Evolutionary Scale
Protein language model research company
EvolutionaryScale is an AI research company that develops protein language models — large models trained on vast collections of protein sequences to predict protein structure, function, and design novel proteins — most notably continuing…
Definition
EvolutionaryScale is an AI research company that develops protein language models — large models trained on vast collections of protein sequences to predict protein structure, function, and design novel proteins — most notably continuing development of the ESM (Evolutionary Scale Modeling) family of models originally created at Meta AI. It positions these models as general-purpose tools for protein science, usable both for understanding existing proteins and generating new ones with desired properties.
Overview
EvolutionaryScale grew out of research originally conducted at Meta AI on protein language models, a class of models that apply the same self-supervised learning techniques used for natural language — predicting masked tokens in a sequence — to protein amino acid sequences instead of words. The underlying premise is that evolution has already run an enormous natural experiment across billions of years and countless species, and that a model trained on the resulting diversity of protein sequences can learn statistical patterns reflecting the physical and functional constraints that shape proteins, without needing explicit physics-based simulation. Mechanically, protein language models like the ESM series are trained on databases containing hundreds of millions of protein sequences drawn from many organisms, learning an internal representation of amino acid sequences that captures information about secondary and tertiary structure, evolutionary conservation, and functional motifs. Once trained, these models can be used in several ways: predicting a protein's structure directly from sequence, generating entirely novel protein sequences with specified structural or functional properties, and producing sequence embeddings useful as input features for other downstream biological prediction tasks. EvolutionaryScale's approach sits alongside, and is often compared with, structure-prediction-focused systems like AlphaFold, but the technical starting point differs: protein language models learn primarily from sequence data and its evolutionary patterns, whereas AlphaFold-style structure predictors were trained with heavier reliance on known solved structures and evolutionary sequence alignments as an explicit input feature. In practice, the two approaches have increasingly converged and can complement each other, with sequence-based models sometimes used to generate candidates that structure predictors then evaluate. In practice, researchers and biotech companies use EvolutionaryScale's models to design novel proteins for applications such as enzymes with new catalytic properties, therapeutic proteins, or research tools, and to analyze existing proteins for functional annotation when experimental characterization is unavailable. The company has released some model versions openly for research use, contributing to a broader ecosystem of open protein modeling tools alongside commercial offerings for organizations seeking more capability or support. The main limitations of protein language models mirror those of the broader field: predictions and generated designs are statistical inferences from training data and can fail for proteins very different from anything well represented in that data, and generated novel proteins still require experimental synthesis and testing to confirm they fold and function as intended. As with other AI biology tools, the models accelerate hypothesis generation but do not replace the wet-lab validation that ultimately determines success.
Key Features
- Develops the ESM (Evolutionary Scale Modeling) protein language model family
- Trained via self-supervised learning on hundreds of millions of protein sequences
- Predicts protein structure and function directly from amino acid sequence
- Generates novel protein designs with specified structural properties
- Originated from protein language model research at Meta AI
- Some model versions released openly for research use
- Produces sequence embeddings usable in downstream biology tasks
Use Cases
Alternatives
Frequently Asked Questions
From the Blog
Economies of Scale: Why Bigger Can Mean Cheaper
Economies of scale happen when producing more of something lowers the average cost per unit, because fixed costs spread across more output. This guide explains how the effect works, its main sources, and where it eventually breaks down.
Read More AI & TechnologyVector Search in Production: Indexes, Filters and Scale
Production vector search is four decisions: which index family you build, how filters interact with that index, how you shard and refresh as the corpus grows, and how you measure recall rather than assume it. Get those right and embedding search stays fast under real traffic; get them wrong and it degrades quietly.
Read More AI & TechnologyWhat Is a Line Graph and When Should You Use One?
A line graph plots data points connected by straight lines to show how a value changes over a continuous scale, usually time. This guide explains how to read one, when it's the right chart choice, and common mistakes to avoid.
Read More AI & TechnologyThe Real Advantages of AI (And Where It Still Falls Short)
AI's biggest advantages are speed, consistency, and the ability to find patterns in data at a scale humans cannot match. This guide explains where AI genuinely helps, where it still needs human oversight, and how to start learning it.
Read More