Influenza

"glycans are added to H1 and H3 HAs at discrete 5-to-7-year intervals, until they reach a functional glycan limit." — Meghan Altman and colleagues, mBio, 2019¹

Summary

Influenza A has been watched longer and sequenced more deeply than any other human pathogen.  This article counts how many virus particles the last century produced, how many mutations they tried, and what the search bought.  Each change is scored as a gain, a loss, a modification or a transfer of function, by the definitions in Measuring Biological Information.

The main findings:

1. The record is 108 years long, not eighty. The 1918 genome was sequenced from preserved tissue and rebuilt by reverse genetics, and human influenza A has been isolated continuously since 1933.^{2 3} NCBI Virus held 1,665,035 influenza A sequences on 27 September 2026, 922,493 from human hosts.⁴
2. The search is the largest in this series. A century of influenza produced 10²¹ to 10²³ virus particles and 10²² to 10²⁴ new mutations.⁵ Every possible single-letter change arises millions of times inside one infected person.⁶
3. Almost none of that search is ever tested between people. The transmission bottleneck is one or two genomes,⁷ so the number of lineages that ever crossed from one person to another is smaller than the number of mutant genomes made by ten to twelve orders of magnitude.⁸
4. Antigenic drift is the cheapest change there is, and is neither a gain nor a loss. The major antigenic changes of H3N2 between 1968 and 2003 were single substitutions at seven reused positions beside the receptor binding site.⁹ The protein keeps binding its receptor and changes which antibodies see it: a modification, shifted.
5. There is a real gain of function, and it is the glycosylation sites. Human H1 haemagglutinin went from one head glycan to five, and H3 from two to seven.^{1 1} A new N-linked glycosylation site is a new protein-processing site, which is on Michael Behe's own 2010 list of gain-of-function mutations.¹⁰ Conceded, and not to be scored as loss.
6. The best-documented multi-step path in influenza is a glycan, not a drug. Three substitutions are needed for full glycosylation of the newest site, and one had already fixed two seasons earlier for an unrelated reason.^{1 1}
7. The record contains a measured ceiling. After the glycan limit was reached, seasonal H1N1 failed to produce a widely circulating six-glycan strain in at least 33 of 35 seasons, although laboratory work shows the protein tolerates eight.^{1 1} The finding comes from mBio, not from creationists.
8. Every pandemic was a transfer. 1957 and 1968 took avian gene segments;¹¹ 2009 combined six segments from a North American swine virus with two from a Eurasian avian-like swine lineage, all already working.¹² The largest events cost the fewest new nucleotides.
9. No new gene and no new protein arose. The genome is about 13,500 bases, almost all coding, with 749 nucleotides read in two frames at once.¹³
10. The genetic-entropy argument about influenza fails on its own primary sources. Carter and Sanford's central inference is that linear accumulation proves the change is not adaptive;¹³ Samir Bhatt and colleagues measured the adaptive part separately and found it linear too.¹⁴ Part 5 agrees with this site's Genetic Entropy article and prints two things in their favour.
11. The ledger is a lower bound, and the extrapolation to large animals is weaker than it looks. Part 6 says so plainly.

Part 1: The record, and the ledger

What has been sequenced

Influenza is the best-observed evolving organism on earth that was not put in a flask on purpose, and three things make it so.

The starting point exists as an object rather than an inference.  In 2005 Jeffery Taubenberger and colleagues finished sequencing the 1918 virus from preserved autopsy tissue, and Terrence Tumpey and colleagues used those sequences to build a live virus carrying all eight 1918 segments.²  It grew without added trypsin, killed mice and chicken eggs, and grew to high titre in human bronchial cells, none of which contemporary human H1N1 does.³

Human influenza A has been isolated continuously since 1933, so the catalogue phrase "eighty years of watching" is wrong in two directions.  Continuous isolation is 93 years to 2026;  the sequenced record reaches back 108.

And the sampling is extraordinarily dense.  On 27 September 2026 NCBI Virus held 1,665,035 influenza A nucleotide sequences, 922,493 of them from human hosts:  527,010 from H3N2 and 370,753 from H1N1.  Influenza B added 220,299.⁴  Those are segment sequences, not genomes:  the influenza genome is eight separate segments, so 922,493 human sequences is at most 115,311 complete genomes, and fewer in practice.¹⁵  GISAID holds more, but publishes no separate influenza figure, so none is printed here.⁴

How many virus particles

The ledger needs three measured numbers:  how many cells one infection uses, how many particles each cell makes, and how many mutations each replicated genome carries.

