Colossal Launches GenoMorph AI to Reconstruct the Dodo from One Genome
Colossal Biosciences built GenoMorph, an AI predicting 2,300 DNA edits to recreate the dodo from living bird data.
Tools · Source: TechCrunch
What happened
Colossal Biosciences announced GenoMorph, a new AI platform designed to accelerate de-extinction when ancient DNA is scarce. The Dallas biotech company is using the tool on its dodo bird project, where researchers have access to only a single viable genome. In contrast, Colossal has around 100 genomes for the woolly mammoth because cold arctic climates preserve tissue far better than tropical island environments.
To compensate for the missing data, Colossal built an avian regulatory atlas from chickens, quail, ducks, and pigeons across six embryonic stages and three facial zones. They fed that developmental data, along with high-resolution 3D CT skull scans and 541 anatomical landmarks from more than 70 living bird species, into a convolutional neural network. The AI maps chromatin accessibility to predict which regulatory switches control beak and skull morphology.
GenoMorph surfaced approximately 2,300 candidate regulatory DNA targets for the dodo's craniofacial structure, including 653 for beak size and 1,698 for thickness and breadth. Alongside the software, Colossal rolled out physical validation tech: a lipid nanoparticle delivery system that injects CRISPR tools directly into bird embryos to skip multi-year breeding cycles. The team is now testing edits in its captive colony of 36 Nicobar pigeons, the dodo's closest living relative.
Key facts
- 2,300 — Candidate regulatory DNA targets identified for dodo craniofacial morphology
- 1 — Dodo genome available to researchers compared to 100 for the mammoth
- 70+ — Bird species mapped with 3D skull scans and 541 anatomical landmarks
- 36 — Nicobar pigeons currently housed in Colossal's surrogate colony
- 0.863 — Internal prediction accuracy AUROC of the convolutional neural network
Why it matters
For AI founders and bio engineers, GenoMorph shows how to solve severe data sparsity by leaning on cross-species relational mapping. Ancient genomes will always be broken, fragmented, and rare. Instead of waiting for magic laboratory yields, Colossal turned living biological diversity into training data, predicting how DNA switches actuate physical geometry across evolutionary trees.
It also bridges the gap between digital generative modeling and physical wet lab execution. Most computational biology startups die in software because wet lab validation takes decades. By pairing ML predictions with lipid nanoparticle delivery in embryos, Colossal creates a tight feedback loop that compresses multi-year animal generation cycles down to embryonic development phases.
For builders
Small data modeling via comparative biology
Builders working in data-starved domains can stop complaining about missing historical records. You can train foundation models on rich adjacent datasets and use evolutionary or structural priors to infer the gaps.
Compressed biological validation loops
Software alone is worthless in biotech without instant physical feedback. Pairing AI predictions directly with lipid nanoparticle embryonic delivery removes years of animal maturation bottlenecks, creating faster lab-to-validation cycles.
Programmable biology venture funding
Tools built for moonshots like de-extinction generate horizontal platforms for agriculture and medicine. Peter Diamandis formed NeoGenesis Capital specifically to fund downstream startups commercializing Colossal's underlying biological stack.
My take
We are not actually bringing back the real dodo bird. We are using machine learning to reverse-engineer living pigeons into a synthetic creature that resembles what colonial sailors ate into extinction. That does not make it useless, but founders should call it what it is: computational design, not pure resurrection.
Original reporting: TechCrunch. This is my rewrite and opinion.