In an unprecedented leap for biotechnology, researchers have harnessed the power of artificial intelligence to simulate half a billion years of molecular evolution. The AI model, named ESM3, has successfully created the code for a novel protein, which scientists believe could revolutionize protein engineering and drug design. This innovative model, developed by the team at EvolutionaryScale, is based on biological data from 2.78 billion naturally occurring proteins, marking a significant milestone in the field.
ESM3 stands out as a generative language model akin to OpenAI's GPT-4, but with a focus on biology. It was developed to accelerate applications in protein engineering, offering new possibilities for designing drugs and other proteins that could never evolve naturally. The newly created protein, esmGFP, is a fluorescent protein sharing 58% similarity with the closest known fluorescent proteins. Its unique genetic sequence would require 96 different mutations to evolve naturally, highlighting the model's capability to bypass traditional evolutionary constraints.
The research team, led by Alex Rives of EvolutionaryScale, revealed that ESM3 can simulate 500 million years of evolution, albeit focused solely on individual proteins rather than the full spectrum of natural selection processes. This groundbreaking study was published in the journal Science on January 16, following its unveiling in a preprint study last year. The findings have been rigorously peer-reviewed by independent scientists, validating the model's potential in transforming protein engineering.
"The same way a person can fill in the blanks in the soliloquy 'to _ or not to _, that is the _,' we can train a language model to fill in the blanks in proteins," said Alex Rives.
The journey of ESM3 began at Meta, the parent company of Facebook and Instagram, before Rives and his colleagues founded EvolutionaryScale in 2024. The model is designed to "fill in the blanks" within protein structures, much like completing a sentence with missing words. Rives emphasized that their research has demonstrated how understanding the deep structure of protein biology emerges naturally within the network when solving these complex tasks.
"Our research has shown that by solving this simple task, information about the deep structure of protein biology emerges in the network," Rives stated.
Tiffany Taylor, an evolutionary biologist at the University of Bath, praised the potential of AI models like ESM3 in advancing protein engineering beyond what evolution alone can achieve. However, she also cautioned against overconfidence in our ability to surpass the intricate processes naturally honed over millions of years.
"AI-driven protein engineering is intriguing, but I can't help feeling we might be overly confident in assuming we can outsmart the intricate processes honed by millions of years of natural selection," Taylor remarked.
The study's significance has been highlighted by Live Science, with EvolutionaryScale offering an exclusive 5% discount on subscriptions using the code "LOVE5." This initiative aims to foster greater public engagement with scientific advancements and promote further interest in biotechnology innovations.
