De-extinction runs into a stubborn problem: DNA doesn't fossilize like a perfect instruction manual. For extinct animals, researchers often have too little usable genetic material to use standard "compare lots of genomes and find the DNA behind this trait" methods. The dodo is a good example. Colossal Biosciences says it has a single dodo genome to work from, while its woolly mammoth project can draw on roughly 100 genomes.
That's the gap Colossal says its AI platform, GenoMorph, is meant to fill. The company's pitch is that it combines comparative genomics, developmental biology, and 3D measurements to rank genetic regions that might relate to the dodo's distinctive facial and cranial features. Colossal also says it has identified about 2,300 potential editing targets so far.
But it's important not to jump steps. Finding promising regions to edit is not the same thing as proving those edits would recreate a dodo - or even produce the exact traits the team is aiming for. GenoMorph is meant to narrow the search. Lab experiments still have to do the heavy lifting. Here's how the platform is described, where AI fits into the biology workflow, and what commercial and ethical questions follow.
What GenoMorph Is Designed to Do
Colossal describes GenoMorph as a platform to help researchers decide which genetic changes might be worth investigating when there's only limited DNA from an extinct animal. It doesn't edit genes itself. Instead, it's supposed to point researchers toward candidate genetic "switches" that could be involved in building a trait.
That distinction matters because a typical comparative-genomics workflow works best when scientists have many genomes - living and extinct - to compare. More examples generally make it easier to tell real genetic signals from everyday variation. With only a handful of complete ancient genomes, it's much harder to separate signal from noise.
Colossal CEO and co-founder Ben Lamm told TechCrunch that the company has one dodo genome available, while the woolly mammoth effort can use around 100 genomes. Colossal attributes the difference to preservation conditions: cold environments tend to protect ancient DNA better than the tropical conditions where the dodo lived. In other words, the evidence base isn't uneven just because of sloppy data collection - it's uneven because of how and where the DNA survives.
Combining development data with morphology
To build GenoMorph, Colossal says it pulled together developmental information from chickens, quail, ducks, and pigeons. It pairs that with genetic data and 3D skull measurements from more than 70 bird species. The goal, as stated by the company, is to link changes in DNA activity during development with differences in beak and facial structure.
In practice, Colossal says the platform looks for patterns across the birds where it has more data, then uses those patterns to suggest which regions of the dodo genome might matter for appearance. Lamm also described it as a "treasure map": not every genetic difference is equally likely to be relevant, and the system is meant to shrink the list.
This is a reasonable way to think about what AI can do in a data-constrained setting. GenoMorph isn't trying to reconstruct a complete dodo blueprint from a single genome. It's using information from related living species to form tighter hypotheses about the extinct one. The value is in prioritizing what to test next - reducing the search space and helping researchers plan experiments.
Colossal says it has identified about 2,300 promising DNA-editing targets tied to dodo cranial and facial morphology. Still, these are candidate targets, not confirmed instructions for producing a dodo-like bird. Even a long target list can include false positives, incomplete explanations, or genetic changes that steer development in unexpected directions.
Why the dodo is a particularly difficult case
Colossal says the dodo's closest living relative is the Nicobar pigeon. That relationship gives the company a living point of comparison, but it doesn't make the species interchangeable. A close relative can help provide biological context, yet it can't fill in every missing detail about how the dodo developed in the first place.
That's one reason GenoMorph is described as combining multiple evidence types. Genome data can point to what genetic differences exist. Developmental data can suggest when and where those differences might matter. And 3D morphology provides a way to compare physical structures. Each source has limits on its own - but together they may offer more than a sparse genome alone.
What the company is essentially claiming is that AI isn't producing "one answer." Instead, it's trying to connect datasets that are hard to interpret in isolation and turn those connections into testable research priorities. Whether the ranked targets hold up will come down to experiments, not to the strength of the software demo.
From AI Predictions to Laboratory Tests
GenoMorph is only one step in a longer process. Colossal still has to test whether selected genetic changes actually affect the traits its models predict - and whether those changes can be introduced and studied in a suitable biological system.
Colossal says it has a method using lipid nanoparticles to deliver gene-editing tools into developing bird embryos. The company says that approach could let researchers observe how specific changes affect development without waiting for an animal to reach sexual maturity and complete a breeding cycle. If it works as described, it could speed up parts of the experimental feedback loop.
Even so, speed isn't the same as certainty. An edit that shifts a feature during embryonic development may behave differently across the full development of an animal. And even if you see a change, that doesn't automatically tell you whether the intervention is safe, stable, or suitable for a broader program.
The difference between a target and a result
It helps to separate the claims into stages:
- Candidate identification: GenoMorph proposes genetic regions that may be linked to dodo-like cranial or facial features.
- Experimental testing: Researchers test whether changing a candidate region produces the expected biological effect.
- Interpretation: Scientists evaluate whether the observed effect supports the initial hypothesis - and whether other changes may be involved.
- Broader development: Further work would require additional evidence about development, health, and what happens when multiple edits are combined.
A result at one stage doesn't guarantee the next. A plausible target doesn't prove an edit will work. Observing an effect in an embryo doesn't, by itself, demonstrate that a resulting animal would be healthy - or reproduce the ecology of an extinct species.
This is where Colossal's other reported progress comes in, but it still shouldn't be treated as proof that GenoMorph's specific predictions were correct. Colossal says its Nicobar pigeon colony now includes 36 birds, and that it has created cells that can develop into eggs and sperm. The company frames these cells as an important part of its strategy because the Nicobar pigeon is the dodo's closest living relative.