Prasith Baccam and colleagues fitted models of virus growth to daily nasal-wash titres from six adults deliberately infected with H1N1.  They fixed the number of target cells in an adult upper respiratory tract at 4 × 10⁸, from a lining area of 160 cm² and 2 to 4 × 10⁻¹¹ m² per cell.¹⁶  One infected cell went on to produce about 22 new productive infections, so a real infection uses a large fraction of that pool.¹⁶

Timo Frensing and colleagues counted what leaves a cell.  They infected canine kidney cells at a multiplicity of ten, so the whole population was infected at once, then replaced the supernatant every four hours and counted both total and infectious particles.  Total output was about 15,000 particles per cell;  infectious output was about 200.¹⁷  Before twenty hours, 2 to 4% of released particles were infectious;  later it fell below 0.3%.¹⁷

Two cautions go with that figure.  It is a cell-culture number at a high multiplicity, not a measurement in a person.  And cells vary enormously:  Frank Heldt and colleagues found an almost thousandfold spread in yield from single infected cells, most releasing only 1 to 75 infectious units and roughly 40% producing none at all.¹⁸  Baccam's paper also closes the easy route, saying that "there is no simple mapping between number of free infectious virions and experimental infectious viral titer"¹⁶  So the ledger multiplies a measured human cell pool by a cell-culture particle yield, and its range is wide on purpose.

How many mutations

The influenza mutation rate was measured properly in 2017.  Matthew Pauly, Megan Procario and Adam Lauring built a test in which reversion to fluorescence in twelve different mutant green-fluorescent-protein reporters reveals each of the twelve possible single-nucleotide changes, instead of the small biased subset that a drug-resistance or antibody-escape screen samples.  They measured 1.8 × 10⁻⁴ substitutions per nucleotide per strand copied for a laboratory H1N1 strain and 2.5 × 10⁻⁴ for a 2014 H3N2, four to thirty-five times the earlier sequence-based estimates.¹⁹  Their summary of what that means is the ledger's third input:

"we estimate that each replicated 13.5 kb genome contains, on average, 2 to 3 mutations"¹⁹

The older figure was an order of magnitude lower.  Rafael Sanjuán and colleagues, putting every published viral rate on a common footing in 2010, gave influenza A a 13.6 kb genome and 2.3 × 10⁻⁵ per nucleotide per cell infection.²⁰  Pauly's method measures all twelve mutation classes;  the earlier ones did not, and the correction runs upward.

The World Health Organization puts seasonal influenza at "around a billion cases" a year, including 3 to 5 million severe cases and 290,000 to 650,000 respiratory deaths.²¹  The century total corrects for the world holding 1.8 billion people in 1918 and 8.2 billion now;  every step is in the note.⁵  For comparison, HIV counts about 6 × 10²² virus particles and 10²² mutations in a century, and Plasmodium falciparum counts 1.5 to 7.5 × 10²¹ parasites and about 10¹⁹ mutations in 75 years.  Influenza's total is the largest of the three.

How thoroughly the genome was searched

There are about 40,500 possible single-letter changes to a 13,500-base genome.  At these rates every one of them arises between 3 million and 440 million times inside a single infected person, and between 3 × 10¹⁵ and 4 × 10¹⁷ times a year worldwide.⁶

Multi-step changes are also cheap.  A specific pair arising in the same replication happens about once in 2 × 10⁸ genomes, or 270 to 27,000 times in one infected person.  A specific triple happens about once in 3 × 10¹² genomes, and 2 × 10⁷ to 2 × 10⁹ times a year worldwide.⁶  So in influenza, mutational supply limits nothing;  whatever constrains its evolution, it is not the waiting time for mutations.

How little of the search survives

Here is what does limit it.  John McCrone and colleagues deep-sequenced 249 specimens from 200 people in a five-season household study covering 6,290 person-seasons, found 43 genetically validated transmission pairs, and fitted the size of the founding population.

"Maximum likelihood optimization of multiple transmission models estimated an effective transmission bottleneck of 1-2 genomes.  Our data suggest that positive selection is inefficient at the level of the individual host and that stochastic processes dominate the host-level evolution of influenza viruses."⁷

The fitted mean was 1.68 genomes on one model and 1.75 on another, with 95% of bottlenecks expected to be three or fewer.⁷  The rest of their data says the same from the other side.  Within-host diversity was low, a median of two minority variants and fewer than ten in 243 of 249 samples,⁷ and the population churns without going anywhere:  in samples a day or more apart only 57% of variants persisted, and 68% of those in the second sample were new or previously undetectable.⁷  Variants in antigenic sites sat no higher than other non-synonymous variants, and the authors concluded that positive selection within a host "rarely amplifies a newly generated variant to a frequency greater than 2%."⁷

Put the two halves together.  Between 10¹¹ and 10¹³ mutant genomes exist inside one infected person, and one or two of them found the next infection.  Over the century about 10¹¹ lineages crossed from one person to another, against 10²² to 10²⁴ mutant genomes made.⁸  Ten to twelve orders of magnitude of the search are thrown away at the doorstep.  Every published argument about influenza, including the young-earth ones, uses the census figure instead.

Part 2: Immune escape, and a measured ceiling

Antigenic drift costs one nucleotide

Derek Smith and colleagues mapped the antigenic evolution of human H3N2 from 1968 to 2003 from haemagglutination-inhibition data, and found that antigenic change tracked genetic change closely but not exactly:  "Antigenic evolution was more punctuated than genetic evolution"²²  Nine years later the same group explained why.  Björn Koel and colleagues tested reconstructed mutants and found the molecular basis:

"From 1968 to 2003, antigenic change was caused mainly by single amino acid substitutions, which occurred at only seven positions in HA immediately adjacent to the receptor binding site.  Most of these substitutions were involved in antigenic change more than once.  Equivalent positions were responsible for the recent antigenic changes of influenza B and A/H1N1 viruses.  Substitution of a single amino acid at one of these positions substantially changed the virus-specific antibody response in infected ferrets."⁹

Thirty-five years of the most-studied antigenic evolution on earth, and the big jumps cost one nucleotide each, at seven positions used over and over.

