Codon degeneracy contributes to divergent fitness effects of rare tRNAs with A-starting anticodons

14 15 Transfer RNA (tRNA) repertoires vary greatly across genomes, shaped by genetic drift and 16 selection. A peculiar pattern across prokaryotes is the near-complete absence of tRNAs with 17 unmodified adenine at the 34th (wobble) position (i.e., tRNAANN). Each of these tRNAs are just 18 a single mutation away from several other tRNAs. Hence, their persistent absence suggests 19 fundamental but hitherto unclear constraints. We engineered 36 Escherichia coli strains 20 expressing tRNAs carrying each theoretically possible ANN anticodon to determine their 21 functionality and fitness effects. Notably, there was no evidence of broad toxicity due to these 22 tRNAs. All five tRNAANN tested underwent post-transcriptional maturation and all seven tested 23 compensated for the deletion of their respective native tRNABNN (carrying G, C or U at the 34th 24 position), demonstrating that tRNA ANN are translationally active. Furthermore, tRNAANN from 25 four-fold degenerate (4D) codon boxes were unmodified and were generally neutral or 26 beneficial, whereas tRNAANN from two-fold degenerate (2D) boxes underwent A34-to-I34 27 modification and were more likely to impair fitness. We suggest superwobbling by tRNAANN — 28 decoding an entire four-codon set — as one mechanism underlying these differential fitness 29 effects. Maximal degeneracy in 4D boxes buffers or exploits tRNAANN superwobbling via 30 synonymous decoding, whereas constrained degeneracy in 2D boxes renders it deleterious, 31 likely through amino acid misincorporation. Thus, these differential fitness effects, sharpens 32 the paradox of neutral or beneficial yet absent 4D tRNAANN, while beginning to empirically 33 unravel underlying causes for the absence of 2D tRNAANN. 34 35 36

tRNA, Mistranslation, Superwobble , tRNA modifying enzymes, Adenine 34 , 37 Inosine, tadA, ADAT, Codon degeneracy 38 39 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint

40 Transfer RNA (tRNA) pools, largely determined by tRNA gene copy numbers (GCN), play a 41 crucial role in maintaining translation rate, translation accuracy, growth rate and overall fitness 42 (Young et al. 1976; Dong et al. 1996; Scott et al. 2010; Scott and Hwa 2011; Klumpp et al. 43 2013; Wilusz 2015; Hu and Lercher 2021; Raval et al. 2023). Within species, tRNA repertoires 44 adapt in response to various environmental factors (Nilsson et al. 2006; Iben et al. 2011; 45 Bedhomme et al. 2019; Ayan et al. 2020; Khomarbaghi et al. 2024), and across species, they 46 echo the diversity of ecological niches (Goodenbour and Pan 2006; Vieira -Silva and Rocha 47 2010; Fujishima and Kanai 2014; Chan and Lowe 2016; Rak et al. 2018) . For a given amino 48 acid, the GCN of tRNAs with distinct anticodons for the same amino acid (‘tRNA isoacceptors’) 49 varies across species. The GCN is often correlated with relative codon usage (Ikemura 1985; 50 Dong et al. 1996; Berg and Kurland 1997; Elf et al. 2003; Rocha 2004; Novoa and Ribas de 51 Pouplana 2012; McDonald et al. 2015), contributing to the diversity of tRNA repertoires. The 52 effective cytosolic tRNA pools can also vary with expression levels from each tRNA gene copy 53 (Kanaya et al. 1999; Dittmar et al. 2004; Dittmar et al. 2006; Bloom -Ackermann et al. 2014; 54 Sagi et al. 2016; Raval et al. 2023) and differential charging across isoacceptors (Elf et al. 55 2003; Dittmar et al. 2005). Another layer of variability is generated by tRNA modifying enzymes 56 (MEs) that can post-transcriptionally modify tRNAs at about a dozen different sites, particularly 57 in the anticodon loop which extends the initially proposed wobble pairing (Crick 1966) to an 58 expanding list of modified wobbles (Agris et al. 2007; Grosjean et al. 2010; Iben and Maraia 59 2012; Novoa et al. 2012; Agris et al. 2017; Maraia and Arimbasseri 2017; Agris et al. 2018; 60 Suzuki 2021) . Despite such numerous mechanisms to specifically generate an immense 61 diversity of functional tRNAs within and across species , tRNAs with specific anticodons are 62 curiously either missing or rare (Maraia and Arimbasseri 2017; Diwan and Agashe 2018; Rak 63 et al. 2018; Torres 2019; Lei and Burton 2020; Ehrlich et al. 2021; Pernod et al. 2021). 64 Of these seemingly ‘prohibited’ tRNAs, perhaps most striking is the lack of those carrying an 65 unmodified adenine at the 34th (wobble) position (henceforth, tRNAANN). A total of eight ANN 66 anticodons are theoretically possible for the eight family-codon boxes with four -fold 67 degeneracy (‘4D’; whereby all four codons within a box, differing only at the 3rd base, encode 68 the same amino acid), and another eight for the split-codon with two-fold degeneracy (‘2D’; 69 where the four codons within a box encode different amino acids). tRNAANN carrying each of 70 these theoretically possible ANN are a single point mutation away from tRNAUNN, tRNAGNN, 71 and tRNACNN (collectively, ‘tRNABNN’) in the same codon box, and should occur frequently in 72 bacterial populations. For instance, with a mutation rate of 9.1×10−11 per base per generation, 73 an Escherichia coli population that grows from ca. 1,000 cells to 1 billion cells (e.g., a 1 mL 74 culture grown in LB medium overnight) will sample at least one mutation leading to a tRNAANN. 75 However, seven 2D codon tRNAANN and one 4D codon tRNAANN are universally absent (Fig. 76 1A). In addition, seven more 4D tRNAANN are absent in prokaryotes, though Leu-tRNAAAG and 77 Thr-tRNAAGT do occur in some bacteria (Andachi et al. 1987; Borén et al. 1993; Inagaki et al. 78 1995; Phillips and de Crécy -Lagard 2011; Diwan and Agashe 2018; Ehrlich et al. 2021) . 79 Curiously, these latter seven 4D codon box tRNA ANN are also absent from eukaryotic 80 organelles of bacterial origin and from bacteriophages (Hatfull 2015; Pope et al. 2015; 81 Morgado and Vicente 2019; Ehrlich et al. 2021) . Hence, comparative genomics alone 82 underscore the persistent rarity of tRNAANN, suggesting that their expression may not be well-83 tolerated. 84 There are some exceptions where tRNAANN are genomically encoded; however , their post-85 transcriptional processing again indicates poor tolerance of unmodified tRNAANN. Both bacteria 86 and eukaryotes encode Arg-tRNAACG, and it is in fact the preferred Arg-tRNA isoacceptor. But 87 in this case, the A34 is post-transcriptionally masked via deamination by the modifying enzyme 88 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint TadA (Karcher and Bock 2009; Diwan and Agashe 2018; Rafels-Ybern et al. 2019; Ehrlich et 89 al. 2021) in prokaryotes or ADAT in eukaryotes, resulting in tRNAs with inosine-34 (tRNAICG) 90 in the cytosol (Gerber and Keller 1999; Wolf et al. 2002; Torres et al. 2014) and in the 91 chloroplast stroma (Delannoy et al. 2009; Karcher and Bock 2009) . Eukaryotes also encode 92 and similarly modify six additional 4D codon box tRNAANN (Torres et al. 2014; Rafels-Ybern et 93 al. 2018) (Fig. 1A). The only remaining 4D codon box tRNA ANN (Gly-tRNAACC) is universally 94 absent, likely because it is a poor substrate for ADAT (Saint-Léger et al. 2016) and TadA 95 (Borén et al. 1993). The A-to-I modification appears to be essential for efficient translation in 96 bacteria (Wolf et al. 2002), eukaryotes (Torres et al. 2021) and chloroplasts (Delannoy et al. 97 2009). Notably, both tRNA ANN and TadA/ADAT appear to be absent in archaea and 98 mitochondria (Ehrlich et al. 2021) . Together, the strict co -occurrence of tRNA ANN and 99 TadA/ADAT and the universal lack of the poor TadA/ADAT substrates tRNAANN point to ancient 100 and strong purifying selection against unmodified tRNAANN, which has likely persisted through 101 the four billion years of life on Earth. 102 Single base mutations should frequently generate tRNAANN from tRNABNN genes, but are 103 presumably removed by selection. A first step towards understanding the intensity of and 104 mechanisms driving such selection is comprehensive in vivo tests of functionality and potential 105 fitness impacts of unmodified tRNAANN. Previous studies have largely focused on a few ANN 106 anticodons in their native backbones or expressed multiple ANN from a single backbone, 107 limiting the ability to generalize. For instance, Gly -tRNAACC is unmodified and functional in 108 Escherichia coli (Borén et al. 1993) and humans (Saint-Léger et al. 2016) and Pro-tRNAAGG 109 does not affect fitness or translational elongation rates in Salmonella (Chen et al. 2002). Both 110 these tRNAANN were generated from the respective native tRNABNN backbones. In another 111 study, all sixteen possible ANN anticodons were generated from a single Methanocaldococcus 112 jannaschii Tyr-tRNA backbone, and these introduced tRNAANN decoded 1-20% of NNU codons 113 in E. coli (Biddle et al. 2016; Schmitt et al. 2018; Schmitt et al. 2024) . These observations 114 suggest that unmodified tRNA ANN are likely to be translationally active and non -lethal. 115 However, the decoding capacities and modifications of tRNAs (including A-to-I) depend on the 116 backbone sequence (Li et al. 1997; Qian et al. 1997; Nakanishi et al. 2005; Schmeing et al. 117 2011; Saint-Léger et al. 2016; Roura Frigolé et al. 2019); and fitness impacts of tRNAs depend 118 on the ecological niche (Bloom-Ackermann et al. 2014; Li et al. 2016; Li and Zhang 2018; 119 Gabzi et al. 2022; Raval et al. 2023). Therefore, validating function in native tRNA backbones 120 and conducting systematic fitness assays across different growth media are both essential to 121 quantify and understand the source of the hypothesized selection against tRNAANN. 122 To this end, we introduced 20 tRNAANN via point mutation in the native tRNA genes of E. coli, 123 and tested their expression, maturation and functionality. Our results provide some of the first 124 empirical evidence for folding, maturation and translational activity of a broad set of tRNAANN 125 in an endogenous context. Most of the tRNAANN were accommodated in the cytosol 126 unmodified, did not alter fitness under tested conditions , and a subset of tested tRNA ANN 127 evidently effectively replaced native tRNABNN. However, of tRNAANN that affected fitness, 2D 128 and 4D tRNAANN differed from each other in that 2D tRNAANN appeared prone to modification 129 and reduced fitness when overexpressed , whereas 4D tRNA ANN appeared unmodified and 130 improved fitness, especially in nutrient poor media . We propose that within-codon box 131 superwobbling (decoding of all four codons) may underlie the observed fitness differences 132 wherein the degeneracy of the code will buffer or profit from superwobbling of 4D tRNAANN but 133 not from 2D tRNAANN due to the limited degeneracy of 2D codon boxes. Overall, our results 134 highlight the lack of any clear overarching source of purifying selection on most tRNAANN and 135 underscore benefits of a majority of 4D tRNAANN, deepening the mystery of their near-universal 136 absence. However, our work also points at mistranslation as a potential source of counter -137 selection for 2D codon amino acids, which merits further research. 138 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint

