EvolutionaryScale's ESM3 simulates 500 million years of evolution

What if you could compress 500 million years of natural evolution into a single prompt on a computer?

EvolutionaryScale—a new venture formed by former Meta AI researchers—introduced ESM3, a frontier generative model for biology. Boasting 98 billion parameters and trained on more than 1x10^24 FLOPS, ESM3 is natively multimodal, reasoning across protein sequence, 3D structure, and biological function simultaneously. To demonstrate its power, the model generated esmGFP, a brand-new green fluorescent protein sharing just 58% sequence identity with any known counterpart in nature. By evolutionary divergence rates, that leap is equivalent to simulating roughly 500 million years of natural selection.

Tokenizing life at the atomic level

Historically, computational biology tools tackled one piece of the puzzle at a time—either predicting folds or suggesting sequence mutations. ESM3 treats biology as a unified discrete language.

To make this work, the team converted complex 3D structures and functional annotations into discrete token vocabularies. This allowed them to train ESM3 using masked language modeling, much like GPT or BERT. The model learns by predicting masked positions across sequence, structural coordinates, and functional labels across billions of proteins scraped from every biome on Earth. To fill in the experimental data gaps, EvolutionaryScale also augmented the dataset with hundreds of millions of synthetic annotations.

Crucially, ESM3 scales with compute. Tasks like atomic coordination—placing amino acids that sit far apart in the sequence but must touch in 3D space—natively emerged only as the model reached frontier scale. The team also incorporated reinforcement-learning-style alignment, enabling the model to iteratively self-critique its designs.

Biology gets a prompt window

Because the modalities share the same vocabulary, researchers can prompt ESM3 with arbitrary combinations of constraints.

In one test, the team prompted the model with the active-site structure of PETase (an enzyme that breaks down PET plastics), along with key amino acids and a functional keyword for hydrolase fold. ESM3 generated viable backbone scaffolds matching those exact operational requirements. What used to take years of directed evolution in Petri dishes can now start as a multi-condition query in an API.

Why it matters

We think ESM3 represents a watershed moment: computational biology is officially transitioning from passive prediction to active, programmable generation.

For wet-lab scientists, biotech startups, and enzyme designers, this changes the playbook. Instead of searching nature for an enzyme and introducing dozens of slow mutations, engineers can specify the functional active site and let the model hallucinate the supporting scaffold. If you have the bench capacity to express and validate novel proteins, tapping into EvolutionaryScale's Forge platform is an immediate no-brainer.

For the broader industry, it raises the stakes for AlphaFold and incumbent pharma AI platforms. Prediction is now a solved commodity; the real battleground is controllable generative design across sequence and function.

That said, the primary bottleneck will simply move downstream. Generating a completely novel protein with 58% homology is incredible on screen, but scaling physical wet-lab validation and synthesis still takes weeks. Furthermore, democratizing the de novo design of functional biomolecules makes biological containment and dual-use oversight urgent operational realities, not distant hypothetical debates.