How much drift is there in total?  Samir Bhatt, Edward Holmes and Oliver Pybus extended a McDonald-Kreitman-style estimator to count adaptive substitutions gene by gene.  For H3N2 over 1977 to 2009, haemagglutinin adapted at 1.52 and neuraminidase at 1.23 adaptive substitutions per gene per year;  for H1N1 the figures were 1.02 and 1.06.¹⁴  Per codon the antigenic HA1 domain gave 0.0040 adaptations per codon per year in H3N2, against 0.0005 for the conserved HA2 domain, and neuraminidase surface residues gave 0.0054.¹⁴  About 80% of fixed non-synonymous changes in HA1 were adaptive, and about 85% in neuraminidase surface residues.¹⁴

Score:  modification of function, shifted.  The protein still folds, still binds sialic acid, still fuses membranes.  What changes is which antibodies recognise it.  Nothing is built and nothing is broken.  Under the conversion rule in Measuring Biological Information a shifted modification counts the substitutions it required, so drift's nucleotide count is roughly one per substitution.  But Koel's reused positions raise a question that article has not settled:  when a position that already produced a functional change produces it again, is that new unique sequence or a revisit?  The answer moves the numerator by an order of magnitude, and it is listed among the open questions below.

Something new: the glycosylation sites

Drift is not the only thing the haemagglutinin does.  It also adds sugar.

Haemagglutinin carries N-linked glycans, each attached at a three-residue motif, and a glycan sitting over an antibody target hides it.  So a substitution that creates a new attachment motif is a different kind of change from one that alters a surface:  it builds a feature that was not there.

Meghan Altman and colleagues measured how many head glycans 72 egg-grown H1N1 and H3N2 viruses carry, by how far the protein moved through a gel rather than by prediction, then mapped the history against every human H1 and H3 sequence in the database.

"H1N1s (triangles) add 3.3 +/- 0.2 kDa per glycan (or 13.1 kDa in total) while increasing from 1 to 5 glycans.  H3N2s (HA1; circles) add 2.5 +/- 0.2 kDa per glycan (or 12.5 kDa in total) while increasing from 2 to 7 glycans."¹

The H3 history is dated:

"The original (1968) pandemic H3 strains had 2 head glycans.  Four times since 1968, glycans were added to H3 HA at regular intervals of 5 to 7 years (N126 by 1974, N246/N122 by 1980, N133-plus-N122 replacement of N276 by 1998, and N144 by 2004)"¹

Seth Zost and colleagues confirmed the most recent H3 addition biochemically.  The clade 3C.2a viruses that emerged in 2014-2015 really are glycosylated at the new site, and antibodies raised against the egg-grown 2016-2017 vaccine strain, which lacks it, neutralise them poorly.²³  The gain had a measured consequence for human immunity:  it "likely contributed to reduced vaccine effectiveness during the 2016-2017 influenza season."²³

Score:  gain of function.  This is conceded, and the reason is Behe's own definition, written in 2010 before any of this was in dispute:

"A "gain-of-FCT" adaptive mutation is a mutation that produces a specific, new, functional coded element while adapting an organism to its environment.  The construction by mutation of a new promoter, intron/exon splice site, or protein processing site are gain-of-FCT mutations."¹⁰

A glycosylation site is a protein-processing site, the third item on his list.  Nine were built in human influenza haemagglutinin in a century, in the wild, on dated records, and one has been rebuilt in a laboratory from its component substitutions.  Under the series' standing rule that a real gain is never scored as a loss, this goes on the ledger as a gain, at one to three nucleotides per site.

The one case where the steps are dated

The 2009 pandemic H1N1 lineage arrived with one head glycan and added a second over the 2015-2016 season.  Altman's group took the case apart by reverse genetics, and the result is the best-documented multi-step adaptive path in the influenza record.

"Surprisingly, through reverse-genetics manipulation of the A/California/07/2009 strain, we found that, despite their having the amino acid sequence necessary for glycosylation, neither the N162 single mutants nor the N162/K163Q double mutants added a glycan to HA in virions.  Rather, three mutations, namely, N162 itself, K163Q, and, crucially, I216T, were required for complete glycosylation of the 162 site"¹

The single mutant that creates the motif is not detectably glycosylated;  the double mutant is partly glycosylated;  only the triple mutant is fully glycosylated.¹

Now the history.  Viruses carrying the first substitution were at low frequency in every season from 2009 and never swept.  The second fixed in 2013-2014, for an unrelated reason:  it made the virus antigenically distinct.  Only when the third appeared did the combination sweep the world, in 2015-2016.^{1 1}

So the enabling step was selected for something else, and the gain followed.  That is the pattern Jesse Bloom found for oseltamivir resistance in Part 4, in a second and independent system, and it is the stronger case, because what arrives at the end is a new coded element rather than the loss of a drug's grip.