139 Generating strains 140 We conducted all tRNA gene manipulations in E. coli MG1655 (the wild type, WT). The native 141 tRNA genes were amplified from the genome and cloned into an IPTG-inducible low copy 142 number plasmid (pACDH, a derivative of pACD with an ACYC origin of replication (Mangroo 143 and RajBhandary 1995; Rao and Varchney 2001) ) or a high copy number plasmid (pUC19 144 (Norrander et al. 1983)). We introduced an B34-to-A34 point mutation (‘B34A’) in the native 145 tRNABNN genes via PCR mutagenesis (Hoa et al. 1989) , followed by Sanger sequencing to 146 confirm the mutation. To introduce tRNAANN on the genome, we utilized a scar-free, two-step 147 allelic exchange method that utilizes the pKOV vector that carries a chloramphenicol 148 resistance marker (Link et al. 1997). Briefly, we amplified the locus with the native tRNA gene 149 of interest (along with ca. 200 -300bp upstream and downstream regions) from the genome 150 while also introducing a B34A substitution in the amplicon through PCR mutagenesis . We 151 cloned the amplicons with mutated alleles into pKOV and transformed the WT with pKOV -152 tRNAANN. We selected colonies with genomic integration of pKOV-tRNAANN on 153 chloramphenicol (20 µg/ml) at 43°C (where the pKOV plasmid replicates very inefficiently) and 154 confirmed the chromosomal insertion via PCR. We selected cells that had removed pKOV, by 155 growing the population first overnight in LB (Lysogeny Broth, Difco) with 5% sucrose at 37°C 156 and then on LB agar with 5% sucrose to isolate single colonies at 37°C. The presence of 157 sucrose in both these steps selects against cells with pKOV since the expression of sacB from 158 pKOV is lethal. proL B34A was introduced on the genome via the Datsenko-Wanner method 159 (Datsenko and Wanner 2000) . Briefly, kanamycin cassette was first inserted 60 bp 160 downstream of proL. The proL::Kan region was amplified through a mutagenesis PCR that 161 introduced B34A via the forward primer. The proL B34A::Kan amplicon was electroporated 162 into the WT carrying plasmid pKD46 where the red recombinase switched the WT proL with 163 proL B34A::Kan. The Kan cassette was cured using PCP20 resulting in the proL B34A strain. 164 We screened the mutant strains for the intended mutation, and to rule out any other mutations, 165 via PCR followed by Sanger sequencing and further confirmed the m via Next Generation 166 Sequencing (Illumina HiSeq PE150, 4-5 million reads per strain, >100x depth). We stored all 167 the strains thus generated as glycerol cryostocks at –80°C. 168 169 Measuring growth parameters 170 We revived strains from cryostocks by streaking onto LB (Lysogeny Broth, Difco) agar plates, 171 with appropriate antibiotics where applicable. We inoculated individual colonies in LB and 172 incubated at 37°C (180 rpm shaking) for 14-16 hours to grow the preculture. We set up growth 173 measurement experiments by sub -culturing 1% v/v of precultures in 48 well microplates 174 (Corning) in 500 µL growth medium: LB or M9 minimal medium (M9 salts, 1mM CaCl2, 2.5 175 mM MgSO4) supplemented with specific carbon and nitrogen sources (“GA”: 0.4% w/v 176 glucose and cas amino acids; or carbon sources alone: glucose (Glu) 0.2% w/v, galactose 177 (Gal) 0.2%, pyruvate (Pyr) 50 mM, or glycerol (Gly) 0.6% w/v) (Raval et al. 2023) and with 178 appropriate antibiotics where applicable. To test for fitness costs that may accumulate over 179 longer periods of time, we passaged the strains with a B34A substitution (six independent 180 lines per strain) for eight days (about 80 generations) by diluting 1% v/v every 24h in 500 µL 181 LB in 48 well microplates incubated at 37°C (180 rpm shaking). We stored cryostocks at fourth 182 and eighth day, revived and measured their fitness as per above. 183 We estimated various growth proxies by measuring optical density at 600 nm (OD600) in a 184 Tecan microplate reader, an automated system (LiconiX incubator, robotic arm and Tecan 185 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint microplate reader) for strains where tRNAANN were introduced from the plasmid , or a 186 microplate reader from Agilent Biotek (Epoch2) for strains with a B34A substitution. We 187 included the WT in every microplate as a reference, to account for differences across plate 188 readers and over time. We inferred growth rate by fitting an exponential equation to 189 approximately the first one third of each growth curve (region where growth was exponential) 190 using Curvefitter software (Delaney et al. 2013). We inferred the lag phase length as the time 191 taken to reach early log phase , defined as approximately 1/3rd of the final OD600 in a given 192 growth medium (OD600 0.2 for LB and M9 GA; 0.15 for M9 Glu, Gal and Gly; and 0.12 for M9 193 Pyr). We normalized fitness of WT carrying tRNAANN, with that of WT carrying the respective 194 tRNABNN control for each case of tRNAANN expression from a plasmid. 195 Finally, we approximated the overall fitness effect of each 2D codon tRNAANN expressed from 196 high copy plasmid by qualitatively scoring its impact on the three growth parameters across 197 six media, assigning +1, 0, and –1 to positive, neutral, and negative effects, respectively. We 198 defined the sum of these scores, across media and across parameters, as the ‘fitness effect 199 index’. We estimated the mistranslation likelihood for each tRNAANN as the ratio of non-cognate 200 to cognate codon usage across the E. coli genome. 201 202 Estimating adenine to inosine modification and mature tRNA levels 203 We measured the relative abundance of each tRNA isotype , for strains with B34A 204 substitutions. Prior work suggests that adenine to inosine modification can be reliably 205 estimated using the frequency of cytosine in the complementary strand, i.e., unmodified A34 206 appears as T34 in reverse-complemented cDNA, whereas I34 appears as C34 (Schmitt et al. 207 2024). We therefore utilized YAMAT -Seq (Shigematsu et al. 2017; Ayan et al. 2020) as 208 optimized previously for E. coli (Raval et al. 2023). Briefly, we grew three biological replicates 209 of each strain in LB till mid-log phase (ca. OD 0.3) and isolated total RNA. After deacylation, 210 we ligated Y-shaped, DNA/RNA hybrid adapters (Shigematsu et al. 2017) (Eurofins) to the 5'-211 NCCA-3' and 3'-inorganic phosphate-5' ends of uncharged tRNAs, reverse transcribed , and 212 further amplified the cDNA products by PCR with sample -specific indices (Illumina). After a 213 quality and quantity check through bioanalyzer (Agilent DNA7500 Series II kit and Agilent 2100 214 Bioanalyzer Systems), we combined equimolar amounts of each sample and further purified 215 the fully assembled cDNA (with adapters and indices ligated at both ends of the tRNA 216 sequence) by separating them from unligated adapters based on size on a 5% poly-acrylamide 217 gel. The final products were sequenced using an Illumina NextSeq 550 High-Output 2x75 bp 218 kit (Single-end, 150 bp reads) at the Max Planck Institute for Evolutionary Biology. YAMAT-219 seq for proL-ANN (with WT control) was performed with size-based separation by magnetic 220 beads (Ampure XP beads) followed by verification of the size of purified products (200 -223 221 bp) using TapeStation DNA D1000 HS Screen Tape and sequencing on the Novaseq platform 222 (paired end, 150 bp) at the National Centre for Biological Sciences. 223 We sorted the combined raw reads into reads derived from individual strains (available from 224 NCBI GEO; accession number GSE328815) based on the unique Illumina barcodes. We 225 downloaded all native tRNA sequences predicted by GtRNAdb 2.0 (Chan and Lowe 2009) for 226 WT and manually added additional tRNAANN sequences by introducing B34A substitution in 227 the sequence of tRNABNN isoacceptors mutated in our study (e.g. to get the gly -tRNAACC 228 sequence, we replaced the C34 in gly-tRNAGCC sequence with A34). We mapped bases 80 -229 151 (expected to tRNA sequences) of the raw reads from each strain against the collection of 230 native and ANN tRNA sequences using Geneious Prime (version 11.1.4). We used the 231 following previously mapping criteria : 10 % mismatch/gap rate, max 2 -bp gap size, max. 5 232 ambiguities; no iterations, discard reads that align equally well to multiple reference sequences 233 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint (Khomarbaghi et al. 2024). On average, at least around 95% of the reads aligned with tRNAs, 234 the number of reads aligned to each tRNA isoacceptor were exported by Geneious and used 235 for calculating within-sample proportions of each tRNA isoacceptor. 236 237 Analyses of tRNA sequences from GtRNA DB and TadA homologues 238 We downloaded all predicted tRNA sequences from all bacteria and archaea available on the 239 Genomic tRNA database, GtRNAdb (Chan and Lowe 2016; Thornlow et al. 2020) including 240 the taxonomy of each species . For 14 genera where at least one tRNA ANN (other than Arg-241 tRNAACG) was present in all of its species (a total of 273 species across genera) we attempted 242 to retrieve protein coding sequences (CDS) as nucleotide sequences from NCBI Genome 243 Datasets (v18.24.0) , using the taxonomic IDs provided on GtRNAdb . We could retrieve 244 genomes of 263 species using this approach for which we retrieved genome AT content from 245 the metadata of the genome assemblies, and calculated codon usage from nucleotide CDS 246 (eight Lactobacilli and two Streptococci species, the genomes could not be fetched readily 247 using the taxonomy IDs). Lastly, out of 555 bacterial genomes from GtRNAdb with at least 248 one tRNAANN (other than Arg-tRNAACG), we could retrieve amino acid CDS for 539 genomes 249 which were used to searched for TadA homologues using the E. coli tadA amino acid 250 sequence as a query (diamond BLAST v2.1.24.178 (Buchfink et al. 2014) ; e-value30%, query coverage >50%). 252 253 Data analysis 254 We used in -house python scripts (except for YAMAT -Seq) for all statistical analysis and 255 generating plots. We used Mann–Whitney U tests to compare the fitness of strains expressing 256 tRNAANN with relevant control strains (WT or strain expressing tRNABNN). To quantify skew in 257 fitness effects relative to zero (neutral), we computed a zero -anchored quantile asymmetry 258 metric (Q-asym0) defined as Q(+) 0.9 - |Q(-)0.1| / Q(+) 0.9 + |Q(-)0.1|, where Q(+) 0.9 is the 90th 259 percentile of positive values of log2(magnitude of fitness) and Q(-)0.1 the 10th percentile of 260 negative values. This metric compares the extent of the positive and negative tails relative to 261 zero, yielding a normalized measure of tail imbalance scaled between −1 ( deleterious skew) 262 and +1 (beneficial skew), with 0 indicating symmetry. While similar in essence to Bowley’s 263 coefficient (a measure based on quantiles and median ), Q-asym0 is zero -anchored. 264 Significance was assessed by bootstrap (10,000 resamples), with percentile -based 95% 265 confidence intervals; skew was considered significant when the interval excluded zero. All raw 266 data used for analysis are provided in the source data files for each figure, along with the 267 statistics; in -house scripts to conduct the analyses and plots are available on Zenodo 268 (https://zenodo.org/records/20180916). 269 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint

270 Generation of an E. coli strain library encompassing all theoretically possible A-271 starting anticodons 272 In theory, sixteen anticodons can start with an adenosine, but fifteen of these are absent from 273 bacterial tRNA repertoires, including that of E. coli MG1655 (the wild type, WT). The WT tRNA 274 repertoire consists of 37 genes for tRNAs with G-, U- or C-starting anticodons (‘tRNABNN’) that 275 decode family or four-fold degenerate codon boxes (‘4D’ codon boxes, where all four codons 276 within a codon box are assigned to the same amino acid) and 45 genes for tRNAs that decode 277 split or two-fold degenerate codon boxes (‘2D’ codon boxes, where the four codons within a 278 codon box are split across two different amino acids). Thus, a total of 82 native tRNABNN genes 279 (Fig. S1) could generate a tRNA with an A-starting anticodon (‘tRNA ANN’) through a single 280 substitution leading to adenosine at the 34th base (‘B34A substitution’) (Fig. 1A, Fig. S1). 281 The functional and fitness effects of these tRNAANN should be determined by a combination of 282 tRNA backbone and the extent of degeneracy of the codons decoded by tRNA ANN through 283 canonical, wobble and superwobble base -paring. Under canonical or wobble pairing, the 284 backbone alone determines amino acid charging, such that a G34 to (‘G34A’) will always 285 retain the original amino acid identity and should hence be tolerated due to degeneracy in 286 both 4D codon and 2D codon boxes. On the other hand, a C34A or U34A substitution will lead 287 to mistranslation in a 2D codon box due to a lack of four-fold degeneracy (Fig. 1A). Hence, 288 the probability of mistranslation is higher in tRNA ANN generated from a tRNA CNN or tRNAUNN 289 backbone of the adjacent 2D codon box, and could potentially explain selection against C34A 290 or U34A substitution s. However, an additional complication is revealed by prior work 291 suggesting that tRNA ANN can superwobble, i.e., read all four codons in a box , in bacteria, 292 mitochondria and eukaryotic cytosol (Sibler et al. 1986; Andachi et al. 1987; Borén et al. 1993; 293 Inagaki et al. 1995; Watanabe et al. 1997; Von Nickisch -Rosenegk et al. 2001; Chen et al. 294 2002; Aldinger et al. 2012; Yokobori et al. 2013; Soma et al. 2023; Kompatscher et al. 2024; 295 Schmitt et al. 2024) . Under this scenario, even within a 2D codon box, a G34 A substitution 296 could result in mistranslation because the tRNA will still be charged with the cognate amino 297 acid but it can decode the non-cognate codons via superwobble (Fig. 1B). We hypothesized 298 that such mistranslation through superwobbling may impair fitness, generating selection 299 against 2D codon tRNAANN derived from a tRNAGNN isoacceptor. In contrast, since a 4D codon 300 tRNAANN remains an isoacceptor regardless of which isoacceptor tRNABNN backbone it arose 301 from (Fig. 1C), 4D tRNAANN should have weaker fitness effects. 