Can computers make better plant-based foods? July/August 2021

By Rebecca Guenard

In This Section

July/August 2021

  • Instead of replacing the protein source in a formulation and then masking the off-flavors of plant-based ingredients, companies are determining what fundamentally makes a product taste good.
  • These companies are collecting vast amounts of data on everything from subjective human responses to precise analytical measurements.
  • They plan to implement statistical modeling and machine learning to determine how to formulate for great tasting plant-based foods.

Until recently, drug development meant addressing the mechanics of a disease. Cancer treatments, for example, were primarily geared toward slowing the rapid division of tumor cells. With the establishment of high-throughput analysis, researchers began approaching drug discovery differently. They turned to producing large volumes of complex and diverse data sets to identify drug targets within individual genes.

As they gathered more genetic data, researchers improved statistical models to predict the location of genome mutations and worked out how to synthesize molecular remedies to correct those errors. They wrote algorithms to train computers how to make such predictions faster. Then they gathered more data to fine-tune their algorithms. Data science is now a core discipline in pharmaceutical research, and food companies are beginning to adapt this approach to their industry, particularly in solving the problem of off-flavors in plant-based foods (Fig 1).

“There is a rational development process to determine what works and what does not to get us closer to our goal,” says Rick Gerkin, associate professor at the University of Arizona in Tempe. “Our goal is ultimately making something that is appealing to a consumer, is healthy and is made from plants.”

Gerkin collaborates with a Berkeley, California start-up called Climax Foods that is using data science to guide the formulation of plant-based creations for seven of the most popular dairy cheeses. The company is collecting data along every step of their cheese production process and comparing it with the sensory perceptions of tasters who comment on the final product. As their data sets grow, they will use them to build statistical models for the formulation of the ideal plant-based cheese that mimics the taste, texture, and nutrition of a dairy cheese. “There is a list of things that are unavailable to us when making a plant-based, all natural product,” says Gerkin. “Under those constraints we have to find new formulations.”

Plant Protein Challenges
FIG. 1. Challenges facing formulators who use plant-based proteins. Source: Nasradadi, M.N., et al., Food Hydrocoll., 118, 106789, 2021

Finding the right protein

Boston-based food ingredient company, Motif Foodworks, is engaged in a similar pursuit. Dilek Uzunalioglu, Motif's head of product applications, says her company is looking for the fundamental factors that make plant-based foods “cravable”. Motif is focused on overcoming the off-flavors and unpleasant textures consumers commonly describe as part of the plant-based food experience. Uzunalioglu says her company is working on a variety of plant-based applications, such as meat and dairy alternatives, but also new forms of plant-based products.

Through collaborations with different experts, Motif is gaining a better understanding of how to select the optimal protein functionality to contribute to specific properties in a finished product. Motif partners with another Boston company called Ginko Bioworks that retrieves genetic information from an extensive library of sequences and uses that information to design new biological products. After screening over 300 proteins in their database, Ginko designs yeast strains that, through fermentation, can make the proteins Motif identifies as potential product ingredients.

This approach helps them select the plant proteins that are best suited for a specific food application and leapfrogs the myriad physical and chemical modifications needed to make some plant proteins more palatable for food applications. “It is really about gathering the insights and then looking for gaps where we can design proteins or any other ingredients to remove the gap,” says Uzunalioglu. “Scientific biology is one of the tools we use, but it is not the only one.” Adding genetic tools to existing analytical techniques, like rheology, helps pinpoint proteins destined to function in food applications while also achieving consumer satisfaction.

Following the drug discovery playbook, Motif is applying high-throughput analysis to characterize a large number of samples at once and screen for the traits that interest them. Automated preparation in multi-well plates followed by simultaneous sample analysis allows pharmaceutical companies to conduct a fast screening of hundreds of potential new product formulations (see Inform, March 2020). Drug developers can calculate and optimize experimental conditions, such as buffers, surfactants, sugars, storage temperature, and mechanical stress, much faster using these methods.

