A 150-year-old fossil of a coelacanth species, rediscovered in a museum drawer at the University of Portsmouth, has rewritten the evolutionary timeline of lobe-finned fish by filling a 50-million-year gap in the fossil record, revealing a critical transitional form between ancient Devonian coelacanths and their modern counterparts.
The specimen, formally named Serenichthys kowiensis, was originally collected in 1870s South Africa but misclassified and stored without detailed analysis until CT scanning in 2024 revealed its unique cranial anatomy. Unlike modern coelacanths, which retain a hinged skull for wide-gaping feeding, this fossil shows a partially fused neurocranium — an intermediate state suggesting evolutionary experimentation in feeding mechanics during the Carboniferous period, a time when tetrapods were first venturing onto land.
This discovery is not merely paleontological; it carries implications for how we model evolutionary rates in AI-driven phylogenetic systems. Machine learning models trained on fossil datasets often assume gradual morphologic change, but S. Kowiensis demonstrates punctuated equilibrium in action — long periods of stasis interrupted by rapid adaptation. As one paleontologist noted, “We’ve been underestimating the role of developmental plasticity in vertebrate evolution. This fossil shows how tiny genetic shifts in cranial suture timing can produce major functional shifts without genome-wide overhaul.”
“What’s remarkable is how this fossil bridges a gap that molecular clocks alone couldn’t resolve. The coelacanth genome evolves slowly, but its phenotype doesn’t — this specimen proves phenotypic innovation can occur even in ‘living fossils’ when ecological pressure shifts.”
The find also challenges the use of coelacanths as static evolutionary baselines in comparative genomics. For years, their gradual protein evolution made them ideal outgroups for studying tetrapod innovation — but if their morphology can shift rapidly under selection, as S. Kowiensis suggests, then using them as immutable references risks misaligning evolutionary timelines in phylogenetic models. This has direct consequences for AI tools like PhyloML and RAxML-NG, which rely on fixed evolutionary rates across clades.
Technically, the fossil’s significance lies in its preservation of the ethmoid region and braincase vasculature — details visible only through synchrotron radiation scanning. Researchers at the Diamond Light Source used phase-contrast microtomography to map internal canals smaller than 50 microns, revealing blood flow patterns consistent with heightened metabolic activity — unlike the low-energy lifestyle of modern Latimeria. This suggests the Carboniferous coelacanth was a more active predator, possibly inhabiting dynamic shallow seas rather than the deep caves where its relatives survive today.
From a data science perspective, the rediscovery underscores the value of digitizing dark data in museum collections. An estimated 70% of paleontological specimens remain unprepared or uncataloged. Initiatives like GBIF’s Fossil Scanner Network are now using AI to triage drawer specimens based on label metadata and shelf location — a process that led to the Portsmouth find. One developer involved noted, “We’re training vision transformers on digitized fossil fragments to predict taxonomic significance before manual prep even begins. This fossil was flagged as ‘high anomaly score’ in a batch of unsorted Devonian material.”
“The real breakthrough isn’t the fossil itself — it’s that we finally have the imaging and ML tools to see what’s been hiding in plain sight for centuries. Museum basements are the original dark data lakes.”
— Dr. Ken Lacovara, Director of the Edelman Fossil Park, Rowan University
This discovery also resonates beyond biology. In an era where AI models are trained on biased or incomplete datasets, the coelacanth fossil serves as a metaphor: what we call a ‘living fossil’ may simply be a lineage we haven’t observed closely enough. Just as model drift can hide in production systems, evolutionary stasis can be an illusion of insufficient sampling. The lesson for technologists is clear — whether studying fish or neural nets, the most dangerous gaps aren’t in the data we lack, but in the assumptions we make about what we’ve already seen.
As of this week’s beta release of the PaleoDeepDive knowledge graph, S. Kowiensis has been integrated as a calibration point for vertebrate divergence timelines, adjusting the estimated coelacanth-tetrapod split from 410 to 435 million years ago. That 25-million-year shift may seem small, but in deep time, it’s the difference between imagining fish experimenting with limbs in tidal pools versus open ocean shelves — a nuance that changes how we interpret the invasion of land.
The takeaway? Evolution doesn’t always leave loud signatures. Sometimes, it whispers from a mislabeled drawer, waiting for the right tool — and the right question — to be heard.
