Synonymous Mutations and Codon Usage

"Synonymous mutations - sometimes called 'silent' mutations - are now widely acknowledged to be able to cause changes in protein expression, conformation and function." — Zuben Sauna and Chava Kimchi-Sarfaty, Nature Reviews Genetics, 2011¹

Summary

A synonymous mutation changes one DNA letter inside a gene without changing the amino acid it codes for.  About a quarter to a third of all new mutations in protein-coding DNA are of this kind.²  They are "generally presumed neutral or nearly so",³ and so serve as the neutral yardstick for measuring mutation rates, selection and divergence times.  The question here is whether that is true.  The answer is that it is mostly true in humans, false for a measurable minority of sites, and not settled for yeast.

1. The code is redundant, and organisms do not use the redundancy at random. Sixty-one codons spell twenty amino acids, so most amino acids have two to six spellings. Every genome prefers some spellings over others. In fast-growing bacteria and yeast the preferred codons match the most abundant transfer RNAs, and selection for fast, accurate translation explains the bias. In humans it does not: the bias tracks the GC content of the surrounding DNA, which is driven by a non-adaptive repair bias, and that model explains 70% of the variation among human genes.⁴
2. Synonymous changes can act in at least four ways, each with a well-measured case. They can break a splicing signal (the single C-to-T change that makes the backup gene SMN2 skip an exon, so that it cannot stand in for a lost SMN1 in spinal muscular atrophy), change how stable or how folded the messenger RNA is, change the pace of translation and so how a protein folds, or create or destroy a microRNA binding site. The most famous folding case, the MDR1 C3435T variant, has a poor replication record in patients.
3. Real but rare in human disease. Our count of ClinVar, the main public database of clinically classified variants, finds 830 synonymous single-letter variants classified pathogenic or likely pathogenic, 0.42% of all such variants, against 72,225 missense and 76,202 nonsense.⁵ The clinical rules count a synonymous variant with no predicted splicing effect as supporting evidence that it is benign,⁶ so this count is biased low. It would have to undercount by a factor of 30 to 45 to put synonymous changes level with missense changes per possible mutation.⁵
4. The 2022 yeast study is a real measurement with an unresolved control problem. Xukang Shen, Siliang Song, Chuan Li and Jianzhi Zhang edited 21 yeast genes and reported that 75.9% of 1,866 synonymous mutants grew significantly worse than the wild type, almost exactly the share for amino-acid changes.³ Leonid Kruglyak and eight colleagues replied that the single wild-type control was built separately from the mutants, and that the growth-rate check failed to confirm small effects.⁷ The Zhang lab answered in 2024 with genome sequencing of 603 strains and rebuilt libraries for 11 genes, each with its own wild type, and again found most mutants non-neutral.⁸ No independent lab has repeated the yeast experiment.
5. Human cells and human populations point the other way. A prime-editing screen in human cells found effects at 0.38% of the synonymous changes it made (409 hits, so roughly 108,000 tested) against 5.32% of amino-acid changes, and "the fitness impact of synonymous mutations didn’t significantly differ from that of the negative controls".⁹ A base-editing screen of 128 essential genes found synonymous effects 37.9 times rarer than missense effects.¹⁰ In 454,668 human exomes, synonymous variants sit at higher frequencies than missense variants, as weak selection predicts.¹¹
6. The population-genetic signal is weak selection on a minority. Human and chimpanzee synonymous sites have diverged 21% to 37% less than intergenic DNA.¹² The often-quoted "39% of synonymous sites are deleterious" is Hellmann and colleagues' 2003 figure for one class of site, non-CpG fourfold sites, measured as a divergence shortfall against intergenic DNA; the authors themselves wrote "It is not clear why".¹³ Ancestral-repeat controls give about 22% to 27% in apes,¹⁴ and fruit flies show strong selection on about 22% of fourfold sites.¹⁵ Jian Lu and Chung-I Wu estimated that over 90% of human-chimp synonymous substitutions are slightly harmful, but with effects "weaker than the effect of genetic drift".¹²
7. What follows for methods. If 21% to 37% of synonymous sites are constrained in apes, dN/dS overstates the amino-acid rate against a true neutral baseline by 1.26 to 1.57 times, and a divergence date computed from synonymous sites with a pedigree mutation rate comes out 21% to 37% too young.¹⁶ A fossil-calibrated clock absorbs a constant bias. None of this moves a date by an order of magnitude, and the correction pushes dates older, not younger.
8. For the design and young-earth arguments. John Sanford's claim that "there are probably no truly neutral nucleotide sites"¹⁷ fits the weak-selection estimates for apes. The stronger claim made by Creation Ministries International and Creation-Evolution Headlines in 2022, that most silent mutations are strongly harmful, rests on the yeast study and fails the human data. The 2017 Theistic Evolution volume's argument that humans and chimpanzees share codons because each codon has a job fails its own test: human-chimp synonymous sites diverge at 79% of the intergenic rate, not near zero.¹⁶ Priced as genetic load, the weakly harmful synonymous mutations of the Lu and Wu estimate would cost at most about 0.1% to 0.2% of fitness over 200 generations.¹⁶

Verdict.  "Silent" is the wrong word for a minority of synonymous sites, and the word "neutral" carries a 20% to 40% error in apes when it is used as a yardstick.  "Mostly strongly harmful" is a claim about one yeast experiment that human cells and human populations do not support.  The design argument gains a modest, real point (redundant code positions carry a second layer of information at some sites) and loses its strong versions.

Part 1: A redundant code, used unevenly

Sixty-one codons for twenty amino acids

DNA is read three letters at a time.  Each three-letter word, a codon, names one amino acid or says "stop".  Four letters give 64 codons;  three are stops, which leaves 61 for 20 amino acids.  Jianzhi Zhang and Wenfeng Qian, in their 2025 review, put the arithmetic plainly:  "Two amino acids — methionine (Met) and tryptophan (Trp) — are each encoded by one codon, whereas the other 18 amino acids are each encoded by two to six different codons."²  A change from one spelling to another is a synonymous mutation.  Most fall on the third letter.  At a "fourfold" site any of the four letters gives the same amino acid;  GGA, GGC, GGG and GGT all mean glycine.

Because every genome has many such sites, a gene could in principle be written in an astronomical number of synonymous ways.  Genomes do not use them evenly.  Every organism studied prefers some codons over their synonyms.  The question for this article is why, and whether the choice matters to the organism.

