It all began with a bold question: What if evolution could be simulated in a lab? At the intersection of artificial intelligence, biology, and evolutionary science, researchers have unveiled a tool that might just rewrite the rules of protein engineering. Named ESM3, this AI model has achieved what once seemed like science fiction: the creation of a protein that nature could evolve through evolution, but didn’t.

Scientists at the EvolutionaryScale and the Arc Institute have developed a new type of green fluorescent protein (GFP). The protein was crafted entirely by an AI model trained to mimic 500 million years of evolutionary change. The experimental protein, named esmGFP, could deepen our understanding of how life evolves at the molecular level.

Rethinking Evolution with AI

Evolution, through countless mutations and adaptations, has shaped life’s molecular machinery. This includes surprising evolutions like proteins that glow in the dark or enable photosynthesis. Now, a new tool, ESM3—a generative AI model not too different from the one powering ChatGPT—has shown it can simulate evolution to create functional proteins that may never have existed naturally.

At its core, ESM3 functions like a translator of the “language” of proteins. Much like predicting the next word in a sentence, the model predicts how proteins might evolve in sequence, structure, and function. Trained on over 770 billion data points of natural protein sequences, ESM3 doesn’t just mimic evolution—it expands upon it, exploring unexplored corners of molecular design. This data deluge is the equivalent of compressing 500 million years of evolutionary history into a single computational framework.

“ESM3 takes a step toward a future of biology where AI is a tool to engineer from first principles, the way we engineer structures, machines, and microchips, and write computer programs,” said Alexander Rives, EvolutionaryScale co-founder and chief scientist.

AI Simulated Millions of Years

To test this remarkable capability, the research team prompted ESM3 to generate a GFP. These proteins, responsible for the vibrant glow of jellyfish and coral, are staples in biotechnology. Their distinctive structure—a coiled alpha helix threaded through a barrel-like framework—enables them to emit light without external chemicals, making them invaluable for tracking biological processes.

Despite decades of engineering, scientists have struggled to create GFPs vastly different from those found in nature. Yet ESM3 somehow broke this boundary. It designed esmGFP, a protein only 58% similar to its closest known counterpart. Achieving such a level of novelty is extraordinary, as even small mutations can cause fluorescent proteins to lose their glow.

To generate esmGFP, the AI began with critical structural prompts, focusing on the residues responsible for GFP’s fluorescence. From there, it iteratively adjusted the protein’s sequence and structure, optimizing the design in a process akin to natural selection. Unlike traditional AI methods, ESM3 used what the researchers called a “chain of thought,” a structured process of generation and refinement.

The protein was then synthesized and tested in the lab. It glowed—a clear sign of success. The AI had effectively charted a course through the vast “protein space,” simulating evolutionary processes that might take nature hundreds of millions of years to accomplish.

Why it Matters

Artificial proteins like esmGFP could open new doors in medicine, environmental science, and synthetic biology. Researchers envision enzymes that break down plastic waste, proteins that offer new treatments for disease, and tools to explore the deepest mysteries of life itself.

ESM3’s ability to explore distant protein possibilities could also help scientists understand how proteins evolve, adapt, and function under entirely new conditions. Evolutionary biology has long debated whether evolution is a matter of chance (contingency) or inevitability (determinism). AIs that simulate evolution might help resolve age-old scientific questions.

“This is a glimpse into the future of biology,” said the research team. “We’re not just studying evolution; we’re simulating it.”

By making ESM3 an open model, EvolutionaryScale hopes to empower scientists worldwide. “We’ve been working on this for a long time, and we’re excited to share it with the scientific community and see what they do with it,” Rives added.

The findings appeared in the journal Science.