How do I generate a new protein with ESM3 step by step?

Use the generate() function with a partial sequence in an ESMProtein object, iteratively sampling tracks as shown in the quickstart; first generate sequence, then structure, with configurable steps and temperature.

ESM3 or AlphaFold for protein structure prediction?

ESM3 predicts structure as part of multimodal generation and is better for novel design, while AlphaFold is optimized for accurate single-sequence structure prediction; choose based on whether you need generative capabilities or precise folding.

What's the cost of using ESM Forge API?

Pricing is not detailed in the README; you must apply for access and check EvolutionaryScale's terms, as costs likely scale with model size (e.g., 98B vs. 1.4B) and usage volume.

Can ESM handle DNA or RNA sequences?

No, ESM is specifically designed for proteins and amino acid sequences; for nucleic acids, you'd need to adapt the model or use other tools, as it's not covered in the documentation.

How to embed a protein sequence locally with ESM C?

Install the esm package, load a model like esmc_300m with ESMC.from_pretrained(), and call encode() on an ESMProtein object, as demonstrated in the ESM C quickstart section.

What are the license restrictions for commercial use?

Licenses vary; ESM3-open is non-commercial, while ESM C offers commercial options via SageMaker under a specific agreement, so review LICENSE.md carefully for compliance.

ESM3 — Multimodal Protein Language Model

What is ESM3?

ESM is a family of protein language models developed by EvolutionaryScale for simulating and designing proteins. It includes ESM3, a multimodal generative model that reasons over protein sequence, structure, and function, and ESM C, a representation learning model for creating protein embeddings. These tools enable researchers to generate novel proteins, predict properties, and analyze biological data at scale.

Target Audience

Computational biologists, bioinformatics researchers, and AI scientists working on protein engineering, drug discovery, and biological sequence analysis.

Value Proposition

Developers choose ESM for its state-of-the-art multimodal reasoning capabilities, scalable architecture, and flexible deployment options, including local inference, cloud API access, and commercial licensing via AWS SageMaker.

Overview

ESM (Evolutionary Scale Modeling) is a family of protein language models developed by EvolutionaryScale. It includes ESM3, a frontier generative model for biology, and ESM C, a representation learning model for creating protein embeddings. These models enable researchers and developers to simulate protein evolution, design novel proteins, and analyze biological sequences with unprecedented scale and control.

Key Features

Multimodal Reasoning — Jointly processes and generates across three protein modalities: amino acid sequence, 3D structure, and functional keywords.
Generative Protein Design — Can be prompted with partial information to iteratively sample and generate complete, novel proteins.
Scalable Architecture — Built on a transformer backbone, with models ranging from 1.4B to 98B parameters, trained on billions of proteins.
Flexible Deployment — Models can be run locally via Hugging Face, accessed via a cloud API (Forge), or deployed commercially on AWS SageMaker.
Efficient Embeddings — ESM C models provide high-performance protein sequence representations as drop-in replacements for earlier ESM2 models.

Philosophy

ESM is developed with a mission to understand biology for human benefit through open, safe, and responsible AI research, guided by a framework that emphasizes risk evaluation and stakeholder collaboration.

Use Cases

Best For

Generating novel proteins with desired structural or functional properties
Predicting protein structure from sequence data
Creating high-quality embeddings for protein sequence analysis
Performing inverse protein folding (designing sequences for given structures)
Research in computational biology and protein engineering
Commercial applications requiring licensed, scalable protein model inference

Not Ideal For

Applications unrelated to proteins or biological sequence analysis, such as general NLP or computer vision tasks
Small-scale academic projects with limited GPU resources or budget for cloud inference
Real-time protein analysis requiring low-latency inference, due to model size and iterative sampling processes
Research prioritizing full model interpretability and explainability over predictive performance

Pros & Cons

Pros

Multimodal Generative Power

ESM3 jointly reasons across sequence, structure, and function tracks, enabling controlled protein design with partial prompts, as shown in the diagram and GFP generation tutorial.