302 303 Figure 1: Expected range of decoding by potential tRNAANN that could arise via single 304 mutations in native tRNABNN. (A) All theoretically possible anticodons (written from 5’-to-3’) 305 organized across 4D and 2D codon boxes in the codon table. ANNs that are universally rare 306 and modified in different domains (summarized from a previous study (Ehrlich et al. 2021)) are 307 color coded as per the key on the bottom. The native BNN tRNAs mutated to ANN in this study 308 (within their respective codon boxes) are indicated in bold; ANNs expressed from a plasmid 309 are underlined, and those expressed from the genome indicated by asterisk s. (B–C) 310 Schematics illustrate decoding in examples of 2D and 4D codon boxes under Watson-Crick 311 (WC; indicated by black lines ), wobble (indicated by dotted black lines) and superwobble 312 (indicated by dotted orange lines) in the wild type (WT) and after B34A substitutions in distinct 313 tRNABNN. In each case, mRNA codons are written from 5’ -to-3’ at the bottom , and 314 complementary anticodons written from 3’-to-5’ on top, with flanking lines and attached single 315 letter codes indicating the tRNA backbone and amino acid after charging, respectively. The 316 34th base is shown in bold and anticodon mutations leading to A34 are marked in red . 317 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint Predicted mistranslation is indicated by red arrows . (B) A codon box split between 318 phenylalanine and leucine, with two codons each decoded by Phe-tRNA and Leu-tRNA in the 319 WT. A G34A substitution would generate Phe-tRNAAAG charged with phenylalanine , which 320 under WC pairing reads the phenylalanine codon UUU but can decode the leucine codons 321 UUA and UUG through superwobbling, resulting in mistranslation. Similarly, U34A or C34A 322 substitutions would result in tRNAAAG charged with leucine, resulting in leucine-to-323 phenylalanine mistranslation. Superwobbling resulting in mistranslation is indicated by red 324 arrows and is expected to be harmful. (C) The leucine 4D codon box, where all four codons 325 are decoded by tRNAGAG, tRNAUAG and tRNACAG isoacceptors. An B34A substitution in any of 326 these isoacceptors should generate Leu-tRNAAAG charged with leucine. Under canonical base 327 pairing, Leu-tRNAAAG decodes the CUU anticodon and under superwobbling it decodes CUC, 328 CUA and CUG. In either case, the lysine charged Leu -tRNAAAG decodes lysine codons, and 329 does not result in mistranslation. Hence, in tRNABNN from 4D codon boxes, B34A mutations 330 are unlikely to mistranslate. Superwobbling by tRNAANN within 4D codon boxes is indicated 331 by green arrows and is expected to be tolerated or beneficial. Such superwobbling is indicated 332 only for a tRNAANN resulting from G34A substitution, but it is equally applicable for any B34A 333 substitution within a 4D codon box. 334 335 336 337 To test these hypotheses, we constructed tRNAANN genes by introducing B34A substitutions 338 in 20 diverse WT tRNABNN genes (Fig. 1A, Fig. S1). From each of the eight 2D codon boxes, 339 we derived one tRNAANN from the native tRNAGNN (mimicking possible transition mutations) to 340 address fitness effects potentially stemming from mistranslation due to superwobbling. Of the 341 eight 4D codon boxes, one is decoded by tRNAACG whereas the remaining seven are decoded 342 by different combinations of tRNABNN isoacceptors with different backbones (Fig. S2). We 343 introduced B34A substitutions in a subset of these tRNABNN isoacceptor genes. We generated 344 one tRNA ANN each from proline, valine, alanine and glycine 4D codon boxes. For serine, 345 leucine and threonine 4D codon boxes, we generated tRNAANN from a subset of isoacceptors 346 differing from each other in their backbone sequences by more than 30%, resulting in two 347 tRNAANN from leucine codon boxes, and three from the threonine and serine codon box. 348 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint This set of 20 tRNAANN genes thus covered all fifteen ANN anticodons absent across bacteria, 349 with serine, leucine and threonine ANN generated from more than one backbone. We 350 integrated five of these tRNA ANN into the WT genome by replacing the corresponding native 351 tRNABNN with its tRNA ANN variant (Fig. 1A), generating strains with tRNA ANN expressed from 352 the genome (individually referred with the gene name followed by ‘ANN’, e.g. pheV-ANN, and 353 collectively referred to as ‘ANN strains’). We also expressed seven tRNAANN through a high-354 copy plasmid in E. coli strains lacking their respective native tRNA BNN, generated previously 355 (Raval et al. 2023) (individually referred with Δ followed by gene name , e.g. ΔpheV, and 356 collectively referred as knockouts, ‘KOs’). In the WT, we additionally expressed five of these 357 tRNAANN from a low-copy plasmid and finally, 19 of the tRNAANN from a high-copy plasmid. The 358 strains thus generated (Table S1) provided a comprehensive system to probe specific aspects 359 of tRNAANN functionality and fitness effects. 360 361 tRNAANN are translationally active 362 A parsimonious mechanism by which tRNAANN can affect cellular fitness is through translation 363 rate and/or accuracy , although evidence for their translational activity remains indirect and 364 limited (Borén et al. 1993; Biddle et al. 2016; Schmitt et al. 2018; Schmitt et al. 2024) . A key 365 prerequisite for translational activity is accurate post-transcriptional folding and maturation of 366 the tRNA transcripts. To quantify the relative abundance of mature tRNAANN within the total 367 tRNA pool, we utilized YAMAT -seq ( Y-shaped Adapter-ligated MAture TRNA sequencing) 368 (Shigematsu et al. 2017; Ayan et al. 2020; Raval et al. 2023) . In this method, two ends of Y-369 shaped adapters are ligated to the conserved 3′-CCA and to 5’ -OH, respectively, in the 370 acceptor arms of fully folded and matured tRNAs . Adapter-ligated tRNAs are then reverse -371 transcribed, PCR -amplified using adapter -specific primers, and the resulting libraries are 372 sequenced and analyzed using a standard small-RNA sequencing work-flow. We quantified 373 the relative abundance of each tRNA species in all five strains where we replaced native 374 tRNABNN genes by tRNAANN variants. For four of these ANN strains, w e also included strains 375 lacking the corresponding native tRNABNN (KOs) (Raval et al. 2023), as a control for potential 376 compensatory upregulation from the remaining gene copies. 377 The 2D codon Phe-tRNAGAA is encoded by two gene copies in E. coli: pheV and pheU. We 378 note that quantitative inferences about Phe-tRNAGAA remain challenging since this is amongst 379 the most difficult tRNA species to detect by YAMAT (Ayan et al. 2020; Khomarbaghi et al. 380 2024). Nonetheless, the ΔpheV strain had decreased Phe-tRNAGAA proportion relative to WT 381 whereas pheV-ANN had a higher Phe-tRNAGAA proportion than ΔpheV (Fig. 2A). This indicated 382 that upregulation from pheU may not be sufficient to restore Phe-tRNAGAA levels in ΔpheV, as 383 also suggested by the fitness defect observed in ΔpheV previously (Raval et al. 2023) . We 384 thus inferred that at least a fraction of Phe-tRNAGAA in pheV-ANN is due to Phe-tRNAAAA 385 undergoing A34-to-inosine 34 modification, which appears in reverse complemented cDNA as 386 C34 and therefore reads as G34 in the gene sequence (Delannoy et al. 2009; Torres et al. 387 2015; Saint-Léger et al. 2016) . In contrast, all four 4D codon tRNAANN that we tested were 388 expressed and detected unmodified among mature tRNAs (Fig. 2A). For instance, the ΔserX 389 strain showed Ser-tRNAGGA levels (encoded by serX and serW) comparable to WT, suggesting 390 potential upregulation of serW. However, serX -ANN showed a significant reduction in Ser -391 tRNAGGA (from the unaltered serW) and significant proportion of Ser-tRNAAGA (from serX-392 AGA). Proportions of Ser-tRNAGGA and Ser-tRNAAGA, were equal, and together matched Ser-393 tRNAGGA levels in the WT (Fig. 2A). Similarly, the Pro -tRNAAGG expressed on the genome 394 (carrying a G34A mutation in the single-copy tRNA gene proL) was also detected unmodified 395 in the cytosolic pool at levels comparable to the native Pro-tRNAGGG in the WT. Likewise, Thr-396 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint tRNAGGU (encoded by thrT and thrV) was lower than WT in both ΔthrT and thrT-ANN strains. 397 However, thrT-ANN expressed both Thr-tRNAGGU and Thr-tRNAAGU, whose combined 398 proportions matched those of Thr-tRNAGGU in the WT. The Gly-tRNACCC (single copy, encoded 399 by glyU) was absent in the tRNA pool of the ΔglyU, whereas glyU-ANN showed Gly-tRNAACC 400 proportions similar to Gly-tRNACCC in the WT. The observation that 4D tRNAANN proportions 401 were similar to WT (for Pro -tRNAAGG and Gly -tRNAACC) or added up to WT together with 402 tRNABNN (for Ser -tRNAAGA and Thr -tRNAAGU) suggested that tRNAANN were tolerated 403 unmodified in cells, without any inhibition of expression from G34A alleles or post-404 transcriptional degradation. Furthermore, in strains carrying tRNAANN, relative levels of 405 isoacceptors or other tRNA species were unaltered (Fig. S 3), suggesting that tRNAANN 406 expression is tolerated without major changes in the overall cytosolic tRNA pools. Taken 407 together, YAMAT-seq confirmed that both 4D and 2D codon tRNAANN are expressed, correctly 408 folded, and fully mature akin to native tRNABNN. The 4D codon tRNAANN remained unmodified, 409 whereas the 2D codon tRNAANN was modified to tRNAINN. 410 We reasoned that if mature tRNAANN are also translationally active, the fitness of strains 411 carrying the G34A substitution should be similar to WT , and tRNA ANN should rescue any 412 deleterious effect of losing the respective tRNABNN. Indeed, whereas ΔpheV and ΔthrT showed 413 lower growth rate than WT, the growth rates of pheV-ANN and thrT-ANN were comparable to 414 WT (Fig. 2B). Strains ΔpheV and ΔthrT also showed a longer lag phase in LB, which was 415 shortened in strains carrying the respective ANNs (Fig. S 4A). The f inal OD was largely 416 indistinguishable (± 5% changes) from WT for most KO and ANN strains, and was rescued in 417 thrT-ANN (Fig. S4B). Better growth parameters of ANN strains as compared to the respective 418 KO strains suggested that tRNAANN functionally replaced the native tRNABNN. Growth rates of 419 serX-ANN, pro-ANN, and glyU -ANN were indistinguishable from WT, suggesting that these 420 tRNAANN were also well tolerated. tRNAANN strains passaged in LB for eight days (about 80 421 generations) also did not show any fitness costs, suggesting a lack of long-term fitness effects 422 of B34A substitutions on the genome (Fig. 2C). These observations further suggested that 423 tRNAANN can successfully replace tRNABNN on the genome without significant fitness costs. 424 To further validate that tRNAANN can compensate for the loss of the respective tRNABNN, we 425 introduced plasmid-borne tRNAANN and tRNABNN into four of the 2D codon and three of the 4D 426 codon tRNA KOs. Across all five growth media and seven tRNAANN tested (35 combinations), 427 tRNAANN improved overall growth of the respective KOs (Fig. S5A). Across all 35 combinations 428 tested, they shortened the lag phase of the respective KOs (statistically significant for 25 429 combinations, Fig. S5B). All strains showed exponential growth in two nutrient-rich media (LB 430 and M9GA0.4) , where complementation by tRNAANN increased growth rate for all 14 431 combinations tested (Fig. 2D) and increased final OD for 10 combinations (Fig. S 5C). 432 Moreover, complementation by tRNAANN and respective tRNABNN increased the fitness of KOs 433 to a similar extent, further corroborating the functionality of tRNAANN. 434 Thus, all five tRNAANN introduced on the genome were folded, matured, and could functionally 435 replace their respective tRNABNN; and seven tRNAANN expressed from a plasmid could also 436 functionally replace their native tRNABNN genes. A priori, a G34A substitution is unlikely to 437 render a tRNA translationally inactive as described earlier, and prior studies also suggest that 438 such tRNAANN have translational activity (Borén et al. 1993; Biddle et al. 2016; Schmitt et al. 439 2018; Schmitt et al. 2024) . Taken together, we concluded that tRNAANN are generally 440 translationally active. 