGA paradox of termination codons. The CpG hypermutabil-
ity-based explanation of the TAG-TGA paradox predicts that the presence or absence of CpG hypermethylation should correlate with the presence or absence of the TAG-TGA distinction in different groups of organisms.
Although information on hypermethylation and hypermutability of CpG in prokaryotes is much more limited than for Eukaryotes, there is evidence for its role in genome evolution of Bacteria and Archaea.It has been shown that CpG-specific DNA methyltrasnsferases do exist in Eubacteria and Archaea and that CpG dinucleotides are significantly underrepresented and TpG dinucleotides overrepresented in some bacteria as expected based on CpG hypermutability 33,46 Analyses of hidden (out-of-frame) codons of bacterial genomes revealed a general strong bias toward TGA but against TAG, with a systematic excess of TGA codons in the organisms studied 47,48 , suggesting that hypermutability of CpG dinucleotide might contribute to termination codon usage bias in prokaryotes.
CpG hypermethylation is widespread in Eukaryotes, nevertheless there is evidence for group-specific differences and that CpG hypermethylation and thus CpG hypermutability has been lost in several lineages 49,50 .Although DNA methyltransferases (DNMTs) and DNA methylation are present in all groups of Ecdysozoa, some groups have apparently lost these enzymes and DNA methylation 51 .Genome-wide and coding levels of DNA methylation are relatively high in Lepidoptera, Coleoptera, Hymenoptera and Hemiptera, with the highest levels observed in Blattodea, but Diptera (including Drosophila melanogaster) have lost CpG methylation 52 .Similarly, members of the DNMT family and DNA methylation are present in Nematoda, but the five Clades of these worms have experienced differential loss of DNA methyltransferases 51 Clade I, including Trichinella spiralis, has retained DNMTs and there is evidence for DNA methylation in this parasitic nematode 53 .In the case of Clade V of nematodes that contains Caenorhabditis elegans, DNMT2 could be detected in 5 species but most species had no DNMTs 51 .The absence of DNMTs in C. elegans is in harmony with ealier observations that this model organism shows zero DNA methylation 54 .Although DNA methylation is present in Fungi, such as Neurospora crassa, it has been lost in Saccharomyces cerevisiae, Schizosaccharomyces pombe, and Aspergillus nidulans 50 .Yeast species do not methylate their DNA and a distinctive CpG transition rate is absent in yeast 55,56 .
Thus, in view of our hypothesis that the TAG-TGA paradox can be best explained by CpG hypermutability, it is of major interest to see how evolutionary gain and loss, presence or absence of CpG hypermethylation had an impact on the stop codon usage of proteomes.To answer these questions, we have examined the termination codon usage of representative groups of organisms using the Codon Statistics Database of Subramanian et al. 57 .
Analysis of the TGA/TAG ratios of species representing major groups of Eukarya, Archaea and Bacteria (Supplementary file 11) has revealed major differences in codon usage bias of the TGA and TAG termination codons (Table 2, Fig. 5).Whereas in the case of Eukaryotes the SD values of TGA/TAG ratios were relatively low, in the case of Archaea and Bacteria there was extreme variation in TGA/TAG ratios (Supplementary file 11, Table 2 Fig. 5).
The TGA/TAG ratios of Bacteria also show marked group-specific differences.In some bacterial taxons (e.g.Deinococcus, Actinobacteria, Proteobacteria and Thermotogae) there was a strong bias for TGA over TAG, whereas in several other taxons (e.g.Bacteroidetes, Cyanobacteria and Firmicutes) there was weak or no preference of TGA over TAG (Supplementary file 11, Table 2, Supplementary Fig. 2).It is noteworthy that Deinococcus radiodurans, with an exceptionally high TGA/TAG ratio, possesses a C-5 cytosine DNA methylase that shows eukaryotic type sequence specificity, preferentially methylating CpG and CpC sequences 58 ; this observation raises the possibility that hypermutability of CpG may have played a role in the high TGA/TAG ratio of Deinococcus.However, significant variation of termination codon usage of Bacteria may stem from the fact that stop codons are frequently interpreted as sense codons.For example, in the case of Mollicutes (e.g.Mycoplasma, Spiroplasma) the TGA codon is read as tryptophan 59,60 and recent studies have shown that in some uncultured bacteria the TGA codon serves as a fifth glycine codon 61,62 .Recently, Belin and Puigbo 5 have pointed out that translational selection has a major impact on stop codon usage of bacteria.These authors have analyzed the frequencies of stop codons in a group of highly expressed genes from 196 prokaryotes under strong translational selection and have shown that the occurrence of the three translation termination codons is highly biased, with TAA being the most prevalent in almost all bacteria.The authors have proposed that readthrough efficiency and context effects explain the prevalence of TAA over TAG and TGA, particularly in highly expressed genes.
Archaea also show significant group-specific differences of TGA/TAG ratios.In some taxons of Archaea (e.g.Methanobacteria) there is only a weak preference of TGA over TAG, whereas in other groups (e.g.Halobacteria, Thermococci) TGA preference is more pronounced (Supplementary file 11, Table 2).Methanosarcinales are exceptional in that species in this group have extremely high TGA/TAG ratios (Supplementary file 11, Table 2, Supplementary Fig. 2).Interestingly, Methanosarcinales are unique among Archaea as they possess a Pyl-coding system that ensures the insertion of pyrrolysine at TAG codons 63,64 , suggesting that the emergence of the Pylcoding system had a strong negative influence on the use of TAG as termination codon.
It seems thus very likely that the use of TGA and TAG as termination codons in prokaryotes is influenced primarily by factors distinct from those operating in Eukaryotes.Whereas in Eukarya the TGA versus TAG bias results from CpG methylation and hypermutability, in the case of prokaryotes changes in the function of stop codons may have a more significant impact on their use as termination codons.
The TGA/TAG ratios of major groups of Eukaryotes vary in a narrower range than in prokaryotes.The lowest values of TGA/TAG ratios were observed in the case of Fungi, whereas the highest values are characteristic of Vertebrates (Supplementary file 11, Table 2).This observation is in harmony with the role of CpG hypermutability in the TAG-TAG paradox: hypermutability and depletion of CpG dinucleotides is more pronounced in vertebrates than plants, fungi or invertebrates 65 .
The mean TGA/TAG ratio for non-chordate Metazoa (Porifera, Cnidaria, Nematoda, Arthropoda, Mollusca and Echinodermata) is significantly lower than those of vertebrates (Supplementary file 11; Table 2), consistent with an overall increase in CpG methylation in the Chordate lineage.In the case of Chordates, the average value of TGA/TAG was 2.2 (Table 2).Interestingly, the invertebrate chordate Branchiostoma had significantly lower values of TGA/TAG (1.3) than vertebrates.It is worthy of note, that although surveys of the amphioxus genome revealed the presence of the DNA methylation machinery, DNA methyltransferases and methyl-CpG-binding domain proteins, analysis of CpG methylation of the amphioxus genome suggests a bimodal distribution of DNA methylation 66 .Domains of methylated DNA are interspersed with domains of unmethylated DNA, a situation radically different from the globally methylated vertebrate genomes.The mosaic pattern of DNA methylation in amphioxus represents the ancestral condition of the chordate genome, while the global pattern of DNA methylation might be a vertebrate innovation.These observations underline a major difference between genomes of invertebrates and vertebrates.Whereas invertebrate genomes tend to be sparsely methylated, and DNA methylation is mostly targeted to a subset of transcription units (gene bodies), vertebrate genomes are generally globally and heavily methylated.The genomes of Cephalochordates and Tunicates represent different stages in the evolutionary transition of promoter and gene body DNA methylation across invertebrate-vertebrate boundary 67 .Significantly, this transition is also paralleled by an increase in the TGA/TAG ratios of the termination codons of the protein-coding genes of Chordates.There were, however, significant variations of CpG methylation within   www.nature.com/scientificreports/ the various taxonomic groups that were also reflected by their TGA/TAG values.For example, in the case of Fungi, Schizosaccharomyces pombe that does not methylate its DNA has no bias for TGA (Supplementary file 11, Table 2).Similarly, in the group of Nematoda, Strongyloides ratti that has lost DNMTs had the lowest TGA/ TAG value (0.89), whereas Trichinella spiralis that methylates its DNA had the highest TGA/TAG values (3.01).Significant variation of TGA/TAG values is also observed in Arthropoda consistent with the fact that CpG methylation also varies in this taxon.Blattodea, Hemiptera and Hymenoptera, insects with the highest levels of CpG methylation 52 had a marked bias for TGA, whereas such a bias was less pronounced in Coleoptera and Lepidoptera that have less significant CpG methylation.Significantly, species of the taxon Diptera, that have lost CpG methylation show practically no such bias (Table 2).In summary, our studies have shown that methylation and hypermutability of CpG dinucleotides is the major source of the bias favoring TGA versus TAG as termination codons of protein-coding genes.The TGA versus TAG bias is generally valid across Eukarya that methylate CpG dinucleotides, however, in the case of Bacteria and Archaea changes in the function of stop codons have a greater influence on termination codon usage of proteins.