How the bias is measured

Two numbers are standard.  The codon adaptation index (CAI) of Paul Sharp and Wen-Hsiung Li compares each gene with a reference set of the organism's most highly expressed genes:  "The index uses a reference set of highly expressed genes from a species to assess the relative merits of each codon, and a score for a gene is calculated from the frequency of use of all codons in that gene."¹⁸  A CAI near 1 means a gene uses the same codons as the most highly expressed genes.  The effective number of codons (Nc or ENC) of Frank Wright needs no reference set.  It "can take values from 20, in the case of extreme bias where one codon is exclusively used for each aa, to 61 when the use of alternative synonymous codons is equally likely."¹⁹  A third measure, the tRNA adaptation index of Mario dos Reis and colleagues, scores codons by how many copies of the matching transfer-RNA gene the genome carries.²⁰

Neither number says why a bias exists.  A gene can have a low Nc because selection favours some codons, or because the DNA around it is rich in G and C, so mutation keeps putting G and C at the third position.  Separating the two is the hard part.

Bacteria and yeast: selection for translation

In fast-growing microbes the evidence for selection is strong.  Ikemura counted transfer RNAs in Escherichia coli and baker's yeast and found that "There exists a strong positive correlation between codon usage and tRNA content in both organisms, and the extent of this correlation relates to the protein production levels of individual genes."²¹  Sharp and Li then showed that the most biased genes also change most slowly at their synonymous sites:  "very highly expressed genes have very biased codon usage and accumulate synonymous substitutions very slowly", and "both the rate of synonymous substitution and the degree of codon usage bias largely reflect the intensity of selection at the translational level."²²  That is the pattern a selected trait leaves:  more bias where more is at stake, and slower change where bias is high.

Selection on codons is weak per site even in microbes.  Population geneticists express it as S = 4Nₑs, the selective effect s scaled by the effective population size Nₑ.  S near 1 means selection and chance are about equal;  S in the hundreds means a harmful change essentially never spreads.  Across 80 bacterial species, Sharp and colleagues found "in 30% of the genomes examined, there was no significant evidence that selection had been effective", and the strength of selection rose with the number of ribosomal-RNA operons and transfer-RNA genes, which marks fast growers.²³  Dos Reis and Wernisch estimated S for ten eukaryotes:  "Fungi genomes present the largest S values (2.24–2.56), whereas multicellular invertebrate and plant genomes present more moderate values (0.61–1.91)."²⁴  Values of 1 to 3 are weak selection.  It is real because these organisms have huge populations, in which even tiny advantages count.

Humans: mostly composition, not translation

Mammals look different.  The same dos Reis and Wernisch paper found that "The large mammalian genomes (human and mouse) show low S values (0.22–0.51) for the most highly expressed genes", and warned that even this "might not be evidence for selection in these organisms as the technique used here to estimate S does not properly account for nucleotide composition heterogeneity along such genomes."²⁴  Urrutia and Hurst corrected for the surrounding DNA in 2,396 human genes.  The patterns that mark translational selection in bacteria were "weakly recovered", two of them "appear to be methodological artifacts", and they concluded:  "We conclude that we find no evidence for selection on codon usage in humans."²⁵

The positive explanation came from GC-biased gene conversion.  When the cell repairs mismatches during meiotic recombination, it slightly favours G and C over A and T.  Over many generations this raises the GC content of DNA that recombines often, coding or not.  Duret and Galtier argued that it "explains the spatial distribution of GC-content in mammalian genomes".²⁶  Pouyet and colleagues then tested it on human codon usage directly:  "We demonstrate here that SCU [synonymous codon usage] is not driven by constraints on tRNA abundance, but by large-scale variation in GC-content, caused by meiotic recombination, via the non-adaptive process of GC-biased gene conversion (gBGC)."  "Overall, the gBGC model explains 70% of the variance in SCU among genes."⁴  The chimpanzee genome consortium reached the same conclusion from divergence:  the constraint it found on human synonymous sites "does not seem to result from selection on the usage of preferred codons, which has been detected in lower organisms such as bacteria, yeast and flies."²⁷

So the visible codon bias of human genes is mostly a by-product of how DNA is repaired, not a code for speed.  That does not make every human synonymous site free to change.  It means that whatever constrains human synonymous sites is mostly something other than translation speed:  splicing signals, RNA structure and the other mechanisms in Part 2.  It also means that GC-biased gene conversion acts on synonymous sites in a way that imitates selection, which is the main trap in Part 4.

Part 2: How a synonymous change can act

A gene is read twice before it becomes a protein.  First its RNA copy is cut and spliced, then the spliced messenger RNA is translated by ribosomes.  A synonymous change leaves the amino acid alone, but it still changes the RNA letter, and several steps read the letters, not the amino acids.  Four mechanisms have measured cases.  Splicing is the strongest by far.

Splicing: the clearest case

Exons carry short signals, called exonic splicing enhancers and silencers, that help the splicing machinery recognise where an exon begins and ends.  Cartegni, Chew and Krainer summarised the consequence in 2002:  "Nonsense, missense and even translationally silent mutations can inactivate genes by inducing the splicing machinery to skip the mutant exons."²⁸

The best-measured example decides a fatal disease.  Humans carry two nearly identical genes, SMN1 and SMN2.  Children who lose both copies of SMN1 develop spinal muscular atrophy, because SMN2 makes mostly a short, unstable protein.  Lorson, Hahnen, Androphy and Wirth tested each of the five letters that differ between the genes.  "Of these nucleotide differences, the exon 7 C-to-T transition at codon 280, a translationally silent variance, was necessary and sufficient to dictate exon 7 alternative splicing."²⁹  The test ran both ways:  putting the SMN1 letter into SMN2 restored full-length product, and the other differences "individually had no detectable effect".²⁹  Cartegni and Krainer then found why:  the change sits inside an enhancer that a splicing protein, SF2/ASF, binds in SMN1, so "SMN2 is only partially active because a translationally silent, single-nucleotide difference in exon 7 causes exon skipping."³⁰

Splicing signals are spread through exons, not confined to one famous site.  Pagani, Raponi and Baralle swapped the seven synonymous differences between the mouse and human versions of CFTR exon 12 into the human exon.  The protein sequence stayed human, yet the exon was nearly lost:  "only 5% of the mature mRNA included the exon", against 80% for the normal human version.  They concluded that "synonymous changes cannot evolve freely, because they are significantly constrained by splicing requirement."³¹  That was seven changes at once in a minigene, so it shows the exon is sensitive, not that any single site is.

