For decades, scientists used a glowing jellyfish protein called Green Fluorescent Protein (GFP) to light up living cells. This protein emits a bright green light when exposed to UV light, and is used to visibly ‘tag’ cells for all sorts of experiments. Then, researchers used AI to generate a protein that serves the same function. And… it worked? The resulting protein, esmGFP, is so different from known natural fluorescent proteins that researchers compare the gap between them to roughly half a billion years of evolution. This protein is widely marketed as ‘AI simulating evolution’; but how much evolution is actually happening here, how does this work, and why is this being AI generated a big deal?
Before this, evolution was the only protein designer we had. A jellyfish glows because evolution happened to stumble onto a protein called GFP that can absorb and emit light (1). Over millions of years, random mutations, that is, changes in DNA, slightly altered proteins. Natural selection, however, kept the versions that still worked. In the same sense, biology uses trial and error to build things, and nature keeps what works.
This means every protein on earth is part of a gigantic, evolutionary family tree. Scientists can compare proteins by looking at sequence identity (2) – the percentage of amino acids that match between two proteins. Usually, closely related proteins have similar sequences and similar functions. The farther apart two proteins are evolutionarily, the more likely the chemistry fails and the protein stops functioning entirely. If you try to mutate or engineer a protein to match the function of a very different, distantly related protein, the chemistry is highly likely to fail. Because their sequences and structures have drifted too far apart, you can't easily swap parts or force them to interact without losing the necessary biochemical function!
Why? Proteins are unbelievably sensitive to shape. They are made up of units called amino acids, and only work if they fold into an extremely precise and stable 3D structure. A few bad mutations can completely ruin the fold. Some examples of this in humans are cystic fibrosis, Alzheimer's, Sickle Cell Anaemia, and Huntington’s (3).
The model researchers used to generate this protein is called ESM3 (4). It was trained on enormous databases of natural proteins, like a large language mode (LLM), but for biology. ChatGPT predicts plausible next words. ESM3 predicts plausible amino acids, structures and functions. Give it part of a protein sequence and it can iteratively fill in the gaps, adjusting the molecule until the chemistry and structure begin to make sense together. One of these generated proteins is esmGFP!
Now this is when stuff starts getting really weird: esmGFP shares around 58% sequence identity with the closest natural fluorescent protein. In evolutionary terms, that is an enormous gap. For context, you share 60% of your DNA with a banana (5). No, I’m not lying to you. This is an astonishing number, and researchers estimate the difference could correspond to roughly half a billion years of evolution.
After reading the frankly ridiculous 58% number, I immediately thought that's weird. It is, in fact, comical to assume that a machine that only has 58% of the parts of another would behave in the same way. It's the same in biology. Normally proteins that are this different are not expected to behave in the same way. Proteins are incredibly dependent on structure, and structure depends on sequence! Change too much and the entire fold usually destabilises – the protein misfolds, clumps together or just stops working.
The fact that esmGFP can glow is even weirder. Forgive the ‘nerd talk’ that will take over the rest of the paragraph, but glowing for a protein is not an easy feat. GFP glows because part of the protein folds inward and creates a tiny chemical structure called a chromophore (6). Inside this pocket, three amino acids – serine, tyrosine, and glycine – react together to form a ring-like structure capable of absorbing and emitting light. In other words, the amino acids all need to be at the right angles, surrounding proteins need to stabilise the structure, the fold has to protect the chromatophore from the outside environment, and more. A model, without simulating any atoms like a physics engine, iteratively predicted something that works out.
This is fascinating because the AI model appears to have internalised some of the deeper rules that connect protein sequence to behaviour – rules that even biologists still don’t fully understand. Moreover, scientists can explain how esmGFP functions chemically. They can map the fold. They can identify the chromophore. They can experimentally confirm that the protein fluoresces. But they can’t explain how artificial intelligence made it to this protein.
Inside models like ESM3, there's still a major interpretability problem for scientists and biologists. The model learned from huge amounts of biological data and somehow developed an internal representation of what ‘working proteins’ look like; however, that representation is mostly hidden from us. The logic the model used to navigate this protein space and propose sequences that evolution may never have encountered at all is a black box (7). Maybe biology has a hidden schema where functional proteins follow deeper statistical or geometric rules. A schema which AI models can learn before we can clearly articulate it ourselves.
Nature is an incredible engineer, but it's also deeply conservative. Evolution does not search for the best possible molecule, only whatever is good enough to survive right now. If a protein helps an organism reproduce, it stays. If not, it disappears. As a result, biology is full of accidents and compromises. Proteins can be inefficient, unstable, or bizarrely complex simply because evolution had no reason, or no viable path, to improve them further.
AI hence creates a different approach to biological design. Instead of searching nature for useful molecules, researchers can generate proteins tailored to human needs: enzymes that break down plastic, proteins that capture carbon more efficiently, medicines aimed at specific targets, or even synthetic underwater adhesives inspired by mussels and barnacles (8) (find the link to this incredibly cool project here !).
The deepest implication of this discovery is philosophical; life on earth is a sample size of one. This means everything we call “biology” comes from the same evolutionary tree, shaped by chance, constraint, and what was “good enough” to survive. esmGFP works, even with a sequence drifted far from known fluorescence proteins. Unsettlingly enough, maybe life isn’t defined by specific molecules at all, but by the rules that make those molecules work.
Thank you for reading and I hope I could convince you that esmGFP matters beyond “AI made a glowing thing.” Check it out and play with esm yourself here!
References
University of Queensland. How the jellyfish revolutionised brain science. 2020. https://qbi.uq.edu.au/brain/nature-discovery/how-jellyfish-revolutionised-brain-science
RCSB Protein Data Bank. Data P. Sequence Similarity Search. 2017. https://www.rcsb.org/docs/search-and-browse/advanced-search/sequence-similarity-search
Valastyan JS, Lindquist S. Mechanisms of protein-folding diseases at a glance. Disease Models & Mechanisms. 2014;7(1):9–14. doi:10.1242/dmm.013474
EvolutionaryScale. Evolutionaryscale.ai. 2024. https://www.evolutionaryscale.ai/
Pfizer. How Genetically Related Are We to Bananas? Pfizer. 2022. https://www.pfizer.com/news/articles/how_genetically_related_are_we_to_bananas
Craggs TD. Green fluorescent protein: structure, folding and chromophore maturation. Chemical Society Reviews. 2009;38(10):2865. doi:10.1039/b903641p
Kosinski M. What is black box artificial intelligence (AI)? IBM. 2024. https://www.ibm.com/think/topics/black-box-ai
Liao H, Hu S, Yang H, Wang L, Tanaka S, Takigawa I, et al. Data-driven de novo design of super-adhesive hydrogels. Nature. 2025;644(8075):89–95. doi:10.1039/b903641p