Scalable Model Family

Offers models from 1.4B to 98B parameters trained on billions of proteins, providing options for different computational needs, as listed in the available models table.

Flexible Deployment Paths

Supports local inference via Hugging Face, cloud API through Forge, and commercial licensing on AWS SageMaker, allowing seamless transition from research to production.

Efficient Embedding Models

ESM C delivers high-performance protein embeddings with reduced memory and faster inference than ESM2, acting as a drop-in replacement for sequence analysis tasks.

Cons

High Computational Overhead

Running large models like the 98B ESM3 locally requires substantial GPU memory and power, which may be inaccessible without high-end hardware or cloud credits.

Complex Commercial Setup

Deploying on AWS SageMaker involves multi-step CloudFormation templates and AWS account management, adding significant setup time and operational overhead.

Vendor Lock-in for API

Access to advanced features via Forge API ties users to EvolutionaryScale's infrastructure, with potential costs, latency, and dependency on external service availability.

What is ESM3?

Target Audience

Computational biologists, bioinformatics researchers, and AI scientists working on protein engineering, drug discovery, and biological sequence analysis.

Value Proposition

Overview

Key Features

Multimodal Reasoning — Jointly processes and generates across three protein modalities: amino acid sequence, 3D structure, and functional keywords.
Generative Protein Design — Can be prompted with partial information to iteratively sample and generate complete, novel proteins.
Scalable Architecture — Built on a transformer backbone, with models ranging from 1.4B to 98B parameters, trained on billions of proteins.
Flexible Deployment — Models can be run locally via Hugging Face, accessed via a cloud API (Forge), or deployed commercially on AWS SageMaker.
Efficient Embeddings — ESM C models provide high-performance protein sequence representations as drop-in replacements for earlier ESM2 models.

Philosophy

Use Cases

Best For

Generating novel proteins with desired structural or functional properties
Predicting protein structure from sequence data
Creating high-quality embeddings for protein sequence analysis
Performing inverse protein folding (designing sequences for given structures)
Research in computational biology and protein engineering
Commercial applications requiring licensed, scalable protein model inference

Not Ideal For

Applications unrelated to proteins or biological sequence analysis, such as general NLP or computer vision tasks
Small-scale academic projects with limited GPU resources or budget for cloud inference
Real-time protein analysis requiring low-latency inference, due to model size and iterative sampling processes
Research prioritizing full model interpretability and explainability over predictive performance

Pros & Cons

Pros

Multimodal Generative Power

ESM3 jointly reasons across sequence, structure, and function tracks, enabling controlled protein design with partial prompts, as shown in the diagram and GFP generation tutorial.

Scalable Model Family

Offers models from 1.4B to 98B parameters trained on billions of proteins, providing options for different computational needs, as listed in the available models table.

Flexible Deployment Paths

Supports local inference via Hugging Face, cloud API through Forge, and commercial licensing on AWS SageMaker, allowing seamless transition from research to production.

Efficient Embedding Models

ESM C delivers high-performance protein embeddings with reduced memory and faster inference than ESM2, acting as a drop-in replacement for sequence analysis tasks.

Cons

High Computational Overhead

Running large models like the 98B ESM3 locally requires substantial GPU memory and power, which may be inaccessible without high-end hardware or cloud credits.

Complex Commercial Setup

Deploying on AWS SageMaker involves multi-step CloudFormation templates and AWS account management, adding significant setup time and operational overhead.

Vendor Lock-in for API

Access to advanced features via Forge API ties users to EvolutionaryScale's infrastructure, with potential costs, latency, and dependency on external service availability.

ESM3

What is ESM3?

Overview

Key Features

Philosophy

Use Cases

Best For

Not Ideal For

Pros & Cons

Pros

Cons

Frequently Asked Questions

Related Projects

Found a gem we're missing?

ESM3

What is ESM3?

Overview

Key Features

Philosophy

Use Cases

Best For

Not Ideal For

Pros & Cons

Pros

Cons

Frequently Asked Questions

Related Projects

Found a gem we're missing?