A team of former Meta scientists has developed a groundbreaking model capable of generating proteins similarly to how chatbots generate text. The model, known as ESM3, was unveiled by EvolutionaryScale in June.
Introducing ESM3: The First Generative Model for Biology
EvolutionaryScale, founded by ex-Meta researchers specialized in AI models for biology, announced ESM3 as “the first generative model for biology.” This model is trained on the sequence, structure, and function of over 2.7 billion proteins, enabling it to generate new proteins based on given prompts.
“We want to build tools that can make biology programmable,” said Alexander Rives, chief scientist at EvolutionaryScale and former lead of Meta’s “AI protein team,” in an interview with Nature.
Significant Funding and Industry Support
In June, EvolutionaryScale raised $142 million in a seed round. Investors include Lux Capital, Amazon Web Services, Nat Friedman, Daniel Gross, and Nvidia’s venture capital arm, NVentures. Lux Capital co-founder and managing partner Josh Wolfe described the company’s achievement as a “ChatGPT moment for biology.”
EvolutionaryScale’s Background and Achievements
While at Meta, Rives and his team developed a database of over 600 million protein structures for drug development. They created earlier versions of the ESM model, such as ESMFold, which used a large language model trained on biological data to predict protein structures. However, the team was disbanded in 2023 during Meta’s “year of efficiency” to prioritize commercial AI products.
Generating New Green Fluorescent Proteins
EvolutionaryScale recently announced a new paper, currently in preview, demonstrating ESM3’s capability to generate new green fluorescent proteins (GFPs). The generated sequence is “only 58% similar to the closest known fluorescent protein,” mimicking proteins responsible for the glow in jellyfish and coral. EvolutionaryScale stated that this achievement is akin to simulating over 500 million years of evolution based on the natural rate of GFP diversification.