How common are such variants?  Chong and colleagues in the Kosuri laboratory put rare variants from the ExAC population database into a splicing reporter, one at a time, across 2,198 exons.  "More than a thousand (3.8%) of these rare genetic variants leads to near complete loss of exon recognition", and "the SDV rate in synonymous mutations, which are usually assumed to be benign, is nearly equivalent to missense variants (3.0% versus 3.1%".³²  So about 3 in 100 rare synonymous variants break splicing badly in this assay, the same rate as amino-acid changes.  A reporter is not a person, and most such variants are rare because selection removes them, which is the point.  A figure often quoted as the share of exonic variants that disrupt splicing, "~10% (513/4,964)", comes from Soemedi and colleagues, but it measured disease-causing nonsynonymous variants and does not apply to synonymous ones.³³

mRNA structure and stability

A synonymous change can alter how the messenger RNA folds on itself.  The best human case is the pain-sensitivity gene COMT.  Nackley, Diatchenko and colleagues found that three common versions of the gene, differing at two synonymous sites and one amino-acid site, differ in enzyme activity, and that "Haplotypes divergent in synonymous changes exhibited the largest difference in COMT enzymatic activity, due to a reduced amount of translated protein."  The stable stem-loop structure in the low-activity RNA was the cause:  "Site-directed mutagenesis that eliminated the stable structure restored the amount of translated protein."³⁴  A rescue experiment like that is strong evidence.  No independent laboratory has repeated it, as far as this article's research found.

In yeast the effect is general.  Presnyak, Coller and colleagues showed that "Substitution of optimal codons with synonymous, non-optimal codons results in dramatic mRNA destabilization, while the converse substitution significantly increases stability."³⁵  In humans the same signal is weak.  Dhindsa and colleagues found that synonymous changes from optimal to non-optimal codons are under much weaker selection than missense changes in 454,668 human exomes.¹¹

In bacteria, RNA folding near the start of a gene dominates.  Kudla, Murray, Tollervey and Plotkin wrote 154 synonymous versions of the gene for green fluorescent protein.  "GFP protein levels varied 250-fold across the library", "but codon bias did not correlate with gene expression.  Rather, the stability of mRNA folding near the ribosomal binding site explained over half the variation in protein levels."³⁶  The same protein, made from the same amino-acid sequence, varied 250-fold in amount because of synonymous letters, and the codon adaptation index of Part 1 did not predict which.

Translation speed and folding

Ribosomes do not read every codon at the same speed.  If a stretch of rare codons slows translation, a protein may fold differently as it emerges.  Yu, Liu and colleagues measured it directly in a cell-free system from the fungus Neurospora, and found that preferred codons speed up elongation of the protein chain while non-optimal codons slow it down.³⁷  In the same fungus, the clock protein FRQ is built from non-optimal codons on purpose.  When Zhou and colleagues rewrote the gene in optimal codons, the protein changed shape and the circadian clock stopped:  "Optimization of frq codon usage abolishes both overt and molecular circadian rhythms."³⁸  In E. coli, Walsh, Clark and colleagues showed "that synonymous mRNA mutations can alter a protein folding mechanism in vivo, leading to changes in cellular fitness".³⁹  This is the mechanism design writers usually have in mind when they speak of a second code for folding.  Its best evidence is in microbes.

The famous human claim is the MDR1 (ABCB1) variant C3435T.  In 2007 Chava Kimchi-Sarfaty, Michael Gottesman and colleagues reported that this synonymous change, together with two nearby variants, made a drug pump with "Similar mRNA and protein levels, but altered conformations".  They proposed that a rare codon "affects the timing of cotranslational folding".⁴⁰  The paper's own data limit the claim.  The single C3435T change alone behaved like the normal gene:  the functions of all single-variant versions "were not distinguishable under standard conditions";  the difference appeared in two- and three-variant combinations, mostly in the presence of inhibitor drugs.⁴⁰  The same laboratory reproduced a folding difference in a second cell system in 2014, but found drug transport "dependent on the drug of choice" and wrote that "the results do not demonstrate the mechanism of the effect of synonymous mutations."⁴¹  Another group measured lower messenger RNA from the 3435T version, which "decreases mRNA stability".⁴²  And in patients, a 2007 review found that studies of this gene's variants and drug response "have produced inconsistent results".⁴³  So the most cited human folding case is real in cell culture, uncertain in mechanism, and unconfirmed in the clinic.

The cystic-fibrosis gene offers a second, designed case.  Bartoszewski and colleagues found that the common ΔF508 mutation also changes the neighbouring codon from ATC to ATT, and that a version keeping ATC "has wild type-like mRNA structure, and enhanced expression levels when compared with native ΔF508."⁴⁴  That is a laboratory construct, not a variant anyone carries, but it shows that the synonymous letter beside the deletion adds to the damage.

MicroRNA sites

MicroRNAs are short RNAs that bind messenger RNAs and lower their output.  A synonymous change can create or destroy a binding site.  Brest, Hofman and colleagues studied a common synonymous variant in the Crohn's disease gene IRGM that earlier work had "classified as non-causative because of the absence of an alteration in the IRGM protein sequence or splice sites."  They found that miR-196 "downregulates the IRGM protective variant (c.313C) but not the risk-associated allele (c.313T)", "thereby implicating a synonymous polymorphism as a likely causal variant."⁴⁵  The disease association has been repeated in a German cohort,⁴⁶ but the variant sits in strong linkage with a nearby deletion, and no second laboratory has tested the microRNA mechanism.

Transcription-factor sites, or "duons"

In 2013 Stergachis, Stamatoyannopoulos and colleagues mapped where transcription factors touch DNA inside exons and reported that "~15% of human codons are dual-use codons ('duons') that simultaneously specify both amino acids and TF recognition sites."  They claimed these were "highly conserved" and that "TF-imposed constraint appears to be a major driver of codon usage bias."⁴⁷  The Theistic Evolution volume cites this paper,⁴⁸ and Jeffrey Tomkins of the Institute for Creation Research called duons a code "utterly defying all naturalistic explanations for its existence".⁴⁹  Xing and He reanalysed the data in 2015 and found the conservation signal was an artefact of base composition:  comparing G/C and A/T sites separately, they found "a similar level of conservation between TF-bound codons and TF-depleted codons, suggesting largely no extra purifying selection provided by the TF occupancy on the codons of human genes."⁵⁰  No published reply from the original authors was found.  The binding is real;  the claimed extra constraint is unconfirmed.  The wider question of DNA read in more than one way belongs to the sibling article Overlapping Genes and Multi-Layer Coding.