Somatic mutation data.
Cancer somatic mutation data were extracted from COSMIC v96 (COSMIC release v96, 31st May 2022), the Catalogue Of Somatic Mutations In Cancer (https:// cancer.sanger.ac.uk/ cosmic/ downl oad) which includes single nucleotide substitutions from targeted and genome wide screens, affecting the coding sequence of human genes.
Since we were interested in the selection forces that operate on the choice of stop codons during tumor evolution, only confirmed somatic, nonsense point mutations that arose during tumor evolution were included in our analyses.Accordingly, for all subsequent analyses we have retained only transcripts containing mutations that were annotated under' Mutation description' as Substitution-Nonsense and under' Mutation somatic status' as Confirmed Somatic, that is confirmed to be somatic in the experiment by sequencing both the tumor and a matched normal tissue from the same patient.As to'Sample Type, Tumor origin': we have excluded mutation data from cell-lines, organoid-cultures, xenografts since they do not properly represent human tumor evolution at the organism level.To eliminate the influence of polymorphisms on the conclusions we retained only somatic mutations flagged 'n' for SNPs.Finally, we have removed redundant data so that each unique nonsense substitution was represented only once in the dataset used in our analyses.
Although the COSMIC files provide information on the nature of the nucleotide substitution, its position in the coding sequence (under MUTATION_CDS e.g."c.3342G > A") and the effect of the substitution on the amino acid sequence of the protein (under MUTATION_AA, e.g."p.R596*"), in most cases the identification of the mutant stop codons (TAA, TAG or TGA) requires the identification of the wild type codon and the position of the substitution within the codon, since in the case of several amino acids (e.g.Lys, Gln, Glu, Tyr, Ser, Leu, Trp) different substitutions may give rise to different nonsense codons (Supplementary table 1).For example, depending on the nucleotide position affected, a G > A substitution of Trp (TGG) converts it to TAG or TGA or, depending on the actual wild type codon used, an A > T substitution at the first position of the codon may convert a Lys (AAA, AAG) to TAA or TAG.To solve this problem, we have downloaded the files (All_COSMIC_Genes.fasta.gz)containing the nucleotide and amino acid sequences of the genes and-using the MUTATION_CDS information-have identified the sequences of the wild type sense codons and the mutant stop codons (see Supplementary file 1).
Germline mutation data.Information on SNPs affecting the coding regions of human genes was obtained from the dbSNP database (https:// www.ncbi.nlm.nih.gov/ snp/).We have analyzed only variants that contained stop mutations that arose through single nucleotide substitution and have identified the sequences of the stop codons using the protocol described for the analysis of the somatic mutation data (Supplementary file 7).