Uzunalioglu says, through a collaboration with researchers at the University of Massachusetts, Motif has developed two-gram samples of protein assays for use in high-throughput measurements. The capability allows them to quickly determine the foaming, gelling, and emulsification properties of any protein they produce through fermentation.

The data gathered from these measured properties can then be associated with the traits that consumers describe in a food, such as moistness and chewiness. “For moistness, we measure water holding capacity and oil holding capacity,” says Uzunalioglu. “For chewiness, we measure water holding capacity and gelling and foaming, things like that.” She says that Motif uses a combination of trained tasters and the general public to gather information on how consumers experience different ingredients.

Motif’s strategy of combining their analytical results with consumer satisfaction is indicative of the reality that, ultimately, human perception decides a product’s success. Formulators know that fragrances and mouthfeel play a crucial role in perception. Many hope to equate human experience with specific molecules, but measuring perception is complicated.

Interpreting sensory perceptions

Unlike with vision, sound, or touch, which researchers can track from an input to a neurological signal in humans, odor perception is harder to pin down. Instead of proceeding directly to the thalamus like other sensory systems, scent signals first travel to brain regions that process emotions and memory. Smell has a profound impact on how we experience food. For instance, those who lost their sense of smell due to the SARS-CoV-2 virus reported that the taste and texture of food had changed (

One way to improve consumer experience with plant-based foods is to identify the source of unpleasantness and avoid ingredients with a similar molecular structure. But that turns out to be a daunting challenge. Vision operates off of four receptors; scent uses 400. Scientists have not yet deciphered the language our olfactory system uses to decode the estimated trillions of different smells the brain understands. However, they have identified enough correlations between physiochemical features and perception to believe the process is structured and not subjective. Therefore, they continue to search for the fundamental molecular features that combine to create a smell.

Earlier this year, University of California, Riverside researchers were able to increase the number of predictable odorant chemicals perceived by the human olfactory system using computational analysis. Genetics, culture, and lived experience all contribute to a sense of smell, but enough similarities in descriptions of smell for the same chemical exist to imply a common physiochemical basis for their perception among humans. The UC researchers set out to train computers to identify the small group of physiochemical traits that were determined in earlier experiments and, using machine learning, see if the computer could then find other traits.

The team used databases from previous experiments that contained odor character profiles—describing smells such as cooked meat, cooked vegetables, or green vegetables— along with the physiochemical description of the compounds. They wrote algorithms to rank chemical features in terms of their contribution to those descriptors. Then they applied machine-learning models to predict the smell humans were most likely to perceive from specific features.

The group reported that using their statistical models, they were able to extend data from what they referred to as “low-throughput, high-cost human studies” and explore new areas of chemical perception. From a library of 440,000 chemicals, the models were able to predict the descriptions of the smells that human participants had given in previous studies. In addition, the researchers introduced 68 million word combinations that describe smells into the model and identified new chemicals that smell like each descriptor (Fig. 2). This chemical discovery aspect of the research could be lucrative to food and ingredient manufacturers. Machine learning may help unveil previously unknown compounds that remind humans of the smell of freshly cooked meat, for example, even for a meatless product.

 How to build perceptual descriptor networks for physiochemical features Diagram
FIG. 2. How to build perceptual descriptor networks for physiochemical features. Source: Kowalewski, J., et al., Chemical Senses, 46, 1-13, 2021.

Managing all the data

At the University of Arizona, Gerkin conducts research similar to the work performed by the UC team (though he uses statistical models to interpret neurological activity in response to odor). He has written that the predictive modeling used by the UC team is a valuable way to find new flavor compounds. His collaboration with Climax Foods could signify that the start-up is interested in making such an attempt. However, when asked, Gerkin said, for now the partnership is focused on “trying to figure out what is the very best way to apply the cutting edge of the academic work into industry.” That means gathering as much data as possible on the fermentation process of cheese.

