Artificial intelligence is becoming part of the design process behind new food ingredients. Researchers are using models originally developed for structural biology, protein engineering and generative design to investigate enzymes, fermentation organisms and alternative proteins.
These systems do not produce a finished food product on their own. They help scientists narrow the number of molecules, sequences and experiments that need to be tested in the laboratory. That can be valuable in food biotechnology, where a single project may involve screening large numbers of protein variants or microbial candidates.
The five tools below are among the most relevant to this emerging workflow. They are not ranked by commercial value. Each addresses a different stage of biological design.
1. AlphaFold and the AlphaFold Database
DeepMind’s AlphaFold changed protein research by making accurate structure prediction available at a scale that was previously difficult to achieve. The AlphaFold Protein Structure Database, operated by EMBL-EBI and Google DeepMind, gives researchers access to predicted structures for a very large number of proteins.
For food bioengineering, structural information can help researchers study enzymes involved in fermentation, plant proteins being evaluated for alternative foods and proteins that influence functionality such as solubility or stability. A predicted structure may also help identify where a sequence could be changed before laboratory testing.
AlphaFold is a starting point rather than a complete design system. Researchers still need to establish whether a protein folds as expected under real processing conditions, behaves safely and performs well in a food formulation.
2. RFdiffusion
RFdiffusion, developed by researchers associated with the University of Washington’s Institute for Protein Design and collaborators, uses diffusion-based generative modeling to propose new protein structures. The method can be used for tasks such as designing proteins around a target shape or functional site.
That capability is relevant to food biotechnology because researchers may want enzymes with a particular activity, greater stability or performance in a defined production process. In principle, generative design can explore protein architectures that would be difficult to find through conventional trial and error alone.
The output is a set of computational candidates, not a validated ingredient. Each candidate must undergo expression, purification, activity testing and, where relevant, food-safety assessment.
3. ProteinMPNN
RFdiffusion and similar systems can propose a protein backbone, but researchers still need amino-acid sequences capable of forming that structure. ProteinMPNN addresses this sequence-design problem by predicting sequences that are compatible with a given protein backbone.
In a food research setting, ProteinMPNN can support the design of enzyme variants or other functional proteins for subsequent testing. It is especially useful as part of a computational pipeline: a model generates or selects a structural concept, ProteinMPNN proposes sequences, and laboratory teams then test whether those sequences express and function as intended.
Its value depends heavily on the quality of the input structure and the experimental validation that follows. Sequence compatibility does not automatically mean that a protein will be manufacturable, stable in a food matrix or acceptable for regulatory use.
4. ESM-3
ESM-3, from EvolutionaryScale, is a generative biological model designed to work across protein sequence, structure and function. The model is part of a broader movement toward protein language models, which learn patterns in biological sequences in a way that resembles how language models learn relationships between words.
For researchers working on food, a model that connects sequence, structure and function could assist with proposing protein variants for enzymes, fermentation systems or novel ingredients. It may also help teams compare design options before committing resources to synthesis and wet-lab screening.
As with other generative models, the central challenge is experimental evidence. A plausible output can still be inactive, difficult to produce or unsuitable for consumption. The model is best viewed as a way to prioritize experiments rather than as an autonomous product-development engine.
5. NVIDIA BioNeMo
NVIDIA BioNeMo is a platform for generative AI and biomolecular modeling. It brings together models, software and computing infrastructure intended to support work with proteins, molecules and related biological data.
For food bioengineering teams, a platform approach can be useful when projects require more than one model. A team may use structure prediction, sequence generation and molecular analysis within a connected computational workflow instead of managing every component separately. This can matter for companies developing enzymes, precision-fermentation ingredients or alternative-protein components.
BioNeMo is not a food-specific validation framework. Its usefulness will depend on the models selected, the data available to the research team and the quality of the laboratory pipeline connected to it.
How these tools fit into food research
The strongest use case is a loop between computation and experimentation. Researchers can use AI to generate or prioritize candidates, synthesize a smaller selection, measure performance in the laboratory and feed the results back into the next design cycle.
That workflow may support several areas of food innovation:
- Precision fermentation: designing or improving enzymes and proteins made by engineered microorganisms.
- Alternative proteins: studying structure and functionality in proteins intended to improve texture, nutrition or processing performance.
- Food enzymes: exploring variants that operate under particular temperatures, acidity levels or manufacturing conditions.
- Ingredient development: screening candidates for stability, solubility and compatibility with a finished formulation.
The commercial bottleneck remains broader than model performance. Teams must also manage fermentation scale-up, ingredient cost, supply chains, food safety, consumer acceptance and regulatory requirements in each target market. AI can reduce some search costs, but it does not remove those obligations.
Why validation still determines the outcome
Protein design models can identify promising candidates, yet biological systems are sensitive to context. A protein may behave differently in a cell, a fermentation vessel or a complex food matrix than it does in a computational prediction.
That is why these tools are most useful when paired with strong experimental design and domain expertise. The winning advantage may belong less to the team with access to a single model than to the one that can connect data quality, laboratory execution and manufacturing economics.
For more coverage of emerging technologies, visit our Innovation and Technology section. Readers following the commercial side of new ingredients can also explore our Food and Health coverage.
