How AI Designed a Fluorescent Protein That Worked in the Lab

A computer can propose a protein sequence, but its usefulness depends on what the resulting molecule does. In a study of the AI model ESM3, researchers generated a fluorescent protein called esmGFP and tested it in the laboratory. Its sequence was substantially different from known fluorescent proteins, yet it retained a measurable biological function. [1]

The peer-reviewed paper by Thomas Hayes and colleagues first appeared online in Science on 16 January 2025, with an issue date of 21 February 2025. It was a laboratory protein-design study. Bacteria served as a system for producing the designed proteins; the experiment did not create a new organism. [1] [7]

The finding at a glance

• ESM3 generated candidates guided by features of fluorescent proteins.
• The team built, tested and refined selected designs.
• esmGFP worked in laboratory fluorescence assays and had 58% sequence identity to its closest known fluorescent-protein match in the study.

Why choose a fluorescent protein?

Fluorescent proteins make biological events easier to observe. A protein that absorbs light and emits it at another colour can act as a visible marker. In 1994, Martin Chalfie and colleagues showed that green fluorescent protein from the jellyfish Aequorea victoria could fluoresce when expressed in Escherichia coli bacteria and the roundworm Caenorhabditis elegans. This helped establish GFP as a tool for monitoring gene expression and protein location in living systems. [4]

Fluorescence requires excitation: light supplies the energy that is subsequently emitted. In the original GFP work, illumination made the signal observable. This matters when reading a headline about a protein that lights up. The relevant test is a measured response to illumination, with appropriate comparisons, rather than an assertion that the molecule generates light by itself. [4]

The sequence must support a working structure

A protein is a chain of amino-acid building blocks. Their order influences how the chain folds and which chemical interactions become possible. A 1996 structural study showed GFP’s characteristic cylinder: eleven strands of beta-sheet surround an internal helix. The light-emitting component, called the chromophore, is protected within this arrangement. Researchers resolved that structure at 1.9 ångströms, illustrating how specific experimental measurements underpin the familiar barrel-shaped drawings. [5]

That architecture helps explain the design challenge. Changing a sequence is easy on a computer; retaining the relationships needed for folding and fluorescence is harder. A distant sequence can still produce a familiar type of structure and function. Sequence novelty therefore needs to be considered alongside experimental performance, rather than treated as a complete measure of success.

How can a language model learn about proteins?

Protein language models learn patterns in sequences rather than sentences. In a 2021 study, Alexander Rives and colleagues trained models on 250 million protein sequences. The resulting representations contained information associated with biochemical properties and protein structure. Their training task included predicting amino acids hidden from an input sequence. Repeatedly solving that problem can capture dependencies between positions, including positions far apart along the chain. [6]

ESM3 extends the approach across sequence, structure and function. The research team represents these kinds of information in forms the model can process together, then trains it to recover missing information. Prompts can supply partial constraints while leaving other features to be generated. Some training information includes predicted structures and functions, so the training collection is not a catalogue in which every property has been measured experimentally. [2] [3]

The design was followed by experiments

For the fluorescent-protein demonstration, the team supplied features associated with the known fluorescence-generating core. The model proposed candidates around those constraints. This was guided generation: scientists chose the task and the information provided. The study’s deposited research materials include bacterial-expression constructs for an initial candidate, B8, and the later esmGFP-C10 design. [2] [7]

The team’s original research account describes two rounds of 96 candidates. In the first, B8 provided a promising but dim fluorescent signal. Researchers then generated another group using that candidate as a starting point. The second round yielded esmGFP, with brightness the team reported as similar to natural GFPs under their assay conditions. This sequence of steps is useful evidence for an iterative design-and-test workflow, rather than a guaranteed outcome from every generated sequence. [2]

What the 58% figure actually measures

The paper reports 58% sequence identity between esmGFP and its closest known fluorescent-protein match in the study. Identity concerns matching amino-acid positions in an alignment. It is a comparison between sequences, not a brightness measurement. The functional result comes from the separate fluorescence experiments. [1] [7]

The word known also matters. A comparison is made against available sequences, not every protein that could exist in nature. The result demonstrates a working design distant from the comparison set. It does not establish that no organism anywhere could contain a related sequence. Nor does one successful candidate reveal the success rate for every future design problem.

Where do “500 million years” come from?

The authors compare esmGFP’s sequence distance with divergence observed among natural fluorescent proteins. Their evolutionary analysis motivates the paper’s 500-million-year description. It is an estimate used to convey distance within this protein family. The AI did not replay a continuous history of ancestral populations, environments and natural selection, and the estimate does not establish a fixed timetable nature would require to produce this exact sequence. [1] [2]

What the experiment leaves open

The demonstrated task was fluorescence in a defined laboratory setting. Extending the result to another protein function requires its own measurements. A proposed enzyme, for example, would need evidence of the intended chemical activity; a plausible-looking fold would not establish that activity. These are different experimental questions. The fluorescent-protein result supports the feasibility of this design approach, while leaving broader performance claims to further work.

It is also important to separate the model’s predictions from observations. A generated sequence is a proposal. A predicted structure is a computational estimate. A protein produced in cells and characterized by an assay supplies another kind of evidence. Useful workflows connect these stages and preserve the ability to trace which conclusions rest on which measurements.

What would make the next result useful?

For research markers, sensible next evaluations would include maturation speed, stability during illumination and behaviour in the intended cellular setting. A protein can perform well in one assay and still be unsuitable for a particular imaging experiment. These are proposed evaluation priorities, not additional benefits established for esmGFP by the result discussed here.

The contribution is a practical connection between computational generation and experimental biology. Scientists can ask for candidate molecules, build selected examples and use the results to guide another round. The most informative question is then both precise and testable: which new sequence performs the biological job, under the conditions where it is needed?

Sources and references

1. Hayes T et al. Simulating 500 million years of evolution with a language model. Science 387(6736), 850–858 (2025). First published online 16 January 2025; issue date 21 February 2025. Peer-reviewed research.
2. Sercu T. ESM3: Simulating 500 million years of evolution with a language model. EvolutionaryScale, original research-team account, 25 June 2024; updated January 2025.
3. American Association for the Advancement of Science. An AI model developed to design proteins simulates 500 million years of protein evolution in developing new fluorescent protein. Original publisher news release, 16 January 2025.
4. Chalfie M, Tu Y, Euskirchen G, Ward WW, Prasher DC. Green fluorescent protein as a marker for gene expression. Science 263(5148), 802–805 (1994).
5. Yang F, Moss LG, Phillips GN Jr. The molecular structure of green fluorescent protein. Nature Biotechnology 14, 1246–1251 (1996). Experimental structure deposited as PDB 1GFL.
6. Rives A et al. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Proceedings of the National Academy of Sciences 118(15), e2016239118 (2021).
7. Addgene. Simulating 500 million years of evolution with a language model. Research-materials record associated with Hayes et al. (2025), including deposited B8 and esmGFP-C10 bacterial-expression plasmids.