In August 2023, as part of Mark Zuckerberg’s “year of efficiency” that led to more than 20,000 layoffs, Meta disbanded a research team of a dozen scientists who had trained a large artificial intelligence language model for biology.
But Alexander Rives, who led the research cohort known as Meta’s “AI protein team,” was undeterred by Meta’s move. He immediately founded a startup with a core group of his former colleagues at Meta, called EvolutionaryScale, to continue his work on building large language models that, instead of generating text, images, or video, generate recipes for entirely new proteins.
The idea is essentially to make biology programmable, with potential applications ranging from drug development and cancer treatments (such as antibodies) to environmental protection techniques (for example, enzymes – which are proteins – can help degrade plastics). Researchers would be able to specify a protein’s function and other attributes, such as its toxicity to humans, as a challenge and have the AI model return the DNA formula for making exactly that protein.
New York and San Francisco-based EvolutionaryScale announced today that it has raised more than $142 million in seed funding, led by Nat Friedman and Daniel Gross, and Lux Capital, with participation from Amazon Web Services (AWS), NVentures (Nvidia’s venture capital arm) and angel investors. The announcement adds Rives and his ex-Meta team to a growing list of Meta alumni — a sort of “Meta AI Mafia” — who have made waves with new startups in the space, most notably Mistral.
In addition to the funding, the company also announced that it has created ESM3, which Rives said Luck is a generative model for biology that has been trained on more computations than any other LLM in the space. Trained on nearly 4 billion proteins from the natural world, this model can simultaneously consider DNA sequence, physical structure, and protein function—three fundamental aspects of protein biology and biochemistry. And in a new paper, EvolutionaryScale showed how he applied ESM3 to create an entirely new fluorescent protein — a type of protein first isolated in glowing jellyfish — that would take millions of years of evolution to create in nature.
He explained that the AI model can process the three-dimensional structure of proteins as a language – like an alphabet of different characters – which can then be invoked like other models including ChatGPT. But in this case, the protein “grammar” allows the model to be prompted for any combination of protein sequence, structure, and function. “We see that the model is able to find very creative solutions to these stimuli,” he said.
Training such a large-scale model requires expertise in both biology and machine learning, and massive amounts of computing power – which explains the early fundraising. “They require a large amount of computation to build and train, similar to other frontier modeling efforts across AI,” Rives said. The fund increase, he said, “really reflects the resources we need to do this.”
EvolutionaryScale is far from the only company focusing on the potential of AI-powered generative biology, or even specifically pursuing an LLM. InstaDeep, a London-based company acquired in 2023 by BioNTech, best known for helping to create Pfizer’s COVID vaccine, has created an LLM in genomics, though not as large as the one EvolutionaryScale is working on. Profluent, an artificial intelligence biotech startup based in San Francisco, is also focused on developing an LLM for designing new proteins. And Google DeepMind’s AlphaFold is a model that predicts protein structure using a generative artificial intelligence model.
Friedman, who led the funding round at EvolutionaryScale with his investment partner Gross, said Rives and other Meta alumni are a “dream team.” (Gross recently teamed up with OpenAI’s Ilya Sutskever and Daniel Levy to launch a new startup, Safe Superintelligence.)
“This was clearly the team that invented protein language modeling and had all the capacity to take it further,” Friedman said. “Alex is a very big thinker. He wants to build a complete multimodal model that captures all the complexity of biology. I was looking for someone who had the ambition, the vision and the breadth of thinking and the expertise to make it happen.”
Rives said ESM3 will have an immediate impact on scientific research, with academics able to use open versions of the model for free. The company will also offer a commercial version for pharmaceutical companies to use in drug discovery and development. This is similar to the model followed by Google’s DeepMind, with a version of AlphaFold available to researchers for free, but with a separate spinout company, Isomorphic Labs, that works with pharmaceutical companies.
As for Meta, Rives said he wasn’t too surprised when the company disbanded his team.
“Meta is not a biotech company,” he said. While Meta’s open research culture made it an “incredible” place to work, he added: “We’ve reached a point where we really wanted to take these models to the next level of scaling. I think building a new company was really the right place to do it.”
It’s an “incredibly talented group of alumni that has come out of” the Met, Friedman said. “They’ve hired some incredible people – and I think the EvolutionaryScale team is very grateful to Meta for incubating their efforts.”