441 442 Figure 2: tRNAANN undergo normal maturation and can compensate for the loss of native 443 tRNAs. (A) For a subset of tRNA genes (pheV, serX, thrT, glyU), we quantified the relative 444 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint proportion of each tRNA species within the mature tRNA pool, for strains lacking the tRNABNN 445 gene (KO, in orange), strains carrying the G34A substitution in the same tRNABNN gene (tRNA-446 ANN, in green), and the wild type (WT, in grey). For proL, we quantified the relative proportion 447 of each tRNA species for proL-ANN and WT. Bar graphs show the relative abundance of 448 tRNABNN and tRNA ANN of interest (n=3, mean ± standard deviation ). ‘Total’ indicates the 449 combined abundance of tRNA BNN and tRNA ANN in ANN strains . Statistical significance is 450 indicated only for relevant comparisons: blue brackets indicate p 0.05. See the source data file for Fig. 2A for all statistical comparisons and 452 Fig. S3 for proportions of all tRNA species. (B) Growth rates relative to WT, for tRNA KO and 453 strains with tRNAANN expression from the genome (as measured in Fig. 2A ). KO strains that 454 are significantly different from WT and ANN strains significantly different from KO are indicated 455 by thick borders. (C) Growth rates of the five strains with tRNAANN expression from the genome 456 in LB after transferring 1% v/v every 24 hours for 8 days. Each cell shows the mean growth 457 rate of six such populations (relative to WT ancestor) for each strain at the fourth and eighth 458 day. None of these growth rates were significantly different from WT (paired t test s). (D) 459 Growth rates relative to WT+pACDH (empty vector control) , for KO+pACDH, KO+pACDH -460 tRNABNN and KO+pACDH-tRNAANN for a subset of genes (asnU, asnV, aspV, tyrV, proL, thrW, 461 glyU). The KO+pACDH strains that were significantly different from WT+pACDH control, and 462 the complementation strains (KO+pACDH -tRNABNN and KO+pACDH-tRNAANN) significantly 463 different from the KO+pACDH , are indicated by thick borders. Asterisks for KO+pACDH-464 tRNAANN cells indicate a significant differen ce from KO+pACDH -tRNABNN (p<0.05, Mann-465 Whitney U test). 466 467 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint 468 469 Overexpression of 2D tRNAANN is often neutral or deleterious 470 The observation that expression of translationally active tRNAANN did not impair fitness of WT 471 (Fig. 2 B-C) was puzzling since their persistent rarity suggest ed otherwise, at least for 2D 472 codon tRNAANN that should be prone to mistranslation (Fig. 1). However, we noticed that all 473 4D codon tRNAANN tested were tolerated unmodified whereas the 2D codon tRNAANN was 474 masked by A34 -to-I34 modification which can potentially restrict supperwobbling (Curran 475 1995; Gerber and Keller 1999; Wolf et al. 2002; Yokobori et al. 2013; Schmitt et al. 2024) and 476 mitigate potential toxicity. Furthermore, while tRNAANN expression from a single copy on the 477 genome closely mimics potential B34A substitutions in natural bacterial populations, the 478 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint tRNAANN thus expressed may also get outcompeted by native tRNABNN and may not contribute 479 substantially to translation . To reconcile a lack of clear fitness effect s (even after ca. 80 480 generations) and the rarity of tRNAANN in nature, we reasoned that the fitness effects of 481 tRNAANN may be attenuated when they are expressed from a single copy but compounded 482 over evolutionary time scale s. If so, tRNAANN overexpressed from a multi-copy plasmid may 483 accentuate fitness effects and allow them to be detected within the timeframe of our 484 experiments. Similarly, temperature stress may also exaggerate weak mistranslation-induced 485 fitness consequences (Rydbn and Isaksson 1984; Thorbjarnardottir et al. 1991; Dahlgren and 486 Rydén-Aulin 2000; Lyu et al. 2023; Romero Romero et al. 2024). 487 To test these hypotheses, we first overexpressed a subset of tRNAANN from a low-copy-number 488 IPTG-inducible plasmid under permissive growth conditions (37°C, LB medium). We observed 489 that both the 2D codon Asn-tRNAAUU and Asp-tRNAAUC reduced relative growth rate (R rel) 490 across all four IPTG concentrations, and Asn-tRNAAUU also lowered the final OD600 (Fig. 3A, 491 Fig. S6A-C). However, the three 4D codon tRNAANN were largely neutral both with respect to 492 growth rate (Fig. 3A, Fig. S6B) and final OD600 (Fig. 3B, Fig. S6C). We next assessed fitness 493 under temperature stress (42° C and 30°C), where we also included Asn-tRNAAUU under its 494 native promoter to allow for native gene regulation. At 42°C Asp-tRNAATC reduced growth rate 495 (Fig. 3C, Fig. S6D), whereas Leu-tRNAAAG lowered the final OD600 (Fig. 3D, Fig. S6E). At 30°C, 496 however, none of the tRNAANN showed any fitness impacts. The 2D codons Asp-tRNAAUC and 497 Asn-tRNAAUU (expressed from their native promoter) lowered the final OD 600 (Fig. 3C-D, Fig. 498 S6D-E). Thus, temperature stress and overexpression indeed revealed some negative fitness 499 impacts, and further suggested that 2D codon box tRNAANN may be more likely to impair fitness 500 than 4D codon box tRNAANN. 501 502 Figure: 3 Overexpression of tRNAANN from a low-copy number plasmid. Heatmaps show 503 growth rates (A) and final OD (B) of strains carrying tRNAANN relative to that of their respective 504 tRNABNN. Throughout the study, we compared fitness effects of overexpression of tRNA ANN 505 with that of respective tRNA BNN to account for fitness effect stemming purely from 506 overexpression. To calculate relative fitness parameters, we divided growth rate or OD of the 507 strain expressing tRNAANN by that of the strain expressing tRNA BNN. Hence values of these 508 parameters above 1 indicate a benefit (shown in green) and below 1 indicate a cost (shown in 509 red) of tRNAANN expression. tRNA genes were expressed on the plasmid pACDH and strains 510 were grown in LB medium at 37°C across a gradient of inducer concentration (IPTG). Cases 511 where tRNAANN were significantly different from tRNA BNN (Mann Whitney test, p<0.05; N=4) 512 with an effect size higher than 5% are further indicated by red or green thick borders. 513 Statistically significant differences of effect sizes smaller than 5% (potentially within the range 514 of noise resulting from detection limits and fitting of the growth equation) are indicated by thick 515 grey borders. Relative growth rates (C) and relative final OD (D) of the same strains under 516 30°C and 42°C, in LB with 0.5 mM IPTG. 517 518 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint 519 520 To test for potential differences between the fitness effects of 2D and 4D codon tRNAANN more 521 systematically, w e next investigated the fitness effects of tRNA carrying all theoretically 522 possible ANN anticodons (Fig. S1) introduced into the WT background from a high copy 523 plasmid, which we expected to further accentuate fitness consequences. In half the tRNAANN-524 media combinations tested, there was no significant fitness effect (Fig. 4A -C, Fig. S7A, Fig. 525 S8). Three out of the eight tested 2D codon tRNAANN were deleterious: P he-tRNAAAA, Cys-526 tRNAACA and His-tRNAAUG significantly prolonged the lag phase and reduced growth rate (Fig. 527 4A-B, Fig. S8A-B) in at least two of six media. Phe -tRNAAAA and Cys-tRNAACA also reduced 528 final OD in four of the six media (Fig. 4C, Fig. S8C). Tyr-tRNAAUA, Asn-tRNAAUU and Asp-529 tRNAAUC prolonged the lag phase and reduced growth rate in at least one medium (Fig. 4A-B, 530 Fig. S8 A). Note tha t overexpression of Phe -tRNAAAA from a high -copy plasmid impaired 531 fitness, in contrast to expression from a single genome -encoded copy (Fig. 2B-C). This is 532 expected because expression from a single copy should result in tRNA levels low enough for 533 the modification system (Fig. 2A) to mitigate fitness effects, whereas a substantial fraction of 534 overexpressed Phe-tRNAAAA is likely to have remained unmodified , reducing fitness. Finally, 535 Ser-tRNAACU improved all three parameters across at least two media (Fig. 4A-C, Fig. S8) and 536 Ile-tRNAAAU increased growth rate (Fig. 4A , Fig. S8B) and OD (Fig. 4C , Fig. S8C) in at least 537 two media. The magnitude of fitness effects (Fig. 4D -F) was skewed, significantly increasing 538 the lag phase length and also lowering final OD; whereas effects on growth rate were less 539 asymmetric and small (i.e., equally likely to be beneficial or deleterious). Overall, these results 540 suggested that the deleterious effects of 2D tRNAANN were larger and more consistent than 541 beneficial effects, and on average, 2D tRNAANN are likely to be neutral or deleterious. 542 543 Figure 4: Overexpression of 2D codon box tRNA ANN from a high-copy plasmid. (A–C) 544 Heat maps show growth parameters of WT strains carrying tRNAANN, relative to those carrying 545 the corresponding native tRNABNN gene on the high copy plasmid pUC19 induced with 0.5mM 546 IPTG at 37°C . Positive effects on growth (tRNA ANN/ tRNABNN > 1) are shown in green and 547 negative impacts (tRNAANN/ tRNABNN < 1) in red. Significant differences between tRNAANN and 548 the corresponding tRNA BNN (Mann-Whitney test, p 5% are 549 highlighted by a thick border. Statistically significant differences with effect size < 5% 550 (potentially reflecting noise from detection limits and fitting of the growth equation) are 551 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint indicated by thick grey borders. (A) Relative length of lag phase, (B) relative growth rate and 552 (C) relative final OD are shown across different media, with values of each parameter given 553 in each cell . The first two rows include nutrient -rich media and the next four rows include 554 nutrient-poor media. In some cases, data for one or more replicates from either the tRNA ANN 555 strain or the respective tRNA BNN variant did not fit the exponential growth equation whereas 556 the other variant grew; e.g., Cys-tRNAANN in Gly 0.6 medium did not grow, unlike Cys-tRNABNN 557 (raw OD600 vs. time curves are shown in Fig. S7). Such cases yielded infinite or infinitesimal 558 values for relative growth rate, so the corresponding cells in the heatmaps are empty; but the 559 fill colours and outlines indicate the qualitative direction of relative growth (green if only 560 tRNAANN grew and red if only tRNABNN grew). Cases where neither tRNAANN nor tRNABNN grew 561 sufficiently well to fit an exponential growth equation or to calculate lag phase (e.g. , AsnT-562 tRNAs in Gly 0.6) are indicated by cells with a cross. The absolute values of growth parameters 563 are shown in Fig. S8 and summarised in the source data file for Fig. S8. (D-F) Distribution of 564 the mean relative fitness effect shown in the heatmaps in Fig. 4A-C, after log2 transformation 565 (bin width= 0.1). Smoothed lines (using bin width 0.4) indicate the underlying probability 566 distribution estimated using the kernel density. Q-asym0 estimates magnitude of fitness effects 567 skew towards beneficial (+1) to harmful ( -1), values with asterisks indicate statistically 568 significant skew; red/green denote harmful/beneficial skew , grey denote no significant skew 569 (see methods for calculation of Q-asym0 and statistical analysis). 570 571 572 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint Overexpression of 4D tRNAANN is often neutral or beneficial 573 We next investigated fitness effects of overexpressing 4D codon tRNAANN, which should be 574 better tolerated due to complete degeneracy in their target codon boxes. Indeed, six of the 575 eleven 4D codon tRNAANN significantly improved lag phase length in at least two media, and 576 four were neutral (Fig. 5A, Fig. S7B, Fig. S8A), in contrast to the results for 2D tRNAANN (Fig. 577 4A, Fig. S7A, Fig. S8A). Exponential growth rate was also improved by five 4D tRNAANN and 578 remained unaffected by overexpression of four other 4D codon tRNAANN (Fig. 5B, Fig. S8B). 579 Three of the 4D codon tRNAANN improved final OD and the remaining eight appeared to be 580 neutral (Fig. 5C, Fig. S8C). Val-tRNAAAC reduced growth rate in three media (Fig. 5B, Fig. S8B) 581 and Gly-tRNAACC (the only universally absent 4D codon tRNAANN) impaired lag phase as well 582 as growth rate (Fig. 5A-B, Fig. S8A-B). Overall, apart from the universally absent Gly-tRNAACC, 583 4D codon tRNAANN either did not affect growth or improved it. Similar to 2D tRNAANN, in half of 584 the media -tRNAANN combinations tested, there were no significant fitness effects. In the 585 remaining half of the cases tested , 4D codon tRNA ANN also showed skewed fitness 586 magnitudes, albeit opposite to that of 2D tRNAANN. Overexpression of 4D tRNAANN skewed 587 each of the three growth parameters significantly towards more beneficial effects (Fig. 5D-F). 588 Thus, we concluded that on average, 4D tRNA ANN are likely to be neutral or substantially 589 beneficial. 590 591 Figure 5: Overexpression of 4D codon box tRNAANN from a high copy plasmid . (A–C) 592 Heat maps show growth parameters of WT strains carrying tRNAANN, relative to those carrying 593 the corresponding native tRNABNN gene on the high copy plasmid pUC19 induced with 0.5mM 594 IPTG at 37°C. Positive effects on growth (tRNA ANN/ tRNABNN > 1) are shown in green and 595 negative impacts (tRNAANN/ tRNABNN < 1) in red. Significant differences between tRNAANN and 596 the corresponding tRNA BNN (Mann-Whitney test, p 5% are 597 highlighted by a thick border. Statistically significant differences with effect size < 5% 598 (potentially reflecting noise from detection limits and fitting of the growth equation) are 599 indicated by thick grey borders. (A) Relative length of lag phase, (B) relative growth rate and 600 (C) relative final OD are shown across different media, with values of each parameter given 601 in each cell. The first two rows include nutrient -rich media and the next four rows include 602 nutrient-poor media. In some cases, data for one or more replicates from either the tRNA ANN 603 strain or the respective tRNA BNN variant did not fit the exponential growth equation whereas 604 the other variant grew; e.g., Ser-tRNAGGA in Gal 0.2 medium did not grow unlike Ser-tRNAAGA 605 (raw OD600 vs. time curves are shown in Fig. S7). Such cases yielded infinite or infinitesimal 606 values for relative growth rate, so the corresponding cells in the heatmaps are empty; but the 607 fill colours and outlines indicate the qualitative direction of relative growth (green if only 608 tRNAANN grew and red if only tRNA BNN grew). The absolute values of growth parameters are 609 shown in Fig. S8 and summarised in the source data file for Fig. S8. (D-F) Distribution of the 610 mean relative fitness effect shown in the heatmaps in Fig. 5A-C, after log2 transformation (bin 611 width= 0.1). Smoothed lines (using bin width 0.4) indicate the underlying probability distribution 612 estimated using the kernel density. Q-asym0 estimates magnitude of fitness effects skew 613 towards beneficial (+1) to harmful ( -1), values with asterisks indicate statistically significant 614 skew; red/green denote deleterious/beneficial skew, grey denote s no significant skew (see 615

