AI Develops New Glowing Protein, Mimicking 500 Million Years of Evolution

The synthesis of new proteins—the fundamental components of biological life—holds great scientific promise. A newly developed AI model offers the potential to generate instructions for creating proteins far beyond what occurs naturally.

Creating Custom Proteins

Scientists in the US have utilized the EvolutionaryScale Model 3 (ESM3) to create a new protein, esmGFP (green fluorescent protein), which shares only 58% of its structure with its closest natural counterpart, tagRFP. This achievement is roughly equivalent to processing 500 million years of evolution through AI, according to the research team. It paves the way for designing custom proteins tailored for specific applications or enhancing the functions of existing proteins.

“Over three billion years of evolution have shaped a biological blueprint embedded within the structure of natural proteins,” the researchers, led by Thomas Hayes, founder of EvolutionaryScale in New York, explain in their published paper.

“In this study, we demonstrate that large-scale language models trained on evolutionary data can generate functional proteins that significantly differ from known proteins.”

ESM3 was trained on a vast dataset consisting of 3.15 billion protein sequences (the arrangement of amino acids in a protein), 236 million protein structures (their 3D shapes), and 539 million protein annotations (descriptive labels).

AI Learning from Data

By identifying patterns in vast datasets, the AI model can learn what works and what doesn’t in protein construction and function, similar to how ChatGPT can generate a new poem that rhymes after analyzing millions of poems written by humans.

What sets esmGFP apart is that it actually works: it fluoresces, just like its relative tagRFP. Fluorescent proteins are responsible for the glow in some ocean organisms and are crucial as markers in medicine and biotechnology.

“We selected fluorescence for its challenging nature, ease of measurement, and because it is one of the most beautiful mechanisms in nature,” the team explains.

The AI reduces much of the trial and error in protein synthesis, while also enabling the exploration of proteins that are vastly different from those currently known.

Proteins Connected by Mutational Pathways

“Proteins can be viewed as existing within an organized space, where each protein is adjacent to others that are one mutational step away,” the researchers write. “The structure of evolution forms a network within this space, linking all proteins through the paths evolution could take between them.”

For evolution to progress, the team explains, each protein must evolve into the next without disrupting the overall functionality of the system it belongs to. A language model understands proteins within this space.

While proteins designed by ESM3 still require validation, synthesis, and testing—processes that take time—the team is confident that further advancements will follow. In the near future, AI could enable the production of proteins for a wide range of applications, from medicines to biomaterials.

“Protein language models don’t explicitly operate within the physical limitations of evolution, but instead can implicitly build a model of the many possible evolutionary paths that could have been taken,” the researchers clarify.

Read the original article on: Science Alert

Read more: AI Assists Doctors in Identifying More Cases of Breast Cancer in the Largest Real-World Study