A startup founded by ex-Meta scientists has used an artificial intelligence protein model called ESM3 to generate a novel fluorescent protein, esmGFP, that the researchers frame as equivalent to 500 million years of natural evolutionary divergence. The achievement rests on green fluorescent protein, or GFP, a 238 amino-acid molecule first pulled from jellyfish decades ago and now so well understood that some scientists argue it should be formally designated as biology’s standard reference protein. If that designation takes hold, it could give AI-driven protein design a shared benchmark, cut duplicated lab work, and accelerate everything from drug development to classroom teaching.
From Jellyfish to Universal Lab Tool
GFP’s origin story begins with the jellyfish Aequorea victoria, collected off the western coast of North America. Scientists first isolated and purified a related bioluminescent protein, aequorin from Aequorea, documenting its biochemical properties and setting the stage for the fluorescence revolution that followed. Unlike aequorin, GFP does not need other proteins or cofactors to glow, a property that made it far more practical as a biological tag. Its chromophore forms autocatalytically during protein folding from three consecutive amino acids, a Ser-Tyr-Gly motif, through cyclization and oxidation deep in the protein’s core. That self-contained chemistry means a researcher can light up a target gene by adding just one extra gene to a cell.
The practical payoff arrived when scientists showed that GFP fluorescence could be produced in E. coli and C. elegans without exogenous cofactors. That experiment proved GFP could work in living organisms far removed from jellyfish, turning it into a universal marker for gene expression. Earlier work had already established GFP as a genetically encodable protein by cloning and sequencing it from Aequorea victoria using cDNA and genomic clones. Together, these milestones converted a curiosity of marine biology into one of the most widely used tools in modern life science, and GFP has since revolutionized fluorescence imaging across cell biology, neuroscience, and developmental research.
Why Biology Needs a Shared Reference Protein
Biology has long relied on model organisms, from E. coli to mice, to standardize experiments across labs. Yet at the protein level, no equivalent consensus exists. Without a shared reference point, study results are hard to compare: two labs can study the same protein under different experimental conditions and reach conclusions that are duplicated and difficult to generalize. The resulting inefficiency slows progress and wastes funding. GFP is a strong candidate to fill this gap precisely because it is already so well characterized: its barrel-shaped structure is known at atomic resolution, its chromophore chemistry is thoroughly documented, and mutations that shift its emission wavelength have been cataloged for three decades.
Designating GFP as a model protein would also reshape how biology is taught. Like classic model organisms, GFP can turn an abstract concept into something students can see in a test tube or under a microscope. Because its fluorescence provides an immediate, visible readout, student experiments yield clear results even when designs are imperfect. That forgiving quality makes GFP an ideal teaching protein, lowering the barrier for undergraduates and early-career researchers to engage with protein biochemistry hands-on rather than only through textbook diagrams. If the field converges on GFP as a reference, the same protocols could train students and benchmark cutting-edge AI tools, creating a shared language between education and research.
AI Protein Design Meets Its Stress Test
The case for GFP as a standard reference gained new urgency when EvolutionaryScale, a company formed by ex-Meta scientists, debuted a gigantic AI model for protein design called ESM3. In a peer-reviewed paper in Science, the team reported that ESM3 generated a novel green-fluorescent protein, esmGFP, by simulating evolutionary change with a language model trained on protein sequences. The result was framed as an analogue to the vast evolutionary divergence that separates distantly related natural fluorescent proteins. Generative modeling of this kind can propose millions of novel proteins that resemble native counterparts, but the hardest part of the process remains proving that a computer-generated sequence can become a properly folded, working protein.
Fluorescent proteins offer a uniquely practical stress test for that verification step. If an AI-designed protein glows, it has folded correctly and formed a functional chromophore, two outcomes that are instantly measurable without expensive assays. If it does not glow, the failure is equally unambiguous. That binary signal makes GFP an efficient filter for evaluating whether a generative model’s designs cross the line from plausible sequence to functional molecule. By repeatedly challenging AI systems to produce new, bright, and stable GFP variants, researchers can quantify how often designs succeed, where they fail, and which architectural changes in the model actually improve real-world performance.
esmGFP and the Promise of an AI Benchmark
In the ESM3 experiments, EvolutionaryScale’s researchers explored an immense space of potential GFP-like sequences and selected candidates that the model predicted would retain fluorescence while diverging substantially from known proteins. The resulting esmGFP sequence differs at many positions from canonical Aequorea GFP yet still folds into a β-barrel structure and forms an internal chromophore that emits green light. By describing this leap as akin to hundreds of millions of years of evolutionary divergence, the team underscored how far AI can roam through sequence space while still landing on a functional protein. For experimentalists, esmGFP is not just a curiosity; it is a concrete data point that demonstrates what current generative models can achieve when pointed at a well-understood scaffold.
Because GFP is so thoroughly characterized, esmGFP’s performance can be dissected in fine detail. Researchers can compare its brightness, photostability, maturation time, and folding efficiency to those of wild-type GFP and widely used engineered variants. Any deficits reveal where the AI’s internal representation of protein physics is still lacking, while unexpected improvements hint at design strategies human engineers may have missed. If future models routinely generate GFP variants that outperform existing tools, those successes will be hard to dismiss as cherry-picked anecdotes. Instead, they will function as reproducible benchmarks that other labs can test, critique, and build upon.
Standardizing AI-Driven Biology Around GFP
For AI protein design to mature into a robust engineering discipline, it needs the kind of shared yardsticks that transformed fields like computer vision and natural language processing. GFP is unusually well suited to play that role. Its fluorescence is easy to measure quantitatively across labs, and its structure–function relationships are well mapped, so subtle changes in sequence can be linked to clear biophysical consequences. If the community agrees on a core set of assays (such as brightness under defined illumination, thermal stability, and tolerance to amino-acid substitutions), then every new model can be evaluated against the same GFP-based metrics. That would make it far easier to distinguish genuine algorithmic advances from gains that stem only from different experimental setups.
Standardization could also extend to how AI models are trained and validated. A curated GFP dataset, including natural variants, engineered mutants, and AI-generated sequences like esmGFP, could serve as a shared testbed for assessing generalization. Researchers might, for example, hold out a subset of rare or counterintuitive mutations and ask whether a model can predict their impact on fluorescence. Because access to some technical reports on ESM3 requires logging in through publisher portals, a community-maintained GFP benchmark would also help level the playing field for academic groups that lack privileged access or industrial-scale resources. In that scenario, GFP becomes more than a teaching tool or imaging aid; it becomes the common currency that anchors an emerging era of AI-guided molecular design.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
