Is the Genetic Code Optimized?

The famous result that the genetic code is one in a million is true and misreported by both sides. In the space actually searched, about two trillion codes beat it, and it sits on the slope of a mediocre peak reachable in a dozen swaps. The first laboratory test of the theory behind it failed to find the predicted effect.

Draft:  September 2026

"Statements like 'the genetic code is one in a million' (or even in 100 million) are technically accurate but can be easily misconstrued should one overlook the fact that there is a huge number of possible codes that are significantly more robust than the standard code that sits on the slope of an unremarkable local peak in an extremely rugged fitness landscape." — Eugene Koonin and Artem Novozhilov.1

"In contrast to this prediction, translation experiments using the non-SGCs and mutagenized libraries revealed that the effects of mutations on protein functions were similar across all genetic codes and reporter genes examined here." — the first laboratory test of error minimization, 2026.2

Summary

Every living thing translates the same table of sixty-four codons into twenty amino acids.  The table is not random:  a single-letter mistake usually gives back either the same amino acid or a chemically similar one.  That pattern is real, and both camps misreport what it shows.

  • "One in a million" is a true statement about a restricted space. Freeland and Hurst generated a million alternative codes that keep the standard code's own block structure and degeneracy, and shuffle only which amino acid sits in which block. One beat it.3
  • That space has about 2.4 × 1018 members, so roughly two trillion codes in it are better. The two figures are the same finding stated two ways, and the second is the one nobody quotes.
  • The code is not even a local optimum. It "is not locally stable, that is, it can be readily 'improved' by a small perturbation," and the peak it sits on "is rather mediocre, with a huge number of taller peaks existing in the landscape."1
  • Starting from a random block-structured code, the standard code's level of optimisation "can be easily achieved with 10–12 evolutionary steps on average."1 A dozen swaps, not an astronomical search.
  • The numbers in circulation measure different things and must never be quoted side by side. The minimisation percentage on the polar-requirement scale is 70 to 78%.1 The probability of picking an equally robust random code is 10-7 to 10-8.4 And the space of unrestricted codes is about 1084. Mixing those denominators is the commonest error on this subject.
  • The first laboratory test of the theory behind all of it failed. Ten non-standard codes were built in a test tube and mutagenised. Mutation effects "were similar across all genetic codes."2 The same paper found the standard code is the multi-property cost optimum among 19,683 candidates.2 Report both halves or neither.
  • The code is neither universal nor forced. A screen of 251,571 genomes found five new reassignments.5 A ciliate decodes UAA and UAG as two different amino acids.6 And all twenty transfer RNA classes have been converted to four-base versions: "the triplet codon code was selected because it is simpler and sufficient, not because a quadruplet codon code is unachievable."7

The verdict.  The code is good, in a measurable sense, and it is not remarkable.  The design argument from its optimality fails on the field's own numbers.  What survives is the older and better question, which nobody has answered:  where the assignments came from.  Koonin's summary after fifty years is that the work "failed to develop a compelling explanation for codon assignments."4

What Freeland and Hurst actually did

The 1998 paper is titled "The Genetic Code Is One in a Million," which is a gift to anyone quoting titles.  Its abstract says:  "if we employ weightings to allow for biases in translation, then only 1 in every million random alternative codes generated is more efficient than the natural code."3

The load-bearing detail is what "random alternative code" means, and it is almost never stated.  These are not arbitrary mappings of sixty-four codons onto twenty-one meanings.  They are codes built by shuffling the twenty amino acids among the standard code's own synonymous codon blocks, keeping the block structure and the degeneracy pattern exactly as they are.  Only the labels move.

That space has 20 factorial members, about 2.4 × 1018.  A hit rate of one in a million therefore means roughly 2.4 × 1012 codes in the same space are better than ours.  Two trillion.

Both camps get this wrong in opposite directions, and Koonin and Novozhilov, who accept the underlying result, name both errors.

The apologetics error is to read "one in a million" as near-uniqueness.  Their sentence is the epigraph above:  technically accurate, easily misconstrued, and there is a huge number of more robust codes.