for calculation of Q-asym0 and statistical analysis). 616 617 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint 618 619 On the whole, a majority of both 2D and 4D tRNAANN appeared to be neutral (Fig. 4-5, Fig. 620 S9A-B). However, when tRNAANN did affect fitness, 4D codon tRNAANN were consistently more 621 likely to improve it than to impair it (Fig. 5D-F, Fig. S9 A-B). 2D codon tRNAANN, in contrast, 622 appeared more likely to impair early growth (lag phase) and less likely to improve later stages 623 (exponential growth and final OD) , compared to 4D codon tRNAANN (Fig. 4D-F, Fig. S9A-B). 624 Recall that the tendency of 2D codon tRNAANN to impair fitness is expected from 625 superwobbling, which can increase mistranslation in 2D codon boxes (Fig. 1). Although direct 626 measurement of mistranslation by 2D codon tRNAANN will require more detailed studies, we 627 indirectly tested for this effect by estimating the genome-wide mistranslation likelihood for each 628 2D codon tRNA ANN (Fig. S9 D), expecting that higher mistranslation likelihood should 629 correspond to lower fitness. 2D codon Phe-tRNAAAA, Cys-tRNAACA, Asn-tRNAATT with high 630 mistranslation likelihood impaired fitness whereas S er-tRNAACU and Ile-tRNAAAU with low 631 mistranslation likelihood improved it. This yielded a weak negative correlation between the 632 overall fitness effect of each tRNAANN and its mistranslation likelihood (Fig. S9E), though we 633 acknowledge that this analysis is constrained because only eight tRNAANN carrying 2D 634 anticodon are theoretically possible. This is also further confounded by potential outliers such 635 as His-tRNAAUG which did not impair fitness despite a high mistranslation likelihood . One 636 explanation could be A34-to-I34 modification, which restricts superwobbling. While prior work 637 shows modification of His-tRNAAUG in an orthogonal tRNA backbone (Biddle et al. 2016; 638 Schmitt et al. 2024), it remains to be directly investigated what fraction of His-tRNAAUG or any 639 other tRNAANN introduced here from a plasmid is modified. Nevertheless, this analysis at least 640 qualitatively suggested that mistranslation resulting from superwobbling may contribute to the 641 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint negative fitness effects of 2D codon tRNAANN. Superwobbling by 4D tRNAANN on the other 642 hand, may be tolerated or even exploited due to the degeneracy of the genetic code. 643 644 B34A substitutions are generally rare, but better tolerated in 4D codon box tRNAs 645 The beneficial and neutral fitness effects of 4D codon tRNAANN were unexpected due to their 646 near-universal absence, with the exception of two 4D codon tRNAANN in Leuconostocaceae 647 (Leuconostoc and Enococcus sp.) reported in previous analyses of a relatively small number 648 of bacterial genomes(Diwan and Agashe 2018; Ehrlich et al. 2021). We re-evaluated the rarity 649 of tRNAANN using the entire prokaryotic tRNA repertoire known to date, which includes 246,393 650 predicted tRNA sequences across 4047 bacteria and 10,517 tRNAs from 220 archaea in the 651 GtRNA database (Chan and Lowe 2016; Thornlow et al. 2020) . Across 10,517 predicted 652 archaeal tRNAs, only three cases of tRNAANN were found, all from 4D codon boxes (Fig. 6A). 653 Of 8383 bacterial tRNAANN genes, 7,688 encoded Arg-tRNAACG, supporting the rarity of other 654 ANN anticodons across bacteria (Fig. 6A). However, the 695 non-ACG tRNAANN sequences 655 did include all other theoretically possible ANN. 81% of these were 4D codon tRNAANN, with 656 Leu-tRNAAAG and Thr -tRNAAGT being most frequent (Fig. 6B). Of the remaining (19%) 2D 657 tRNAANN sequences, half were Phe -tRNAAAA and His -tRNAAUG, which are potentially 658 compatible with the A-to-I modification system. 659 Firmicutes encoded the highest number of tRNAANN (Fig. 6B, Fig. S10), and within this phylum, 660 all Lactobacillus, Lactococcus and Streptococcus species encoded Leu-tRNAAAG (Fig. 6C). 661 These three genera lacked Leu-tRNAGAG, suggesting a G34A substitution dating back at least 662 until their common ancestor (Fig. 6C). Thr-tRNAAGT also showed a similar pattern, albeit for 663 fewer genera. The presence of an ANN anticodon in lieu of a GNN is similar to Arg -tRNAs 664 where ACG is preferred and GCG is absent, and suggests that these genera not only encode, 665 but potentially prefer these specific 4D tRNAANN. Interestingly, these two ANNs only replaced 666 GNNs, but not other BNN isoacceptors, suggesting that potential superwobbling by 667 unmodified ANNs may be suboptimal even in 4D codon boxes (e.g., due to slower translation 668 rate). In contrast, 2D-codon tRNAANN were present sporadically across species from different 669 genera (Fig. S10) and always in the presence of tRNAGNN (Fig. S11). We also found some 670 tRNAANN genes in Proteobacteria (genera Salmonella, Escherichia) and Tenericutes 671 (subphylum Mollicutes), although their occurrence was scattered within the subphyla or 672 genera, suggesting lineage specific acquisition or loss (Fig. 6B, S10). 673 674 Figure 6: Occurrence of tRNA ANN across prokaryotes. (A) Fractions of tRNA ANN and 675 tRNABNN within all bacterial and archaeal tRNAs reported in the GtRNA DB . The number of 676 tRNA genes i s indicated in parentheses. (B) Number of bacterial genomes carrying each 677 tRNAANN gene, grouped by phylum. 4D and 2D tRNA ANN are labelled in green and red, 678 respectively. For tRNAANN present in fewer than 50 species, refer to the right hand Y axis. (C) 679 Bacterial genera (colored by phylum, see panel B) where at least one tRNAANN is found in all 680 species (number of species is indicated in parentheses). The coloured portion of each cell 681 represents the fraction of species where the tRNAANN is present. 4D and 2D tRNA ANN are 682 labelled in green and red, respectively. For comparison, presence of the respective tRNAGNN 683 is also shown (labelled in black). Instances where a tRNANN potentially replaced a tRNAGNN 684 (indicated by a concomitant lack of the respective tRNAGNN) are highlighted by black squares. 685 (D) For the genera shown in panel C, average genome AT content (1 indicating 100% AT) , 686 fraction of U-ending codons out of all codons used in the genome , and fraction of U -ending 687 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint codons within the codon box where the tRNAANN is present (averaged across all species within 688 a genus). 689 690 691 692 Next, we analysed genomic features of species that do have tRNAANN. The 15 genera with at 693 least one tRNAANN found in all species ( i.e., candidates for acquisition by their respective 694 ancestors) all had high genomic AT content ranging from 60-70% (Fig. 6D) and 30-40% of all 695 codons in these genomes were U-ending. In genera such as Streptococcus and Lactococcus, 696 U-ending codons were also preferred over their synonyms in the same codon box where 697 tRNAANN was pres ent. However, in other genera, this preference was only moderate , 698 suggesting that factors other than codon usage contribute to the retention of tRNAANN in some 699 lineages. Lastly, 88% of the species with at least one non -ACG tRNAANN encode a TadA 700 homologue (Fig. S10) which may modify these tRNAANN (e.g., Leu-tRNAGAG is modified in 701 Oenococcus (Rafels-Ybern et al. 2019) and Streptococcus (Wulff et al. 2024) ). Such large-702 scale genome-based predictions are somewhat limited due to prediction errors and a lack of 703 direct evidence for gene expression and functionality. Nonetheless, this analysis of 704 exceptional non -ACG tRNAANN suggests that 4D codon tRNAANN (Leu-tRNAGAG and Thr-705 tRNAAGT) are retained across evolutionary timescales and potentially even preferred in some 706 genera; and that overall, other 4D codon tRNAANN are generally better tolerated than 707 unmodified 2D codon tRNAANN, as shown by our experimental results. 708 709 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint

710 “It seems likely that inosine will be formed enzymically from an adenine in the nascent sRNA. 711 This may mean that A in this position will be rare or absent, depending upon the exact 712 specificity of the enzyme(s) involved.” (Crick 1966) 713 Anticipation of constraints on anticodon space dates back to the 1960s when only a handful 714 of tRNAs were sequenced (Ingram 1963; Holley R.W. 1965; Crick 1966) . These sequences 715 already showed adenosine 34 to inosine 34 (A-to-I) modification and Crick argued that I34 at 716 the first or wobble base of the anticodon can allow pairing with codons ending in U, C or A and 717 drive amino acid misincorporation (mistranslation) during decoding of two -fold degenerate 718 (2D) codon boxes. He speculated in a footnote (see above) that A34 may therefore be rare 719 among tRNAs decoding 2D codon boxes (Crick 1966). In essence, selection for translational 720 fidelity can constrain the anticodon space. While his was an argument against wobbling by 721 I34, subsequent research showed that unmodified A34 itself can decode all four anticodons 722 in a codon box (supperwobble) in bacteria, mitochondria and eukaryotic cytosol (Sibler et al. 723 1986; Andachi et al. 1987; Borén et al. 1993; Inagaki et al. 1995; Watanabe et al. 1997; Von 724 Nickisch-Rosenegk et al. 2001; Chen et al. 2002; Aldinger et al. 2012; Yokobori et al. 2013; 725 Soma et al. 2023; Kompatscher et al. 2024; Schmitt et al. 2024) .Therefore, unmodified A34 726 may be more harmful than I34 in 2D codon boxes, though both are expected to be suboptimal 727 and rare in 2D codon boxes. 728 Genome sequencing over the last three decades revealed that A34 in tRNA genes decoding 729 two-fold degenerate codon boxes (‘2D codon tRNAANN’) are indeed extremely rare, and that 730 when A34 is observed in tRNAs decoding four -fold degenerate codon boxes (‘4D codon 731 tRNAANN’) it typically co-occurs with A-to-I modifying enzymes (MEs) (Chan and Lowe 2016; 732 Diwan and Agashe 2018; Ehrlich et al. 2021) . The near-universal absence of unmodified 2D 733 and 4D codon tRNAANN implies strong purifying selection, arguably acting since early cellular 734 evolution (Fig. 7A). Although numerous sources of purifying selection can be speculated (Fig. 735 7B), a key first step towards explaining the absence of tRNA ANN is to directly determine the 736 functionality and fitness effects of expressing such tRNAs. Our results show that tRNAANN in 737 their native backbones are tolerated, folded, matured and translationally active in E. coli. A 738 majority of the 4D codon tRNAANN were also tolerated in an unmodified state and showed 739 neutral or positive fitness effects when overexpressed. However, 2D codon box tRNAANN were 740 more likely to be deleterious when overexpressed and their magnitude of fitness effect 741 appeared to scale with a coarsely estimated likelihood of mistranslation. 742 Orthogonal tRNAANN, introduced from a medium -copy number plasmid and designed to 743 incorporate tyrosine against each sense codon , reduced growth rate by 10-50% (Schmitt et 744 al. 2018). Similar global mistranslation, albeit less severe in magnitude, likely contributed to 745 the fitness cost s of 2D codon tRNA ANN. Furthermore, all tested 4D codon tRNAANN were 746 tolerated unmodified, whereas the only tested 2D codon Phe -tRNAAAA was modified to Phe -747 tRNAIAA. The INN anticodon can decode only three codons and thus it is likely to mis -748 incorporate amino acids at fewer codons than unmodified ANN. A prior study tested all 749 tRNAANN from a non-native backbone and found that only a 2D codon His-tRNAAUG underwent 750 such modification (Biddle et al. 2016; Schmitt et al. 2024) . The enzyme responsible for 751 modification of these “novel” bacterial tRNAANN substrates for I34 modification remains 752 unknown, though a parsimonious explanation is TadA which canonically modifies Arg -753 tRNAACG and was recently shown to modify numerous mRNAs in E. coli (Arad et al. 2026) . 