H5N1 is an influenza A virus of birds. It is called highly pathogenic because of what it does to poultry, not because of what it does to people [1]. It has infected humans sporadically since 1997, sometimes severely. Since 2021 a particular lineage has spread widely in wild birds and has repeatedly infected mammals, and in 2024 it turned up in American dairy cattle — a genuinely new development that is worth understanding precisely rather than vaguely [2].
The single most important fact about H5N1 is a negative one: it does not transmit efficiently between humans. That is not a reassurance that can be banked permanently, because it is a property of the virus that could change. But it is the current state of affairs, and understanding why it matters — and what specifically would have to change — is the difference between informed watchfulness and undirected alarm. This review tries to lay out that logic.
Start here: what H5N1 is, and what the barrier is made of
Influenza A viruses are fundamentally waterfowl viruses. Wild aquatic birds are the natural reservoir, carrying an enormous diversity of subtypes largely without disease [3] [4]. Occasionally one crosses into poultry, and occasionally a poultry-adapted lineage acquires the ability to cause severe systemic disease in chickens — that is what "highly pathogenic" designates.
H5N1 in its modern form emerged in southern China, first infecting humans in Hong Kong in 1997, when the virus that infected people was shown to be closely related to a highly pathogenic avian virus [5] [6]. The lineage's genesis and its subsequent diversification into multiple sublineages across Asia have been traced in detail [7] [8] [9].
The barrier to human transmission is not a single wall. It is at least three things, and this matters because each is a separate thing to watch.
The first is receptor binding. Influenza attaches to sialic acid on host cells, and avian viruses prefer a different linkage from human viruses — avian-type alpha-2,3 versus human-type alpha-2,6 [10]. The structure of the hemagglutinin from a Vietnamese H5N1 isolate showed a clear avian alpha-2,3 preference; notably, mutations that convert avian H2 and H3 hemagglutinins to human receptor specificity, when placed on the H5 framework, did permit binding to a natural human alpha-2,6 glycan — which the authors described as suggesting a path for the virus to gain a foothold in the human population [11]. This is the same switch that has been reconstructed for past pandemic viruses [12] [13].
The second is where in the airway the virus lands. Human H5N1 infection has been associated with attachment to the lower respiratory tract rather than the upper airway [14]. A virus that replicates deep in the lung can cause severe disease while being poorly positioned to be coughed into someone else's nose — which is part of why the historical human cases have been both serious and non-transmissible.
The third is replication in a mammalian host at mammalian body temperature, which involves the viral polymerase, and where changes in the PB2 subunit are the classic markers of adaptation [4].
None of these is hypothetical, and none has fully occurred in nature in a way that produced sustained human transmission. That is the current position.
Pillar 1: measurement, diagnosis and surveillance
What a human case has looked like
Historically, severe. A series of ten patients in Vietnam described a serious respiratory illness [15], and fatal outcome has been associated with high viral load and an intense cytokine response [16] [17], with unusual presentations including a child presenting with diarrhoea and coma [18].
That picture requires an important qualification, and this review would be misleading without it. Cases are detected because someone is ill enough to be tested, so the severity of detected cases overstates the severity of infection. There is direct evidence of this: in a retrospective cohort after the 1997 Hong Kong outbreak, 3.7 percent of 217 exposed health care workers were H5N1-seropositive against 0.7 percent of 309 unexposed workers — infections that were identified serologically, not because those workers became critically ill [19]. Milder presentations are also well documented for other avian subtypes: an H7N7 outbreak in the Netherlands caused mostly conjunctivitis alongside one fatal case [20], conjunctivitis was the presentation in an early avian influenza case [21], and H7N3 in British Columbia behaved similarly [22]. The dairy-associated human cases in 2024 fit that milder pattern rather than the severe one [23].
The laboratory
Diagnosis is molecular. Influenza A is confirmed and then subtyped by RT-PCR; universal primer sets allow full-length amplification of all influenza A viruses for sequencing [24], and reverse genetics from plasmids allows a virus to be reconstructed and studied [25]. Sequence sharing through GISAID is what makes global comparison possible in near real time [26].
Serology answers a different question: not "is this person infected now" but "was this population infected at all". That is how subclinical infection in exposed workers gets counted [19], and it is the measurement that would tell us whether quiet human infection is more common than case reports suggest.
Surveillance
Surveillance is the actual front line, and it runs across species. In birds, it means monitoring wild reservoirs [3] and poultry flocks. In mammals, the last few years have supplied an unusual amount of evidence: H5N1 clade 2.3.4.4b in farmed mink in Spain [27], in marine mammals and seabirds in Peru [28], in seals in New England [29], and then in dairy cattle and cats in the United States [2] [30], with ongoing situation reporting [31]. Influenza A in cattle was not previously a prominent concern [32].