The sceptic's error is to conclude there is nothing to explain.  The same authors:  "no attempt to explain the observed robustness of the code to translation errors without invoking at least some extent of selection has been convincing.  So it does appear that selection for translation error minimization played a substantial role in the evolution of the code to the standard form."1

The numbers, and why they cannot be mixed

Four different figures circulate.  They are not in conflict.  They answer different questions.

Figure What it measures
1 in 106 hit rate among block-preserving shuffles, with translation bias weighting
10-7 to 10-8 probability of picking an equally robust random code, across several cost functions
70 to 78% minimisation percentage: how far the code has travelled from an average random code toward the best found
~1084 size of the unrestricted code space, in which "there is a huge number with a higher degree of error minimization"

The last two sit in one sentence of Koonin's 2017 review:  the code "is indeed highly resilient, with the probability to pick an equally robust random code being on the order of 10-7–10-8…  Obviously, however, among the ~1084 possible random codes, there is a huge number with a higher degree of error minimization than the SGC."4

The minimisation percentage is the least flattering figure and the most rarely quoted.  On the polar-requirement scale the standard code has covered about 70% of the distance from an average random code to the best code found by optimisation, or 78% measured against one specific optimised code.1  A student who scores 70% has not scored a global optimum.

The code is not a peak

The strongest single result against a design argument from optimality is that the code is not even locally optimal.

Novozhilov, Wolf and Koonin ran a greedy optimisation over the code landscape.  Two findings, both quoted here from Koonin and Novozhilov's review of that work.

The code can be improved by a single swap.  "It should be emphasized that, under this model, the standard code is not locally stable, that is, it can be readily 'improved' by a small perturbation (an additional swap)."1  Ours sits "about half way (measured in the number of codon series swaps) along an upward evolutionary trajectory from a random code to the summit of the respective local peak.  Moreover, this peak is rather mediocre, with a huge number of taller peaks existing in the landscape."

And getting to it from a random code is cheap.  "Rather unexpectedly, starting from a random code, the level of optimization of the standard code can be easily achieved with 10–12 evolutionary steps on average."1

Ten to twelve swaps.  That is the number that decides the question, because a design argument from optimality needs the code to be hard to reach and hard to beat.  On the field's own cost function it is neither.

One honest caveat, and it is Koonin's own.  "It cannot be ruled out that the fitness functions employed in modeling selection for error minimization… are far from being an accurate representation of the 'real' optimization criterion.  Should that be the case, the general assessment of the entire field of code evolution would have to be particularly somber because that would imply we have no clue as to what is important in a code."1  Every number in this article rests on a cost function that somebody chose.

He adds a second methodological warning that applies to every optimality figure ever published on this subject.  These analyses rearrange the existing code table, which "allows formal … estimation of robustness but not the evolutionary processes that lead to it."4

An optimality claim that collapsed when its inputs were checked

This one is worth following, because it shows how such figures are built and how they fail.

The coevolution theory says the code's structure reflects the order in which amino acid biosynthesis pathways appeared, so precursor and product amino acids should sit in related codons.  Tze-Fei Wong proposed it in 1975 using eight precursor-product pairs.8  Rubén Amirnovin then tested those pairs against random codes.  The results, as Koonin reports them:

Input set Fraction of random codes that do better
Wong's original eight pairs 0.1%
Removing Gln-His and the questioned Val-Leu pair 3.6%
Precursor-product pairs from E. coli's characterised metabolism 34%

The signal all but disappears once the inputs stop being assumed.1  Any optimality figure on this subject should be checked for the same problem before it is quoted, including the ones this article prints.

The first experimental test

Every number above is computational.  In July 2026 a Japanese group did the experiment.

Ryota Miyachi and Norikazu Ichihashi built ten non-standard genetic codes in a reconstituted, transfer-RNA-free translation system, by reassigning alanine, serine and leucine to vacant codons of a twenty-one-tRNA minimal code.  Then they mutagenised three reporter genes and measured how much function was lost.  Error minimization predicts that a code closer to the standard one loses less.

