EvolutionaryScale has introduced ESM3, a large-scale AI model built to generate new functional proteins from the patterns found in evolution’s own data. The model is designed to work across protein sequence, three-dimensional structure, and biological function, giving it a broader view of proteins than systems that only process text-like sequence data.
The company’s demonstration centers on esmGFP, a green fluorescent protein created by ESM3. According to the researchers, the protein differs enough from known fluorescent proteins that a comparable change would have taken over 500 million years to happen naturally.
What ESM3 Was Built to Learn
ESM3 was trained on a very large collection of protein information. The dataset included 2.78 billion natural protein sequences, 236 million protein structures, and 539 million proteins with functional annotations. In total, the model processed 771 billion tokens during training.
That training mix matters because proteins are not defined by sequence alone. A protein’s biological role depends on how its amino acid chain relates to its three-dimensional form and the function that form enables. ESM3 was built to treat those dimensions as connected signals rather than separate problems.
Instead of learning only from written symbols, the model uses discrete tokens that represent three kinds of protein information:
- the amino acid sequence of a protein
- the three-dimensional structure of a protein
- the biological function associated with a protein
The researchers describe proteins as part of an organized space where each protein sits next to other proteins that differ by a single mutation event. From that view, protein evolution can be understood as movement through many possible paths, with each step changing the protein while potentially preserving useful function.
Why Three-Dimensional Structure Matters
ESM3 uses a novel architecture with geometric attention, which is meant to help it process protein structures efficiently. That structural understanding is central to the model’s purpose because a protein’s shape is closely tied to what it can do.
Older models may learn useful relationships from protein sequences, but ESM3 is designed to connect sequence with structure and function in a more direct way. This lets the model reason across the organized space of possible proteins and generate new candidates that still fit a desired biological goal.
The source article describes the system as implicitly constructing a model of possible evolutionary pathways connecting proteins without losing the function of the higher-level system. In plain language, ESM3 is not just assembling random protein-like strings. It is using learned relationships among sequence, structure, and function to propose proteins that can still work.
The esmGFP Demonstration
To show what ESM3 could do, the researchers asked it to generate a new green fluorescent protein. They gave the model the sequence and structure of key residues that determine fluorescence. From that starting point, ESM3 gradually filled in the rest of the protein sequence and structure.
The result was esmGFP. The protein showed high luminosity, while having only a 58% sequence similarity to the closest known fluorescent protein. The team said that such a large difference would have taken over 500 million years to occur naturally.
This is the central claim behind the demonstration: ESM3 can use evolutionary training data to move through protein design space in ways that nature might reach only over very long timescales. The model is not replacing evolution with a simple shortcut. It is learning from the products of evolution, then using those learned patterns to generate new functional proteins.
Who Is Behind the Model
EvolutionaryScale was founded by former members of the Meta-FAIR protein group. The source article notes that they were involved in ESMFold, among other projects. Meta disbanded the department in August 2023, while others like Alphabet continue to work in this field with Deepmind's AlphaFold 3.
That background is relevant because ESM3 builds on the broader idea that transformers can capture complex biological relationships. The study is presented as another example of transformer models being used to understand proteins and generate new functions.
According to the EvolutionaryScale team, ESM3 supports a program-driven approach to protein design. The goal is to bridge the gap between human specifications and the complexity of biology. The source article says this technique could eventually enable numerous applications in biotechnology and medicine.
Access and Responsible Use
The researchers also stress that models with this level of capability need responsible handling. They are releasing an open version of ESM3 for researchers to use. According to the team, the model has been tested for safety by experts, and the experts concluded that the positive effects of the release outweigh the risks.
The complete ESM3 models are to be made available via an API with free access for academic research. That access model points to a two-track approach: broader research use through an open version, and fuller model access through an API.
For now, the key takeaway is that ESM3 shows how protein design can be framed as a generative AI problem grounded in evolutionary data. Its esmGFP demonstration gives a concrete example: a functional fluorescent protein that is substantially different from the closest known fluorescent protein, generated from a model trained on sequence, structure, and function at large scale.