One qualification matters.  An alternative one-step route exists:  a different residue in the same region needs only a single mutation to be glycosylated, and viruses using it were also present at low frequency throughout.¹  The three-step path is not the only door, so a design argument treating it as an isolated improbability overstates the case.

The ceiling

The same paper contains the finding that matters most to this series, and it is a measurement rather than an argument.

Glycans were added at five-to-seven-year intervals until a limit was reached, then swapped rather than added, at roughly double the interval, 10.3 ± 0.8 years for the re-emergent H1N1 lineage.^{1 1}  And then the additions stopped:

"During at least 33/35 seasons after the glycan limit was reached, sH1N1 failed to evolve a widely circulating strain with 6 glycans.  This history is consistent with a functional glycan limit."¹

The limit is not a limit of the protein but a limit in life:

"it has been shown that H1 HAs can accommodate at least 8 head glycans while maintaining in vitro fitness.  However, sH1 circulated for over 3 decades after reaching 5 glycans without fixation of a sixth glycan.  Interestingly, at 9 to 10 years after reintroduction in 1977, sH1N1 strains with 6 glycans did briefly circulate before being replaced by strains with only 4 glycans.  These findings point to in vivo glycan-based fitness costs involved in maintaining the glycan limit"¹

The costs are named:  glycan addition can alter receptor binding for the worse, make the virus easier for innate immune lectins to neutralise, or stress the machinery that folds the protein.¹

Put this beside Part 1.  The mutation is manufactured millions of times a year, the protein tolerates eight glycans in a dish, and thirty-three seasons of the most intense immune selection there is produced no sixth glycan that spread.  The wall is not mutational supply and not protein chemistry.  It is what the whole organism has to do at once.

Part 3: The pandemics were transfers

Four pandemics fall inside the record, and three were made by putting existing gene segments together in a new combination, which is called reassortment.  Under the scoring procedure, a sequence that arrived from another lineage is a transfer, and the nucleotide count in the receiver is zero.

1918 is the exception.  The 1918 polymerase proteins differed from the avian consensus at only a handful of residues each, and Taubenberger's group concluded that the virus "was not a reassortant virus (like those of the 1957 and 1968 pandemics), but more likely an entirely avian-like virus that adapted to humans."²  Across the whole record, "ten amino acid changes in the polymerase proteins consistently differentiate the 1918 and subsequent human influenza virus sequences from avian virus sequences."²  Score:  modification of function, shifted, host range.  Ten positions, and a bird virus works in a mammal.

1957 and 1968.  Yoshihiro Kawaoka, Scott Krauss and Robert Webster built phylogenies of one polymerase gene and found that the 1957 pandemic strain took it from an avian virus and kept it until 1968, when the 1968 strain took a different avian one:  "transmission of the PB1 gene from avian to mammalian species is a relatively frequent event."¹¹  Score:  transfer.

2009.  Gavin Smith and colleagues reconstructed the history from complete genomes plus fifteen newly sequenced swine viruses.

"the polymerase genes, plus HA, NP and NS, emerged from a triple-reassortant virus circulating in North American swine.  The source triple-reassortant itself comprised genes derived from avian (PB2 and PA), human H3N2 (PB1) and classical swine (HA, NP and NS) lineages.  In contrast, the NA and M gene segments have their origin in the Eurasian avian-like swine H1N1 lineage."¹²

Eight segments, four source lineages, nothing new, and the assembly happened long before anyone noticed:  the common ancestor of the outbreak and its nearest sampled swine relatives dates 9.2 to 17.2 years earlier, so "the ancestors of the epidemic have been circulating undetected for about a decade."¹²  Score:  transfer.  Zero new nucleotides.

That is the shape of the record.  The events that killed the most people and changed the most about human immunity cost the fewest new nucleotides, because they made nothing.  They dealt an existing hand differently.

Crossing into a new mammal

The cattle outbreak that began in March 2024 is the newest case.  Xiao Hu and colleagues sequenced H5N1 from dairy cattle, two cats, six wild birds and a skunk.  All were nearly identical and formed a new genotype inside an existing clade, after two reassortment events since 2023.  Their reading is that the cattle viruses "exhibit critical mutations in HA, M1, and NS genes but lack critical mutations in PB2 and PB1 genes, which enhance virulence or adaptation to mammals," while one human case linked to cattle had picked up a single long-known polymerase marker.²⁴  Score:  transfer by reassortment, plus one substitution in one human case.