The prediction failed.  "In contrast to this prediction, translation experiments using the non-SGCs and mutagenized libraries revealed that the effects of mutations on protein functions were similar across all genetic codes and reporter genes examined here."2

And the same paper found the opposite result on a different measure.  Combining polar requirement, molecular volume and hydropathy into a single index, "the near-SGC showed the lowest overall cost when compared with all 19,683 candidate non-SGCs."2  The standard code is the multi-property cost optimum in their system, and the robustness that optimum is supposed to buy could not be detected.

Report both halves or neither.  The authors are careful about reach and this article is too:  the finding covers individual reporter proteins in a reconstituted system, and only three amino acids could be reassigned.  The codes tested span 18 to 51% of the cost range covered by a million randomly sampled codes, so they are not extreme alternatives.

One trap for anyone citing this paper.  Its editorial assessment reads "The data provide solid support for the claim of having provided experimental verification of the error minimization theory," which sounds like confirmation and, read against the body, means the method is sound rather than the prediction confirmed.  Anyone quoting the assessment line alone will get the paper backwards.2

Could the code have been otherwise?

"The code is universal, so something fixed it at the start" is a common argument.  The evidence answers it in three directions.

The code is not universal, and the exceptions are not curiosities.  Yekaterina Shulgina and Sean Eddy screened "251,571 assembled genomes from GenBank" and "discovered five new reassignments of arginine codons (AGG, CGA, and CGG), representing the first sense codon changes in bacteria."5  Twenty-seven translation tables are now catalogued.  And a ciliate has been found in which UAA and UAG, which everyone treated as a matched pair of stop codons, encode two different amino acids.6

A triplet code is not the only workable kind.  Erika DeBenedictis, Dieter Söll and Kevin Esvelt converted transfer RNAs "from all 20 canonical isoacceptor classes" into functional four-base versions.  Their conclusion:  "the triplet codon code was selected because it is simpler and sufficient, not because a quadruplet codon code is unachievable."7  The limit they found is on the enzyme side:  "only nine of the twenty aminoacyl tRNA synthetases tolerate quadruplet anticodons."

But changing a code is expensive, and that is the other half.  Ákos Nyerges, in George Church's laboratory, built Escherichia coli strains carrying up to 45.8% of a synthetic genome under a deliberately deleterious 57-codon code, and reports flatly that "synonymous codon replacement — recoding — is frequently lethal."9

Put the three together and the picture is a fitness barrier rather than a wall or a free parameter.  The code changes, and it changes only where something has already driven a codon nearly out of use.  Shulgina and Eddy find the reassignments following the codon-capture route, in which "the codon is first driven to near extinction by pressures unrelated to reassignment, such as biased genomic GC content or genome reduction."5  Reassignment is editing, not authorship.

Where the assignments came from

Four families of theory compete.  Stereochemical affinity between a codon and its amino acid.  Coevolution with biosynthesis pathways.  Error minimisation.  And Crick's frozen accident, in which the assignments are arbitrary and fixed because changing them is lethal.

Each explains part of the pattern.  None explains the actual assignments, and the field's most-cited reviewer says so.  Fifty years of work has produced "notable features of the standard code, such as high robustness to errors, but failed to develop a compelling explanation for codon assignments."4

Where real progress has been made is on the order in which amino acids joined the code, and one finding there embarrasses a textbook.  Sawsan Wehbi's group points out that the standard recruitment order rests on "potentially biased criteria, such as the absence of sulfur-containing amino acids from the Urey-Miller experiment which lacked sulfur."  Inferring the order from ancestral protein sequences instead puts "metal-binding and sulfur-containing amino acids earlier than previously thought."10  A 2026 study reaches a converging chronology from sixteen molecular-complexity metrics, which is independent evidence because it is derived from chemistry rather than from any apparatus.11

And there is sequence evidence that the code was still being built after protein synthesis was working.  The pre-LUCA ancestor of the tryptophan and tyrosine synthetases "was shown to be devoid of tryptophan amino acids despite later versions of both synthetases containing tryptophan at several sites," and the common ancestor of the leucine, isoleucine and valine synthetases "did not discriminate between these three amino acids."12  Neither camp's simple story accommodates that.  It is not a code installed complete, and it is not a code assembled before translation began.