754 Our study thus suggests potential substrate flexibility of this ancient and essential enzyme 755 (Wolf et al. 2002; Delannoy et al. 2009; Diwan and Agashe 2018; Torres et al. 2021) that may 756 mitigate the deleterious effects of other tRNA ANN and confer additional advantages, though 757 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint this remains to be confirmed. Intriguingly, the two tRNAANN that are compatible with 758 modification also constituted 50% of the rarely found 2D codon tRNAANN across bacteria (Fig 759 6B), suggesting that A-to-I modification might mitigate mistranslation and fitness impairment 760 by such tRNAs. Taken together, selection for translational fidelity hence emerges as one 761 plausible explanation for the absence of 2D codon tRNAANN. 762 763 Figure 7: A deep mystery of scarce tRNAANN: (A) The near-universal absence of unmodified 764 tRNAANN suggests purifying selection on tRNAANN genes and the evolution of mechanisms to 765 mitigate deleterious effects of tRNAANN that are retained and expressed. (B) Potential sources 766 of negative selection (in the red box) could involve translation and include error-prone or 767 inefficient decoding by unmodified tRNAs (A-U pairing is also the least efficient (Pernod et al. 768 2021)) ribosomal stalling or disassembly. However, they may also include non-translational 769 sources of selection such as toxic effects of fragments resulting from tRNA degradation 770 (suggested for some tRNABNN in eukaryotes (Magee and Rigoutsos 2020; Polacek and Ivanov 771 2020)), interreference with non -translational processes (suggested for some tRNA BNN in 772 eukaryotes (Seligmann 2010; Katz et al. 2016; Balasubramaniam et al. 2017; Hamdani et al. 773 2019; Su et al. 2020; Ehrlich et al. 2021)), or more speculatively, DNA structure-level effects. 774 Potential mechanisms for removing or managing tRNAANN genes (in the green box) include 775 substitutions in the anticodon loop (e.g., A34B) or elsewhere in the gene , causing 776 pseudogenization, loss or silencing of the gene. Known mitigation mechanisms (in the yellow 777 box) are likely ancient and include changes in the tRNA backbone (e.g., pairing between bases 778 32 and 38, which reduces miscoding) and post-transcriptional modifications. Substrate tRNAs 779 and respective modifying enzymes (MEs) always co -occur, and beyond mitigating the 780 translational effects of unmodified anticodons (e.g. , of tRNAANN and tRNAGNN), modification 781 may also confer additional advantages that drive selection favouring substrate tRNAs and 782 counter-selection against poor substrates. MEs (e.g., TadA) introduced into the archaeal host 783 by endosymbiosis likely allowed eukaryotic tRNA repertoires to expand (e.g., tRNAANN). 784 Overall, while translational fidelity appears to be a key factor explaining why tRNAANN are 785 absent, alternative sources of negative selection remains to be investigated. 786 787 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint 788 789 In contrast to 2D codon tRNAANN, superwobbling by a tRNA ANN within a 4D codon box is not 790 expected to cause mistranslation. Indeed, due to four-fold degeneracy, all 4D tRNAANN (except 791 Gly- tRNAACC) were neutral or substantially beneficial. 4D tRNAANN were also more likely to be 792 retained or even preferred as compared to 2D codon tRNAANN among the exceptional cases 793 of A34 occurrences across bacteria. Nevertheless, seven of eight 4D tRNAANN are missing 794 from 99.99% of predicted bacterial tRNA repertoires. Both bacteria and eukaryotes decode 795 the arginine 4D codon box with Arg -tRNAACG; whereas for the other seven 4D codon boxes, 796 bacteria prefer tRNAUNN and eukaryotes use tRNAANN (Novoa et al. 2012). Each of these 4D 797 codon tRNAs is modified via A34-to-I34 or U34-to-cmo5U34 (Novoa et al. 2012; Diwan and 798 Agashe 2018) and tRNAGNN alternatives are absent. Codons recognized by these modified 799 tRNAs are enriched in highly expressed genes, and these tRNAs (Novoa et al. 2012) and A-800 to-I modification appear to be essential (Wolf et al. 2002; Delannoy et al. 2009; Torres et al. 801 2021), further underscoring a crucial role of tRNA modification in decoding 4D codon boxes. 802 Indeed, it has been proposed that coevolution between MEs and tRNA repertoires shaped the 803 ancient preference for modifiable 4D codon tRNAUNN in bacteria and tRNAANN in eukaryotes 804 (Novoa et al. 2012) . The sources of selection against unmodified 4D codon tRNA ANN and 805 tRNAGNN remain incompletely understood, although unmodified 4D codon tRNAGNN (absent 806 across eukaryotes and present with G34 -to-queuosine modification in most bacteria (Diwan 807 and Agashe 2018) ) makes the second and third bases error -prone, causing mistranslation 808 outside the cognate codon boxes , and driving cellular toxicity (Pernod et al. 2021) . This 809 inherent propensity to error and toxicity is mitigated by a canonical base pairing between the 810 32nd and 38 th bases of the tRNA , which respectively mark the beginning and end of the 811 anticodon loop (Pernod et al. 2021). This additional base pairing changes the conformation of 812 the anticodon loop and lowers miscoding by tRNAGNN. This underscores the idea that specific 813 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint anticodons — potentially also ANN — in specific backbones are inherently error-prone, unless 814 mitigated by additional features of the backbone or anticodon modifications . Thus, selection 815 for translational fidelity may also contribute to the absence of 4D tRNAANN (Fig. 7B). 816 Our observation that both 2D and 4D tRNAANN showed lower growth rates at 42°C than at 817 30°C hints at such mistranslation, exacerbated by high temperature. While mistranslation is 818 generally thought to be deleterious (O’connor et al. 1992; Beebe et al. 2003; Bacher et al. 819 2004; Kohanski et al. 2008; Berg et al. 2019; Kelly et al. 2019), it can also be beneficial under 820 specific conditions, and especially under stress (Fan et al. 2015; Samhita et al. 2020; Samhita 821 et al. 2021; Samhita 2022) . The extent of mistranslation also varies across tRNAANN due to 822 differences in stability across superwobble base -pairs (e.g., A-A is least stable). 823 Consequences of amino acid misincorporation on protein structure and fitness also depends 824 on the physico-chemical differences between the exchanged amino acids ; thus, mis-825 incorporation by some tRNAANN (e.g., replacing Lys with Asn) may be less detrimental than 826 others (e.g., Arg to Ser) (Sengupta et al. 2007). Fitness effects of mistranslation are therefore 827 likely to vary across specific tRNAANN and environment s, and our suggestion that 828 mistranslation contributes to the absence of tRNAANN remains to be directly tested. A first step 829 would be to measure the extent of mistranslation in cells expressing tRNAANN. Strains with 830 high-fidelity ribosomes (Ruusala et al. 1984; Chumpolkulwong et al. 2004; Agarwal et al. 2015) 831 and downregulation of tadA could also be used to test mistranslation -driven fitness defects, 832 which should be alleviated and exacerbated in the respective strain backgrounds. Likewise, 833 antibiotics reducing ribosomal fidelity should also amplify such fitness defects. Lastly, we note 834 that despite our normalization of fitness effects of overexpressed tRNA ANN with that of 835 overexpressed tRNABNN, we cannot rule out that overexpression of tRNAANN caused extremely 836 high mistranslation or resulted in other effects qualitatively different from those observed in 837 nature. Thus, such overexpression may not have amplified natural fitness effects as we had 838 intended. We hope that future studies using more sensitive fitness assays and long term 839 evolution may provide further insights. 840 Overall, while mistranslation is an attractive hypothesis, given these causes of substantial 841 variation in its effects across specific tRNAANN and environments, it is unlikely to be the sole 842 explanation for the rarity of all tRNAANN across prokaryotes. Therefore, we speculate important 843 contributions from other mechanisms, likely as universal as core features of the translation 844 machinery. For instance, correct pairing at the A site in bacterial ribosomes causes 845 A1492/A1493 to flip out from 16S rRNA into the minor groove of the first two codon-anticodon 846 base pair and promotes EF-Tu GTP hydrolysis and amino acylated tRNA accommodation 847 (Schmeing and Ramakrishnan 2009) . It is possible that an adenosine at the first anticodon 848 position causes steric hindrance (due to the presence of two additional As in the vicinity) and 849 impairs this essential step inside the decoding center, which is conserved across the tree of 850 life (Rodnina and Wintermeyer 2009; Schmeing and Ramakrishnan 2009; Dever et al. 2018; 851 Tirumalai et al. 2021) . However, an adenosine is tolerated on the mRNA (e.g. , in A-ending 852 codons), suggesting that any steric hindrance is likely asymmetric and specifically unable to 853 accommodate tRNA with A34. While this argument remains to be tested via structural or 854 molecular dynamics analyses, it serves as a useful example for the nature of potential 855 fundamental explanations for the ancient and persistent rarity of tRNAANN. 856 A34 may also be rare because tRNA genes at large are highly conserved despite several 857 times higher mutations rates than the rest of the genome (Thornlow et al. 2018). For instance, 858 out of 62,941 mutations observed across 50,000 generations of adaptive evolution in twelve 859 E. coli evolution populations, only 50 were in tRNA genes and none in the anticodons (Table 860 S2). Analysis of ca. 700 mutations from E. coli mutation accumulation lines (Sane et al. 2025) 861 also captured only one insertion at the first base of the leuW tRNA gene, suggesting that even 862 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint under a regime largely dominated by drift, mutations inside tRNA genes are rare. A saturation 863 mutagenesis study of the yeast Arg-tRNACCU also found most substitutions to be detrimental, 864 with those in the anticodon loop reducing fitness by more than 25% (Li et al. 2016) . Hence, 865 overall strong purifying selection suggests that any disadvantage of B34A substitutions (e.g., 866 low levels of mistranslation) may be penalized especially severely and contribute to the 867 scarcity of A34 over evolutionary timescales. Apart from substitutions, the presence or 868 absence of entire tRNA genes can also be under selection, e.g., due to ecological factors such 869 as nutrient availability (Raval et al. 2023). Even in the current study, tRNAANN were more likely 870 to affect fitness in nutrient poor media (Fig. 4,5, Fig. S9A-C); whereas in nutrient-rich media, 871 most tRNAANN, particularly 4D codon tRNA ANN, were neutral. It is therefore expected that if 872 some tRNAANN genes enhance translation capacity , they may be retained due to drift or 873 positive selection following nutrient upshifts . Fitness assays during nutrient fluctuations, 874 including competition with WT, might reveal further fitness effects of unmodified tRNAANN. 875 In closing, t he near-complete absence of unmodified tRNAANN remains intriguing, more so 876 given the largely neutral and sometimes beneficial effects that we observed. At a population 877 size of a billion and assuming a 10% fitness advantage (i.e., selection coefficient s=0.1), a 878 mutation creating a tRNAANN from a tRNABNN has about 17% chance of reaching fixation in an 879 average of 410 generations, as inferred from 1000 discrete -time Wright–Fisher transitions 880 (Fig. S12). Even a tRNAANN with a 10% disadvantage can persist for tens of generations, and 881 a tRNAANN with a disadvantage smaller than 1% can persist for over a thousand generations. 882 Nonetheless, tRNAANN are remarkably rare. Although cases with negative fitness effects of 2D 883 tRNAANN suggest that selection for translational fidelity likely contributes to their absence, the 884 neutral and beneficial effects observed for 4D tRNAANN suggests that other fundamental 885 constraints on these molecules, operating in nature, remain hidden. 886 887 DATA AVAILABILITY 888 Supplementary figures are available with this submission. Raw reads for YAMAT -seq are 889 available on from NCBI GEO; accession number GSE328815 . Additional data are available 890 on Zenodo (10.5281/zenodo.20180916) which includes, but is not limited to, the following 891 supplementary data and in-house scripts: 892 Source data for the main and supplementary figures 893 Table S1 (Primers and strains from this study) 894 Table S2 (tRNA gene mutations in long term evolution experiment) 895 Scripts, Source files for the scripts, Output files and plots. 896 897

