EvolutionaryScale: Advancing Life Sciences with Frontier AI

EvolutionaryScale is a public benefit company pioneering frontier AI for life sciences, with ESM3—a groundbreaking generative language model for biology—capable of simultaneously understanding and designing proteins by analyzing sequence, structure, and function. This powerful AI enables scientists to imagine and create novel proteins beyond natural evolution, such as esmGFP, which was designed in a fraction of the time it would take through natural processes. Trained on 2.78 billion natural proteins and 771 billion unique tokens using 1 trillion teraflops of compute, ESM3 features emergent reasoning, multimodal prompting, and self-improvement via feedback mechanisms. Available in small, medium, and large sizes through API, AWS Sagemaker, Omics, Bedrock, NVIDIA BioNemo, and NVIDIA NIM microservices, it supports both commercial and open research through its ESM3-open model, released under a non-commercial license with full weights and source code on GitHub. With a commitment to scientific rigor, responsible AI, and community collaboration, EvolutionaryScale is accelerating innovation in biomedicine, synthetic biology, and sustainable science.

Key Features:

• ESM3 Protein Modeling: State-of-the-art AI model that reasons over protein sequence, structure, and function simultaneously.
• Generative Protein Design: Capable of creating novel proteins like esmGFP, designed in a fraction of the time natural evolution would require.
• Large-Scale Training: Trained on 2.78 billion natural proteins and 771 billion unique tokens using 1 trillion teraflops of compute.
• 98 Billion Parameters: One of the largest evolutionary-scale models developed for biological data.
• Multimodal Prompting: Enables precise control over protein design by incorporating sequence, structure, and function constraints.
• Self-Improvement via Feedback: Uses RLHF-inspired mechanisms to refine protein generations based on lab data or internal evaluations.
• Synthetic Data Augmentation: Enhances training with synthetic data to compensate for limited experimental annotations.
• Multiple Deployment Options: Accessible via API, AWS Sagemaker, AWS Bedrock, Omics platform, NVIDIA BioNemo, and NVIDIA NIM microservices.
• ESM3-open Model: A small, open, and safe 1.4B parameter model with weights and source code available on GitHub under a non-commercial license.
• Public Beta Access: Scientists in academia and industry can access the API through Forge for early testing and development.
• Responsible AI Development: Prioritizes transparency, risk evaluation, mitigation, and stakeholder collaboration.

Pricing: EvolutionaryScale offers public beta access to its ESM3 API for scientists, with full pricing details available through its partner platforms and API access. The ESM3-open model is free for non-commercial use, while commercial access and larger model variants are available via subscription or usage-based plans through AWS and NVIDIA.

Conclusion: EvolutionaryScale’s ESM3 represents a transformative leap in AI-driven biological discovery, empowering researchers to design life-changing proteins with unprecedented speed and accuracy—making it an essential tool for advancing human health and scientific innovation.