Solving a Ten-Year Marine Murder Case: What Actually Killed Five Billion Sea Stars

The Case, Reopened

June 2013, Olympic National Park. A diver counts sea stars on a rock wall. One of them has a lesion. A white crater in the body wall, about the size of a thumbprint. Two days later the animal's arms are walking in different directions. A sea star has no central brain. Each arm runs on its own nerve ring, and once the connective tissue between them softens the arms stop agreeing. Within a week the star has pulled itself apart, and what the rock holds afterwards is a smear.

Within eighteen months the same thing was happening from Baja California to Alaska. More than twenty species of sea star were affected [1]. The hardest hit was the sunflower sea star Pycnopodia helianthoides. Any diver in Puget Sound used to see one on every dive. Almost nobody sees one now. A range-wide analysis pooling 48,810 surveys put the declines past 99.2% from Baja California to Cape Flattery in Washington, and past 87.8% from the Salish Sea up into the Gulf of Alaska [2]. The IUCN assessment drew on 61,043 surveys. Overall decline 90.6%. Critically endangered [3]. Counted in animals, the losses run into the billions. The 2025 paper's own summary says the epidemic "killed billions" of sunflower sea stars. It calls the event the largest documented marine epidemic in a non-commercial species [1].

For ten years nobody could name the agent.

An odd position for a modern disease investigation. We can sequence everything in a sample in an afternoon. Telling the microbes apart is the hard part. One causes a disease. Another arrives because the animal is already dying. A third sat there the whole time and never mattered. That distinction is the entire subject of this field note.

What the Disease Does to the Body

Before the detective work, the victim. Forget the sea star pressed flat against an aquarium wall. The sunflower star is the largest sea star in the world. Up to a metre across. As many as twenty-four arms. It moves at a speed you can watch in real time, close to a metre a minute on a good surface. It hunts. Urchins run from it. Urchins are not built for running.

Inside, mostly water. The body cavity is the coelom. It holds coelomic fluid, roughly seawater with cells floating in it. Those cells are the coelomocytes, and they are the animal's immune system. They engulf invaders. They clot a wound shut. A sea star has no blood and no liver, so the coelomic fluid does a great deal of the work that in you would be shared out among several organs. Remember that fluid. The case gets solved there. Everyone had failed to look in it carefully enough.

Wasting runs through stages you can score from a photograph, which is how the field tracked it for a decade without knowing its cause. First, arm twisting. One or more arms curl or knot. The animal loses its grip on the rock. Then autotomy. The star drops an arm at a breakage plane, the trick a lizard uses with its tail. A healthy star survives the loss and regrows the arm. In a wasting star the dropped arm is the beginning of the end. Then the body wall loses integrity, the animal deflates, and it dies. The challenge experiments described below timed the whole sequence. Exposure to death ran between five and sixteen days [1].

One problem trapped the field for a decade. Sea stars twist their arms for other reasons. Warm water does it. So does low oxygen. So does rough handling. In the 2025 paper's own low-dose experiment, a seasonal temperature spike to 13.9 °C caused transient arm twisting in 43% of the control animals, none of which went on to lose an arm or die [1]. So the most visible sign of the disease is not specific to the disease. Without a case definition, you cannot test a cause.

Ten Years of the Wrong Suspect

In 2014 a team led by Ian Hewson published a candidate in PNAS: a densovirus, a small single-stranded DNA virus, named sea star associated densovirus. The logic was clean and, on its face, convincing. Grind up tissue from a sick star. Push the slurry through a filter with pores of 0.22 micrometres. Small enough to hold back bacteria and cells. Wide enough to let viruses through. Inject the filtrate into a healthy star. The healthy star gets sick. Viruses pass filters. Bacteria do not. Therefore a virus [4].

The paper was enormously influential. For several years the public simply called it "the sea star virus". Then it came apart. Follow-up work by the same group could not hold the association together: the virus turned up in healthy animals as readily as in sick ones, and the experiments stopped reproducing [5]. By 2021 the same laboratory was arguing for something quite different: a boundary-layer effect, in which ordinary non-pathogenic microbes bloom on organic matter at the animal's surface and starve its skin of oxygen [6]. Hewson reviewed the whole affair in 2025, and the result is one of the more honest documents in disease ecology, with a title that says exactly what it holds: lessons learned [7].

