An AI ahead: it has generated what evolution has not yet conceived 🧬

A team of researchers has recently made a significant breakthrough in the field of synthetic biology thanks to artificial intelligence. With this AI, they were able to outpace natural evolution.

They used a generative language model, ESM3, to design a completely new fluorescent protein, opening up unprecedented possibilities in protein engineering.


ESM3 stands out for its ability to integrate sequence, structure, and function of proteins, offering a novel holistic approach. This model, trained on billions of protein data points, allows for the simulation of biological evolutions over vast timescales, up to 500 million years.

The creation of this fluorescent protein, whose genetic sequence radically differs from all known proteins, illustrates the potential of ESM3. This success paves the way for various applications, ranging from medicine to environmental remediation, and the design of innovative materials.

ESM3 is accessible in a public beta version via an API, enabling scientists to leverage this tool for protein engineering. This accessibility fosters increased collaboration among researchers and accelerates discoveries in the field.

The training of ESM3 on a vast corpus of protein data, including sequences, structures, and functional annotations, has achieved unprecedented accuracy. This model, capable of handling up to 98 billion parameters, represents a major advancement in protein modeling.

The implications of this technology are vast, offering powerful tools to explore the immense diversity of proteins. ESM3 not only allows for a better understanding of natural proteins but also enables the creation of proteins with unique properties for specific applications.

This innovation, published in Science, marks a turning point in the use of AI for synthetic biology. It demonstrates how generative language models can transform our approach to protein design, simulating complex evolutionary processes to generate molecules with novel functionalities.

How does ESM3 model proteins?

ESM3 uses an innovative approach to model proteins by integrating sequence, structure, and function into a generative language model. Unlike previous models, ESM3 represents these three aspects through discrete token alphabets, enabling more precise and holistic protein generation.

The model is trained on a vast dataset, including billions of protein sequences, millions of structures, and functional annotations. This wealth of data allows ESM3 to simulate complex evolutionary processes, offering a deep understanding of natural proteins and the ability to design new proteins.

ESM3 can handle up to 98 billion parameters, making it one of the most powerful models for protein modeling. This capability enables precise and detailed simulations, opening new perspectives for research in synthetic biology.

What are the potential applications of ESM3?

ESM3 opens up unprecedented possibilities in various fields, including medicine, where it could enable the design of therapeutic proteins with unique properties. In environmental science, it could contribute to pollution remediation by creating enzymes capable of degrading specific pollutants.

In the field of materials, ESM3 could be used to design proteins with particular mechanical or optical properties, useful for the creation of new materials. These applications illustrate the transformative potential of ESM3 for science and technology.

The accessibility of ESM3 via a public beta API facilitates its adoption by the scientific community. This enables increased collaboration and accelerates discoveries by providing a powerful tool for protein engineering to a wide range of researchers.