The Wild Type is a weekly look at the most interesting thing happening in biotech, written for the curious rather than the credentialed. I am Riya, a biology student, and each issue is one idea, explained properly, in about five minutes. No background required.
This week: why almost every tool in biology was fished out of an animal, and the machine that has just broken the habit.
Photo by Jane Ta on Unsplash
The Lead
In the early 1960s, a chemist named Osamu Shimomura started scooping jellyfish out of the water off Washington State. He went back every summer for thirty years and collected over 850,000 of them, all to solve a simple mystery: why did they glow?
What he pulled out was Green Fluorescent Protein, or GFP. Its utility is hard to overstate. Once you can make a molecule glow, you can track it. Attach GFP to a cancer cell, and you can watch a tumor migrate through a living mouse. Attach it to a neuron, and you can watch a thought travel. It won the Nobel Prize in 2008 and sits in labs on every continent.
But GFP arrived the way almost everything in biotechnology historically has. We didn't invent it. We found an animal that had already spent millions of years perfecting it, took a copy of its gene, and grew it in vats.
The enzyme that makes PCR tests possible came out of a microbe in a Yellowstone hot spring. CRISPR came out of the immune systems of bacteria. For decades, the greatest tools of modern medicine have been brilliant acts of salvage.
Then, in June 2024, a startup in New York asked a computer to imagine a brand new glowing protein from scratch. It did. They built it, it worked, and nobody went anywhere near the sea.
The protein itself, called esmGFP, is a bit of a dud. It is green, which is the one colour we have plenty of, and its glow takes days to mature. No lab is going to buy it. But what makes it worth five minutes of your time is the sheer mathematical improbability of its existence. Because the last time the scientific community wanted a glowing protein we didn't already own, it took five years of failure and a boat.
The Big Picture
The colour we wanted back then was red.
One colour lets you watch one thing, but biology is a game of interactions. To see how two molecules affect each other, you need to tag one green, tag the other red, and watch them collide. Red light is also highly prized because it travels much further through living tissue without scattering.
In the 1990s, researchers tried to engineer Shimomura's jellyfish protein to turn red. They spent five years tweaking it, swapping out links in its molecular chain. They managed to make it blue, cyan, and yellow. But red never came.
The barrier is physics. A protein is a chain of amino acids that folds into a complex three dimensional shape. The shape is what does the work. The jellyfish protein is nearly two hundred and forty links long, folded into a tight barrel of eleven strands with a tiny thread running down the middle. Three specific links on that inner thread react together to create the light source.
If you want to change how that light source behaves, you have to mutate the links around it. But proteins are delicate. Swap just one link, and the entire barrel might fail to fold. Instead of a tool, you get a useless clump of microscopic sludge.
Because every part of the folded protein presses on every other part, you cannot calculate the physics on paper. You have to feel your way through a dark maze, one mutation at a time, and every single step has to land on a molecule that still folds and still glows.
Red finally arrived in 1999 because a Russian team led by Sergey Lukyanov gave up on the lab, went diving in the Indo Pacific, and found a mushroom anemone on a coral reef that naturally glowed red. They called the protein DsRed.
DsRed shared only 26% of its amino acids with the jellyfish protein. To get from green to red, nearly three quarters of the protein's sequence had to be different.
For a human designer, trying to guess your way through that many changes is absurd. It is like trying to rewrite 75% of the words in a paragraph while keeping the exact same sentence structure. We couldn't build it, so we had to wait for nature to hand it to us.
Startup Spotlight
A startup called EvolutionaryScale exists because Meta shut down its AI protein team during layoffs in 2023. The team's leader, Alexander Rives, took his colleagues, raised 142 million dollars, and built a model called ESM3. In November 2025 the Chan Zuckerberg Initiative bought the company,
ESM3 works like a language model, but instead of reading Wikipedia, it read the sequence of nearly three billion natural proteins. It didn't learn human grammar, it learned the handwriting of evolution.
To prove the model's power, they gave it a massive challenge. They kept only the three link light source at the center of the jellyfish protein and asked ESM3 to write a completely new, unrecognised barrel around it. What came back was esmGFP.
Its sequence differs from the closest natural glowing protein by 96 amino acids. Out of 229 amino acids, it shares only 58% sequence similarity with the closest fluorescent protein found in nature.
To put that in perspective, in traditional protein engineering, changing even five or ten amino acids at random almost always breaks the molecule. Changing 96 links simultaneously is the molecular equivalent of demolishing 96 random load bearing walls in a skyscraper and expecting it to stay standing.
If you tried to find a working protein with 96 mutations by random guessing, the math is brutal. The number of possible combinations is 2096, a number vastly larger than the number of atoms in the observable universe.
Yet, the model generated a sequence that actually kept the skyscraper standing.
The Gap
The machine did not achieve this in a single perfect leap. It still had to use an iterative design loop. In the first round, the team synthesised 96 of the AI's designs and put them in a plastic tray. Most did nothing. The first prototype draft, located in a well labeled B8, was a deeply flawed proof of concept. It gave off a faint glow that was 50 times dimmer than a wild jellyfish and took an entire week to light up.
In the past, a result that weak would be a dead end. But because the AI understood the underlying grammar of the fold, the scientists were able to feed this weak B8 protein back into the model as a new starting point. They generated a second set of 96 proteins. On that second tray, in well C10, they found the winner that glowed as brightly as any wild jellyfish.
The AI didn't eliminate the laboratory bench. Humans still had to physically mix the liquids, run the plates, and find the survivor. But the miracle is that they only had to test 192 candidate proteins.
We didn't bypass biology's experimental nature, we just made the guessing highly intelligent. The search space shrank from the size of the universe to two plastic trays.
On the Radar
While the glowing protein is a neat parlour trick, the implications of this shift are starting to appear across the field:
Designing the code of life: The Arc Institute in Palo Alto is building models that design entire genomes rather than single proteins. They recently used an AI model called Evo to design brand new bacteriophages, which are viruses that hunt bacteria. Out of 285 machine written viral genomes, 16 successfully propagated and targeted bacteria. It is the first time machine written DNA of this scale has successfully booted up to create functioning, replicating entities.
The designer gene editors: A Berkeley startup called Profluent used its own models to design an entirely new gene editing mechanism hundreds of mutations away from any CRISPR system found in nature, proved it successfully edited human cells, and open sourced it.
The oceans are now training data: The Tara Oceans expedition spent years filtering global seawater and returned with 40 million genes, over half of which belong to microbes we have never identified. We are still going to the ocean with boats, not to scavenge tools, but to harvest training data.
For the last seventy years, biotechnology has run on a simple loop where we borrow what nature invents. If we wanted an enzyme that could break down plastic, or survive acid, or target a cancer cell, our first instinct was to pack a bag, go outside, and search.
By transitioning from copying to writing, artificial intelligence is effectively turning biology from an observational science into an information science. We are entering an era where we no longer have to wait for evolution to solve our problems. Instead of scanning the horizon for a rare organism that happens to possess the tool we need, we can simulate eons of evolutionary trial and error inside a server farm in an afternoon.
The next time we need a molecular tool that doesn't exist, we won't look to the reefs.
We will look to a prompt.
This video provides an excellent visual overview of ESM3, demonstrating how the model uses a multimodal transformer to simulate millions of years of evolution and produce the glowing esmGFP protein.