Careful here. Tell this wrong and it becomes a story about somebody being stupid. The more useful version: the 2014 experiment was a good experiment resting on one load-bearing assumption, that a 0.22 µm filter sorts infectious agents into viral and non-viral, and nobody had ever tested that assumption on this system. Nobody ran the obvious control. Nobody asked whether the filtered preparation caused disease as reliably as the unfiltered one. Eleven years later, somebody did.

A Filter, and What Went Through It

The 2025 study ran seven independent exposure experiments between 2021 and 2024 [1]. Follow the procedure step by step. Unglamorous work, and the reason the answer holds.

Divers collected sunflower sea stars in Washington State and moved them to the USGS Marrowstone Marine Field Station. Each animal got its own tank, inside a shared water bath for temperature stability, on single-pass seawater that had been sand-filtered, particle-filtered to 10 to 25 micrometres, and passed twice through ultraviolet sterilisation. Every star spent at least two weeks in quarantine. Wasting kills in under two weeks, so a star still clean after quarantine was almost certainly not infected. Food was mussels and clams, frozen for at least 24 hours first, so that live microbes did not ride in on dinner.

Now the experiment. Take a star that is visibly wasting. Blend its tissues with water from its tank. Spin the mix gently to drop out the large pieces. Keep the supernatant. Inject it into a healthy star. The healthy star wastes and dies. So far this repeats 2014.

Then change one thing. Pass the same inoculum through a 0.22 µm filter first. Inject that. Nothing happens. No disease signs [1].

They repeated the comparison with coelomic fluid in place of ground tissue. Coelomic fluid has the consistency of seawater. Tissue homogenate clogs a filter. The fluid does not. Same result. Raw coelomic fluid from a diseased star transmits the disease, injected as a single 150 µl dose into the coelomic cavity through the armpit between two arms. Push it first through a 1.0 µm membrane, then a 0.22 µm membrane. Nothing transmits.

The whole reversal fits in one sentence. Whatever causes the disease will not pass a 0.22 µm filter. So it is not a virus.

The formal version of that experiment used eight stars per group. All eight exposed to unfiltered coelomic fluid showed disease signs and died. Arm twisting appeared 3 to 12 days post exposure. Arm dropping 5 to 13 days. Death 6 to 13 days. The eight controls took the same fluid after ten minutes in boiling water. None died [1]. Heat treatment makes a neat control, since it leaves the chemistry of the fluid broadly intact while killing anything alive in it. A toxin or an immune trigger in the fluid would have shown up in the boiled version. The boiled version did nothing.

Putting a Name on It

A cellular agent narrows the field to perhaps a few thousand candidates. To get a name, the team sequenced.

They sampled coelomic fluid from every star in the experiment. Before injection, from all sixteen. From the controls five days in. From exposed animals once more than a quarter of their arms had twisted, and again after each dropped its first arm. From those samples they co-extracted RNA and DNA, then built two datasets. The RNA gave a metatranscriptome, a record of which genes were actually being transcribed by whatever was living in the fluid, bacteria and protists and viruses alike. The DNA gave 16S ribosomal RNA gene amplicons, a standard bacterial census that reads one slow-changing gene present in every bacterium and uses it as a name tag.

Both datasets pointed at one organism. In diseased stars the microbial reads belonged largely to Vibrio pectenicida. In apparently healthy stars the 16S data found it under 1% relative abundance, or often not at all. The metatranscriptome did not assemble it [1].

Then the step that matters. Coelomic fluid from two diseased stars at Friday Harbor went onto marine agar plates. Incubated at 21 °C, five to seven days. Colonies were picked and restreaked. Each plate ended up holding the descendants of a single cell. Two of those pure cultures, labelled FHCF-3 and FHCF-5, had 16S genes confirming both as Vibrio pectenicida.

Three healthy stars per group then took an injection: about 10⁷ colony-forming units of FHCF-3, the same dose of FHCF-5, or a heat-killed version of each. All six stars given live bacteria developed disease signs and died. All six controls survived the four-week experiment. Bacteria re-isolated from the coelomic fluid of the dead animals sequenced as the same species [1].