What the mechanisms add up to

Every mechanism above is real, and splicing is the best documented.  None of them, by itself, says what fraction of synonymous sites matter.  Most of the vivid cases involve many changes at once (whole-gene recoding, seven-site swaps, haplotypes), stressed conditions, or sites at exon edges.  The fraction question needs the counts in Parts 3 and 4.

Part 3: The 2022 yeast study and the dispute over it

What the study did and found

In June 2022 Xukang Shen, Siliang Song, Chuan Li and Jianzhi Zhang of the University of Michigan published the most direct test so far.  They chose 21 genes of baker's yeast with a range of functions and expression levels.  In each they deleted a 150-letter stretch, then used CRISPR editing to put it back from a pool of templates carrying every possible single-letter change.  That gave 8,341 mutant strains:  1,866 synonymous, 6,306 nonsynonymous and 169 nonsense.  The strains of each gene were grown together in rich medium with a wild-type strain, and each mutant's fitness was read from how its share of DNA sequencing reads rose or fell over the competition.³

The result surprised the authors.  "The median fitness of the 169 nonsense mutants is 0.940 … As expected, the corresponding value for the 6,306 nonsynonymous mutants is much higher, reaching 0.988 … Surprisingly, the median fitness of the 1,866 synonymous mutants is 0.989, much closer to that of nonsynonymous mutants than to the neutral expectation of 1".³  Counting mutants one by one:  "Among synonymous mutations, 75.9% are significantly deleterious while 1.3% are significantly beneficial.  The corresponding values are 75.8% and 1.6% for nonsynonymous mutations."³  Both kinds of mutation often lowered the amount of the gene's messenger RNA, and the size of that drop partly predicted the fitness cost.³

The authors stated the limits of their own claim.  Their abstract ends:  "The strong non-neutrality of most synonymous mutations, if also true for other genes and in other organisms, would require reexamining numerous biological conclusions about mutation, selection, effective population size, divergence time, and disease mechanism that rely on the neutral assumption of synonymous mutations."³  They also noted that their assay cannot see effects much below 0.5% ("generally below 5×10−3"), and that in nature, where selection can act on far smaller effects, the two classes could still differ.³

The replies

The objection came within five weeks, as a preprint, and was published in Nature in April 2023.  Leonid Kruglyak, Craig Kaplan and seven co-authors argued that the claim "is not supported by the data reported by Shen et al." for three reasons:  "First, there is an absence of appropriate matched wild-type (WT) controls for strain background with which to compare variant effects.  Second, technical replicates are treated as biological replicates in competition-based fitness assays. … Third, the lack of biological replicates in independent growth assays intended to validate competition-based fitness measurements precludes the interpretation that most synonymous variants are deleterious in yeast."⁷

The first point is the heart of it.  Each mutant pool went through two rounds of CRISPR editing, a gene deletion and a repair, while the single wild-type control was a separately chosen healthy colony.  Any damage picked up during editing would make every mutant, synonymous or not, look a little less fit than the control.  "Selection of a single WT strain is especially sensitive to potential ascertainment bias, as any obviously unhealthy strain would not be selected but, within the variant pools and among pools of the same variant, there is no such filter."⁷  Their sharpest numerical point concerns the one independent check Shen and colleagues ran, a growth-rate test of rebuilt strains.  Those strains were chosen for large effects, and "When only variants with a sequencing-based fitness of ≥0.981 are considered (75% of all generated missense and synonymous variants have a measured fitness effect smaller than this), the average growth-based fitness of the reconstructed strains clusters around 1 (median = 0.997)."⁷  In other words, the independent check did not confirm the small effects that make up most of the 75.9%.

Ryan Dhindsa, Leonid Kruglyak, Daniel MacArthur, Matthew Hurles, Slavé Petrovski and others answered the human-disease claim in the American Journal of Human Genetics.  They argued "that there is insufficient evidence to overturn the large, coherent body of knowledge establishing the predominant neutrality of synonymous variants in the human genome."¹¹  Their evidence was population data, which Part 4 returns to, and disease association:  across 4,911 traits in the UK Biobank exomes, "only two genes were significantly associated with at least one phenotype in the synonymous model, compared to 27 genes in the missense model".¹¹  They did not claim synonymous variants never matter:  they "should still be considered as rare contributors to severe disease".¹¹  Several of the authors work for drug companies, which the paper declares;  their argument rests on public data anyone can check.

Nature published no reply from Shen and colleagues alongside the Matters Arising.⁵¹  Their answer came as a new experiment in Molecular Biology and Evolution in October 2024.  They "sequenced the genomes of 603 yeast strains", found "zero mutations" at the sites where CRISPR was most likely to cut by mistake, but did find "secondary mutations outside potential CRISPR/Cas9 off-target sites in 26.7% of the 580 constructed mutants sequenced".⁸  For seven genes where at most 12.5% of mutants carried such a mutation, they bounded the bias at "no more than 0.014 × 12.5% = 0.00175" in fitness and still found 63.0% of 658 synonymous mutations significantly non-neutral.⁸  For 11 genes they rebuilt each library four times, each with its own wild-type control made alongside the mutants, and found that "86.5% of synonymous mutations and 85.5% of nonsynonymous mutations are significantly nonneutral;  the corresponding values were 83.4% and 84.0% in the original data."⁸

That answers the control objection for those 11 genes, and it is a real answer.  Two gaps remain.  The 2024 figures count significant effects in either direction, not harm alone.  And no laboratory outside Zhang's has repeated a yeast experiment of this kind.  Zhang and Qian's 2025 review summarises the dispute in one line:  the original control "raised doubts about their results  However, new experiments in which the control and mutants were constructed simultaneously in the same test tube confirmed the original finding".²  That is the authors' own judgement of their own work.

