This is cross-posted from the EA Forum.
I’m writing this in a personal capacity, and am not representing the views of my employer. Many thanks to everyone who provided feedback and ideas.
In previous posts, I set out how artificial intelligence brings huge risks, and huge opportunities, for all animals, and outlined some ways to help ensure AI represents animals’ interests. The present post explores one of the greatest opportunities AI brings for animals: accelerating the development and mainstreaming of tasty, nutritious, affordable, and convenient alternatives to animal products.
AI’s potential for alternative proteins is being explored by many companies and other corporate bodies (such as Animal Alternative Technologies, BioCraft, Climax Foods, CULT Food Science, Eternal, Equinom, Eat Just, KraftHeinz, NotCo, Protein Industries Canada, and Shiru, to name a handful), a few non-profit organizations (such as GreenProtein and New Harvest), and at least one national government, with the UK’s Food Standards Agency publishing a 2022 report that notes:
Combining artificial intelligence and machine learning and other emerging technologies such as 3D printing [...] could lead to an entirely new food production system, in which any desired molecule or food could be synthesised from any feedstock input. This has the potential to profoundly change the types of food we consume in the future, and the way most of our food is produced [...] This may bring considerable benefits for food system resilience and food security, and at the same time help to reduce food waste.
Nonetheless, research in this area is still in its infancy, in part because many of the companies working on this are keeping their cards very close to their chest. As such, the opportunities listed below are generally avenues where further exploration seems particularly promising, rather than tried-and-tested applications. All corrections, additions, and points of disagreement are very welcome.
Optimizing the extrusion process
Extrusion is the process that converts plant proteins into a food product with the required shape, structure, and texture. This is far more complex than it might appear at first glance. Critical variables include temperature, pressure, moisture content, feed rates (i.e., the speed at which the proteins are fed into the extrusion mechanism), screw speed (i.e., the speed at which the extrusion mechanism operates), and die design (i.e., the shape and size of the molds or ‘dies’ through which the protein material is forced). Another variable is the ingredients themselves: proteins are complex inputs (especially protein types from non-soy sources, which we know comparatively little about), further complicated by the varying effects of the different ‘isolation’ methods used to extract the proteins from their original plant source.
A change to any one of these parameters can yield significant variations in the final product, often in highly unpredictable ways. Running trials to test which combinations of these parameters are most likely to yield the desired results can be prohibitively expensive, especially for smaller start-ups.
In principle, AI could radically streamline this process by running simulations using a wide array of ingredients and other parameters to find the most promising combinations. While this is unlikely to entirely replace real-world extrusion trials, at least in the short-term, it can help to whittle down the set of possible combinations so that fewer real-world trials are necessary. This is a key objective of GreenProtein.ai, which is developing an AI-powered extrusion model that companies can use to streamline their processes and help turn the ‘black box’ of extrusion from an art into a science.
Mapping unique combinations of plant proteins
All the amino acids found in animal-derived products can also be found in plants. As such, in principle, AI can be used to map the protein structures of different animal proteins, then replicate this molecular structure using only plant-based ingredients. (With up to 30,000 edible plant species in the world, there are many such ingredients to choose from.)
However, matching the taste, texture, and nutrition of regular meat products requires more than just matching their amino acid profile. It also depends on factors such as the arrangement of those amino acids, the structure of the final plant-based product (and the methods used to achieve that structure), and the reaction with other components (like sugars and carbohydrates) during cooking. Also, different combinations of plant proteins are needed for different processes, such as the extrusion process for plant-based meat and emulsification (the process of combining liquids to create a stable mixture) for plant-based dairy products.
How can AI systems approximate taste and texture? Often, this is based on human feedback on successive iterations of a product. The models then use this feedback to modify the ingredients and process to better align with the flavors and textures being sought.
Machine learning models also appear to be getting better at predicting textures themselves. Texture is typically analyzed based on the Texture Profile Analysis (TPA) methodology, with a reference conventional meat product (such a boiled chicken breast) for comparison. Researchers have used machine learning to predict ‘hardness’ and ‘chewiness’ using this TPA framework, with some success. Machine learning has also been used in broader food science applications to detect different aromas and sort these into ‘good’ versus ‘bad’ smells. A more sophisticated model could in principle distinguish, for example, ‘meaty’ versus ‘artificial’ smells.