Each mammalian spillover matters less as a threat in itself than as an opportunity for the virus to sample mammalian adaptation. That is the reason to count them.
Centerpiece: a simple simulatable model of the epidemic threshold
The pandemic question has a precise mathematical form, and stating it precisely is the best defence against both complacency and alarm.
In the standard SIR framework, an infection's fate in a susceptible population is governed by the basic reproduction number R0 — the average number of further infections caused by one infected individual. Two standard results follow, neither with any free parameter.
The probability that a single introduction starts a self-sustaining outbreak, under a branching-process approximation with a geometric offspring distribution, is zero for R0 at or below one, and 1 − 1/R0 above it. And the final epidemic size Z — the fraction of the population eventually infected — is the root of the classic final-size equation Z = 1 − exp(−R0·Z), which has only the root zero at or below one, and a positive root above it.
The threshold at R0 = 1 is not a convention or a fitted breakpoint. It is the value at which both expressions change character. Below it, every chain of transmission terminates. Above it, chains can run away.
Two grounded anchors give the picture scale. The 2009 H1N1 pandemic — a virus that did cross — was estimated at R0 of 1.4 to 1.6 by three epidemiological analyses, with a genetic analysis giving a central estimate of 1.2 [33]. That is barely above one, and it is enough: at R0 = 1.4 the final-size equation gives an eventual attack rate around 51 percent. Meanwhile, simulation work on containing an emerging strain at source found that targeted antiviral prophylaxis had a high probability of containment if R0 was below 1.60, rising to 2.1 with pre-vaccination and 2.4 with quarantine added [34] [35] [36]. Those are the dotted ceilings in the figure, and the 2009 band straddles the lowest of them.
Where does H5N1 in humans sit? Below one — but the honest answer is that no reproduction number has been estimated for it, which is why the figure shades a region rather than marking a point. What exists is the observational evidence that chains terminate. In 2004, at least 44 people were infected across eight Asian countries, killing 32, and most had close contact with poultry; a family cluster in Thailand appeared to involve person-to-person transmission to a mother and an aunt who gave unprotected nursing care, but no additional chains of transmission were identified, and sequencing found no change in the receptor-binding site of the hemagglutinin [37]. Probable limited person-to-person transmission has been reported in China as well [38]. Transmission is not zero. It does not sustain.
The right-hand panel is where that observation becomes useful. Below the threshold, every introduction dies out — but not immediately. The expected total number of cases descending from one introduction is 1/(1 − R), which is 2 at R = 0.5, 5 at R = 0.8, and 10 at R = 0.9. Cluster size is a magnifying lens on R: it barely moves while R is small and climbs steeply as R approaches one. A reported family cluster of three, taken as a mean, would correspond to R of about 0.67 — though one observed cluster is emphatically not a mean, and should not be read as an estimate.
That is the teaching point, and it is the operational one. The whole pandemic question is whether adaptation pushes R across one — and because cluster sizes grow before the threshold is crossed, growing chains of human cases are a signal that is visible in advance. This is precisely why surveillance counts human clusters and screens sequences for transmissibility markers, rather than waiting for something to be obviously happening.
Three honest limits. The geometric offspring distribution is a modelling convenience; influenza superspreading is overdispersed, which lowers the probability of a major outbreak at the same R0 without moving the threshold. The final-size equation assumes a homogeneously mixing, fully susceptible population, which no real population is. And R0 taken alone is a poor summary of an epidemic — a point the epidemiological literature makes about itself [39] [40].
The same threshold logic governs the other transmissible diseases in this collection, which is why it is worth learning once.
Pillar 2: countermeasures, honestly
Antivirals
Two classes exist. The adamantanes interfere with viral uncoating, work only against influenza A, and are compromised by rapid, stable, transmissible resistance. The neuraminidase inhibitors — oseltamivir and zanamivir — block release of progeny virus from infected cells, are much less toxic, and must be given early because viral replication peaks 24 to 72 hours after symptom onset [41].
The evidence for how much they help is weaker than their stockpile status implies, and it is important to say so plainly. A Cochrane review that obtained 107 clinical study reports from regulators and manufacturers found that oseltamivir shortened time to first alleviation of symptoms in adults by 16.8 hours (95 percent confidence interval 8.4 to 25.1), that prophylaxis reduced symptomatic influenza, but that the treatment trials did not settle whether complications such as pneumonia are reduced, and that oseltamivir increased nausea, vomiting, psychiatric and renal adverse events [42]. The same programme documented that 60 percent of patient data from phase 3 oseltamivir treatment trials had never been published [43]. These drugs are not useless and they are not what the stockpiling rhetoric of the 2000s suggested.