898 We thank Saurabh Mahajan, Laasya Samhita, S upratim Sengupta and members of the 899 Agashe lab for discussion and critical comments on the manuscript; the NCBS NGS facility for 900 help with genome and YAMAT sequencing; Gaurav Diwan and Joshua Miranda for setting up 901 and maintaining our automated growth measurement system; Gunda Dechow-Seligmann and 902 Sven Künzel for helping with data collection for YAMAT-seq; George Stoletov, Raagini Biswas, 903 Adrita Chakraborty and Pratibha Sanjenbam for helping with the experiments; and the NCBS 904 laboratory kitchen staff for their crucial support. PKR acknowledges the use of ChatGPT 905 (OpenAI) for assistance with writing Python scripts. We acknowledge funding and support 906 from the National Centre for Biological Sciences (NCBS-TIFR) and the Department of Atomic 907 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint Energy, Government of India (Project Identification No. RTI 4006) to DA, a CSIR-UGC-NET 908 June/2018/430 fellowship to PKR, the Max Planck Society (JG and SL), and the International 909 Max Planck Research School for Evolutionary Biology (SL). 910 911 AUTHOR CONTRIBUTIONS 912 PKR: Conceptualization, Experimental design, Methodology, Investigation, Data curation, 913 Validation, Formal analysis, Visualization, Writing - original draft, review and editing. SL: 914 Investigation, Data curation, Formal analysis JG: Experimental design, Funding acquisition, 915 Resources, Methodology, Writing - review and editing. DA: Conceptualization, Project 916 administration, Supervision, Experimental design, Funding acquisition, Resources, Writing - 917 review and editing. 918 919

920 Agarwal D, Kamath D, Gregory ST, O’Connor M. 2015. Modulation of decoding fidelity by 921 ribosomal proteins S4 and S5. J. Bacteriol. 197:1017–1025. 922 Agris PF, Eruysal ER, Narendran A, Väre VYP, Vangaveti S, Ranganathan S V. 2018. 923 Celebrating wobble decoding: Half a century and still much is new. RNA Biol. 15:537–924 553. 925 Agris PF, Narendran A, Sarachan K, Väre VYP, Eruysal E. 2017. The Importance of Being 926 Modified: The Role of RNA Modifications in Translational Fidelity. Enzymes (Essen). 927 41:1–50. 928 Agris PF, Vendeix FAP, Graham WD. 2007. tRNA’s Wobble Decoding of the Genome: 40 929 Years of Modification. J. Mol. Biol. 366:1–13. 930 Aldinger CA, Leisinger AK, Gaston KW, Limbach PA, Igloi GL. 2012. The absence of A-to-I 931 editing in the anticodon of plant cytoplasmic tRNAArgACG demands a relaxation of the 932 wobble decoding rules. RNA Biol. 9:1239–1246. 933 Andachi Y, Yamao F, Iwami M, Muto A, Osawa S. 1987. Occurrence of unmodified adenine 934 and uracil at the first position of anticodon in threonine tRNAs in Mycoplasma 935 capricolum (codon recognition/unmodified anticodon/biased mutation pressure). Proc. 936 Nati. Acad. Sci. USA 84:7398–7402. 937 Arad D, Fargeon O, Levin L, Svenningsen S Lo, Aspit L, Bar-Yaacov D. 2026. Landscape 938 and dynamics of TadA-dependent RNA editing in Escherichia coli reveal a role in 939 nutrient-rich growth.Hinton D, editor. mBio. 940 Ayan GB, Park HJ, Gallie J. 2020. The birth of a bacterial tRNA gene by large-scale, tandem 941 duplication events. Elife 9. 942 Bacher JM, Rie De Cré Cy-Lagard V, Schimmel PR. 2004. Inhibited cell growth and protein 943 functional changes from an editing-defective tRNA synthetase. Proc Natl Acad Sci U S 944 A . 102:1697–1701. 945 Balasubramaniam M, Reis RJS, Ayyadevara S, Wang X, Ganne A, Khaidakov M. 2017. 946 Involvement of tRNAs in replication of human mitochondrial DNA and modifying effects 947 of telomerase. Mech. Ageing Dev. 166:55–63. 948 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint Bedhomme S, Amorós-Moya D, Valero LM, Bonifaci N, Pujana MÀ, Bravo IG, Gonzalez J. 949 2019. Evolutionary Changes after Translational Challenges Imposed by Horizontal 950 Gene Transfer. Genome Biol. Evol. 11:814–831. 951 Beebe K, Ribas de Pouplana L, Schimmel P. 2003. Elucidation of tRNA-dependent editing 952 by a class II tRNA synthetase and signi®cance for cell viability. EMBO J. 22:668–675. 953 Berg MD, Zhu Y, Genereaux J, Ruiz BY, Rodriguez-Mias RA, Allan T, Bahcheli A, Villén J, 954 Brandl CJ. 2019. Modulating mistranslation potential of tRNASer in saccharomyces 955 cerevisiae. Genetics 213:849–863. 956 Berg OG, Kurland CG. 1997. Growth rate-optimised tRNA abundance and codon usage. J. 957 Mol. Biol. 270:544–550. 958 Biddle W, Schmitt MA, Fisk JD. 2016. Modification of orthogonal tRNAs: Unexpected 959 consequences for sense codon reassignment. Nucleic Acids Res. 44:10042–10050. 960 Bloom-Ackermann Z, Navon S, Gingold H, Towers R, Pilpel Y, Dahan O. 2014. A 961 comprehensive tRNA deletion library unravels the genetic architecture of the tRNA 962 pool. PLoS Genet. 10:e1004084. 963 Borén T, Elias P, Samuelsson T, Claesson C, Barciszewska M, Gehrke CW, Kuo KC, Lustig 964 F. 1993. Undiscriminating codon reading with adenosine in the wobble position. J. Mol. 965 Biol. 230:739–749. 966 Buchfink B, Xie C, Huson DH. 2014. Fast and sensitive protein alignment using DIAMOND. 967 Nat. Methods 12:59–60. 968 Chan PP, Lowe TM. 2009. GtRNAdb: a database of transfer RNA genes detected in 969 genomic sequence. Nucleic Acids Res. 37:D93-7. 970 Chan PP, Lowe TM. 2016. GtRNAdb 2.0: An expanded database of transfer RNA genes 971 identified in complete and draft genomes. Nucleic Acids Res. 44:D184–D189. 972 Chen P, Qian Q, Zhang S, Isaksson LA, Björk GR. 2002. A cytosolic tRNA with an 973 unmodified adenosine in the wobble position reads a codon ending with the non-974 complementary nucleoside cytidine. J. Mol. Biol. 317:481–492. 975 Chumpolkulwong N, Hori-Takemoto C, Hosaka T, Inaoka T, Kigawa T, Shirouzu M, Ochi K, 976 Yokoyama S. 2004. Effects of Escherichia coli ribosomal protein S12 mutations on cell-977 free protein synthesis. Eur. J. Biochem. 271:1127–1134. 978 Crick FHC. 1966. Codon—anticodon pairing: The wobble hypothesis. J. Mol. Biol. 19:548–979 555. 980 Curran JF. 1995. Decoding with the A:l wobble pair is inefficient. Nucleic Acids Res. 23. 981 Dahlgren A, Rydén-Aulin M. 2000. A novel mutation in ribosomal protein S4 that affects the 982 function of a mutated RF1. 82:683–691. 983 Datsenko KA, Wanner BL. 2000. One-step inactivation of chromosomal genes in Escherichia 984 coli K-12 using PCR products. Proc. Natl. Acad. Sci. U. S. A. 97:6640–6645. 985 Delaney NF, Kaczmarek ME, Ward LM, Swanson PK, Lee M-C, Marx CJ. 2013. 986 Development of an Optimized Medium, Strain and High-Throughput Culturing Methods 987 for Methylobacterium extorquens. PLoS One 8:1–10. 988 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint Delannoy E, Ret M Le, Faivre-Nitschke E, Estavillo GM, Bergdoll M, Taylor NL, Pogson BJ, 989 Small I, Lmbault P, Gualberto JM. 2009. Arabidopsis tRNA adenosine deaminase 990 arginine edits the wobble nucleotide of chloroplast tRNAArg(ACG) and is essential for 991 efficient chloroplast translation. Plant Cell 21:2058–2071. 992 Dever TE, Dinman JD, Green R. 2018. Translation elongation and recoding in eukaryotes. 993 Cold Spring Harb. Perspect. Biol. 10. 994 Dittmar KA, Goodenbour JM, Pan T. 2006. Tissue-specific differences in human transfer 995 RNA expression. PLoS Genet. 2:2107–2115. 996 Dittmar KA, Mobley EM, Radek AJ, Pan T. 2004. Exploring the regulation of tRNA 997 distribution on the genomic scale. J. Mol. Biol. 337:31–47. 998 Dittmar KA, Sørensen MA, Elf J, Ehrenberg M, Pan T. 2005. Selective charging of tRNA 999 isoacceptors induced by amino-acid starvation. EMBO Rep. 6:151–157. 1000 Diwan GD, Agashe D. 2018. Wobbling forth and drifting back: The evolutionary history and 1001 impact of bacterial tRNA modifications. Mol. Biol. Evol. 35:2046–2059. 1002 Dong H, Nilsson L, Kurland CG. 1996. Co-variation of tRNA abundance and codon usage in 1003 Escherichia coli at different growth rates. J. Mol. Biol. 260:649–663. 1004 Ehrlich R, Davyt M, López I, Chalar C, Marín M. 2021. On the Track of the Missing tRNA 1005 Genes: A Source of Non-Canonical Functions? Front. Mol. Biosci. 8:1–15. 1006 Elf J, Nilsson D, Tenson T, Ehrenberg M. 2003. Selective charging of tRNA isoacceptors 1007 explains patterns of codon usage. Science 300:1718–1722. 1008 Fan Y, Wu J, Ung MH, De Lay N, Cheng C, Ling J. 2015. Protein mistranslation protects 1009 bacteria against oxidative stress. Nucleic Acids Res. 43:1740–1748. 1010 Fujishima K, Kanai A. 2014. tRNA gene diversity in the three domains of life. Front. Genet. 1011 5:1–11. 1012 Gabzi T, Pilpel Y, Friedlander T. 2022. Fitness Landscape Analysis of a tRNA Gene Reveals 1013 that the Wild Type Allele is Sub-optimal, Yet Mutationally Robust. Mol. Biol. Evol. 39. 1014 Gerber AP, Keller W. 1999. An Adenosine Deaminase That Generates Inosine at the 1015 Wobble Position of tRNAs. Science (1979). 286:1146–1149. 1016 Goodenbour JM, Pan T. 2006. Diversity of tRNA genes in eukaryotes. Nucleic Acids Res. 1017 34:6137–6146. 1018 Grosjean H, de Crécy-Lagard V, Marck C. 2010. Deciphering synonymous codons in the 1019 three domains of life: co-evolution with specific tRNA modification enzymes. FEBS 1020 Lett. 584:252–264. 1021 Hamdani O, Dhillon N, Hsieh T-HS, Fujita T, Ocampo J, Kirkland JG, Lawrimore J, 1022 Kobayashi TJ, Friedman B, Fulton D, et al. 2019. tRNA Genes Affect Chromosome 1023 Structure and Function via Local Effects. Mol. Cell. Biol. 39. 1024 Hatfull GF. 2015. Dark Matter of the Biosphere: the Amazing World of Bacteriophage 1025 Diversity. J. Virol. 89:8107–8110. 1026 Hoa SN, Hunt HD, Horton RM, Pullen JK, Peasea LR. 1989. Site-directed mutagenesis by 1027 overlap extension using the polymerase chain reaction. Gene 77:51–59. 1028 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint Holley R.W. AJ, EGA, MJT, MM, MSH, PJR, ZA. 1965. Structure of a ribonucleic acid. 1029 Science (1979). 147:1462–1465. 1030 Hu XP, Lercher MJ. 2021. An optimal growth law for RNA composition and its partial 1031 implementation through ribosomal and tRNA gene locations in bacterial genomes. 1032 PLoS Genet. 17. 1033 Iben JR, Epstein JA, Bayfield MA, Bruinsma MW, Hasson S, Bacikova D, Ahmad D, 1034 Rockwell D, Kittler ELW, Zapp ML, et al. 2011. Comparative whole genome sequencing 1035 reveals phenotypic tRNA gene duplication in spontaneous Schizosaccharomyces 1036 pombe la mutants. Nucleic Acids Res. 39:4728–4742. 1037 Iben JR, Maraia RJ. 2012. tRNAomics: tRNA gene copy number variation and codon use 1038 provide bioinformatic evidence of a new anticodon:codon wobble pair in a eukaryote. 1039 RNA 18:1358–1372. 1040 Ikemura T. 1985. Codon usage and tRNA content in unicellular and multicellular organisms. 1041 Mol. Biol. Evol. 2:13–34. 1042 Inagaki Y, Kojima A, Bessho Y, Hori H, Ohama T, Osawa S. 1995. Translation of 1043 Synonymous Codons in Family Boxes by Mycoplasma capricolum tRNAs with 1044 Unmodified Uridine or Adenosine at the First Anticodon Position. J. Mol. Biol:486–492. 1045 Ingram VM& SJA. 1963. Studies on the structure of purified alanine and valine transfer RNA 1046 from yeast. In: synthesis and structure of macromolecules. Cold. Spr. Harb. Symp. 1047 Quant. Biol. 28:133–138. 1048 Kanaya S, Yamada Y, Kudo Y, Ikemura T. 1999. Studies of codon usage and tRNA genes of 1049 18 unicellular organisms and quantification of Bacillus subtilis tRNAs: gene expression 1050 level and species-specific diversity of codon usage based on multivariate analysis. 1051 Gene 238:143–155. 1052 Karcher D, Bock R. 2009. Identification of the chloroplast adenosine-to-inosine tRNA editing 1053 enzyme. RNA 15:1251–1257. 1054 Katz A, Elgamal S, Rajkovic A, Ibba M. 2016. Non-canonical roles of tRNAs and tRNA 1055 mimics in bacterial cell biology. Mol. Microbiol. 101:545–558. 1056 Kelly P, Backes N, Mohler K, Buser C, Kavoor A, Rinehart J, Phillips G, Ibba M. 2019. 1057 Alanyl-tRNA synthetase quality control prevents global dysregulation of the escherichia 1058 coli proteome. mBio 10. 1059 Khomarbaghi Z, Ngan WY, Ayan GB, Lim S, Dechow-Seligmann G, Nandy P, Gallie J. 2024. 1060 Large-scale duplication events underpin population-level flexibility in tRNA gene copy 1061 number in Pseudomonas fluorescens SBW25. Nucleic Acids Res. 52:2446–2462. 1062 Klumpp S, Scott M, Pedersen S, Hwa T. 2013. Molecular crowding limits translation and cell 1063 growth. Proc. Natl. Acad. Sci. U. S. A. 110:16754–16759. 1064 Kohanski MA, Dwyer DJ, Wierzbowski J, Cottarel G, Collins JJ. 2008. Mistranslation of 1065 Membrane Proteins and Two-Component System Activation Trigger Antibiotic-Mediated 1066 Cell Death. Cell 135:679–690. 1067 Kompatscher M, Bartosik K, Erharter K, Plangger R, Juen FS, Kreutz C, Micura R, Westhof 1068 E, Erlacher MD. 2024. Contribution of tRNA sequence and modifications to the 1069 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint decoding preferences of E. coli and M. mycoides tRNAGlyUCC for synonymous glycine 1070 codons. Nucleic Acids Res. 52:1374–1386. 1071 Lei L, Burton ZF. 2020. Evolution of life on earth: TRNA, aminoacyl-tRNA synthetases and 1072 the genetic code. . 10. 1073 Li C, Qian W, Maclean CJ, Zhang J. 2016. The fitness landscape of a tRNA gene. Science 1074 (1979). 