This combined machine learning/human feedback approach is used by the Giuseppe platform created by NotCo, the company behind NotMilk, NotChicken, and the NotBurger. This platform contributed to the non-intuitive inclusion of cabbage and pineapple in the recipe for their dairy-free milks, and the inclusion of broccoli, goji berries, mushrooms and a mystery nut in their chocolate recipe. (This platform incorporates Giuseppe Biagio, a user interface that allows chefs to trial product formulations suggested by the Giuseppe platform, then provide sensorial reviews on those formulations that the model then learns from.) IBM’s Chef Watson system likewise draws on existing recipes to generate thousands of ingredient combination ideas, then uses models of human perception to select the best ones, while a similar model reportedly alerted Eat Just to the suitability of mung beans for their scrambled ‘egg’ products.
Tailored optimization of plants for specific products
AI can be used to map specific plants throughout the entire supply chain, from seed to plate. By monitoring how different crops compare in terms of ease of processing, the taste and nutrition of the final product, and other such metrics, companies can ensure they are using the source ingredients that best suit their requirements. Such requirements vary significantly between products: for example, while some products require arresting flavors, others call for a more neutral taste profile. Similarly, needs for protein content, water solubility, and viscosity will depend on the nature of the target product.
For instance, Equinom’s Manna platform has information about millions of different seeds, which it can use to select the most suitable breeding combinations of seeds for alternative protein companies’ needs. Their development of Sabra hummus can help illustrate this: first, Sabra requested high-quality sesame that could be locally grown. Manna then identified the biochemical traits needed to meet Sabra’s sensory requirements and recommended some optimal breeding combinations from the millions of seeds on their AI platform. They then bred several lines of sesame seeds, each of which Sabra then used to make their hummus. Sabra’s Research & Development team then tested each hummus on metrics such as taste, texture, and nutrition, and finally selected the best one.
Another example is Korean food tech company The PlantEat, which has used AI to develop a new soybean variation called ‘Hayoung’ that is reportedly tastier, more easily digestible, and less allergenic.
Improving reliability of crop production
AI systems can optimize the growth of the crops needed for alternative protein products. For example, in-field sensors can be used to to gather data about the health, size, and quality of crops throughout the growing season, helping producers promptly address real-time risks (such as the spread of disease) and seize future opportunities for improvement (such as altering light levels, spacing, the combination of different kinds of plant, and make-up of soil nutrients). This lends itself particularly well to indoor farming models that allow for closer control of the harvest. A consortium led by Protein Industries Canada is taking this approach for pea and canola crops, to ensure that alternative protein ingredient processors have a more consistent supply of protein to work with.
Optimizing cell culture media
The cost of cell culture media is often regarded as the biggest limiting factor to the large-scale production of affordable cultivated meat (though there are some signs that significant cost reductions should be possible in the near future, even without large technological leaps.) This is therefore an obvious contender when considering the highest-priority applications of AI for cultivated meat production. Fortunately, it also seems to be a problem for which AI is uniquely well-suited.
For animal cells, cell culture media typically consists of around 30 to 50 nutrients (including energy sources, amino acids, vitamins, salts, and metals), growth factors (proteins that play a crucial role in regulating cell growth, proliferation, and differentiation), and other substances that regulate variables such as pressure and osmosis. ‘Undefined’ nutrient sources, such as animal serum and hydrolysates (complex mixtures obtained from breaking down biological materials), are often used because of their growth benefits: their ‘undefined’ nature means that the exact composition and concentration of many of the nutrients they contain are not precisely known, and they can interact with each other in complex and unpredictable ways that may significantly vary depending on the types of cells being cultivated. Identifying which of these ingredients are actually needed, and in what quantities and concentrations, is crucial to improving the media’s yield and affordability. Matters are further complicated by the importance of eliminating ethically problematic animal-derived components from the media. The most significant of these is probably fetal bovine serum (FBS), whose use is still widespread in the industry despite many companies having eliminated FBS from their supply chains.
Given the countless possible combinations of nutrients, growth factors, and other substances, theoretical scientific understanding combined with trial and error approaches can only take one so far. Incorporating machine learning algorithms can make this process radically more efficient by substituting real-world trials with virtual simulations (and also greatly helping researchers to synthesize and apply the sizable existing literature on the topic).