Other organisms: real effects, rarely a majority

Earlier and later experiments in bacteria show that synonymous changes can matter a great deal in particular genes.  In Salmonella, Lind, Berg and Andersson found that of 126 mutations in two ribosomal-protein genes, "most mutations (120 out of 126) are weakly deleterious", and "The DFEs for synonymous and nonsynonymous substitutions are similar".⁵²  In the bacterium Methylobacterium extorquens, rewriting a whole enzyme gene in rare codons "decreased fitness by 40%, whereas an allele coded entirely by frequent codons decreased fitness by more than 90%";  these alleles changed many codons at once.⁵³  In E. coli growing on a poor food source, Shelley Copley's laboratory found single synonymous mutations near the start of a weak enzyme's gene that "increase growth rate by 41 and 67%", while other synonymous changes at the same codons "completely abolishes growth".⁵⁴  In Pseudomonas fluorescens, Lebeuf-Taylor and colleagues measured 39 synonymous mutants of one gene and found "highly variable fitness effects, both deleterious and beneficial, resembling those of nonsynonymous mutations in the same gene."⁵⁵

The largest bacterial survey is Copley's 2024 study of more than 50,000 CRISPR edits in E. coli, each carrying two to six synonymous "immunizing" changes.  "Twenty-two percent of edits containing only immunizing mutations had high impacts on fitness during growth of E. coli in 50 mM acetate", but "fewer edits affected fitness during growth in glucose, a preferred carbon source".⁵⁶  Effects depend on conditions, and a cluster of changes is not one change.

Human cells: the direct test the yeast study called for

Two large screens have now edited synonymous sites inside human cells, where most of the disease and dating questions live.  Niu, Wei and colleagues used prime editing on a library of 297,900 guide RNAs:  "A total of 1,917 mutations impacting cell fitness were identified … Among these, 409 were synonymous mutations, and 1,508 were nonsynonymous mutations.  In our screen, the occurrence rate of synonymous mutations was only 0.38%, compared to 5.32% for nonsynonymous mutations.  While the fitness impact of synonymous mutations didn’t significantly differ from that of the negative controls, nonsynonymous mutations, including missense and nonsense types, exhibited clear effects on cell fitness".⁹  Most of the synonymous hits they could explain broke splicing.⁹  Their library included the human versions of the yeast genes, and those "didn’t show significant enrichment".⁹

Rao, Sokirniy, O'Brien and Pritchard used base editing on 8,558 possible synonymous mutations in 128 highly essential genes, deliberately leaving out sites next to splice junctions.  "We find that synonymous mutations rarely have fitness effects on growth, occurring around 37.9-fold (95% CI: 22.16–81.48-fold) less frequently than missense mutations."¹⁰  Their synonymous hit rate, 0.95%, was "statistically indistinguishable from negative controls".¹⁰

Saturation editing of the breast-cancer gene BRCA1 shows where the exceptions sit.  Findlay and colleagues in Jay Shendure's laboratory found that "98.7% of synonymous SNVs >3 bp from splice junctions scored above −1.25", that is, worked normally, while 22.9% of protein-neutral changes in the splice region next to exon boundaries were non-functional.⁵⁷  The non-functional synonymous changes "tended to markedly reduce mRNA levels (median 5.4-fold reduction)".⁵⁷

These human screens have limits that the yeast side can fairly press.  Human cells carry two copies of each gene, so a harmful change in one copy is partly masked;  prime editing changed only about 15% of target copies on average;⁹ and a screen for cell growth cannot see effects of a fraction of a percent, the same blind spot the yeast authors admitted.  What the screens show is that large synonymous effects on human cell growth are rare, between a fourteenth and a thirty-eighth as common as missense effects, and that splicing is their usual cause.  They do not show that the remaining sites have effects of exactly zero.

Part 4: How much selection acts on synonymous sites in nature

Laboratory screens ask whether a change hurts a cell in a dish.  Population genetics asks a different question:  has selection, over many generations in real populations, removed some synonymous changes?  It can detect effects far too small for any growth assay, because a population "measures" fitness across millions of individuals.  It has two tools, and each rests on an assumption the reader should keep in view.

1. Divergence. Compare how much two species differ at synonymous sites with how much they differ at a stretch of DNA assumed to be neutral, such as intergenic DNA, introns, pseudogenes or old transposable-element copies. A shortfall at synonymous sites means some changes were removed. This tool assumes the two species share an ancestor and that the reference DNA is itself unconstrained.
2. Polymorphism. Within one species, compare the frequencies of synonymous variants with those of a reference class. Harmful variants stay rare. This tool needs no assumption about common ancestry, only about the reference.

Humans and chimpanzees

The figure most often quoted on the design side comes from a 2007 review by Michael Pheasant and John Mattick:  "Comparison between protein-coding and intergenic regions in human and chimp indicate that ∼39% of synonymous sites are deleterious and subject to negative selection".⁵⁸  The source Pheasant and Mattick cite is Ines Hellmann and colleagues, 2003, and what Hellmann measured is narrower.  At one class of site, fourfold-degenerate sites not in CpG pairs, human-chimp divergence was 39% lower than in intergenic DNA:  "The extent of this reduction indicates that 39% of mutations have been eliminated, presumably as a result of purifying selection."¹³  Hellmann's own discussion then says:  "It is not clear why as much as 39% of 4d sites might be under purifying selection in humans."¹³  The same paper found that fourfold sites carry more variation within humans than the genome average, 111% against 86% for divergence, which the authors read as "slightly deleterious variants segregate in the human population".¹³  So "39%" is a real measurement of one site class against one reference, with the mechanism unknown.  It is not a count of synonymous sites with known function.

The chimpanzee genome paper of 2005 repeated the test on the whole genome and found the same direction with a smaller size.  "Although overall divergence rates are very similar at fourfold degenerate and intronic sites, direct comparison is misleading because the former have a higher frequency of the highly mutable CpG dinucleotides (9% compared with 2%).  When CpG and non-CpG sites are considered separately, we find that both CpG sites and non-CpG sites show markedly lower divergence in exonic synonymous sites than in introns (~50% and ~30% lower, respectively)."²⁷  The difference is highly significant, P < 10⁻²⁷.²⁷

Jian Lu and Chung-I Wu used the X chromosome as a lever.  Harmful recessive changes are exposed on the single X of males, so if synonymous changes are slightly harmful, the X should accumulate fewer of them.  It does.  Across the autosomes, synonymous divergence was 0.791 of intergenic divergence, or 0.635 with repeats and CpG sites masked out.¹²  Fitting a population model to the X-to-autosome contrast, they concluded that "At least 90% of the synonymous substitutions in human and chimpanzee are estimated to be deleterious, but the fitness effect is weaker than the effect of genetic drift", with 2Nₑs between 0.5 and 0.8.¹²  With the usual human Nₑ of about 10,000, that is a selection coefficient of 2.5 to 4 × 10⁻⁵ per mutation, roughly 300 to 400 times smaller than the median yeast effect in Shen 2022.¹⁶  Lu and Wu's model assumes a degree of recessiveness, so the 90% is a model estimate.  It is also the closest thing in the literature to a measured confirmation of John Sanford's "nearly-neutral" picture for these sites, and it comes from mainstream evolutionary geneticists.