352:837–840. 1075 Li C, Zhang J. 2018. Multi-environment fitness landscapes of a tRNA gene. Nat. Ecol. Evol. 1076 2:1025–1032. 1077 Li J-N, Esberg B, Curran JF, Bjo GR, Rk È. 1997. Three Modified Nucleosides Present in the 1078 Anticodon Stem and Loop Influence the in vivo aa-tRNA Selection in a tRNA-dependent 1079 Manner. J. Mol. Biol 271:209–221. 1080 Link AJ, Phillips † Dereth, Church GM. 1997. Methods for Generating Precise Deletions and 1081 Insertions in the Genome of Wild-Type Escherichia coli: Application to Open Reading 1082 Frame Characterization. J. Bacteriol. 179:6228–6237. 1083 Lyu Z, Wilson C, Ling J. 2023. Translational Fidelity during Bacterial Stresses and Host 1084 Interactions. Pathogens 12. 1085 Magee R, Rigoutsos I. 2020. On the expanding roles of tRNA fragments in modulating cell 1086 behavior. Nucleic Acids Res. 48:9433–9448. 1087 Mangroo D, RajBhandary UL. 1995. Mutants of Escherichia coli initiator tRNA defective in 1088 initiation. Effects of overproduction of methionyl-tRNA transformylase and the initiation 1089 factors IF2 and IF3. Journal of Biological Chemistry 270:12203–12209. 1090 Maraia RJ, Arimbasseri AG. 2017. Factors that shape eukaryotic tRNAomes: Processing, 1091 modification and anticodon-codon use. Biomolecules 7. 1092 McDonald MJ, Chou CH, Swamy KBS, Huang H Da, Leu JY. 2015. The evolutionary 1093 dynamics of tRNA-gene copy number and codon-use in E. coli. BMC Evol. Biol. 15:1–1094 10. 1095 Morgado S, Vicente AC. 2019. Global in-silico scenario of tRNA genes and their organization 1096 in virus genomes. Viruses 11. 1097 Nakanishi K, Fukai S, Ikeuchi Y, Soma A, Sekine Y, Suzuki T, Nureki O. 2005. Structural 1098 basis for lysidine formation by ATP pyrophosphatase accompanied by a lysine-specific 1099 loop and a tRNA-recognition domain. Proc. Natl. Acad. Sci. U. S. A. 102:7487–7492. 1100 Von Nickisch-Rosenegk M, Brown WM, Boore JL. 2001. Complete Sequence of the 1101 Mitochondrial Genome of the Tapeworm Hymenolepis diminuta: Gene Arrangements 1102 Indicate that Platyhelminths Are Eutrochozoans. Mol. Biol. Evol 18:721–730. 1103 Nilsson AI, Zorzet A, Kanth A, Dahlströ S, Berg OG, Andersson DI. 2006. Reducing the 1104 fitness cost of antibiotic resistance by amplification of initiator tRNA genes. Proc. Natl. 1105 Acad. Sci. U. S. A. 103:6976–6981. 1106 Norrander J, Kempe T, Messing J. 1983. Construction of improved Ml3 vectors using 1107 oligodeoxynucleotidedirected mutagenesis. Gene 26:935–7335. 1108 Novoa EM, Pavon-Eternod M, Pan T, Ribas De Pouplana L. 2012. A role for tRNA 1109 modifications in genome structure and codon usage. Cell 149:202–213. 1110 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint Novoa EM, Ribas de Pouplana L. 2012. Speeding with control: Codon usage, tRNAs, and 1111 ribosomes. Trends in Genetics 28:574–581. 1112 O’connor M, Ulrich Goringer H, Dahlberg AE, Planck M-. 1992. A ribosomal ambiguity 1113 mutation in the 530 loop of E.coli 16S rRNA. Nucleic Acids Res. 20:4221–4227. 1114 Pernod K, Schaeffer L, Chicher J, Hok E, Rick C, Geslain R, Eriani G, Westhof E, 1115 Ryckelynck M, Martin F. 2021. The nature of the purine at position 34 in tRNAs of 4-1116 codon boxes is correlated with nucleotides at positions 32 and 38 to maintain decoding 1117 fidelity. Nucleic Acids Res. 48:6170–6183. 1118 Phillips G, de Crécy-Lagard V. 2011. Biosynthesis and function of tRNA modifications in 1119 Archaea. Curr. Opin. Microbiol. 14:335–341. 1120 Polacek N, Ivanov P. 2020. The regulatory world of tRNA fragments beyond canonical tRNA 1121 biology. RNA Biol. 17:1057–1059. 1122 Pope WH, Bowman CA, Russell DA, Jacobs-Sera D, Asai DJ, Cresawn SG, Jacobs William 1123 R J, Hendrix RW, Lawrence JG, Hatfull GF, et al. 2015. Whole genome comparison of 1124 a large collection of mycobacteriophages reveals a continuum of phage genetic 1125 diversity.Kolter R, editor. Elife 4:e06416. 1126 Qian Q, Bjo GR, Rk È. 1997. Structural Alterations Far from the Anticodon of the tRNA Pro 1127 GGG of Salmonella typhimurium Induce +1 Frameshifting at the Peptidyl-site. J. Mol. 1128 Biol. 273:978–992. 1129 Rafels-Ybern À, Torres AG, Camacho N, Herencia-Ropero A, Frigolé HR, Wulff TF, Raboteg 1130 M, Bordons A, Grau-Bove X, Ruiz-Trillo I, et al. 2019. The expansion of inosine at the 1131 wobble position of tRNAs, and its role in the evolution of proteomes. Mol. Biol. Evol. 1132 36:650–662. 1133 Rafels-Ybern À, Torres AG, Grau-Bove X, Ruiz-Trillo I, Ribas de Pouplana L. 2018. Codon 1134 adaptation to tRNAs with Inosine modification at position 34 is widespread among 1135 Eukaryotes and present in two Bacterial phyla. RNA Biol. 15:500–507. 1136 Rak R, Dahan O, Pilpel Y. 2018. Repertoires of tRNAs: The Couplers of Genomics and 1137 Proteomics. Annu. Rev. Cell Dev. Biol. 34:239–264. 1138 Rao AR, Varchney U. 2001. Specific interaction between the ribosome recycling factor and 1139 the elongation factor G from Mycobacterium tuberculosis mediates peptidyl-tRNA 1140 release and ribosome recycling in Escherichia coli. EMBO J. 20:2977–2986. 1141 Raval PK, Ngan WY, Gallie J, Agashe D. 2023. The layered costs and benefits of 1142 translational redundancy. Elife 12:e81005. 1143 Rocha. 2004. Codon usage bias from tRNA’s point of view: Redundancy, specialization, and 1144 efficient decoding for translation optimization. Genome Res. 14:2279–2286. 1145 Rodnina M V., Wintermeyer W. 2009. Recent mechanistic insights into eukaryotic 1146 ribosomes. Curr. Opin. Cell Biol. 21:435–443. 1147 Romero Romero ML, Poehls J, Kirilenko A, Richter D, Jumel T, Shevchenko A, Toth-1148 Petroczy A. 2024. Environment modulates protein heterogeneity through transcriptional 1149 and translational stop codon readthrough. Nat. Commun. 15:4446. 1150 Roura Frigolé H, Camacho N, Coma MC, Fernández-Lozano C, García-Lema J, Rafels-1151 Ybern À, Canals A, Coll M, De Pouplana LR. 2019. tRNA deamination by ADAT 1152 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint requires substrate-specific recognition mechanisms and can be inhibited by tRFs. RNA 1153 25:607–619. 1154 Ruusala T, Andersson D, Ehrenberg M, Kurland CG. 1984. Hyper-accurate ribosomes inhibit 1155 growth. EMBO J. 3:2575–2580. 1156 Rydbn SM, Isaksson LA. 1984. A Temperature-Sensitive Mutant of Escherichia coli that 1157 Shows Enhanced Misreading of UAG[A and Increased Efficiency for Some tRNA 1158 Nonsense Suppressors. Mol Gen Genet 193:38–45. 1159 Sagi D, Rak R, Gingold H, Adir I, Maayan G, Dahan O, Broday L, Pilpel Y, Rechavi O. 2016. 1160 Tissue- and Time-Specific Expression of Otherwise Identical tRNA Genes. PLoS 1161 Genet. 12. 1162 Saint-Léger A, Bello C, Dans PD, Torres AG, Novoa EM, Camacho N, Orozco M, 1163 Kondrashov FA, De Pouplana LR. 2016. Saturation of recognition elements blocks 1164 evolution of new tRNA identities. Sci. Adv. 2:e1501860. 1165 Samhita L. 2022. Re-reading the genetic code: The evolutionary potential of frameshifting in 1166 time. J. Biosci. 47:49. 1167 Samhita L, K Raval P, Stephenson G, Thutupalli S, Agashe D. 2021. The impact of 1168 mistranslation on phenotypic variability and fitness. Evolution (N. Y). 75:1201–1217. 1169 Samhita L, Raval PK, Agashe D. 2020. Global mistranslation increases cell survival under 1170 stress in Escherichia coli. PLoS Genet. 16:1–21. 1171 Sane M, Parveen S, Agashe D. 2025. Mutation bias alters the distribution of fitness effects of 1172 mutations. PLoS Biol. 23(7):e3003282. 1173 Schmeing TM, Ramakrishnan V. 2009. What recent ribosome structures have revealed 1174 about the mechanism of translation. Nature 461:1234–1242. 1175 Schmeing TM, Voorhees RM, Kelley AC, Ramakrishnan V. 2011. How mutations in tRNA 1176 distant from the anticodon affect the fidelity of decoding. Nat. Struct. Mol. Biol. 18:432–1177 437. 1178 Schmitt MA, Biddle W, Fisk JD. 2018. Mapping the Plasticity of the Escherichia coli Genetic 1179 Code with Orthogonal Pair-Directed Sense Codon Reassignment. Biochemistry 1180 57:2762–2774. 1181 Schmitt MA, Tittle JM, Fisk JD. 2024. Codon decoding by orthogonal tRNAs interrogates the 1182 in vivo preferences of unmodified adenosine in the wobble position. Front. Genet. 15. 1183 Scott M, Gunderson CW, Mateescu EM, Zhang Z, Hwa T. 2010. Interdependence of Cell 1184 Growth and Gene Expression: Origins and Consequences. Science (1979). 330:1099–1185 1102. 1186 Scott M, Hwa T. 2011. Bacterial growth laws and their applications. Curr. Opin. Biotechnol. 1187 22:559–565. 1188 Seligmann H. 2010. Mitochondrial tRNAs as light strand replication origins: Similarity 1189 between anticodon loops and the loop of the light strand replication origin predicts 1190 initiation of DNA replication. BioSystems 99:85–93. 1191 Sengupta S, Yang X, Higgs PG. 2007. The mechanisms of codon reassignments in 1192 mitochondrial genetic codes. J. Mol. Evol. 64:662–688. 1193 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint Shigematsu M, Honda S, Loher P, Telonis AG, Rigoutsos I, Kirino Y. 2017. YAMAT-seq: An 1194 efficient method for high-throughput sequencing of mature transfer RNAs. Nucleic Acids 1195 Res. 45:e70. 1196 Sibler AP, Dirheimer G, Martin RP. 1986. Codon reading patterns in Saccharomyces 1197 cerevisiae mitochondria based on sequences of mitochondrial tRNAs. FEBS Lett. 1198 194:131–138. 1199 Soma A, Kubota A, Tomoe D, Ikeuchi Y, Kawamura F, Arimoto H, Shiwa Y, Kanesaki Y, 1200 Nanamiya H, Yoshikawa H, et al. 2023. yaaJ, the tRNA-Specific Adenosine 1201 Deaminase, Is Dispensable in Bacillus subtilis. Genes (Basel). 14:1515. 1202 Su Z, Wilson B, Kumar P, Dutta A. 2020. Noncanonical Roles of tRNAs: tRNA Fragments 1203 and Beyond. Annu. Rev. Genet. 54:47–69. 1204 Suzuki T. 2021. The expanding world of tRNA modifications and their disease relevance. 1205 Nat. Rev. Mol. Cell Biol. 22:375–392. 1206 Thorbjarnardottir S, Bjornsson A, Amundad6ttir L, Eggertsson G. 1991. Temperature 1207 Sensitivity Caused by Missense Suppressor supH and Amber Suppressor supP in 1208 Escherichia coli. J. Bacteriol. 173:412–416. 1209 Thornlow BP, Armstrong J, Holmes AD, Howard JM, Corbett-Detig RB, Lowe TM. 2020. 1210 Predicting transfer RNA gene activity from sequence and genome context. Genome 1211 Res. 30:85–94. 1212 Thornlow BP, Hough J, Roger JM, Gong H, Lowe TM, Corbett-Detig RB. 2018. Transfer 1213 RNA genes experience exceptionally elevated mutation rates. Proc. Natl. Acad. Sci. U. 1214 S. A. 115:8996–9001. 1215 Tirumalai MR, Rivas M, Tran Q, Fox GE. 2021. The Peptidyl Transferase Center: a Window 1216 to the Past. Microbiol Mol Biol Rev 85:1055–1067. 1217 Torres AG. 2019. Enjoy the Silence: Nearly Half of Human tRNA Genes Are Silent. 1218 Bioinform. Biol. Insights 13. 1219 Torres AG, Piñeyro D, Filonava L, Stracker TH, Batlle E, Ribas De Pouplana L. 2014. A-to-I 1220 editing on tRNAs: Biochemical, biological and evolutionary implications. FEBS Lett. 1221 588:4279–4286. 1222 Torres AG, Piñeyro D, Rodríguez-Escribà M, Camacho N, Reina O, Saint-Léger A, Filonava 1223 L, Batlle E, Ribas De Pouplana L. 2015. Inosine modifications in human tRNAs are 1224 incorporated at the precursor tRNA level. Nucleic Acids Res. 43:5145–5157. 1225 Torres AG, Rodríguez-Escribà M, Marcet-Houben M, Santos Vieira HG, Camacho N, Catena 1226 H, Murillo Recio M, Rafels-Ybern À, Reina O, Torres FM, et al. 2021. Human tRNAs 1227 with inosine 34 are essential to efficiently translate eukarya-specific low-complexity 1228 proteins. Nucleic Acids Res. 49:7011–7034. 1229 Vieira-Silva S, Rocha EPC. 2010. The systemic imprint of growth and its uses in ecological 1230 (meta)genomics. PLoS Genet. 6:e1000808. 1231 Watanabe Y-I, Tsurui H, Ueda T, Furusihima-Shimogawara R, Takamiya S, Kita K, 1232 Nishikawa K, Watanabe K. 1997. Primary sequence of mitochondrial tRNA Arg of a 1233 nematode Ascaris suum: occurrence of unmodified adenosine at the first position of the 1234 anticodon. Biochim. Biophys. Acta 1350:119–122. 1235 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint Wilusz JE. 2015. Controlling translation via modulation of tRNA levels. Wiley Interdiscip. 1236 Rev. RNA 6:453–470. 1237 Wolf J, Gerber AP, Keller W. 2002. tadA, an essential tRNA-specic adenosine deaminase 1238 from Escherichia coli. EMBO J. 21:3841–3851. 1239 Wulff TF, Hahnke K, Lécrivain AL, Schmidt K, Ahmed-Begrich R, Finstermeier K, 1240 Charpentier E. 2024. Dynamics of diversified A-to-I editing in Streptococcus pyogenes 1241 is governed by changes in mRNA stability. Nucleic Acids Res. 52:11234–11253. 1242 Yokobori SI, Kitamura A, Grosjean H, Bessho Y. 2013. Life without tRNAArg-adenosine 1243 deaminase TadA: Evolutionary consequences of decoding the four CGN codons as 1244 arginine in Mycoplasmas and other Mollicutes. Nucleic Acids Res. 41:6531–6543. 1245 Young RY, Bremer H, Programs B. 1976. Polypeptide-Chain-Elongation Rate in Escherichia 1246 coli B/r as a Function of Growth Rate. Biochem. J 160:185–194. 1247 1248 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted May 15, 2026. ; https://doi.org/10.64898/2026.05.15.725136doi: bioRxiv preprint