ESM
ESM3 is a generative language model for biology developed by EvolutionaryScale.
Publisher review
ESM3 is a generative language model for biology developed by EvolutionaryScale. It is designed for scientists and researchers in academia and industry who need to understand, imagine, and create proteins. The model simultaneously reasons over the sequence, structure, and function of proteins, making it a tool for protein design and engineering. ESM3 is trained across the natural diversity of the Earth, including billions of proteins from various organisms and biomes, from the Amazon rainforest to hydrothermal vents. It is available in three sizes—small, medium, and large—through the company's API and partner platforms, with a free limited-time preview through Forge for academic and commercial use.
ESM3 works by taking user inputs that mix sequence, structure, or function data simultaneously, prompting the model to explore a vast space of protein possibilities. It was trained with over 1x10^24 FLOPS and 98 billion parameters, making it one of the most computationally intensive biological models ever created. A key demonstration is the generation of esmGFP, a novel green fluorescent protein with a sequence only 58% similar to the closest known natural fluorescent protein. This generation is estimated to simulate over 500 million years of evolution. The model can also design proteins for specific functions, such as carbonic anhydrase for carbon capture, PETase for plastic degradation, and antibodies for new medicines.
ESM3 competes directly with AlphaFold and Boltz in the protein modeling space. While AlphaFold is primarily focused on structure prediction and Boltz on molecular dynamics, ESM3 differentiates itself as a generative model that can create new proteins from scratch, not just predict existing ones. Its ability to reason over sequence, structure, and function simultaneously is a unique capability not offered by competitors. However, ESM3 is a newer entrant, and its generative outputs require experimental validation, whereas AlphaFold has a longer track record of validated predictions.
The honest trade-offs with ESM3 include its high computational cost for training and inference, which may limit accessibility for smaller labs. The model's generative outputs, while impressive, are not guaranteed to be functional without experimental testing. The free preview through Forge is time-limited, and full API access may incur costs that are not yet publicly detailed. Additionally, the model's large size (98 billion parameters) may require significant hardware resources for local deployment, though cloud API access mitigates this. The Gantt chart view in the associated ESM Strategy software is not exportable, which may be a limitation for project management use cases.
How it works
-
Generative protein design
Creates new proteins from scratch, such as esmGFP, a green fluorescent protein only 58% similar to natural ones.
-
Simulates 500 million years of evolution
Generates proteins that would take nature an estimated 500 million years to evolve, based on diversification rates.
-
Multi-modal reasoning
Simultaneously reasons over protein sequence, structure, and function from user-provided inputs.
-
Trained on Earth's biodiversity
Trained on billions of proteins from diverse biomes, including Amazon rainforest, oceans, and hydrothermal vents.
-
Three model sizes
Available in small, medium, and large sizes to accommodate different computational resources and use cases.
-
High computational scale
Trained with over 1x10^24 FLOPS and 98 billion parameters, among the largest biological models.
-
Chain-of-thought prompting
Uses chain-of-thought prompting to generate proteins, as demonstrated with the esmGFP design.
Strengths and trade-offs
Strengths
- ESM3 can generate a new green fluorescent protein (esmGFP) that is only 58% similar to the closest known natural fluorescent protein, demonstrating significant novelty.
- The model simulates over 500 million years of evolution in a single generation, as estimated from natural GFP diversification rates.
- It is trained with over 1x10^24 FLOPS and 98 billion parameters, making it one of the most computationally intensive biological models available.
- ESM3 simultaneously reasons over sequence, structure, and function, a capability not offered by competitors like AlphaFold or Boltz.
Trade-offs
- The free limited-time preview through Forge is time-restricted, and full API pricing is not publicly disclosed, potentially limiting access.
- Generative outputs require experimental validation, as the model does not guarantee functional proteins without lab testing.
- The high computational cost of training and inference may be prohibitive for smaller research labs without cloud credits or institutional support.
- The associated ESM Strategy software has a Gantt chart view that is not exportable, which may hinder project management workflows.
Pricing context
Available through the company's API and partner platforms with a free limited-time preview through Forge for academic and commercial use. Full pricing details are not publicly disclosed.
Getting started with ESM
-
Sign up for ESM3
Visit the EvolutionaryScale website and create an account to access the free limited-time preview through Forge. Provide your academic or commercial affiliation details during registration to activate your access.