For example, a group of researchers recently used machine learning algorithms to design media with 181% more cells than a common commercial variant with a similar economic cost, while doing so using 38% fewer experiments than the standard experimentation method. This is still a far cry from the radical improvements needed, but it’s at least somewhat promising.
Another group of researchers used predictive modeling by neural networks to predict the global warming potential, cost, and growth rate of different reduced-serum media formulations for a zebrafish cultivated meat production system. (Reducing the amount of serum is beneficial largely because of the serum’s high costs and the difficulty of predicting how it will react with other elements of the cell culture media, as well as the ethical considerations mentioned above.) The model was able to accurately predict those variables with a success rate of 98%.
Image segmentation and classification
Microscopic analysis can help assess the number and orientation of cells in a cell culture, as well as individual cells’ health, behavior, and lineage (i.e., what kind of cell they are). Monitoring these features is crucial for cell culture optimization. This task consists of both image segmentation (i.e., separating dense images of many cells into the constituent cells) and image classification (i.e., identifying the individual features of each cell).
Currently, researchers tend to perform this analysis manually, as the computer systems designed for such analysis (which rely on relatively crude measurements of variables such as cross-sectional area, circularity, and pixel intensity) are comparatively rudimentary and unreliable. In principle, Machine Learning algorithms would be well-suited to this kind of task, learning from labeled data-sets to classify images of cell cultures with increasing accuracy. Such algorithms are already being put to similar use in fields such as pathology, radiology, and medicine.
The current limiting factor here is data. Though many cells have been analyzed in this way, annotated cell images have rarely been uploaded into the public domain. This is often because researchers either haven’t seen the value of doing so, or because they (or their employers) have seen the value, and therefore wish to keep such data confidential. There are some exceptions, such as the public repositories Image Data Resource and TissueNet, but a much greater number of relevant images are still needed.
Growing the necessary cell cultures for cultivated meat is extremely costly and resource-intensive. AI-powered predictive analytics could radically streamline the process by monitoring and controlling parameters such as the temperature and pH levels within bioreactors (in addition to the supply of nutrients and growth factors in the cell culture media, as mentioned above), ensuring that these are optimally tailored to cells’ requirements at different phases of the process. This could bring benefits not only in the form of higher quality, greater speed, fewer inputs, and lower costs, but also a smaller environmental footprint, which will be important both in its own right and to secure stakeholder support.
We built our AI to both fine-tune the cell proliferation process and enhance the nutritional value of our cultivated meat. AI is helping us identify and understand connections between biochemical cues (such as nutrients in our growth media) and the outcomes they elicit (such as cell growth). This information helps us greatly narrow down the hands-on experimentation we need to do in the lab, which gets us to the outcomes we want far quicker, and at far less expense.
Animal Alternative Technologies takes a similar approach. They have apparently used AI to create an entire manufacturing system named the ‘Renaissance Farm’ that will help willing food businesses produce cultivated meat at scale. This allegedly goes beyond products such as mince or snack-size bites to include whole-cut products such as steak, the structure of which is notoriously hard to replicate.
Discovering new cell lines
AI can be used to select, and engineer, the most productive and stable cell lines for cultivated meat production, and to predict the taste and texture of the end product based on the cell lines and their cultivation conditions. (A cell line is a population of cells grown from a single cell or small group of cells and grown in a cell culture.) AI models can gather data on cell lines such as their growth rates, nutritional needs, genetic makeup, and reaction to different cultivation conditions, then identify how these traits correlate with characteristics such as growth and stability. It can then use this information to identify specific genetic modifications that might enhance those characteristics, and test these predictions using simulations of various cultivation conditions.
Examining gene networks
Gene networks are collections of proteins and molecules within a cell that interact with one another, and with other substances in the cell, to ultimately determine the cell’s function (e.g., whether a cell remains a stem cell or develops into a muscle cell or nerve cell). Understanding these networks in greater detail could help address various roadblocks in cellular agriculture, such as ‘the need to maintain proliferation over long periods of time, differentiation potential into preferred tissues, robustness in bioreactor environments, genetic stability, and survival in [...] culture media’. Machine learning methods could help researchers understand how these gene networks operate under different conditions, and therefore help optimize the cell lines and conditions to ensure that they produce the desired outputs.