Substitution metrics.
The 61 sense codons can undergo 549 single base substitutions and, depending on the wild type and mutant codon, substitutions can be assigned to the silent, missense or nonsense mutation category.Codons, however, differ significantly in the probability that their mutation would lead to nonsense mutation and whether the point mutation generates TAA, TAG or TGA.For example, single nucleotide substitutions of synonymous codons of Lys, Gln, Glu and Tyr can give rise to only TAA or TAG, whereas substitutions of codons of Arg, Gly and Cys can generate only stop codon TGA (Supplementary table 1).Since amino acids and synonymous codons do not occur with the same frequency in the coding region of human genes this may have a significant influence on the expected probability and choice of nonsense mutation.Furthermore, different classes of substitutions do not occur with equal probability; tumor tissues show a characteristic spectrum of substitutions classes 68,69 .
Substitutions are assigned to six classes (C > A, C > G, C > T, T > A, T > C, and T > G) referred to by the pyrimidine of the mutated Watson-Crick base pair.It is of crucial importance to take differences in the probability of the six mutation classes into account since-due to the unique structure of the genetic code-the six types of substitutions differ markedly in the probability that they would lead to TAA, TAG or TGA nonsense mutation.To take into account the influence of codon frequencies and mutation bias on the frequency of the three nonsense mutations of human proteins in the absence of selection, we have followed the procedure described previously 22 .
For these calculations, we have downloaded the coding sequences of human protein coding genes (All_COS-MIC_Genes.fasta.gz)from the COSMIC database (https:// cancer.sanger.ac.uk/ cosmic) and their codon usage and amino acid composition were determined using the SMS server (https:// www.bioin forma tics.org/ sms2/ codon_ usage.html) 70 .