Eöry, Halligan and Keightley used ancestral transposable-element copies as the neutral reference.  In apes "constraint is more than twice as high as in murids", and "more than about one-fifth of mutations at 4-fold sites are effectively selected against in hominids".¹⁴  Their table gives 0.215 to 0.272 for apes and 0.108 to 0.117 for mice and rats.¹⁴  That ranking is odd for a selection story, since mice have the larger populations and should purge more efficiently.  The authors suggest more complex gene structure in apes.  A simpler possibility, discussed below, is GC-biased gene conversion acting differently in the two lineages.

The trap: conserved is not the same as functional

A 2025 study of 240 mammal genomes by Matthew Christmas and colleagues found that "Around 20% of synonymous sites are conserved in placental mammal genomes", and concluded that "a combination of GC-biased gene conversion, functional constraint, and mutation rate variation has likely shaped synonymous site evolution."⁵⁹  GC-biased gene conversion, introduced in Part 1, pushes G and C into recombining DNA whether or not it helps.  Fanny Pouyet and colleagues showed that together with background selection it affects "as much as 95% of the variants" of the human genome, and that "synonymous sites and non-transcribed regions are also affected, albeit to different degrees.  Their use for demographic inference can lead to strong biases."⁶⁰  So some of the synonymous "constraint" in Parts 4 and 5 is not function at all;  it is a repair bias that mimics selection.  This cuts against both sides that would like a clean number:  the design side cannot count all of it as function, and the neutral side cannot treat synonymous sites as an untouched yardstick.

Fruit flies: strong selection on a fifth

Fruit flies have far larger populations than humans, so selection can act on far smaller effects.  David Lawrie, Dmitri Petrov and colleagues sequenced 168 fly strains and compared the frequency spectrum at fourfold sites with that in short introns, which are thought to be nearly neutral:  "22% of four-fold synonymous (4D) sites in Drosophila melanogaster evolve under very strong selective constraint while few, if any, appear to be under weak constraint."¹⁵  The fitted strength was 4Nₑs of about −283.¹⁵  Heather Machado and colleagues later found that selection on preferred codons "ranges from weak (|Nes| ∼ 1) to strong (|Nes| > 10), with strong selection acting on 10–20% of synonymous sites in preferred codons."⁶¹  Flies, like yeast and bacteria, do select on codon choice.  Even there it is a minority of sites, and the authors name GC content as the chief confound.

Within humans: the frequency test

The polymorphism test needs no shared ancestry.  If most human synonymous mutations cost a percent or so of fitness, as most yeast mutants did, selection would keep them as rare as missense variants.  They are not.  In the UK Biobank, missense and protein-truncating variants "appear at much lower allele frequencies than synonymous variants", even for synonymous changes that swap a well-used codon for a poorly used one;  "These data firmly demonstrate that synonymous variants are under significantly weaker constraint than nonsynonymous variants".¹¹  The gnomAD project, which catalogues variation in 141,456 people, uses synonymous variants as its yardstick for this reason:  "For analyses of selection on coding variation, synonymous variation provides a convenient baseline, controlling for other potential population genetic forces".⁶²  The phrase is "convenient baseline", not "neutral", and Part 5 prices what that convenience costs.

Putting the numbers together

Every natural-population estimate lands between about one tenth and two fifths.  The yeast laboratory figure is two to four times higher.  Zhang and Qian, the senior authors of the yeast study, concede the gap for apes:  the human-chimp ratio "suggest[s] that the fraction of synonymous mutations that are deleterious is most likely below 50% in great apes, contrasting the estimate of at least 75% in yeast from lab-based DFE quantification, although these fractions may not be directly comparable".²  They are not directly comparable:  a laboratory measures effects above about 0.5% in one environment, while divergence integrates every effect large enough for selection to notice, in every environment, over the whole history of the lineage.  Even so, the arithmetic is unforgiving for a Shen-sized effect in apes.  A mutation that costs 1.1% of fitness in a population of 10,000 has a chance of fixing about 10⁻¹⁸⁹ times that of a neutral one;  if 76% of synonymous mutations were like that, human-chimp synonymous divergence would be at most about 24% of intergenic divergence.  It is 63% to 79%.¹⁶

Part 5: What follows for methods and arguments that assume silence

Shen and colleagues listed what would need re-examining if their result held generally:  tests of selection that compare synonymous with nonsynonymous change, mutation rates and population sizes inferred from synonymous variation, and the fact that "synonymous substitution-based dating of evolutionary divergences may be unjustifiable in some cases."³  Each of these can be priced with the natural-population figures of Part 4, which are the relevant ones for any species other than laboratory yeast.  The arithmetic is in the article's research script.¹⁶

dN/dS and tests for selection

The ratio dN/dS compares the rate of amino-acid change with the rate of synonymous change.  A ratio well below 1 is read as purifying selection on the protein, and a ratio above 1 as positive selection.  The method treats dS as the neutral rate.  If a fraction of synonymous sites is constrained, dS is too low and the ratio too high.  With human-chimp synonymous divergence at 0.635 to 0.791 of intergenic divergence, dN/dS overstates amino-acid change against a true neutral baseline by 1.26 to 1.57 times.  The genome-wide human-chimp ratio is 0.20 to 0.23,²⁷ so for a typical gene the bias does not change the verdict of purifying selection.  It matters for genes with ratios between about 0.6 and 1, which a corrected baseline could push across the line in either reading.  For splicing signals alone, Parmley, Chamary and Hurst estimated that "constraints on synonymous evolution within ESEs causes the true mutation rate to be underestimated by not more than approximately 8%."⁶³

Molecular clocks

There are two ways to turn synonymous divergence into a date, and the bias behaves differently in each.  A clock calibrated against fossils or other dated splits measures the rate on the same kind of site it then applies, so a constant constraint is built into the rate and cancels.  A clock calibrated with a mutation rate measured in families does not cancel:  family studies count new mutations at all sites, while the constrained synonymous sites accumulate fewer substitutions than mutations.  Treating synonymous divergence as neutral then makes a split look 21% to 37% more recent than the same mutation rate applied to neutral DNA would.¹⁶  Sharp and Li saw a version of this in bacteria in 1987, where genes differ so much in codon selection that "separate molecular clocks of synonymous substitution might be required for different genes."²²

The direction matters for a young-earth reader.  Correcting for synonymous constraint makes mutation-rate dates older, not younger, and by tens of percent, not by the factor of a thousand or more that separates mainstream dates from a young-earth chronology.  The general problems of clock calibration are treated in Molecular Clocks;  synonymous constraint is a modest entry in that ledger.