The claims, audited

Claim Verdict
"The genetic code is one in a million" True and misleading. Two trillion codes in the same space are better
"The code is optimal" False as stated. Not even locally optimal; one swap improves it
"The code is unreachable by chance" Not supported. 10-12 swaps from a random block-structured code
"The code is universal, so it was fixed at the start" False. Twenty-seven tables, five new reassignments in one screen, a ciliate with two meanings for the old stop codons
"Only a triplet code could work" False. All twenty transfer RNA classes converted to quadruplets
"The code can change freely" False. Recoding "is frequently lethal"; reassignment needs a nearly unused codon first
"Nobody has explained the codon assignments" Sound, and it is Koonin's own summary after fifty years

Objections

1. You are using one cost function

Yes, and so is everyone.  Koonin says so himself, in the passage quoted above about having "no clue as to what is important in a code" if the function is wrong.  That cuts against the optimality argument first, because the impressive-sounding figures are the ones computed from it.  The results that survive a change of function are the qualitative ones:  the code is good, the code is beatable, and the code is reachable.

2. The variant codes are all derived from the standard one, so they prove nothing about origins

Correct, and the article says so.  Every known variant is an edit of the standard code, and every edit follows a codon nearly falling out of use.  What the variants refute is the narrower claim that the code cannot change, which is the form the universality argument usually takes.

3. The 2026 experiment tested only three amino acids

True, and the authors say so.  It is one experiment in a reconstituted system with three reporter proteins.  What it establishes is that the predicted effect was not detectable under conditions designed to detect it, which is worth more than another computation and less than a settled refutation.

4. Quadruplet codes need engineered enzymes, so they are not natural alternatives

Fair.  Only nine of the twenty synthetases tolerated quadruplet anticodons, and the rest needed engineering.  The claim the result supports is the one its authors make:  a quadruplet code is achievable, so the triplet code is not chemically forced.

5. This does not touch the real argument, which is that the code is a code

That is the stronger argument and it belongs in the translation article, which is about the interdependence of the code and the machinery that reads it.  Optimality and codehood are different claims and they should not be run together.  This article is only about optimality, and on that question the design argument fails on the field's own numbers.

6. Retiring this argument gives ground to naturalism

It gives up an argument that does not work.  The alternative is to keep quoting a 1998 title against readers who can find the two-trillion figure in the same literature in ten minutes.  The remaining question, where the assignments came from, is unanswered after fifty years by the field's own account, and it is a better place to stand.

What this article claims, and what it does not

It claims that the code's error robustness is real, that it is measured within a restricted space of codes, that about two trillion codes in that space are better, that the code is not locally optimal, that it is reachable in about a dozen swaps, that the first experimental test of error minimization failed to find the predicted effect, and that the code is neither universal nor chemically forced.

It does not claim the code is unremarkable in every sense;  the same 2026 experiment found it is the multi-property cost optimum among 19,683 candidates.  It does not claim the assignments are explained;  they are not, and that is the field's own verdict.  It has not opened Freeland and Hurst's body, Freeland 2000, Crick 1968, or Amirnovin's papers, and prints no number resting on those alone.  One correction is recorded here rather than buried:  this project's own research file attributed the "10–12 evolutionary steps" and "not locally stable" sentences to Novozhilov, Wolf and Koonin 2007;  they are in Koonin and Novozhilov's 2009 review, which is where this article cites them from.  Those gaps are listed in the literature review.

Sources

  1. 1.

