Machine learning is making fruits and vegetables more delicious
There’s a reason so much of the produce sold in the grocery store often tastes like cardboard.
Actually, there are several reasons. Most of them stem from the fact that tastiness is far down on the list of what the food industry encourages plant breeders to prioritize when developing new varieties — called “cultivars” — of produce.
When they do want to focus on taste, breeders don't have good tools for quickly sampling the fruit from thousands of cultivars. In a surprising new paper, researchers at the University of Florida describe a new method for "tasting" produce based on its chemical profile.
They also stumbled on a big surprise. For more than a century, breeders have focused on sweetness and sourness when they tried to develop tastier cultivars. The new research shows that the tried-and-true approach ignores roughly half of what makes a tasty fruit or veggie so delicious.
Agricultural scientist Patrico Muñoz, one of the paper’s co-authors, told IE that his team determined that in blueberries, for example, “only 40 percent [of how well people like a fruit] is explained by sugar and acid.” The rest is explained by chemicals called volatile organic compounds that we perceive with receptors in our noses, not our mouths.
That find — and the method they used to get there — could change the future of agriculture.
Untangling the connection between chemicals and flavor
The researchers behind this study focused on dozens of varieties of tomatoes and blueberries, including commercial cultivars sold in supermarkets, heirloom varieties more likely to be found at farmers markets and farm-to-table restaurants, and newly developed strains that recently graduated from breeding programs.
They had two types of data for each cultivar. First, a chemical profile detailing what and in what amount dozens of chemical compounds are found in its fruit. Second, they had results from consumer panels in which hundreds of real people had rated each tomato or blueberry cultivar on measures like how sweet it was and how much they liked it overall.
Combining these two datasets allowed the researchers to tease out, say, how much the taste of the different tomato cultivars was influenced by the ratio of 2-Methylbutyl acetate to 1-nitro-2-phenylethane.
Some of the cultivars in the study are GMOs, but even wild fruits are chock full of these kinds of chemicals. One reason is that plants — which generally spend their lives firmly rooted wherever their seeds happen to land — have figured out how to make and use compounds to control the world around them in order to survive and reproduce.
“In fruit, plants make [these types of chemicals] to attract animals that are going to [eat the fruit and] spread the seeds,” Muñoz says.
Yes, basically every food contains tons of chemicals most people can’t pronounce.
When all of this data was assembled, the researchers used machine learning to build models aimed at explaining how the chemical makeup of a fruit is related to what that fruit tastes like.
For some fruits, this relationship is relatively straightforward.
“In raspberries, there is one single compound that’s the raspberry flavor,” Nuõz says. But other fruits and vegetables — including blueberries and tomatoes — flavor is the product of a complex interplay among dozens and dozens of chemicals.
Their model revealed how much the various chemical components correlated with the human tasters' ratings for each of the varieties. Surprisingly, the sugars and acids in the fruits only accounted for roughly half the variation in the tasters' preferences from one variety to the next.
"That means that for the last 100 years, we have made progress in less than half of [the variables] that explain [preferences]," Muñoz says.
Breeders will be able to "taste" a lot more produce
Plant breeders don’t have direct control over which chemicals are contained in a certain cultivar’s fruit. Instead, they influence the genes, which code for the metabolic pathways that manufacture the compounds that ultimately determine what a fruit tastes like. Even with modern technologies, it’s an unwieldy process that’s typically done at a large scale.
Marcio Resende, another co-author on the study, tells IE “breeding still relies heavily on field experimentation,” just as it has for centuries.
What separates today’s breeders from their forerunners is the technologies they use to measure traits that tell them “which varieties go on to the next stage in the funnel of breeding,” Resende says.
They use tools like drones and autonomous robots to “quantify whatever traits are important” in a process called high-throughput phenotyping, he says. Existing technologies have been up to the task of measuring traits like how much fruit a cultivar produces and what color those fruit are.
Measuring taste has proven far more difficult because there were really just two options: the breeder could sample fruit themself or they could assemble a panel of testers. For breeders testing thousands of cultivars, that choice is a big tradeoff. Sampling is highly subjective, and systematic testing with people is expensive.
“If you assemble a traditional consumer sensory panel and bring 100 people to a room... you can't feed 1,000 varieties in the same day,” Resende says.
This new research is “a proof of concept that shows we can now build models to do the same thing” by measuring chemicals, he says.
This kind of research will never produce a perfect version of any particular fruit. For one thing, flavor preferences vary across time and culture. Since machine learning models can only make predictions based on the data they were trained on. The panels in Resende and Muñoz’s data were done in the United States, so there’s a good chance consumers in other markets would have different preferences.
Although they are staple foods in cuisines globally, many commercial fruit varieties have become progressively less flavorful over time. Due to the cost and difficulty associated with flavor phenotyping, breeding programs have long been challenged in selecting for this complex trait. To address this issue, we leveraged targeted metabolomics of diverse tomato and blueberry accessions and their corresponding consumer panel ratings to create statistical and machine learning models that can predict sensory perceptions of fruit flavor. Using these models, a breeding program can assess flavor ratings for a large number of genotypes, previously limited by the low throughput of consumer sensory panels. The ability to predict consumer ratings of liking, sweet, sour, umami, and flavor intensity was evaluated by a 10-fold cross-validation, and the accuracies of 18 different models were assessed. The prediction accuracies were high for most attributes and ranged from 0.87 for sourness intensity in blueberry using XGBoost to 0.46 for overall liking in tomato using linear regression. Further, the best performing models were used to infer the flavor compounds (sugars, acids, and volatiles) that contribute most to each flavor attribute. We found that the variance decomposition of overall liking score estimates that 42% and 56% of the variance was explained by volatile organic compounds in tomato and blueberry, respectively. We expect that these models will enable an earlier incorporation of flavor as breeding targets and encourage selection and release of more flavorful fruit varieties.