Mutation rates and population sizes

Population sizes inferred from synonymous diversity assume the diversity is neutral.  Background selection and GC-biased gene conversion both disturb it, and Pouyet and colleagues found that using such sites "for demographic inference can lead to strong biases."⁶⁰  Shen and colleagues argued that, if synonymous mutations are strongly harmful, "N e inferred from synonymous polymorphisms in natural populations is likely substantially underestimated".³  For humans the polymorphism data of Part 4 bound the size of this effect:  synonymous variants behave much more like neutral variants than like missense variants, and within humans fourfold sites carry more diversity than the genome average, not less.¹³

Genetic entropy and mutation load

John Sanford's Genetic Entropy argues that "there are probably no truly neutral nucleotide sites in the genome.  Therefore there is no way to change any given site without some biological effect, no matter how subtle.  While most sites are probably nearly-neutral, very few, if any, should be absolutely neutral."¹⁷  He lists codon preferences among the genome-wide patterns that show "the very subtle role that single nucleotides play".¹⁷  Robert Carter of Creation Ministries International went further in 2022, reading the yeast study as "the final nail in the coffin?" for neutral evolution and asking:  "So, you need selectively neutral alleles?  What neutral alleles?"⁶⁴  Carter reported the Kruglyak objection fairly and added that "even if the study is eventually retracted, this does not mean that ‘wobble’ bases are neutral."⁶⁴  David Coppedge of Creation-Evolution Headlines wrote that the yeast result made "Sanford’s prediction look overly optimistic!"⁶⁵

The evidence divides these claims.  Sanford's claim, that most synonymous sites are nearly neutral but not exactly neutral, is close to what Lu and Wu estimated for apes.  Carter's and Coppedge's stronger reading, that most synonymous mutations are strongly harmful, is the yeast result carried over to humans, and the human-cell screens and allele frequencies of Parts 3 and 4 do not support it.

The two readings also predict opposite things for the load argument.  New mutations per person in human coding DNA number about 0.78 per generation, at the commonly used rate of 1.20 × 10⁻⁸ per letter per generation,⁶⁶ of which a quarter to a third are synonymous:  about 0.19 to 0.26 synonymous mutations per child.¹⁶

1. If they were Shen-sized (76% harmful, median cost 1.1%), they would be strongly selected: about 0.15 to 0.2 such mutations per child, each removed within a few generations. That is a steady cost of about 0.16% to 0.22% of fitness per generation in selective deaths, but nothing accumulates, so it does not drive decline.
2. If they are Lu-and-Wu-sized (90% harmful, 2.5 to 4 × 10⁻⁵ each), selection barely sees them, and they behave as Sanford describes. If every one were fixed, with nothing restored by back-mutation, fitness would fall by at most about 0.1% to 0.2% over 200 generations, roughly the 6,000 years of a young-earth chronology, and by at most about 4% to 9% over 10,000 generations.¹⁶

Either way, synonymous sites are a small term in the genetic-entropy account.  The argument stands or falls on amino-acid changes and the much larger noncoding genome, treated in Genetic Entropy.

The codon-identity argument for design

Chapter 15 of the 2017 volume Theistic Evolution, by Ann Gauger, Ola Hössjer and Colin Reeves, answers the common-descent argument from shared codons this way:

"Synonymous codon use.  One of the evidences offered in favor of common descent is the “reuse” of synonymous codons.  Some codons—triplet nucleotide sequences in DNA—specify the same amino acid.  For example, lysine may be encoded by four different codons in RNA—CUU, CUC, CUA, or CUG.  The assumption has been that these codons truly are synonymous and ought to be able to be used interchangeably.  Yet when there is a choice among possible codons, chimpanzee and human genomes almost always use the same codon.  It is therefore argued that this is because the sequences share a common ancestor."⁴⁸

They reply that codons have other jobs, so "Codons may therefore be identical because they have to perform multiple jobs that are specific to one codon—none of the others will do.  The fact that both humans and chimpanzees use the same codons may be due to this rather than to their ancestry.  In addition, this is a strong argument for design:  the sophistication required to have codes within codes would not arise by simple mutation and natural selection."⁴⁸

There is a slip in the example:  CUU, CUC, CUA and CUG code for leucine, not lysine, which has two codons, AAA and AAG.²  The argument itself can be tested.  If human and chimpanzee codons matched because each codon had a job only it could do, synonymous sites would differ between the species far less than DNA with no job.  They do not.  Human and chimpanzee synonymous sites differ at 1.12% of positions, against 1.41% for intergenic DNA:  they change at 79% of the intergenic rate, and match at 98.9% against 98.6%.¹²  The matching is almost entirely the general closeness of the two genomes, with about a fifth to a third extra from constraint.  A design reply could hold that both genomes were made similar for shared function and have since drifted;  but then codon identity stops being evidence for design.  What remains for design is the minority of constrained sites, which Part 2 shows to be real.

This site's Functional DNA article

The published article Functional DNA is Evidence of Design uses a different route.  From Freedman and colleagues' 2011 table of disease-associated variants in coding DNA, 79 nonsynonymous and 26 synonymous, it infers that synonymous sites are "83% as functional as non-synonymous sites", and from that that 62% of synonymous mutations are harmful.⁶⁷  The data here suggest that figure is too high, for three reasons.  Genome-wide association hits are mostly markers for nearby causal variants rather than causes, as the sibling article GWAS and Noncoding Variants shows.  Hindorff and colleagues, counting the same kind of hits, found them "significantly overrepresented only in nonsynonymous sites" and promoters, not in synonymous ones.⁶⁸  And the direct measures of Parts 3 and 4 put synonymous constraint in apes at about 20% to 40%, and synonymous effects in human cells and clinics at a thirtieth to a fortieth of the missense rate per possible mutation.  The design case in that article does not depend on this one estimate, but the estimate should come down.

