EvolutionaryScale Unveils ESM3 Generative AI Model for Advanced Protein Design

EvolutionaryScale has introduced its latest generative AI model, ESM3, which leverages NVIDIA H100 GPUs to enhance protein research and discovery. The new model, which is the third-generation of EvolutionaryScale’s ESM series, integrates the sequence, structure, and functions of proteins, offering a sophisticated tool for protein discovery engineers.

Programmable Biology with ESM3

Emerging from Meta’s Fundamental AI Research (FAIR) unit, EvolutionaryScale has secured funding from Lux Capital, Nat Friedman, Daniel Gross, and NVIDIA. This financial backing supports the company’s mission to advance programmable biology. The ESM3 model’s capabilities can be pivotal in targeting cancer cells, finding alternatives to harmful plastics, and driving environmental mitigations, among other applications.

The ESM3 model, containing 98 billion parameters, represents a significant leap from its predecessor, ESM2. Utilizing approximately 25 times more FLOPS and 60 times more data, ESM3 is trained on nearly 2.8 billion protein sequences, sourced from various organisms and biomes. This extensive training allows for greater accuracy in identifying and validating new proteins.

Innovations in Protein Design

ESM3’s generative nature and its “all to all” model architecture enable the integration of structure and function annotations as inputs, rather than merely as outputs. This innovation allows researchers to prompt the model to generate new proteins with provided scaffolds, self-improve based on feedback, and design proteins with specified functionalities. The model acts like a virtual researcher, iterating on protein designs in response to user prompts.

Tom Sercu, co-founder and VP of engineering at EvolutionaryScale, highlighted the model’s capabilities in solving complex protein design problems, such as creating a novel Green Fluorescent Protein. Sercu expressed optimism about ESM3’s potential to accelerate scientific research and open new avenues in life sciences.

See also: 14 Companies Pioneering AI Foundation Models in Pharma and Biotech

ESM3’s Integration with NVIDIA BioNeMo

ESM3 is set to be available on NVIDIA BioNeMo, a generative AI platform for drug discovery. EvolutionaryScale will offer an API for closed beta access and release code and weights for a smaller, non-commercial version of ESM3. The full ESM3 family of models will be available to select customers as an NVIDIA NIM microservice, optimized for run-time in collaboration with NVIDIA, and supported by an NVIDIA AI Enterprise software license.

The training of ESM3 was conducted using the Andromeda cluster, which employs NVIDIA H100 GPUs and NVIDIA Quantum-2 InfiniBand networking. This robust computational infrastructure underscores the increasing computing power necessary for training advanced biological models.

Topics:
Emerging Technologies