-
Connect to the API
Obtain your API key from the Forge dashboard after signing up. Configure your environment by setting the API key as an environment variable or including it in your request headers for authentication.
-
Choose a model size
Select one of the three model sizes—small, medium, or large—based on your computational resources and task complexity. Specify the size in your API call parameters to balance performance and cost.
-
Generate a novel protein
Submit a prompt that mixes sequence, structure, or function data to the ESM3 API. For example, provide a partial sequence and a desired function to generate a new protein like esmGFP, then review the output.
-
Validate outputs experimentally
Download the generated protein sequences from the API response. Plan and conduct laboratory experiments to test the functionality of the designed proteins, as computational outputs require validation.
Frequently Asked Questions
What is ESM3 and what does it do?
ESM3 is a generative language model for biology developed by EvolutionaryScale. It helps scientists understand, imagine, and create proteins by reasoning over sequence, structure, and function simultaneously. It can design new proteins from scratch, like the novel green fluorescent protein esmGFP.
How does ESM3 compare to AlphaFold?
AlphaFold focuses on predicting protein structures, while ESM3 is a generative model that creates new proteins from scratch. ESM3 reasons over sequence, structure, and function at once, a unique capability. However, AlphaFold has a longer track record of validated predictions, whereas ESM3 outputs need experimental testing.
Is ESM3 free to use?
ESM3 offers a free limited-time preview through Forge for academic and commercial use. Full API access is available through the company's API and partner platforms, but detailed pricing is not publicly disclosed yet. The free preview is time-restricted, so users should check current availability.
What can ESM3 create, like esmGFP?
ESM3 can generate novel proteins such as esmGFP, a green fluorescent protein only 58% similar to the closest known natural one. It can also design proteins for carbon capture, plastic degradation, and new medicines. This generation is estimated to simulate over 500 million years of evolution.
What are the limitations of ESM3?
ESM3's generative outputs require experimental validation, as they are not guaranteed functional without lab testing. The model has high computational costs for training and inference, which may limit access for smaller labs. The free preview is time-limited, and full pricing is not yet public.
How was ESM3 trained and what makes it powerful?
ESM3 was trained on billions of proteins from diverse biomes like the Amazon rainforest and hydrothermal vents. It used over 1x10^24 FLOPS and 98 billion parameters, making it one of the most computationally intensive biological models. It comes in small, medium, and large sizes for different needs.
Alternatives
- ServiceNow ↗
- BMC Helix ↗
- Freshservice ↗
How ESM compares
Direct head-to-head against 3 competitors. Picked by 7wData.
ESM
- Pricing
- Available through the company's API and partner platforms with a free limited-time preview through Forge for academic and commercial use. Full pricing details are not publicly disclosed.
- Target
- ESM3 is a generative language model for biology developed by EvolutionaryScale.
- Strength
- ESM3 can generate a new green fluorescent protein (esmGFP) that is only 58% similar to the closest known natural fluorescent protein, demonstrating significant novelty.
- Watch for
- The free limited-time preview through Forge is time-restricted, and full API pricing is not publicly disclosed, potentially limiting access.
ServiceNow
- Pricing
- Custom pricing (contact sales)
- Target
- Large enterprises needing deep extensibility and workflow orchestration across departments
- Deployment
- Cloud, on-premises
- Strength
- Deepest ITIL workflow coverage and no-code mobile app builder
- Watch for
- Steep learning curve, costly and lengthy deployment, pricing escalation
BMC Helix
- Pricing
- Custom pricing (separate licensing)
- Target
- Hybrid cloud enterprises needing unified ServiceOps and multi-cloud support
- Deployment
- Cloud, on-premises
- Strength
- HelixGPT AI and 25+ language support
- Watch for
- Complex setup, overwhelming UI for non-tech users
Freshservice
- Pricing
- Starter: $19/agent/month; Growth: $49; Pro: $99; Enterprise: Custom
- Target
- Mid-market to large enterprises seeking unified ITSM, ITOM, ITAM, and ESM
- Deployment
- Cloud
- Strength
- Freddy AI automation and 100+ pre-built workflows
- Watch for
- Needs customization for niche use cases, limited advanced reporting filters
User reviews
No user reviews yet. Be the first to write one.
Sources
Reporting on this tool draws on these publicly available sources.