Nvidia and Amazon-backed EvolutionaryScale raises $142M for Protein-generating AI
EvolutionaryScale has raised $142 million to develop AI models, including ESM3, for designing new proteins to advance scientific research, with significant funding from notable investors and collaborations with Amazon and Nvidia.
EvolutionaryScale, a fairly new company, has raised a huge amount of money to build AI models that will make new proteins for scientific research.
Nat Friedman, the former CEO of GitHub, Daniel Gross, and Lux Capital, led a $142 million seed round that EvolutionaryScale announced today.
Amazon and NVentures, Nvidia’s corporate venture arm, also participated. The company also released ESM3, an AI model that can make proteins for use in drug discovery and materials science.
Alexander Rives, co-founder and chief scientist of EvolutionaryScale, stated that ESM3 is a step towards a future in biology in which AI is used to engineer from first principles. This is similar to how we design structures, machines, microchips, and computers.
When Rives, Tom Secru, and Sal Candido worked at Meta‘s FAIR AI research lab in 2019, they started making generative AI models to study proteins. After their team broke up, Rives, Secru, and Candido left Meta to continue working on the projects they had started.
Characterizing proteins can help us understand how a disease works and find ways to stop or reverse it. On the other hand, making proteins can lead to completely new types of drugs, tools, and treatments.
Regrettably, the current method of designing proteins in the lab incurs significant costs, encompassing both computer resources and human labor.
To design a protein, you first have to think of a structure that might be able to do a job in the body or on a product. Then you have to find a protein sequence, which is the order of amino acids that make up a protein that is likely to “fold” into the structure.
For proteins to perform their intended functions, they must correctly fold into three-dimensional shapes.
According to Rives, ESM3’s training on a set of 2.78 billion proteins enables it to “reason over” the sequence, structure, and function of proteins. This means that the model can make new proteins like Google DeepMind‘s AlphaFold.
EvolutionaryScale is introducing a scaled-down version of the 98-billion-parameter model for offline use. Through its cloud-based developer platform, you can use the full model for free.
EvolutionaryScale says it created a new type of green fluorescent protein (GFP) using ESM3. GFP is the protein that makes jellyfish and coral glow in various colors. On its website, the company has a preprint paper that explains what it does.
“This is something we’ve been working on for a long time, and we can’t wait to share it with scientists and see what they do with it,” Rives said.
Of course, Evolutionary Scale isn’t a charity. The 20-person company told that it plans to make money through partnerships, usage fees, and revenue sharing.
For instance, EvolutionaryScale could collaborate with drug companies to integrate ESM3 into their operations or divide the profits with researchers when ESM3 enables the public release of breakthrough discoveries.
EvolutionaryScale says it will soon offer ESM3 and its variations to some AWS customers through AWS’ SageMaker AI dev platform, Bedrock AI platform, and HealthOmics service.
Some customers who use NVIDIA’s NIM microservices and have an NVIDIA enterprise software license will also be able to use ESM3.
Evolutionary Scale says that customers of both AWS and Nvidia will be able to use their data to fine-tune ESM3 if they want to.
EvolutionaryScale Faces Long Road to Profitability Amid Stiff Competition
EvolutionaryScale might not start making money for a while. Forbes got a copy of EvolutionaryScale’s pitch deck in August of last year, and it repeatedly said that it could take ten years for generative AI models to help design therapies.
The company will also have to deal with rivals such as Isomorphic Labs, a spinoff from DeepMind that already has deals with big drug companies, Insitro, Recursion, and Inceptive, all of which are publicly traded.
To create a general-purpose AI model for use in biotech, EvolutionaryScale is investing heavily in expanding its model training to include data other than proteins.
Bigger models, bigger data sets, and more powerful computers are driving the amazing speed at which AI is developing, according to a spokesperson for EvolutionaryScale. I concur that this is also true in the field of biology.
In the past five years, the ESM team has done research on scaling in biology. As language models grow, they learn biology and understand how things work and are structured.