A second experiment tested dose dependence, one of the things that separates a pathogen from a coincidence. Seven stars per group. Injected with roughly 10⁵ or roughly 10³ colony-forming units. Heat-killed controls for each. Thirteen of the fourteen exposed animals sickened and died. The high dose killed faster: 6 to 11 days. The low dose took 11 to 16. One low-dose star twisted a single arm on day 12 and then recovered. No control star lost an arm or died [1]. Figure 1 lays out the ladder.

The genome was sequenced separately and announced in 2025: 4,368,354 base pairs, 41.5% GC content, assembled into five contigs, carrying three genes for proteins resembling aerolysin, a pore-forming toxin family that punches holes in host cell membranes [8]. A plausible mechanism, not a demonstrated one. Nobody has yet shown that these toxins are what kills a sea star.

The species itself was not new. Vibrio pectenicida was described in 1998 from moribund great scallop larvae in a hatchery in Brittany, where it had been causing losses for years [9]. Its known trick there was killing haemocytes with a heat-stable toxin. Haemocytes are the scallop equivalent of the sea star's coelomocytes. The North American strain closest to FHCF-3 came out of a sick geoduck hatchery in Washington State in 2000. So the accused was already on file, twenty-five years deep in the shellfish-pathology literature, and nobody connected it to sea stars because nobody was looking in coelomic fluid.

Postulate one stays partial. The authors say so themselves rather than being caught at it. In their October field sampling, V. pectenicida sequences turned up in about 74% of grossly normal stars at sites where the disease was active, and in about 16% of stars at sites where it was not [1]. Koch ran into the same trouble with cholera and with typhoid, which is why every modern restatement of the criteria formally relaxes the strict first postulate [10]. Carriage without illness is normal for bacterial pathogens, and that same carriage is the opening through which the main objection arrives.

Notes From the Club Table

Meeting minutes, not prose. We include them because the arguments taught us more than the conclusions did.

The Arithmetic of a Reef That Lost Its Predator

Our own calculation. Two state variables, one imposed forcing. Every parameter but one is chosen to make a pre-outbreak reef look approximately like a real one. The exception is the sea star attack rate, taken from a published feeding study: a sunflower star eats 0.68 purple urchins per day, or 248 per year [11].

Kelp biomass \(K\) grows logistically and is grazed. Urchin density \(U\) grows on what it eats and is eaten by stars. Grazing and predation both saturate, because a full animal stops feeding:

Sea star density \(P(t)\) is imposed, not solved for. The model holds no pathogen. Wasting enters only as a collapse in \(P\).

Integrate for 400 years. Hold stars at 0.0168 per square metre. Call it 1.68 stars per 100 square metres. The model settles at 9.22 kg of kelp per square metre. Urchins, 0.51 per square metre. Each star then eats 50.4 urchins a year. Stars between them remove 0.846 urchins per square metre per year. Urchins between them eat 1.316 kg of kelp per square metre per year. The residual derivative is 4.43 × 10^-13. Converged.

Now apply the disease. Mortality of 2.05 per year for 1.6 years removes 96.2% of the stars. Star density falls from 0.0168 to 0.000632 per square metre. It stays there. The model has no recovery term.

Kelp drops below a tenth of its old value 2.7 years after onset. Urchins rise from 0.51 per square metre to 41.79, a factor of 82. Kelp settles at 0.0003 kg per square metre. A loss of 100.0% to the printed precision. The final state arrives by about year 20 and does not move afterwards.

Is the barren a second stable state, or merely a slow transient? The model answers directly. Set stars to zero. Run 300 years from a forested start: kelp 9.0, urchins 0.5. Run it again from a barren start: kelp 0.2, urchins 9.0. Both end at kelp 0.0003 and urchins 42.083. Without predation, one outcome. Put predation back at the historical density and the forest holds. So the system carries two stable states across a range of star densities. The disease moved it from one to the other. Removing the disease does not move it back.

How Fast a Disease Walks Down a Coastline

A different question now. Our second model is ours again, not the paper's. What kind of transmission empties 3,000 km of coast in about a year?

Chop the coast into 150 patches of 20 km each. Each patch holds susceptible, infectious and dead stars. Transmission between patches falls off exponentially with distance. Length scale 45 km. The rows of that kernel are normalised, so the basic reproduction number comes out as exactly the transmission rate over the recovery rate. The infectious period is fixed at 12 days from the anchor paper's challenge experiments, since mean time from exposure to death runs about that and essentially nothing recovers.