These are separate pieces. Building a larger colony and doing reproductive-cell work can support future experiments. GenoMorph, meanwhile, is aimed at candidate edits. The approaches may converge over time, but each has to demonstrate its own function.
What would count as meaningful validation?
A strong validation story would go beyond announcing a large number of targets. It would spell out how candidates were chosen, which experiments were run, which predictions were confirmed or rejected, and how uncertainty was handled. Negative results matter here too. If an edit doesn't show the predicted effect - or produces something different - the outcome should refine the model and guide the next round of testing.
So for readers trying to judge the platform, the key questions are practical:
- How often do predictions point to regions that turn out to matter experimentally?
- How frequently do the proposed edits produce the predicted developmental effect?
- Can the team distinguish direct effects from changes caused by other genetic or environmental factors?
- Are results reproducible across experiments?
Colossal's announcement doesn't answer those questions. It explains the platform, describes its inputs, and reports its target count, while placing laboratory validation as the next step. That's fair for an early research program - but it isn't yet evidence that the AI has solved the biological problem.
The Business Case for Programmable Biology
Colossal's announcement also ties the work to a broader investment thesis. Peter Diamandis - founder of XPRIZE and an investor in Colossal - announced NeoGenesis Capital, a new venture firm meant to back companies applying Colossal's programmable-biology platform across health, agriculture, and conservation.
The distinction matters: the fund is described around Colossal's programmable-biology platform, not necessarily GenoMorph alone. The announcement doesn't say how much money the fund plans to deploy, which companies it will invest in, or what the terms might look like. Those details are necessary if you want to assess the fund's scale and likely impact.
That said, the logic is fairly clear. A tool built to tackle one difficult biology problem could have broader uses if it reliably helps teams interpret genetic data and prioritize what to test. In health, agriculture, or conservation, the potential value would depend on whether the system improves decision-making enough to justify its cost, data needs, and validation burden.
Where the value could come from
In biology, expensive work often happens before results: assembling relevant data, choosing candidate interventions, running experiments, and learning from failures. A platform that improves the quality of those early choices could reduce wasted lab effort. It could also help researchers use data that might otherwise be too fragmented or limited for straightforward comparisons.
That's still a hypothesis, not a settled business track record. For it to become a durable advantage, Colossal would need to show consistent performance across projects, species, and datasets. If the approach only works in a narrow scenario, it could still be valuable - but the market and economics wouldn't look the same as a broadly reusable platform for programmable biology.
There's also a difference between building a model and building a product that others can actually use. A one-off system tuned to Colossal's internal data might help its own scientists. A platform meant for outside companies would require clearer documentation, data access, reproducibility, security, and support for different research workflows.
Ultimately, investors will want evidence that the technology changes outcomes, not only that it produces interesting predictions. Useful measures might include fewer experimental cycles to reach a validated result, better success rates when selecting candidates, or evidence that the approach transfers to new datasets. Without metrics like that, "programmable biology" can sound more like an aspiration than a defined performance claim.
Limits, Safety and Questions That Remain
AI doesn't erase uncertainty in ancient-DNA research. It can help organize evidence and generate hypotheses, but its conclusions are only as good as the data and assumptions feeding into the model. The living birds used in training are related to the dodo, but they aren't the dodo. Differences between species, incomplete records, and the complexity of development all limit what can reasonably be inferred.
There's also a risk in treating genetic associations like simple switches. Traits like beak shape and facial structure can involve multiple genes, timing during development, and interactions between biological systems. A model might flag a region worth testing without capturing every factor that ultimately shapes the trait.
Lamm told TechCrunch that GenoMorph is intended to make genetic engineering more precise, controlled, and testable. He also emphasized that identifying a possible edit doesn't mean the company will introduce it into an animal right away - or assume it's safe. That's an important distinction: prioritization isn't authorization, and a prediction isn't a safety assessment.
Safety involves more than the edit itself
Safety is bigger than whether a specific genetic change produces a visible trait. Researchers would need to consider effects on development and health, how multiple edits interact, and what happens if an animal is bred or released. Each step raises different technical and governance questions.
Conservation adds another layer. Recreating some traits associated with an extinct species wouldn't automatically restore its ecological role. Habitat conditions, food sources, disease pressures, and interactions with other animals matter. Any real conservation claim would need to be evaluated against those factors, not inferred from a genetic or morphological result alone.
There's also a governance question that can't be handed off to an AI model: who decides which traits are desirable, what level of evidence justifies moving forward, and how risks are reviewed? The model can help set priorities for research, but responsibility for experimental design, interpretation, and deployment stays with people and institutions.
For now, the most defensible reading of the announcement is cautious. GenoMorph may help Colossal focus limited dodo genetic data into a more manageable set of hypotheses. The reported 2,300 targets are a starting point for lab work, not a count of verified edits or a timeline for bringing the dodo back. The next signal to watch is whether GenoMorph's predictions survive experimental testing - and how openly the company reports both successes and failures.
Conclusion
GenoMorph is aimed at a real bottleneck: researchers have far less genetic evidence for the dodo than for some other extinct species, so they need a disciplined way to prioritize what to test. Colossal's approach combines AI with biological and physical data, then relies on experiments to see whether the predictions hold up.
The next milestone worth watching isn't just a bigger target list. It's validation that's transparent about what worked, what didn't, and what the results imply for safety and development. If you're tracking AI in biotechnology, focus on that evidence - and be careful to separate a useful research lead from a demonstrated biological outcome.