    Koonin, Eugene V., and Artem S. Novozhilov.  “Origin and Evolution of the Genetic Code: The Universal Enigma.”  IUBMB Life 61(2):99–111.  2009.  The minimization figures:  “The minimization percentage of the standard code has been estimated at ~70% when the polar requirement scale is used as the measure of amino acid exchangeability (96, 101).  Figure 2 shows an example of a code that was optimized for robustness to translation errors by swapping codon assignments…  With respect to this code, the minimization percentage of the standard code is 78%.”  The review also states the randomisation procedure that makes "one in a million" mean what it means:  “the one most frequently used involves shuffling of amino acid assignments between the synonymous codon series that are intrinsic to the standard code, so that 20! ≈ 2.4 · 1018 possible codes are searched;  different random code generators can produce substantially different results.”  And the warning about the phrase itself:  “Statements like 'the genetic code is one in a million' (or even in 100 million) are technically accurate but can be easily misconstrued should one overlook the fact that there is a huge number of possible codes that are significantly more robust than the standard code that sits on the slope of an unremarkable local peak in an extremely rugged fitness landscape.”  A review of how the genetic code originated and evolved, by two researchers at the US National Center for Biotechnology Information.  The article takes from it the field's own optimality figures and its warnings about how those figures are quoted.  Mirror:  PMC ↩

    1. 1a.

      On the coevolution theory, with Wong's original eight precursor-product pairs, the review reports that.  “only 0.1% of random codes perform better”.  Removing Gln-His and Val-Leu, it reports that.  “the proportion of better random codes rises to 3.6%”.  Koonin attaches the doubt to Val-Leu alone, writing that.  “the pairs Gln-His and Val-Leu are removed (the validity of the latter pair has been questioned)”.  The review also reports.  “if the precursor-product pairs are taken from the well-characterized metabolic pathways of E. coli, the proportion that a random code shows a stronger correlation reaches 34%.”  The review also notes a wobble recount by Ronneberg and colleagues that removes the statistical support altogether.  The tests were Amirnovin's, as the review reports them. ↩

    2. 1b.

      On the code's position in the landscape, the review says:  “the standard genetic code appears to be a point that is located about half way (measured in the number of codon series swaps) along an upward evolutionary trajectory from a random code to the summit of the respective local peak.  Moreover, this peak is rather mediocre, with a huge number of taller peaks existing in the landscape (Fig. 3).  It should be emphasized that, under this model, the standard code is not locally stable, that is, it can be readily 'improved' by a small perturbation (an additional swap).”  The underlying computation is the 2007 paper's: Novozhilov, Artem S., Yuri I. Wolf and Eugene V. Koonin, "Evolution of the genetic code: partial optimization of a random code for robustness to translation error in a rugged fitness landscape," Biology Direct 2 (2007): 24.  Neither sentence is in that paper; both are in this 2009 review, which summarises it. ↩ ↩

    3. 1c.

      The review's words on the same model:  “Under this model, the code fitness landscape is extremely rugged, so that almost any random code yields its own local maximum.  Rather unexpectedly, starting from a random code, the level of optimization of the standard code can be easily achieved with 10–12 evolutionary steps on average, and often, optimization can be continued to reach the level that is attainable when the optimization starts from the standard code.”  The swaps are of whole codon series.  The 2007 paper's own peer reviewers note that each would require several physical mutations, so twelve swaps are not twelve point mutations.  The figure is quoted as the field gives it, and this note carries the qualification. ↩ ↩

    4. 1d.

      “The minimization percentage of the standard code has been estimated at ~70% when the polar requirement scale is used as the measure of amino acid exchangeability (96, 101).  Figure 2 shows an example of a code that was optimized for robustness to translation errors by swapping codon assignments for amino acids to minimize the value of the cost function given by formula (I).  With respect to this code, the minimization percentage of the standard code is 78%.” ↩ ↩

    5. 1e.

      A caveat that applies to the optimality figures:  “Of course, it cannot be ruled out that the fitness functions employed in modeling selection for error minimization (Eq (I) and similar ones) in the evolution of the code are far from being an accurate representation of the 'real' optimization criterion.  Should that be the case, the general assessment of the entire field of code evolution would have to be particularly somber because that would imply we have no clue as to what is important in a code.”  The next sentence in the review declines that possibility:  “However, this does not seem to be a particularly likely possibility.” ↩

    6. 1f.