Two laboratory experiments in 2012 asked how cheap a host jump is, by selecting for airborne transmission between ferrets.  Sander Herfst and colleagues reported that viruses which transmitted through the air consistently carried four substitutions in haemagglutinin and one in a polymerase protein.  No recipient ferret died, and the transmissible viruses stayed sensitive to oseltamivir and were still recognised by antisera raised against vaccine strains.²⁵  Masaki Imai and colleagues obtained droplet transmission from an H5 haemagglutinin carrying four mutations, on the other seven segments of a 2009 pandemic H1N1 virus.  Their own caveat must be quoted:

"however, we do not know whether the four mutations in the H5 HA identified here would render a wholly avian H5N1 virus transmissible.  The genetic origin of the remaining seven viral gene segments may also critically contribute to transmissibility in mammals."²⁶

So the honest reading is four or five substitutions, in a selection where an experimenter chose the direction, and in one of the two cases on a backbone already fully adapted to mammals.  Score:  modification of function, shifted, host range.  A receptor-binding site that preferred one sugar linkage now prefers another;  a polymerase that worked at one temperature now works at another.  Nothing is built.  The substitutions are named in both papers and are deliberately not listed here;  the reason is under the open questions.

Part 4: The drugs

Four drug classes have been used against influenza.  All were answered, three of them by a single substitution, and each answer is dated in the record.

Adamantanes

Amantadine and rimantadine block an ion channel the virus needs.  Rick Bright and colleagues screened more than 7,000 isolates sent to the WHO collaborating centre at the US CDC between October 1994 and March 2005.  Resistance rose "from 0.4% in 1994-1995 to 12.3% in 2003-2004", with 61% of resistant isolates since 2003 from Asia.²⁷

The next season it stopped being a trend.  Of 209 H3N2 viruses collected from 26 US states between October and December 2005, 193, or 92.3%, carried the substitution S31N.²⁸  The authors recommended that the drugs "not be used for the treatment or prophylaxis of influenza in the United States until susceptibility to adamantanes has been reestablished among circulating influenza A isolates"²⁸  Susceptibility never was.  A whole drug class was retired by one nucleotide.

Score:  modification of function.  The drug's grip on a working channel is lost and the channel keeps working.  The direction of the channel's own activity was not measured, so it is left unmeasured rather than guessed.

Oseltamivir, and the permissive mutations

Oseltamivir blocks the neuraminidase enzyme, and one substitution weakens the drug's binding.  It is written H274Y in one numbering convention and H275Y in another.  For eight years it was judged clinically unimportant, because it made the virus sick.  Then it swept the world.

"The His²⁷⁴->Tyr²⁷⁴ (H274Y) mutation confers oseltamivir resistance on N1 influenza neuraminidase but had long been thought to compromise viral fitness.  However, beginning in 2007-2008, viruses containing H274Y rapidly became predominant among human seasonal H1N1 isolates. ... The evolution of oseltamivir resistance was therefore enabled by "permissive" mutations that allowed the virus to tolerate subsequent occurrences of H274Y."²⁹

Jesse Bloom, Lizhi Ian Gong and David Baltimore found what the cost was and what paid for it.  The resistance substitution reduces how much neuraminidase reaches the cell surface.  Two earlier substitutions, V234M and R222Q, restore it, and both were already fixed on the branch where resistance spread.  Neither is anywhere near the resistance site in the folded protein.²⁹  In culture, the resistant virus alone grew to about a hundredfold lower titre than wild type;  with both enabling substitutions it grew as well as wild type, and under oseltamivir it beat the single mutant by more than a hundredfold.²⁹

Four points about this case are usually reported badly, in one direction or the other.

It is a real case of required enabling steps.  Resistance "attenuates seasonal H1N1 unless there are permissive secondary mutations that maintain adequate surface NA expression."²⁹  The path had an order, and the order was not optional.

The enabling steps were not harmful and were not waiting for the resistance mutation.  Bloom's own list of what may have driven them is "stochastic drift, genetic hitchhiking, or selection for antigenic change or tuning of the NA/hemagglutinin balance."²⁹  Nothing had to cross a valley.  Two substitutions that were useful or neutral for other reasons opened a door.  The resistance step itself may not have been driven by the drug either:  "the fact that initial geographic prevalence of H274Y was not correlated with oseltamivir usage is generally viewed as evidence against direct selection for resistance"²⁹

The whole path was climbed inside a decade, which is what Part 1 predicts:  a specific pair of substitutions is manufactured hundreds to tens of thousands of times inside a single patient.

Score:  modification of function, shifted.  The enzyme keeps its catalytic job;  what changes is its grip on a drug and how much of it reaches the cell surface.

Baloxavir

Baloxavir blocks a different enzyme and was licensed in 2018.  Resistance appeared inside the licensing trial:  "polymerase acidic protein variants with I38T/M/F substitutions conferring reduced susceptibility to baloxavir occurred in 2.2% and 9.7% of baloxavir recipients in the phase 2 trial and phase 3 trial, respectively."³⁰  About one treated patient in ten, from a single dose, at three alternative substitutions in one codon.  Leah Lee and colleagues then put resistant and wild-type viruses into direct competition in ferrets:  the H3N2 resistant virus lost within-host fitness but transmitted about as well, and the 2009-lineage H1N1 resistant virus matched within a host but transmitted much less well.³¹  Score:  modification of function, down.  It is also the Bloom pattern waiting to happen:  a costly resistance change looking for something to pay for it.