Discovery of promising strains and ‘designer enzymes’
AI systems can significantly enhance the efficiency of ‘high-throughput screening’, which involves the rapid analysis of a large number of biological samples or compounds, to identify the most effective strains for producing specific proteins. (‘Strains' refer to microorganisms, such as bacteria or yeast, which are selected or engineered for their ability to produce target proteins in large quantities.) Complementing this, AI can also help to identify the most effective feedstock composition for a particular strain: i.e., finding the right mix of nutrients and materials that these microorganisms need to grow and produce proteins efficiently.
In addition to aiding the discovery of promising existing strains, AI-assisted synthetic biology approaches can help researchers to design new ‘designer enzymes’ that function more efficiently than existing ones. AI algorithms can use genetic data to predict which genetic modifications are likely to most increase protein expression levels (i.e., the amount of a specific protein that the microorganism is able to produce) and overall cellular productivity (which encompasses factors such as the microorganism’s health, its efficiency at converting feedstock, its ability to withstand variations in environmental conditions, and its longevity).
For example, Vivici uses precision fermentation to produce dairy proteins that can be used in dairy alternatives. They recently partnered with cell programming company Ginkgo Bioworks, who will use their generative AI platform to provide Vivici with the most promising candidate strains for their purposes. Using these strains, Vivici is expected to launch its ‘nature-equivalent’ whey protein beta-lactoglobulin product in early 2024. The major whey protein in most mammals’ milks (though not humans’), beta-lactoglobulin is a versatile ingredient that can be used to enhance the texture and ‘mouthfeel’ of many different foods and beverages.
Applications of AI for fermentation are greatly helped by broader AI-powered breakthroughs in protein structure research. The main breakthrough has been the creation of AlphaFold, an AI model developed by Google’s DeepMind that can accurately predict a protein’s structure based on the sequence of its amino acids. (The amino acid sequence dictates how the protein will fold into its unique 3D shape, which in turn determines how the protein behaves and interacts within the body. In the pharmaceutical field, for example, this helps researchers to understand the structures of harmful proteins involved with diseases and design new drugs to target those harmful proteins, including by using AI simulations to test out candidate drugs.)
The tech-driven alternative proteins company Shiru is leveraging AlphaFold to identify proteins that can replace traditional ingredients: ‘We use [AlphaFold’s] protein structure models to help us determine protein function, enabling us to identify the best protein ingredients for all food categories based on their inherent chemical and structural identity.’ This underpins their AI model Flourish, which identifies naturally occurring proteins that are likely to meet the specifications for a particular product, uses host microorganisms to produce those proteins, tests those proteins in the laboratory and real-world applications, then brings the successful novel proteins to market.
As with cultivated meat production, AI systems can analyze critical variables in fermentation bioreactors (such as temperature and pH levels) to minimize time and costs while maximizing growth and protein yield. For example, the company Eternal reports using machine learning in this vein to optimize their fermentation processes in order to develop their fungus-based Mycofood. Meanwhile, POW.BIO uses AI-powered process optimization and bioreactor management in their development of a continuous fermentation process that sustains cells at peak productivity while avoiding contamination by unwanted microbes.
A recent paper applied machine learning models to the fermentation of E.coli and found that the model successfully highlighted fermentation processes that were deviating from expectations, which in practice could allow operators to rapidly detect problems in specific bioreactors and either make the necessary adjustments to get the fermentation process back on track, or halt the process to save unnecessary time and costs.
AI models can help distill a huge array of scientific literature and make new connections between studies, greatly helping researchers pinpoint the information they need and reach conclusions about the most promising steps to take. This could increasingly take the form of semi-autonomous AI assistants such as Coscientist, ‘an artificial intelligence system driven by GPT-4 that autonomously designs, plans and performs complex experiments by incorporating large language models empowered by tools such as internet and documentation search, code execution and experimental automation’.
Understanding food chemistry
The relation between flavor of ingredients before and after cooking is poorly understood. AI predictive tools could identify what flavor would emerge after a certain cooking process and provide options for different ingredients and processes to use to attain a specific final flavor profile. This is an application being explored by IBM’s Chef Watson: in addition to its algorithms for selecting the best flavor combinations, it also uses algorithms to identify optimal ingredient proportions and recipe steps.
Predictive maintenance of bioreactors
In addition to optimizing bioprocesses within bioreactors, as mentioned above, AI can also help with equipment maintenance, predicting failures and thereby minimizing wasted time and money.