      On the coevolution theory, reporting Amirnovin's tests, with Wong's original eight precursor-product pairs:  “only 0.1% of random codes perform better”.  Removing Gln-His and Val-Leu:  “the proportion of better random codes rises to 3.6%”.  Koonin attaches the doubt to Val-Leu alone, writing that.  “the validity of the latter pair has been questioned”.  The review also reports.  “if the precursor-product pairs are taken from the well-characterized metabolic pathways of E. coli, the proportion that a random code shows a stronger correlation reaches 34%.” ↩

  2. 2.

    Miyachi, Ryota, and Norikazu Ichihashi.  “Experimental verification of the error minimization theory using non-standard genetic codes constructed in vitro.”  eLife 15:RP111164.  2026.  Section eLife Assessment:  “The data provide solid support for the claim of having provided experimental verification of the error minimization theory”.  Read against the body, the assessment is about the method being sound, not about the prediction being confirmed.  Anyone citing it from the assessment line alone will get it backwards.  A laboratory experiment (version of record dated 13 July 2026) that built ten non-standard genetic codes in a test tube, in a reconstituted translation system with no cell tRNAs, by reassigning alanine, serine and leucine to empty codons of a 21-tRNA minimal code.  The authors mutated three reporter genes and measured the loss of activity, and the tested codes span 18.4% (polar requirement), 37.6% (molecular volume) and 50.8% (hydropathy index) of the cost range spanned by one million randomly sampled Freeland–Hurst-style codes.  The article uses it as the first laboratory test of the error minimization theory, and it found that mutations did not hurt more under codes with higher cost, which the theory predicts they would. ↩

    1. 2a.

      Section Discussion:  “The error minimization theory predicts that the deleterious effect of mutations on protein function becomes more severe when using genetic codes with higher mutational cost (Freeland et al., 2003;  Freeland and Hurst, 1998;  Haig and Hurst, 1991;  Massey, 2008).  In contrast to this prediction, translation experiments using the non-SGCs and mutagenized libraries revealed that the effects of mutations on protein functions were similar across all genetic codes and reporter genes examined here, suggesting that the mutational robustness of protein activity remained largely unchanged within at least the ranges of mutational cost tested in this study.”  The authors state their own limits:  “It should be noted that this conclusion is limited to the activity of individual reporter proteins translated in a reconstituted in vitro system.  Therefore, whether similar trends would be observed at the level of cellular fitness or long-term evolution remains an open question.”  A second limit:  “only three amino acids (Ala, Ser, and Leu) can be reassigned to different codons.” ↩ ↩ ↩

    2. 2b.

      “In both integrated indices, the near-SGC showed the lowest overall cost when compared with all 19,683 candidate non-SGCs (Figure 5—figure supplement 7), indicating that no candidate non-SGC exhibited a lower combined cost than the near-SGC when the three physicochemical properties were considered jointly.” ↩ ↩

  3. 3.

    Freeland, Stephen J., and Laurence D. Hurst.  “The Genetic Code Is One in a Million.”  Journal of Molecular Evolution 47:238–248.  1998.  “In addition, if we employ weightings to allow for biases in translation, then only 1 in every million random alternative codes generated is more efficient than the natural code.”  The load-bearing detail is the sampled set, and it is almost never stated.  The random alternative codes are not arbitrary mappings of 64 codons onto 21 meanings.  They are codes built by shuffling the twenty amino acids among the standard code's own synonymous codon blocks, keeping block structure and degeneracy fixed, a space of 20! ≈ 2.4 × 1018 codes.  A hit rate of 10-6 therefore implies roughly 2.4 × 1012 codes that beat the standard code.  A 2000 paper by Freeland, Robin D. Knight, Laura F. Landweber and Laurence D. Hurst, "Early Fixation of an Optimal Genetic Code," Molecular Biology and Evolution 17 (2000): 511-518, which says the code is.  “at or very close to a global optimum”.  That claim is conditional in the 2000 abstract, which says.  “if theoretically possible code structures are limited to reflect plausible biological constraints.”  A study that compared the standard genetic code with random alternative codes, asking how many of them do better.  The article uses it for the "one in a million" figure, which is true only of the restricted space that was sampled. ↩ ↩

  4. 4.