Part 5: The genetic-entropy claim

What Carter and Sanford argued

In 2012 Robert Carter and John Sanford published the young-earth movement's main paper on influenza, in a secular journal.  They aligned 3,755 human and 351 swine complete H1N1 genomes and counted differences from the reconstructed 1918 genome.  Their reported result:

"At the time of its disappearance in 2009, the human H1N1 lineage had accumulated over 1400 point mutations (more than 10% of the genome), including approximately 330 non-synonymous changes (7.4% of all codons).  The accumulation of both point mutations and non-synonymous amino acid changes occurred at constant rates (mu = 14.4 and 2.4 new mutations/year, respectively), and mutations accumulated uniformly across the entire influenza genome."¹³

Their conclusion:

"We suggest H1N1 has been undergoing natural genetic attenuation, and that significant attenuation may even occur during a single pandemic.  This process ... has apparently contributed to the exponential decline in mortality rates over time, as seen in all major human influenza strains."¹³

Their argument has four legs:  the linearity of the accumulation, the extinctions of human H1N1 in the 1950s and in 2009, the decline in mortality since 1918, and a drift of codon usage away from human, swine and duck preferences.  The reasoning that ties the first leg to the conclusion is one sentence:

"We feel that the 15% divergence must be primarily non-adaptive because adaptation should occur rapidly and then reach a natural optimum.  Yet, we see that divergence increases in a remarkably linear manner."¹³

Where it fails

The linearity argument is answered by a direct measurement, made a year earlier.  Bhatt, Holmes and Pybus measured the adaptive component of influenza's change separately from the rest, using a method built for exactly that purpose.  What they found:

"Importantly, all genes have accumulated adaptive substitutions at an approximately constant rate;  we found no evidence for periods of selective activity and stasis in either subtype"¹⁴

Adaptive change in influenza is linear in time, so linearity does not separate adaptive from non-adaptive change, and the inference fails.  The reason is that influenza's target is not fixed:  human immunity moves every year, so adaptation never reaches an optimum and stops.  Neither paper cites the other.

No fitness was measured.  Fitness in virology means replicative or transmission success, measured by competition assay.  The paper runs none;  its two proxies are case-fatality rate and codon bias.

The mortality proxy has a large documented cause that has nothing to do with the virus's genome.  David Morens, Jeffery Taubenberger and Anthony Fauci examined lung sections from 58 autopsies of 1918 deaths and reviewed 109 published autopsy series covering 8,398 autopsies.  The samples "uniformly exhibited severe changes indicative of bacterial pneumonia", and the published series "clearly and consistently implicated secondary bacterial pneumonia caused by common upper respiratory-tract bacteria in most influenza fatalities."³²  In one series of about 9,000 patients followed to autopsy, sterile cultures grew pneumococci or streptococci from 164 of 167 lung samples.³²  1918 was the last influenza pandemic before antibiotics.

The extinction evidence describes something real and misreads its cause.  Human H1N1 did vanish after 1957 and again after 2009, and both times it was displaced by a virus carrying a surface protein the population had no immunity to.  H2N2 replaced H1N1 in 1957 after 39 years;  H3N2 replaced H2N2 in 1968 after 11.  A lineage dying of load at 11 while another survived to 39 is not following a clock.

And there is a clean control, which points the other way.  One influenza lineage really did go extinct inside the observation window.  B/Yamagata has scarcely been detected since March 2020.  Wenjie Han and colleagues combined molecular, antigenic and epidemiological data with simulation and found the cause:

"the probable extinction of B/Yamagata was driven by reduced transmission due to nonpharmaceutical interventions (NPIs) and a depleted susceptible population caused by conserved antigenicity and the 2017/2018 outbreak.  Specifically, B/Yamagata exhibited slower antigenic evolution ... consistent with its weaker positive selection pressure."³³

It went extinct because it changed too little, not too much, and in simulation it would have survived had it drifted substantially.  Extinction in influenza is not diagnostic of mutational load.

One thing this article does not charge them with.  This site's Genetic Entropy notes say Carter removed reassortant sequences and then presented their absence as decay.  That could not be confirmed:  the methods say only that "non-human/swine, non-H1N1, and incomplete genomes" were removed,¹³ and the paper itself states that the 2009 virus "was due to a reassortment between two swine viruses" and "was not derived from the 1976 re-introduction virus."¹³  They knew.  The accurate charge is narrower and still fatal:  they count the replacement of one circulating lineage by an unrelated one as an extinction caused by mutational load.

What survives

The genome really has almost no room.  Carter and Sanford's own description is accurate and is a genuine constraint:

"The influenza genome consists of eight RNA segments totaling over 13,100 nucleotides.  These code for up to eleven distinct proteins, two with alternate reading frames and one through alternate splicing. ... Almost all of the influenza genome is protein coding, and 5.7% (749 nucleotides) of the genome codes for two protein products simultaneously"¹³

And the mainstream concedes the mechanism they appeal to.  Taubenberger's group, in Nature, on one polymerase gene:

"In humans, the PB1 proteins experience linear change over time.  Indeed, PB1 in humans acquires <0.4 amino acid changes per year.  As there is such strong antigenic selection on human viruses, it is possible that although the observed changes in PB1 are selectively beneficial with respect to antigenicity, they are mildly deleterious to enzyme function.  Such complex fitness trade-offs are thought to be commonplace in RNA virus evolution."²

Pauly's group, in eLife, on where influenza sits:

"This mutation rate clearly places influenza close to a theoretical maximum rate, and we and others have shown that small increases in the virus' mutation rate lead to considerable losses in genome infectivity"¹⁹

Between 28 and 31% of random mutations in influenza are lethal, and only 34 to 49% of replicated genomes can replicate again.¹⁹  Mutational load in influenza is real, quantified and mainstream.  What the mainstream denies is that the load is heading anywhere.  Purifying selection removes it at every step, and the 1968 H3N2 lineage has accumulated change at the same steady rate for 58 years without going anywhere.

