EvolutionaryScale

EvolutionaryScale is a public benefit company that trained protein language models, the best known of them ESM3. Its team came from the group at Meta that built the earlier ESM models. It announced itself on June 25, 2024, with ESM3 and a seed round of $142 million.^[1][2] On November 6, 2025, its law firm announced a strategic transaction under which the team would join Biohub, the medical research organization founded by Mark Zuckerberg and Priscilla Chan, with Alexander Rives, the company's co-founder and chief scientist, becoming Biohub's head of science.^[3]

Molecular Reality Corporation does not name EvolutionaryScale in the column of five companies it calls brain labs.^[4]

A protein language model is trained like a text model, but its words are amino acids. A sequence is shown with some residues hidden, and the network learns to fill them in. Nothing about structure or function is supplied, and the claim of a 2019 preprint, published in PNAS in April 2021, is in its title: "Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences". Its lead author was Rives, then at Facebook AI Research.^[5]

The next step was to ask a language model to fold. ESM-2, described in Science on March 17, 2023, was trained at sizes up to 15 billion parameters, and the structure of a protein could be read out of its internal representations at the resolution of individual atoms. Folding straight from one sequence, without building an alignment of related sequences first, made the method up to 60 times faster than the state of the art. That speed let the authors fold 617 million metagenomic sequences, of which about 225 million were predicted with high confidence.^[6] They published the result as the ESM Metagenomic Atlas. None of its structures was measured.

ESM3 reasons over the sequence, structure and function of a protein together, and a prompt can combine them.^[7] The company released the code and weights of a 1.4-billion-parameter version. The largest model, at 98 billion parameters, was offered through an API.^[1]

The test in the paper was a fluorescent protein. The authors prompted ESM3 to produce a green fluorescent protein "with a chain of thought", synthesized a set of the generations, and found a bright one. The protein, which they named esmGFP, is 229 amino acids long and shares 58 percent identity with the closest known fluorescent protein. The paper says that natural fluorescent proteins this far apart are separated by over five hundred million years of evolution.^[7][1]

The announcement of November 6, 2025, gave no terms and no founding or funding figures. It described the company as "a frontier research lab building AI models for biological research" and Biohub as "a 501c3 medical research organization".^[3] The company's blog now says that it is part of Biohub.^[1]

The company's argument for its own sensors begins with the data behind the best-known protein model: "AlphaFold 3 learned mostly from the half-century-old Protein Data Bank".^[8] The Protein Data Bank holds structures that people measured. ESM-2 learned from sequences alone, a store that the ESM-2 authors describe as growing exponentially, to "billions of protein sequences revealed by large scale gene sequencing experiments".^[6] AlphaFold and ESM therefore start from archives that grow at different speeds.

• AlphaFold
• Arc Institute
• Brain lab
• Protein Data Bank
• Protein sequencing