    Koonin, Eugene V.  “Frozen Accident Pushing 50: Stereochemistry, Expansion, and Chance in the Evolution of the Genetic Code.”  Life 7(2):22.  2017.  “Extensive quantitative analyses that employed cost functions differently derived from physico-chemical properties of amino acids have shown that the code is indeed highly resilient, with the probability to pick an equally robust random code being on the order of 10-7–10-8 [14,15,61–68].  Obviously, however, among the ~1084 possible random codes, there is a huge number with a higher degree of error minimization than the SGC.  Furthermore, the SGC is not a local peak on the code fitness landscape because certain local rearrangements can increase the level of error minimization;  a quantitatively, the SGC is positioned roughly halfway from an average random code to the summit of the corresponding local peak [15] (Figure 1).”  A concept paper, which proposes ideas rather than reporting an experiment, by a researcher at the US National Center for Biotechnology Information.  It reviews the code's evolution since Francis Crick's 1968 frozen accident idea, and it finds that the code's robustness to error is well established while its codon assignments remain unexplained.  The article uses it for the figures on how robust the code is, and for its point that analyses of robustness do not show how the code evolved. ↩ ↩

    1. 4a.

      On the error minimization theory, the review says:  “The typical conclusion of the error minimization theorists is that the code evolved under selective pressure for robustness.  However, this conclusion is not necessarily justified.  Most if not all code optimization analyses focus on rearrangements of the standard code table, which allows formal (and in itself, valid, given the cost function) estimation of robustness but not the evolutionary processes that lead to it.” ↩

    2. 4b.

      “The subsequent 49 years of code studies have elucidated notable features of the standard code, such as high robustness to errors, but failed to develop a compelling explanation for codon assignments.” ↩ ↩

  5. 5.

    Shulgina, Yekaterina, and Sean R. Eddy.  “A computational screen for alternative genetic codes in over 250,000 genomes.”  eLife 10:e71402.  2021.  Section Abstract:  “discovered five new reassignments of arginine codons (AGG, CGA, and CGG), representing the first sense codon changes in bacteria”.  Section Results:  “251,571 assembled genomes from GenBank”.  A computational study that built a tool, Codetta, to predict from genome sequence which amino acid each codon encodes, and ran it on over 250,000 bacterial and archaeal genomes from GenBank.  It found five new reassignments of arginine codons (AGG, CGA and CGG), the first sense-codon changes found in bacteria, and the CGA and CGG reassignments sit in genomes with low GC content, meaning few guanine and cytosine bases.  The article uses it to show that the genetic code changes, but only through a narrow route. ↩ ↩

    1. 5a.

      The authors, describing the codon-capture model they find the reassignments follow:  “the codon is first driven to near extinction by pressures unrelated to reassignment, such as biased genomic GC content or genome reduction, which then minimizes the impact of reassignment on protein translation.”  That is the mechanism that reconciles "codes do change" with "changing a code is usually lethal."  Code change, on this model, is not evidence that the code is free to vary.  It is evidence that the code varies only through a narrow door. ↩

  6. 6.

    McGowan, Jamie et al.  “Identification of a non-canonical ciliate nuclear genetic code where UAA and UAG code for different amino acids.”  PLOS Genetics 19(10):e1010913.  2023.  “UAA has been reassigned to encode lysine, while UAG has been reassigned to encode glutamic acid.”  A genome and transcriptome study of an uncultured ciliate, a single-celled protist.  The organism reassigns two of the three stop codons to amino acids, and the reassignment is supported by suppressor tRNA genes with matching anticodons.  The article uses it because UAA and UAG nearly always have the same translation, which suggested that their evolution is coupled. ↩ ↩

  7. 7.