Resistance is a live concern specific to H5N1: oseltamivir resistance has been documented emerging during treatment of H5N1 infection [44]. Encouragingly, the airborne-transmissible ferret-adapted H5N1 viruses remained sensitive to oseltamivir [45].
Vaccines and the manufacturing problem
The structural difficulty is timing. Vaccine production by conventional methods cannot be carried out at the speed required to halt a new strain, so vaccine would likely not be available for the first wave [41]. Stockpiling candidate vaccines against H5 strains partially addresses this, and the ferret-transmissible viruses reacted well with antisera raised against H5 vaccine strains — evidence that stockpiled antigens are not obviously mismatched [45].
Two lines of work attack the timing problem directly. mRNA platforms were demonstrated against avian subtypes well before COVID-19: lipid-nanoparticle mRNA vaccines encoding H10N8 and H7N9 hemagglutinins produced robust responses in mice, ferrets and non-human primates, protected mice from lethal challenge, and gave very high seroconversion rates in a phase 1 human study, with production speed and scale explicitly cited as the motivation [46]. Separately, chimeric-hemagglutinin designs aim at broad, subtype-spanning immunity [47], as does work on broadly neutralising antibodies against group 1 and group 2 hemagglutinins [48] [49].
Prevention, which is the part that actually works
Most of the risk reduction is unglamorous: poultry and dairy biosecurity, culling and movement controls, protective equipment for exposed workers, and not drinking unpasteurised milk from affected herds — the cats that died in the 2024 dairy outbreak had been fed raw colostrum and milk [2]. Modelling of containment at source and of layered community mitigation describes what would be attempted if transmission did begin [34] [50] [51] [52], and ordinary measures such as hand hygiene have measurable effects on influenza transmission [53] [54].
Pillar 3: the news, and how to read it
The dairy-cattle outbreak
In 2024, H5N1 clade 2.3.4.4b was reported in dairy cattle in Kansas and Texas, reflecting continued spread of a lineage that entered the United States in late 2021. Infected cattle had non-specific illness, reduced feed intake and an abrupt drop in milk production; fatal systemic infection developed in domestic cats fed raw colostrum and milk from affected cows. Cow-to-cow transmission appeared to have occurred, since infections were seen on Michigan, Idaho and Ohio farms that had received cows from infected herds. The authors' framing is the right one: the detection of virus in unpasteurised bovine milk is a concern because of potential cross-species transmission, and continued surveillance is needed to prevent mammal-to-mammal transmission [2] [30].
What this does and does not mean is worth being careful about. Mammal-to-mammal transmission in cattle is a real change, and it enlarges the surface over which the virus can sample adaptive mutations. It is not human-to-human transmission. The human cases arising from it have been few and mild [23].
What adaptation would actually look like
Two experiments define the field's understanding, and both were controversial. In one, H5N1 was modified by site-directed mutagenesis and then serially passaged in ferrets until it became airborne-transmissible between them; four amino acid substitutions in hemagglutinin and one in PB2 were consistently present in the transmitted viruses, and none of the recipient ferrets died [45]. In the other, an H5 hemagglutinin was experimentally adapted to confer respiratory-droplet transmission on a reassortant H5/H1N1 virus in ferrets [55].
How to read these matters. They establish that the required adaptation is reachable without reassortment in an intermediate host — a genuinely important and sobering finding. They do not establish that it is likely, and ferret transmissibility is a model of human transmissibility, not a measurement of it. Their practical value is the list of markers they produced: specific changes that surveillance can now look for in sequences from birds, mammals and people. That is how a laboratory result becomes an early-warning system.
Precedent, in both directions
Two subtypes have run this course partway. H7N9 in China produced a substantial number of human infections with defined epidemiology [56] [57]. The 2009 swine-origin H1N1 completed it: a virus whose emergence and pandemic potential were recognised in weeks [58] [59] [60] [61], and which turned out to be far milder than 1918 [62] while still causing substantial global mortality [63] [64]. The lesson runs both ways. Crossing the threshold does not require a virus to become more dangerous, and a pandemic virus need not be a catastrophic one.
Dig deeper in lmmol
The threshold model in the centerpiece is the shared spine of every transmissible disease in this collection, and it is worth seeing in more than one setting. Measles is the instructive opposite of H5N1: a virus with one of the highest reproduction numbers known, where the entire public-health problem is holding population immunity high enough to keep the effective reproduction number below one. Dengue shows the same threshold with a different lever — there the reproduction number depends on a mosquito vector and rises and falls with temperature, so the question is where transmission becomes possible rather than whether adaptation occurs. Malaria shares the vector-borne form of the same arithmetic. For contrast with a slow pathogen whose reservoir is latent human infection rather than an animal one, see tuberculosis. Severe influenza kills through pneumonia and its consequences, which connects to sepsis, and it falls hardest on people with existing lung disease — see COPD and asthma. The full collection is at health.