Combination with 3D printing
There is precedent for alternative proteins being produced using 3D printers, such as the world’s first 3D-printed vegan salmon filet that hit stores last September. There is also precedent for AI being used in conjunction with 3D printing in bioprinting, drugs, gels, biopolymers, and smart materials. Perhaps most relevantly, the two technologies have been used together to print tissues and organs for transplantation, with neural networks predicting the cell viability, mechanical properties and cell proliferation of 3D printed tissues.
Streamlining supply chains
AI technology can help to streamline supply chains, helping companies bring down costs and expand into new areas. Supply chain tasks could include analyzing and predicting market trends, optimizing distribution routes and schedules, and automating inventory management. For example, plant-based meat producer Heura recently commissioned a provider of AI-assisted supply chain optimization software to improve the planning speed and accuracy of their supply chains to help them expand into new regions.
AI is increasingly being used to generate personalized nutritional guidance. This is especially useful for people with specific nutritional needs, such as pregnant women, school children, athletes, and people with various diseases. This could help people to identify which alternative proteins are most likely to meet their health and nutritional needs, especially in a world where alternative proteins can be specifically tailored to meet such needs (for example, by adjusting the amount of protein, fat, or fiber, or fortifying with vitamins and minerals that specific demographics are more likely to lack), potentially with the aid of personal or local 3D printers.
Tailored sales and advertising
AI algorithms can be used to collect detailed information about internet users and tailor advertising to match the interests and behaviors evinced by their online histories. The same product could therefore be presented to one potential consumer based on its health benefits, to another on sustainability grounds, and another on animal welfare grounds, depending on each consumer’s predicted user profile. With organizations increasingly using AI to drive up their business-to-business sales, alternative protein companies could also leverage AI to increase take-up of their products by distributors.
Beyond just emulating the limited textures and flavor profiles of conventional animal products, AI can help establish alternative proteins as a fundamentally different, and superior, product. For example, machine learning algorithms can continually monitor consumer preferences and market trends to provide alternative protein companies with real-time suggestions for novel products. This approach is already being rolled out in the broader food and drinks industry, such as with Tastewise’s TasteGPT model. Largely free from the immensely inefficient biological constraints that face the conventional animal product industry, alternative protein companies would be able to respond to such suggestions at a rapid pace, designing and marketing totally new products from scratch before the zeitgeist shifts.
Lack of substance
AI-washing ‘occurs when companies claim that their offerings involve AI technology, but in reality, they either don’t or their connection to AI is minimal and not core to the product’s functionality.’ Its purpose ‘is to make products or services appear more cutting-edge than they actually are, with the goal of drawing investors’ and customers’ attention.’
This, combined with the general lack of transparency mentioned above, can make it hard to judge when a company is using AI in a meaningful way, and how. AI-washing and unrealistically optimistic claims about what a company will achieve, or has already achieved, using AI, without solid evidence to back it up, could also lead to skepticism about the entire field from investors, governments, and other stakeholders.
Lack of affordable computing power
Handling vast datasets will require a huge amount of computing power. For small companies, this could prove prohibitively expensive. While larger companies typically have more capital to invest in the technology required, the overall financial and time investment may be significantly larger if they need to replace or adapt existing systems throughout their operations. Quantum computing has been noted as a potential game-changer in this regard (though there’s still plenty of uncertainty about when reliable quantum computers will be developed, let alone become affordable to alternative protein start-ups).
Lack of data
All of the possible applications listed in this post rely on having a huge amount of relevant data that the AI systems can learn from. Generally speaking, this data is still extremely limited. This is in part due to the infancy of the alternative proteins field (especially for cultivated meat), and in part due to the lack of data-sharing between the companies working in this space (which is detailed in the section below). There are also relatively few research papers in the public domain about machine learning and alternative proteins, especially regarding cellular agriculture. As such, many companies currently need to rely to a great extent on their own sets of data that are collected using costly, time-consuming, often coarse-grained methods. Fortunately, as outlined in the following section, there are various organizations working to address this issue.
Lack of collaboration
Of course, some of the earlier, larger companies in this space (like Impossible Foods and Beyond Meat) do now have a lot of data to work with. The issue is that most of the funding in the cultivated meat space is in the form of private investments in start-up companies, rather than public funding for open-source academic work. This leads to a lot of siloing and a relative lack of data sharing. Some organizations are working to address this problem, such as New Harvest (including through their OpenCellAg Repository and Open Source Bioreactor) and Good Food Institute (including through their cell line repository and open-access tools for seafood data). There are also some examples of companies banding together to share knowledge and expertise; for example, Protein Industries Canada has fostered the creation of a partnership of four Canadian companies to mutually develop AI technology for farmers and ingredient processors, and the Cultivated Meat Modeling Consortium facilitates collaboration between organizations on the development of computational modeling approaches to optimize the development of cultivated meat.