Tune the transmission rate until the simulated front moves at 3,000 km per year. Call it 8.22 km per day. The fit gives a transmission rate of 0.1307 per day. Reproduction number, 1.568. The front then travels at 8.22 km per day, with an R² of 0.99937 on arrival time against distance, and crosses the whole coast in 365 days.

Now the uncomfortable part. In the classic epidemic model of Kermack and McKendrick, the fraction of the population that eventually falls ill, \(Z\), satisfies \(Z = 1 - e^{-R_0 Z}\) [12]. Invert it.

Feed in a body count. Our fitted R₀ of 1.568 predicts a final epidemic size of 62.42%. The model's actual coastwide mortality is 62.42%, agreeing to 0.000 percentage points. The code is consistent. Consistent is not right. The real epidemic killed far more than 62% of anything.

The wave speed says 1.57. The body count says somewhere between 2.6 and 4.9. Two answers to one question. The gap is the interesting output of this model. Several explanations are available, and they do not exclude each other. The infectious period might run considerably longer in the field than in a tank, since a dying star in the open sheds into moving water for as long as it takes to fall apart. The coast is also not a one-dimensional string of identical patches: long-range jumps carried by currents would move a front faster than our kernel allows at any given R₀, which would let a modest reproduction number produce a fast wave. And constant transmission may simply be the wrong assumption. Aalto and colleagues argued exactly that in 2020, after showing that the real spread did not look like a travelling wave at all and probably needed environmental drivers, temperature in particular, to explain its timing [14].

We also checked how much of the answer is an artefact of the two numbers we guessed, the kernel length and the infectious period. Table 2 is that scan.

Across those nine combinations the implied R₀ runs from 1.39 to 4.01. Median 1.57. The wave speed pins down one product: transmission rate times kernel length. Many different reproduction numbers reproduce the same 8.22 km per day. Call it an identifiability problem, not an error in our code. Notice also the three rows at the widest kernel. All three return the same reproduction number of 1.386, and two produce no measurable speed at all. Those rows are the bisection converging on the value where the front barely forms, and they should not be read as physical results. We left them in because the script printed them, and hiding them would be dishonest.

The Best Case Against It

Now the counter-case. We argue against the paper as hard as we can, because a field note that only assembles supporting evidence is advertising.

Objection one, and the strongest: no microscopy. In July 2026 Thierry Work and five co-authors published a formal comment in the same journal, and their point is narrow and hard to dismiss [15]. Koch's postulates were built for a world in which you could see the organism in the lesion. Prentice and colleagues show that the bacterium causes death, and that it swarms in coelomic fluid, but they do not show where in the animal it sits or what it does to the tissue. Without histopathology, a slice of the lesion under a microscope showing the pathogen in contact with the damage, you cannot separate a bacterium that destroys tissue from a bacterium that overwhelms an animal already failing for some other reason. Work and colleagues add a practical argument: a case definition resting on gross signs alone will keep misclassifying heat stress and handling stress as disease, the exact problem that sank the last decade of work.

The authors replied in the same issue [16]. They have a real defence available. V. pectenicida has form for being invisible. In infected scallop larvae, immunostaining confirmed that the cells were infected while histology found no bacterial cell walls at all [1]. If the organism behaves the same way in sea stars, then the absence of visible bacteria across hundreds of diseased sea star tissue samples examined early in the epidemic stops being evidence against a bacterium and becomes a prediction that came true. An elegant argument. Also the kind of argument that explains away the missing evidence rather than supplying it, and the club split on how much credit to give it.

Objection two: the carriage problem. The bacterium turns up in stars that look perfectly healthy, at around 16% of individuals even at unaffected sites in October, and around 74% of grossly normal animals at sites where the disease was active [1]. A pathogen you can find in three-quarters of the well animals is a pathogen whose presence does not predict illness. The authors answer that these stars had almost certainly been exposed, that the epidemic is running everywhere they sampled, and that low levels may be tolerated until conditions shift. All of it plausible. None of it tested.

