NVIDIA BioNeMo Accelerates Protein Design 50x for Custom Enzyme Engineering

Published 2025-02-01 · Health AI

NVIDIA has released an updated version of its BioNeMo generative AI platform that reduces the time required to design novel functional enzymes from months to days, achieving what the company describes as a 50-fold acceleration over traditional computational protein design methods. The platform, announced at the JP Morgan Healthcare Conference in January 2025, is already being used by 14 biotechnology companies and three academic consortia to engineer enzymes for applications ranging from pharmaceutical synthesis to plastic degradation.

The release positions NVIDIA at the centre of a rapidly evolving intersection between AI and synthetic biology — a field where the ability to design proteins with specific functions has historically been constrained by the enormous combinatorial complexity of amino acid sequences and the difficulty of predicting how sequence changes affect three-dimensional structure and catalytic activity.

What BioNeMo Does

BioNeMo is a cloud-based platform that combines several AI models for protein engineering into an integrated workflow. The core model, called ProtGen-2, is a 3.2-billion-parameter generative transformer trained on the UniProt protein database (over 250 million sequences) and augmented with structural data from the AlphaFold Protein Structure Database and the Protein Data Bank. The model generates candidate protein sequences that are predicted to fold into stable structures with desired functional properties.

The workflow begins with a user specifying a target function — for example, "an enzyme that catalyses the hydrolysis of PET plastic at 65 degrees Celsius and pH 8.0." The system then generates thousands of candidate sequences, predicts their structures using a built-in AlphaFold-based module, evaluates their predicted stability and activity using physics-informed scoring functions, and ranks the candidates by a composite "designability" score. The entire process takes between four and twelve hours on NVIDIA's DGX Cloud infrastructure, compared with the weeks or months required for traditional computational design approaches.

Dr. Anima Anandkumar, NVIDIA's director of AI research and a professor at Caltech, described the acceleration as a consequence of generative modelling: "Traditional protein design is essentially a search problem — you screen massive libraries of random variants hoping to find one that works. Generative AI inverts that process. Instead of searching for a needle in a haystack, you train a model to understand what needles look like and have it draw one for you."

Validation Through Experimental Testing

NVIDIA partnered with the Baker Lab at the University of Washington — the group behind the Rosetta protein design software — to experimentally validate BioNeMo's outputs. In a blind test, the platform designed 50 candidate enzymes for five distinct catalytic reactions: a ketoreductase for pharmaceutical intermediate synthesis, a lipase for biodiesel production, a cellulase for biomass processing, a nitrilase for chemical manufacturing, and a cutinase for PET plastic degradation.

Of the 250 designed sequences (50 per reaction), 143 (57%) were experimentally confirmed to express as soluble proteins, and 89 of those (62% of expressed proteins) displayed measurable catalytic activity. For three of the five reactions, the best BioNeMo-designed enzyme matched or exceeded the activity of the best known natural enzyme for the same reaction.

The PET degradation result drew particular attention. The best BioNeMo-designed cutinase achieved a PET hydrolysis rate 3.4 times higher than LCC-ICCG, the enzyme currently used in the Carbios industrial plastic recycling process, at comparable temperature and pH conditions. If this result holds up at industrial scale, it could significantly reduce the cost of enzymatic plastic recycling — currently estimated at $2,000 to $4,000 per tonne of PET processed, compared with approximately $1,000 per tonne for virgin PET production.

Industrial and Therapeutic Applications

Several of the 14 biotech companies using BioNeMo have disclosed their application areas. Ginkgo Bioworks is using the platform to engineer enzymes for sustainable aviation fuel production. Twist Bioscience is integrating BioNeMo into its synthetic DNA production pipeline to improve the efficiency of gene synthesis for therapeutic applications. And Recursion Pharmaceuticals is applying the platform to design enzymes that can synthesise drug candidates using biocatalytic routes that avoid the toxic solvents and heavy-metal catalysts required by conventional organic chemistry.

On the therapeutic side, the platform's ability to design binding proteins — not just enzymes — opens possibilities for biologics development. BioNeMo includes a separate model, BindGen-1, that designs minibinders: small, stable proteins that bind to specific target proteins with high affinity. In a validation study, BindGen-1 designed minibinders targeting the SARS-CoV-2 spike protein, the influenza haemagglutinin stem, and the cancer-associated receptor HER2. The top candidates showed nanomolar-range binding affinities in surface plasmon resonance experiments — a level that is therapeutically relevant, though still below the picomolar affinities achieved by the best monoclonal antibodies.

Comparison With AlphaFold and Other Tools

BioNeMo's release has invited comparisons with DeepMind's AlphaFold 3, which focuses on predicting protein-drug interactions rather than designing novel proteins. The two platforms are complementary rather than competitive: AlphaFold predicts how existing proteins and molecules interact, while BioNeMo generates new protein sequences with desired properties. A complete computational protein engineering pipeline could use BioNeMo for design and AlphaFold for validation — predicting how the designed protein would behave in context before committing to synthesis.

The Baker Lab's Rosetta software, the incumbent tool for computational protein design, remains widely used and continues to be developed. David Baker, who won the 2024 Nobel Prize in Chemistry alongside Hassabis for protein design work, has publicly welcomed BioNeMo as a complementary approach and noted that his lab is integrating generative AI components into the Rosetta framework. The competition between generative AI approaches and physics-based methods is likely to accelerate progress in both camps.

Computational Cost and Accessibility

BioNeMo runs on NVIDIA's DGX Cloud platform, with pricing that starts at approximately $37,000 per month for a dedicated instance. This cost places the platform beyond the reach of most academic labs and small startups. NVIDIA has announced an academic programme that provides subsidised access to BioNeMo through a competitive application process, with 50 grants of $100,000 in cloud credits each planned for 2025.

The pricing model has drawn criticism from open science advocates who argue that critical tools for biological research should not be controlled by a single hardware vendor. NVIDIA has responded by publishing the architecture details and training methodology for ProtGen-2 in an open-access paper, enabling other groups to replicate the model on alternative hardware. However, the computational requirements for training a 3.2-billion-parameter protein model remain substantial — estimated at approximately $400,000 in cloud GPU costs — creating a significant barrier to independent replication.

The open science concern connects to broader questions about the concentration of AI capability in a handful of technology companies. As tools like BioNeMo and AlphaFold become central to drug discovery and biotechnology, the entities that control these platforms wield disproportionate influence over which research directions are pursued and which are not. The WHO's recent guidelines on AI in health explicitly address this concern, recommending that critical health AI tools be developed with open-source licensing and transparent governance structures.

Looking Forward

NVIDIA is developing the next generation of BioNeMo, which will incorporate multi-modal training on protein sequences, structural data, and experimental functional measurements simultaneously. The company is also exploring the use of reinforcement learning to optimise designed proteins for properties beyond catalytic activity — including thermostability, expression yield, and immunogenicity — based on feedback from experimental testing.

The long-term vision, according to Dr. Anandkumar, is "a closed-loop system where AI designs a protein, a robotic lab synthesises and tests it, the results feed back into the model, and the cycle repeats autonomously." Several groups, including the Baker Lab and a consortium at the University of Manchester, are building exactly this kind of self-driving lab infrastructure. The convergence of generative protein design, automated experimentation, and closed-loop optimisation could compress what has historically been a multi-year drug or enzyme development programme into a matter of weeks.

For more on AI applications in drug discovery and clinical development, explore our AI in Health research repository. Related coverage includes AlphaFold 3's drug-target interaction predictions and Mayo Clinic's AI-driven heart failure detection.

← Back to Publications