“You want to connect the variable that you can control in your process to something you can measure—either by a panel of people evaluating the cheese telling you what they like, or by things you can measure quantitatively,” says Gerkin. Like Motif, they are gathering a variety of data from rheology to GCMS and LCMS. He emphasizes the importance of then knowing what to do with all that data to make it valuable in formulations. Data science is most useful, he says, when you can visualize how all the pieces of information fit together, when you know the uncertainty of all your measurements, and when you can effectively program computers to retrieve valuable predictions from everything you have gathered.

Uzunalioglu says Motif is in the process of data collection that will eventually be used to build predictive algorithms. She confirmed that is the end goal, but for now they are collecting as wide a variety of data as possible. Along with biotechnology and tasting panels, Motif is gathering research on psychology and oral processing (see “Improving mouthfeel data”). Uzunalioglu says, incorporating a breadth of information is part of her company’s development philosophy.

During the pandemic, sales of plant-based foods increased due to meat shortages that occurred when outbreaks hit processing plants. Matt Roszell, head of Motif’s marketing communications, says that was an opportunity to gain new customers in the plant-based space, but many did not adopt the products permanently because they did not love their experience. “Plant-based foods are never going to be more sustainable or better for people’s health if they just do not eat them,” he says.

To get to that next level of customer acceptance, food scientists need more options than covering up odd flavors or textures with added sugar and salt to hide the earthiness of plant proteins. And, they want to know if anything other than gums and hydrocolloids can replicate the performance of animal fats.

Uzunalioglu and Roszell say they expect that the arsenal of data and computational power Motif has applied to determining what piques consumers’ taste buds will lead to an indistinguishable eating experience between plant-based foods and native meats within one or two years.

Improving mouthfeel data

Plant-based food manufacturers often want to balance high protein in formulation with low fat and low sugar. Unfortunately, this combination typically leads to a product that consumers describe as chalky or gritty. Over the past decade, researchers have been investigating the components of mouthfeel hoping they might reveal a clue to avoiding unpleasant textures, but a recent review of saliva-protein interaction experiments found more experiments are needed.

The sensory perception and after taste a person experiences when eating is related to saliva’s interaction with the food. The bio-lubricant coats mouth surfaces and assists with food processing. Saliva is mostly water, but contains proteins and ionic compounds that influence our perception of texture, especially when interacting with other proteins.

A team of researchers at the University of Leads, Leads, UK, performed what they believe is the first systematic review of protein-saliva studies confirming that these interactions are dominated by electrostatic charges strongly influenced by pH. However, the findings were exclusively for dairy proteins and, according to the researchers, the literature did not yet contain an analysis of plant protein–saliva interaction to predict mouthfeel perception.

Furthermore, the systematic review indicates that the current body of work on protein-saliva interactions is flawed. The review authors suggest that some of the methodology does not reflect a physiologically relevant saliva-to-protein ratio and is, therefore, an inaccurate simulation. In addition, there was a lack of similarity across studies making comparisons difficult. They suggest that in the future, standardization be applied to protein-saliva experiments to improve research quality and enable comparisons.

As data scientists become confident in the application of predictive models, mouthfeel data is likely one of the components they will be interested in incorporating in their statistics. First, they must be certain that data is relevant and reliable.

About the Author

Rebecca Guenard is the associate editor of Inform at AOCS. She can be contacted at


Modification approaches of plant-based proteins to improve their techno-functionality and use in food products, Nasradadi, M.N., et al., Food Hydrocoll 118: 106789, 2021.

Protein–saliva interactions: a systematic review, Brown, F.N., et al., Food Funct. 12: 3324–3351, 2021.

A System-Wide Understanding of the Human Olfactory Percept Chemical Space, Kowalewski, J., B. Huynh, and A. Ray, Chemical Senses 46: 1–13, 2021.

Parsing sage and rosemary in time: the machine learning race to crack olfactory perception, Gerkin, R.C., Chemical Senses 46: 1–5, 2021.

Attend an AOCS webinar
Sign up for the INFORM TOC