    DeBenedictis, Erika A. et al.  “Measuring the tolerance of the genetic code to altered codon size.”  eLife 11:e76941.  2022.  Section Abstract:  “the triplet codon code was selected because it is simpler and sufficient, not because a quadruplet codon code is unachievable”.  “from all 20 canonical isoacceptor classes”.  “only nine of the twenty aminoacyl tRNA synthetases tolerate quadruplet anticodons”.  The synthetase limit, read with the conclusion, is that a quadruplet code is buildable but.  “sharply limits the chemical alphabet available to a nascent all-quadruplet code.”  A laboratory study in Escherichia coli that rebuilt transfer RNAs, the molecules that carry amino acids to the ribosome, so that they read four-base codons.  It tested converted versions from all twenty amino acid classes, and its limit lies in the enzymes that load amino acids onto those RNAs.  The article uses it to show that the triplet code was chosen because it is simpler and sufficient, and because the experiment runs in a modern bacterium it says little about origins. ↩ ↩

  8. 8.

    Wong, J. Tze-Fei.  “A co-evolution theory of the genetic code.”  Proceedings of the National Academy of Sciences 72:1909–1912.  1975.  A theoretical paper that proposes the co-evolution theory of the genetic code.  The theory is described here as Koonin and Novozhilov's 2009 review describes it, and the tests of it are Amirnovin's, as reported there.  The article uses it as the origin of the theory it then tests against random codes. ↩

  9. 9.

    Nyerges, Ákos et al.  “Probing the limits of genetic recoding using multi-omics-guided evolution.”  Nature Communications 17:5311.  2026.  Section Abstract:  “synonymous codon replacement—recoding—is frequently lethal”.  A laboratory study that built Escherichia coli strains carrying up to 45.8% of a synthetic genome under a deliberately deleterious 57-codon code, which was chosen to probe the limits of recoding.  The experiment is evidence that recoding is costly, not that every recoding is lethal, since it also mitigates the defects it finds.  The article uses it for the cost of changing a genetic code. ↩

  10. 10.

    Wehbi, Sawsan et al.  “Order of amino acid recruitment into the genetic code resolved by last universal common ancestor's protein domains.”  Proceedings of the National Academy of Sciences 121:e2410311121.  2024.  Their further claim is an inference from reconstructed frequencies, not a recovered ancient code: ancestral frequency patterns.  “provide hints about earlier, alternative genetic codes”.  and they hedge it themselves with.  “if at least some of these sequences predate the current code.”  Section Abstract:  “potentially biased criteria, such as the absence of sulfur-containing amino acids from the Urey–Miller experiment which lacked sulfur.”  Section Significance:  “metal-binding and sulfur-containing amino acids earlier than previously thought”.  A computational study that infers the order in which amino acids joined the genetic code directly from how often each amino acid appears in reconstructed ancient protein sequences.  Earlier orders were inferred from a consensus of forty metrics, many of which reflect abiotic abundance on ancient Earth.  The article uses it because it places metal-binding and sulfur-containing amino acids earlier than the textbook order does. ↩

  11. 11.

    Hashmi, Syeda Ameena et al.  “Molecular Complexity Constrained Early Amino Acid Recruitment into the Genetic Code.”  Genome Biology and Evolution 18:evag012.  2026.  “we reconstruct a complexity-based amino acid chronology by integrating 16 molecular complexity metrics from chemical graph and information theory.”  A computational study that ranks the amino acids by molecular complexity, scoring each with 16 measures drawn from chemical graph and information theory, to reconstruct the order in which they joined the code.  Earlier chronologies rest on prebiotic experiments, biosynthetic pathways or genomic trends, so this one is independent of them, and the article uses it as corroboration.  A June 2026 correction to the paper exists, so treat its detail as provisional, although Wehbi et al. independently corroborate the direction of the result. ↩

  12. 12.

    Goldman, Aaron D. et al.  “Universal paralogs provide a window into evolution before the last universal common ancestor.”  Cell Genomics 6:101140.  2026.  A Perspective, a review that reports two earlier reconstructions of early protein evolution rather than new results of its own.  Readers who build on it should go to its references 70 and 75, which are the primary papers.  The article uses it for sequence evidence that the genetic code was still being built after protein synthesis was working. ↩