Part 6: What cannot be priced, and what the ledger adds up to

Gains that cannot be priced

Whatever nobody looked for.  Drug resistance is found because patients stop getting better, and drift because a vaccine stops working.  A change that made influenza slightly better at surviving in air would leave no such signal.  Bhatt's method found significant adaptation in genes nobody had flagged, including the nucleoprotein, a non-structural protein and the ion channel, and the causes "are unclear."¹⁴  Those substitutions are real, counted and unexplained, so they cannot be scored.

The transfers.  Reassortment scores zero new nucleotides in the receiver, which hides real work done somewhere else.  Every segment of the 2009 virus was originated somewhere, and the debt moves to birds and pigs rather than being cancelled.

The 1918 host jump.  Nobody knows how large the avian population was that produced the 1918 virus, or over how long, so its ten polymerase positions cannot be priced.

The ledger below is therefore a lower bound, higher than what is listed by an unknown amount.  That weakens the argument this series is making, and the data that exist cannot remove the weakness.

The ledger in one table

One gain, one loss, seven modifications, three transfers and one class that cannot be scored.  No new gene, no new protein and no new binding between two proteins, in a century of the densest sampling any organism has had.

The unit of account

The series measures unique functional sequence:  nucleotides that had to be originated rather than copied, per organism born.  The numerator can be counted two ways, and the answers differ by an order of magnitude.

Counting every adaptive substitution that fixed.  Bhatt's rates give 88 adaptive substitutions in H3N2 haemagglutinin and neuraminidase across 1977 to 2009, and 67 in H1N1.³⁴  Of the roughly 325 amino-acid changes Carter and Sanford counted in the human H1N1 lineage from 1918, Bhatt's adaptive fractions make about 100 to 200 adaptive.  So the numerator is of order 100 to 300 nucleotides per lineage per century.

Counting distinct positions ever used.  Seven positions produced the major antigenic changes of H3N2 and were reused, nine glycosylation sites were built, four positions carry the drug resistances, and ten separate human from avian polymerases.  The distinct count is about thirty.

Divided by the 3 × 10²¹ to 3 × 10²³ virus particles made since 1918, the rate is between 10⁻²² and 10⁻¹⁹ nucleotides of new functional sequence per particle.³⁴  The same calculation gives about 8 × 10⁻²⁰ per particle for HIV and 2 × 10⁻²¹ to 7 × 10⁻²⁰ per parasite for Plasmodium falciparum.  Three organisms with different genomes, mutation rates and life cycles land in the same band.  That convergence belongs to Benchmarking Evolution.

Two cautions.  The figure is a lower bound.  And it counts sequence, which is not function:  a new glycan site is a real new feature of a protein and it cost one to three nucleotides.

How far this carries to large animals

Not as far as the ledger looks.

Influenza shows that in an organism with an enormous population, a very high mutation rate and about the strongest selection nature applies, a century of search produced a hundred-odd useful substitutions, nine new glycosylation sites, no new gene, and a ceiling on the one class of gain it did produce.  If supply were the limiting factor for novelty, influenza should have produced far more.

The limits the record shows are three, and none of them is supply.  Selection must favour each step in the host it is in.  The transmission bottleneck discards almost everything.  And functional trade-offs put a wall in front of the one gain that was accumulating.

Whether those limits scale to mammals is not settled here, since a mammal genome has room to add and influenza's has almost none.  So the argument is not "mammals could not have done it" but something narrower:  where the search is large enough to see clearly and the record complete enough to audit, the output of gains is small, the classes are mostly not gains, and one accumulating gain hit a measured ceiling that mutation could not push through.

Objections and responses

"The virus is under stabilising selection for a job it already does well, so of course it makes nothing new.  Novelty comes from new niches."  This is the strongest objection and it partly succeeds.  Influenza's environment changes in one way, by human immunity, and the answer is on the surface of two proteins.  But influenza does get new niches:  a new host species in 1918, 1957, 1968, 2009, and cattle in 2024.  Three of those five were transfers of existing segments and the other two were handfuls of substitutions that shifted an existing binding site.  When influenza got a new niche it did not build for it;  it recombined and tuned.