Figure 1 .
Figure 1.Comparison of the observed and expected spectra of nonsense substitutions on tumor suppressor genes (TSG), passenger genes (PG), oncogenes (OG) and tumor essential genes (TEG).(A) The light grey columns represent the expected fTAA, fTAG and fTGA values, the dark grey columns show the values observed in cancer tissues.(B) The figure shows the ratios of the observed and expected fTAA, fTAG and fTGA values for the four groups of genes.

Figure 2 .Figure 3 .
Figure 2.Frequencies of nonsense mutations of different stopogenic sense codons in tumor suppressor genes (TSG), passenger genes (PG), oncogenes (OG) and tumor essential genes (TEG) in cancer.The ordinates show fN representing the fraction of stopogenic codons that experienced a nonsense mutation at least once, corrected for codons that can give rise to two nonsense codons (e.g.TAC, TGG).The abscissa lists the stopogenic codons.In each panel the horizontal lines indicate the average values for the eighteen stopogenic codons.

Figure 4 .
Figure 4. Evolutionary origin of the termination codons of protein-coding genes.The upper part of the figure shows a modified version of the model proposed by Carvunis et al. 26 for the gradual emergence of proteincoding genes in non-genic sequences via proto-genes.A key aspect of the model is that the short ORFs of protogenes utilize pre-existing termination codons, increase in length over evolutionary time and this process also involves changes in the position of termination codons.The lower part of the figure shows different scenarios for changes in the termination codons of ORFs during the evolution of genes.(a) The termination codon is lost through nonsense to sense mutations and pre-existing hidden nonsense codon downstream of the original ORF serves as the new termination codons of the C-terminally extended ORFs.(b) The original termination codon is lost as a result of mutations causing frame-shift (indels, changes in splicing) and pre-existing nonsense codons (out of frame of the original ORF) serve as the new termination codons of the mutant ORFs.(c) Sense to nonsense mutations upstream of the actual termination codon creates the new termination codons of the mutant ORFs.

Figure 5 .
Figure 5. Variation of the TGA/TAG ratios of proteomes of species from different groups of organisms.A. Boxplots of TGA/TAG ratios of Bacteria, Archaea, Plants, Fungi and Metazoa.N: number of species in the taxon; Mean: mean of TGA/TAG values of the species in the taxon; SD: standard deviation of TGA/TAG values.

Table 1 .
Nonsense codon usage bias in different datasets of protein-coding genes.Dataset Proteins

Table 2 .
TGA/TAG ratios of different groups of organisms.