Objections and responses

"The 2022 Nature study proved most synonymous mutations are harmful.  Critics just dislike the result."  The critics named specific design flaws, and the strongest of them is checkable:  the one independent growth test did not confirm effects below about 2%, which is most of the 75.9%.⁷  The authors' 2024 rebuild answers the control objection for 11 genes, and their result held.⁸  So the yeast result is defensible.  What fails is extending it to humans:  two independent human-cell screens and the human allele-frequency data all find synonymous effects far rarer than missense effects.^{9 10 11}

"Human cells in a dish are not people.  Small effects that matter over generations are invisible to a growth screen."  True, and the same is true of the yeast screen, which cannot see effects below about 0.5%.³  That is why Part 4 matters:  population genetics sees effects of 10⁻⁵ and finds them at a fifth to two fifths of ape synonymous sites, not three quarters.  The skeptic and the creationist should both accept that tiny effects exist;  the dispute is about how many sites carry large ones.

"Divergence between humans and chimpanzees assumes common ancestry.  A young-earth creationist need not accept it."  Correct, and the article says so.  The within-human tests do not assume it:  synonymous variants are far more common than missense variants in the UK Biobank,¹¹ human fourfold sites carry above-average diversity,¹³ and the cell screens test the mutations directly.  All three point the same way as the divergence data.

"Intergenic DNA is not neutral either, so the 0.635 to 0.791 ratios understate synonymous constraint."  Partly right.  If some intergenic DNA is constrained, the true neutral rate is higher than Kᵢ, and the synonymous shortfall is larger.  For the shortfall to reach Shen's 76%, the neutral rate would have to be about 2.6 to 3.3 times the intergenic rate, which would mean most intergenic DNA was under strong constraint.  Ancestral-repeat controls, which avoid most of that problem, give a similar answer, about 22% to 27% for apes.¹⁴

"Conservation proves function.  If a fifth of synonymous sites are conserved, they are functional."  Some are.  But GC-biased gene conversion produces conservation-like patterns at synonymous sites without any function, and the authors of the largest mammal survey name it as one of three causes.⁵⁹  "Conserved" is an upper bound on "functional" here, not a measure of it.

"Codon bias in humans shows design:  the genome prefers certain spellings."  The bias exists, but in humans it follows the GC content of the surrounding DNA, including introns, and a repair-bias model explains 70% of its variation among genes.⁴  The same bias in bacteria and yeast does track translation, and there it is evidence of fine tuning by selection or by design, depending on one's starting point.

"Calling these mutations 'silent' slowed the search for function, as Carter argued."  Some of that is fair.  The SMN2 splicing change, the COMT RNA structure and the IRGM microRNA site were each, at first, assumed harmless because they changed no amino acid;  Brest and colleagues note the IRGM variant had been "classified as non-causative".⁴⁵  But the neutral assumption was never universal:  Sharp, Li and Ikemura were measuring selection on synonymous codons in the early 1980s.^{21 22}

"If synonymous constraint biases clocks, molecular dates are unreliable."  The bias is real and has a size:  21% to 37% for mutation-rate dates in apes, and none for a fossil-calibrated clock with constant constraint.¹⁶  Correcting it makes dates older.  It cannot carry a young-earth argument.

"Most synonymous sites are neutral, so this is a non-issue for evolutionary biology."  Not quite.  A fifth to two fifths of ape synonymous sites show constraint, and Lu and Wu estimated that most substitutions at them were slightly harmful.¹²  Laurence Moran wrote that "most (but not all) synonymous mutations are neutral or nearly neutral".⁶⁹  "Nearly neutral" is Sanford's category too.  The disagreement left is about whether slightly harmful changes accumulate fast enough to matter, which is the genetic-entropy question, not this article's.

Issues and unanswered questions

1. No independent repeat of the yeast experiment. Every yeast measurement on this scale comes from one laboratory. A rebuild by another group, with matched wild-type controls built the same way as the mutants and a library of wild-type-only edits to set the noise floor, would settle it.
2. The MDR1 C3435T case is unresolved. The folding effect appears in two cell systems from one laboratory, the messenger-RNA effect is disputed, and an editor's note on the 2014 follow-up appeared in 2025 that this article could not read.⁴¹
3. Human splicing rates for synonymous variants rest on reporters. The 3.0% figure comes from a reporter assay of rare variants.³² A genome-wide measure in the native context would be better. The prime-editing screen points the same way: most of its synonymous hits broke splicing.⁹
4. How much synonymous "constraint" is GC-biased gene conversion? No study yet splits the 20% to 40% figure into function and repair bias for humans. That split is the number the design argument would need.
5. The ClinVar count is our own. It counts classifications, not mechanisms, and is biased low by the benign-by-default rule.⁶ A published, mechanism-by-mechanism tally of pathogenic synonymous variants was not found.
6. The Lu and Wu estimate depends on a model of recessiveness. If harmful synonymous effects are not partly recessive, the 90% figure would change. It has not been repeated with later genomes, as far as this research found.
7. Hominid constraint exceeds murid constraint at fourfold sites,¹⁴ the opposite of what larger mouse populations would predict. Neither side has a settled explanation.

Verdict

Synonymous mutations are not all silent.  A measured minority alter splicing, RNA structure, translation and folding, and several hundred are classified as causing disease.  In apes, a fifth to two fifths of synonymous sites show the marks of selection, partly from function and partly from a repair bias that imitates it, and most substitutions at those sites appear slightly harmful rather than strongly harmful.

The 2022 claim that most synonymous mutations are strongly harmful is a real measurement in yeast, defended by its authors with new data and still unrepeated by others.  It does not carry over to human cells or human populations, and its own authors now concede the ape figure is below half.

For methods, treating synonymous sites as neutral biases dN/dS and mutation-rate clocks by 20% to 60%, which is worth correcting and changes no date by an order of magnitude.  For the design debate, Sanford's "nearly neutral" picture of these sites is closer to the evidence than the claim that most are strongly harmful.  The argument from shared human-chimp codons as evidence of design fails its own test.  The modest, defensible point is that redundant code positions carry a second layer of information at some sites, and that point stands on splicing, not on the yeast study.

Sources