"Thirteen thousand five hundred bases with overlapping reading frames leaves no room to add anything, so the absence of a new gene proves nothing."  This succeeds, and Carter and Sanford make the point better than the mainstream does.¹³  So the article leans not on "no new gene" but on the glycan ceiling, which is not a constraint of space:  there is room for a sixth glycan and thirty-three seasons of selection did not fix one.¹

"The 10²² figure is arithmetic, not a measurement."  Correct, and the article says so.  Its weakest input is the particle yield per cell, which is a cell-culture figure at high multiplicity when cells vary a thousandfold.¹⁸  The range spans two orders of magnitude for that reason, and every step is in the note.⁵

"If almost the whole search is discarded at the bottleneck, the 10²² figure does not support the design argument either."  This is right, and it is the most important thing here that cuts both ways.  A young-earth argument that says "influenza tried 10²² mutations and got almost nothing" overstates the effective search by ten orders of magnitude.  The effective figure, about 10¹¹ lineages that crossed a host boundary, is still very large compared with anything available to mammals, but it is not 10²², and nobody on either side has done the calculation properly.⁸

"The glycan sites are just antigenic escape with extra steps, not a real gain."  No.  A glycan attachment site is a functional coded element by Behe's own published definition, and its creation is on his own list of gain-of-function mutations.¹⁰  Redefining it after the fact to exclude the case would be exactly the move this site criticises others for.  It is a gain, it was watched, and it happened nine times.

"Then the ceiling is just a fitness cost, and fitness costs are everywhere."  Partly right.  A fitness cost is a local wall, not a universal one, and the paper's own candidate explanations are specific to haemagglutinin:  receptor binding, innate immune lectins, protein folding.¹  What makes the case worth the space is that all three usual explanations for a missing adaptation are absent.  The mutation is supplied constantly, the protein tolerates the change in a dish, and the selection pressure is unrelenting.  The adaptation still did not happen, for thirty-three seasons.  That is a cleanly measured instance of the whole organism vetoing an improvement to one part, which is the pattern the design argument predicts and the one comparative genomics cannot see.

Issues and unanswered questions

1. The mammalian-transmissibility substitutions are deliberately not listed. Both 2012 ferret papers name them and the citations here point straight at those papers. This article prints only the counts the authors put in their own abstracts, and scores from those counts. A compiled, explained list of the changes that make a highly pathogenic avian influenza virus transmissible between mammals is the one part of this topic that could do harm assembled in one place for a general readership, and no argument above needs it. The research notes record the same decision.
2. Reused positions. Whether a second functional change at an already-used position counts as new unique sequence is not settled by Measuring Biological Information, and the answer moves this numerator by an order of magnitude.
3. The effective search has not been computed. The inputs all exist in McCrone's paper: the bottleneck distribution, the within-host frequency distribution, and a model for the chance that a variant at a given frequency passes a bottleneck of a given size. Nobody has run it.
4. No competition assay between early-lineage and late-lineage influenza exists. That is the experiment that would settle Carter and Sanford's claim either way, and reverse genetics has been able to do it since 2005.
5. Seven primaries were read in abstract only for access reasons: Smith 2004, Koel 2013 and Tumpey 2005, paywalled at Science with the Internet Archive offline on the day of writing, plus Bright 2005, Bright 2006, Nakajima 1978 and Hayden 2018. Every number taken from them is inside the quoted abstract. The seven H3 positions Koel identified were not read from his paper, and nothing here depends on the list.
6. The particle yield per cell is the weakest link in the ledger. No measurement of genomes replicated per infected human exists in the literature read, and Baccam's paper says the available human measurement cannot be converted.¹⁶
7. The count of nine glycosylation sites is assembled from the H1 series (one to five head glycans), the H3 series (two to seven) and the 2009 lineage's addition, with swaps not counted. A stricter count requiring each site to be shown glycosylated biochemically rather than predicted would be smaller; two are confirmed that way.
8. The loss row in the ledger is the one row with no source, and is marked unverified in the table. Shi 2012 on the accessory protein PA-X is the primary to read; the count of one loss is provisional until it is.

Verdict

A century of influenza produced between 10²² and 10²⁴ new mutations, the largest measured search in this series.  Twelve adaptive changes can be scored:  one gain, one loss, seven modifications and three transfers, plus a class of counted but unexplained adaptation that cannot be scored.  No new gene, no new protein, no new protein-protein interaction.

The gain is real and is conceded without hedging.  Nine new glycosylation sites were built in haemagglutinin, one of them from three substitutions with the enabling step selected two seasons earlier for a different reason.  A glycosylation site is a functional coded element on Behe's own published list, and this site will not redefine it to win an argument.

What makes the record worth the space is the measured ceiling.  After the glycan limit was reached, thirty-three of thirty-five seasons produced no further addition, although the mutation is manufactured millions of times a year and the protein carries more glycans in a dish.  The wall is not mutational supply and it is not protein chemistry.  It is what the whole virus has to do at once, measured by virologists who were not arguing about anything.

The young-earth argument about influenza, as published, fails.  Carter and Sanford's central inference is refuted by a measurement made a year before their paper, their mortality proxy has a documented bacterial cause, and the one lineage that really did go extinct went extinct for the opposite reason.  Two of their factual claims stand, and the mainstream concedes their mechanism while denying their conclusion.

The extrapolation to large animals is narrower than the ledger looks.  Influenza's genome has almost no room to add anything, so the absence of a new gene there proves little.  The glycan ceiling proves more, because space was not the constraint.

Sources