Objection three: the saprobe hypothesis. Ian Hewson proposed the densovirus and then dismantled it. In August 2025 he posted a preprint arguing that Vibrio pectenicida FHCF-3 behaves like a saprobe, an organism that flourishes on decaying tissue rather than starting the decay [17]. His evidence: the bacterium was not detected in body-wall samples of other affected sea star species collected at the same time as the Pycnopodia die-off; in ochre stars its abundance on the surface runs inverse to its presence in coelomic fluid; and it enriches strongly whenever organic matter is added. This one needs care in two directions at once. The preprint has not been peer reviewed, and it comes from the man whose previous candidate did not survive. Equally, someone who has publicly retracted his own cause of death is not the man you would expect to be careless about a second one. The claim about other host species is testable. Somebody should test it.

What the objections do not touch. All three argue about interpretation and scope. None explains away the central experiment. A pure culture, grown from a single colony on a plate, injected into healthy quarantined animals at three doses, killed them in a dose-ordered sequence while heat-killed controls from the same culture survived. To defeat that you would need the colony contaminated, or the controls mishandled, or the deaths a stress response to injection. The paper rules out the last: the injection procedure itself never induced disease signs in any healthy star.

Getting the Kelp Back Is Harder Than Losing It

Suppose a cure tomorrow. What then?

Here the identification earns its keep, and here our model has something to say, so we will separate the two carefully. The paper's contribution is practical. With a named organism you can build a test, screen wild populations, screen captive breeding stock, and check whether a candidate release site is currently hot before you put critically endangered animals into it. Before August 2025 none of that was possible.

Our model asks the blunter question. Restock the barren at some multiple \(\varphi\) of the pre-outbreak star density. Hold them there. Run forty years. Count it as recovery if kelp gets back above half its original value.

The threshold, found by bisection, is 1.551. The model needs 55% more sea stars than the reef held before the outbreak to buy recovery within forty years. Restore the historical density exactly, a multiple of 1.00. Kelp sits at 0.0004 kg per square metre after four decades. That asymmetry is the practical meaning of two stable states, and it is why ecologists worry about them.

We also tested whether a slow trickle works as well as a sudden restocking. Adding 0.20 stars per 100 square metres per year for ten years fails. Adding 0.40 per year for ten years succeeds. In the second case the standing density climbs past the threshold while urchins are still being knocked down. In the first it never does.

Read those three numbers as a shape, not as advice. The threshold of 1.55 depends directly on the drift-algae subsidy we set by hand, and moving that one parameter moves the threshold. What survives the choice of parameters is the asymmetry itself. In any model with two stable states, coming back costs more than leaving did. Real restoration work, including the urchin-culling and star-rearing programmes now running on the west coast, is planned on much more careful numbers than ours.

What We Would Want to See Next

Four things. In the order we would spend money on them.

Get a microscope onto a lesion. Work and colleagues are right about the missing piece, and it is not an expensive one. Fluorescent in situ hybridisation on a wasting lesion would show whether FHCF-3 sits at the tissue damage or somewhere else entirely. One experiment would settle the saprobe objection, in place of a decade of argument.

Test the other species. Wasting hit more than twenty asteroid species. The challenge experiments used one. If FHCF-3 is the agent in ochre stars and leather stars too, the case gets enormously stronger. If not, then "a causative agent of sea star wasting disease", the careful phrasing the authors chose for their title, turns out to have been the right phrasing all along, and the epidemic was more than one disease wearing the same symptoms.

Put temperature in the model. Every strand of this story points at warm water. Vibrio species have been called the microbial barometer of climate change, because their pathogenic members expand with warming seas. The 2013 to 2015 epidemic overlapped the north Pacific marine heatwave. The controls in the anchor paper twitched during a 13.9 °C spike. Nobody has run the challenge experiment at two temperatures. Doing so would say more about the next decade than anything else on this list.

And find out whether the urchins ever let go. Kelp is why this matters to people who do not care about sea stars. Bull kelp canopy fell by more than 90% along 350 km of northern California, purple urchins went up sixty-fold in a single year, the abalone fishery closed [18]. Those barrens are still there. Whether they are a temporary state or a durable one is a question about alternative stable states in the real ocean rather than in our arithmetic [19] [20], and a solved murder case does not answer it.

References