In the plant-based space, Protein Industries Canada has also partnered with the Alberta Machine Intelligence Institute to launch the Data Readiness Improvement Program, which aims to help plant-based protein companies use AI effectively. GreenProtein is also seeking to address this issue by pooling, and standardizing, anonymized extrusion data from willing research institutes so that they can benefit from each others’ work: reportedly the first attempt at carrying out such an exercise, and one that could be replicated across the alternative protein industry.
Though AI’s actual uses in the alternative proteins space so far appear pretty small-scale, these could become much more game-changing as the technology progresses. If we see improved data-sharing efforts and robust support from governments, investors, and other stakeholders – though this is far from a given, especially with regards to cellular agriculture – AI could eventually lift the alternative protein industry into spaces inaccessible to conventional animal farming. And of course, the opportunities listed here are just a subset of the ones that we’re currently able to identify; if, as most experts seem to agree, there’s at least a 50% chance of transformative AI systems being developed in the next 50 years, that could open up fundamentally different opportunities that are currently impossible to predict.
All in all, while there are a whole lot of ways that future AI advances could exacerbate animals’ exploitation and suffering, this is at least one area where there appears to be considerable cause for hope.
For context, the Good Food Institute (GFI) describes plant-based protein production as follows:
‘The general method used to produce plant-based meat involves three primary steps. First, we grow crops as a source of raw materials. Second, we process these crops to get rid of the parts of the plants we don’t want. At this stage, we end up with the proteins, fats, and fiber ingredients that will become our plant-based meat product. Finally, we put together the desired mixture of ingredients. This ingredient mixture then goes through a manufacturing process to create the muscle-like texture needed for meat.’
Mentioned 1 hour, 21 minutes into the video.
For context, the Good Food Institute (GFI) describes cultivated meat production as follows:
‘The manufacturing process begins with acquiring and banking stem cells from an animal. These cells are then grown in bioreactors (known colloquially as cultivators) at high densities and volumes. Similar to what happens inside an animal’s body, the cells are fed an oxygen-rich cell culture medium made up of basic nutrients such as amino acids, glucose, vitamins, and inorganic salts, and supplemented with growth factors and other proteins.
Changes in the medium composition, often in tandem with cues from a scaffolding structure, trigger immature cells to differentiate into the skeletal muscle, fat, and connective tissues that make up meat. The differentiated cells are then harvested, prepared, and packaged into final products. This process is expected to take between 2-8 weeks, depending on what kind of meat is being cultivated. Some companies are pursuing a similar strategy to create milk and other animal products.’
Though it's worth noting that while an estimated 2 million calves are still used for FBS each year, most of this FBS is for use in biomedical research, rather than cellular agriculture.
See 21 minutes into the video.
See 50 minutes into the video.
For context, the Good Food Institute (GFI) describes the protein fermentation process as follows:
‘Traditional fermentation uses intact live microorganisms to modulate and process plant-derived ingredients. [...] Examples are using the fungus Rhizopus to ferment soybeans into tempeh, as well as using various lactic acid bacteria to produce cheese and yogurt.
Biomass fermentation leverages the fast growth and high protein content of many microorganisms to efficiently produce large quantities of protein. The microbial biomass itself can serve as an ingredient, with the cells intact or minimally processed — for example, the cells can be broken open to improve digestibility or enrich for even higher protein content. [...] Examples of biomass fermentation are Quorn’s and Meati’s use of filamentous fungi as the base for their products.
Precision fermentation uses microbial hosts as “cell factories” for producing specific functional ingredients. These ingredients typically require greater purity than the primary protein ingredients and are incorporated at much lower levels. [...] Precision fermentation can produce enzymes, flavoring agents, vitamins, natural pigments, and fats. Examples include Perfect Day’s dairy proteins, Clara Foods’ egg proteins, and Impossible Foods’ heme protein.’
Though based on feedback from some alternative protein researchers, it’s debatable whether this is actually a major limitation in practice.