Open-source AI lab aims to close taste and texture gaps in sustainable alt-proteins

US-based Food System Innovations (FSI) has launched its Food Intelligence Lab to build an open-source AI infrastructure that can accelerate more sustainable protein development. 

The initiative responds to a major challenge in the food system: while AI systems are increasingly used to speed up more sustainable F&B NPD, they can lack the integrated data needed to reliably predict key consumer outcomes like taste and texture. This can limit consumer adoption of these products.

The Food Intelligence Lab aims to solve this gap by using AI to improve product formulation and sensory prediction through large-scale datasets, including sensory data and other instrumental measurements. The information will help food scientists better predict consumer responses and formulate products closer to animal-based benchmarks in taste, mouthfeel, and functionality.

Reducing dependence on animal protein is important for decarbonizing the food system, as livestock supply chains contribute heavily to global climate emissions, such as methane (44%) and nitrous oxide (53%), according to FSI.

FSI is a program from the NGO Humane America Animal Foundation. The program has received a US$2 million grant from the Bezos Earth Fund for this new lab. The philanthropic impact platform works to support the development of sustainable food systems by funding, creating, and scaling science-based solutions, especially around sustainable proteins and AI-powered food innovation.

With the Food Intelligence Lab, the NGO aims to shorten commercialization timelines for better-tasting, more sustainable products using AI.

Combining sensory panels and AI

Anna Thomas, director of Machine Learning at the Food Intelligence Lab and a computer scientist at Stanford University, US, tells Food Ingredients First that general-purpose AI models cannot reliably predict the human sensory experience of food.

Anna Thomas: AI can free food scientists to focus on the big strategic decisions such as the taste and cost during the design process.“Drawing upon multiple datasets, we are building models that could approximate similarity to the target animal product.”

The lab will use sensory data from non-profit research initiative Nectar, which will help FSI train AI tools to make better-tasting products faster.

“In addition to Nectar sensory panel data, our models will use instrumental measurements of sustainable protein products, such as TPA (texture profile analysis), pH, and shear tests, as well as molecular composition data, to predict sensory properties,” Thomas says.

AI optimization for plant-based dairy sensory performance

Thomas explains that the initiative’s key insight is that food design is an “optimization problem” — maximizing alignment with consumer preferences, subject to constraints on cost and nutrition.

“We have designed and implemented algorithms to efficiently solve this optimization problem. On two plant-based dairy categories in initial tests, we found that augmenting human expertise with Bayesian optimization algorithms improved sensory satisfaction, as assessed by a trained human panel, by 29% and 26% in one week,” she tells us.

This work was done in collaboration with AI-powered formulation company Proxy Foods AI. The AI-optimized yogurt matched the animal-based benchmark on consistency, creaminess, and tanginess, showing how AI can reduce trial-and-error and speed up commercialization.

“Expert-Guided Bayesian Optimization (EGBO) is an algorithm for black-box function optimization in which a domain expert, for example, a human or LLM, selects a small subset of variables for Bayesian optimization and may expand it over time,” Thomas explains.

Boosting F&B sensory analysis

The Food Intelligence Lab is also developing TasteBench, a public benchmark and competition for predicting sensory similarity between sustainable protein products and animal-based products.

Early results show AI models can perform competitively with human panelists, suggesting promising potential for reducing the time and cost of sensory testing.

AI can accelerate NPD, but sustainable protein needs better data infrastructure to predict key consumer aspects like taste and texture.However, as AI increasingly complements R&D teams, Thomas emphasizes that sensory panels will remain critical to assess consumer liking for products. However, “specialized AI models could efficiently screen formulations for similarity to animal-based benchmarks.”

This will also reduce the number of human evaluations required, in turn removing a key barrier to success for sustainable protein start-ups with limited resources, she adds.

“Our initial experiments show that our methods exceed the performance of a professional human food scientist working without AI tools and given the same time budget by 17%.”

She describes the team’s Bayesian optimization methods as “experimentation algorithms” that can help companies reduce the number of trials needed to achieve sensory targets.

“We see these algorithms as a partner to food scientists, so that they can focus on strategic decisions, such as the high-level objective and constraints (cost, nutrition) of the formulation design process,” says Thomas.

What’s next for Food System Innovations?

FSI will present its EGBO and TasteBench papers at the upcoming AI for Scientific Discovery Workshop at the 2026 International Conference on Machine Learning in Seoul, South Korea, next month.

The team is now working to increase the understanding of how its tool performs relative to traditional workflows in follow-up experiments.

“We’re excited about these initial results and what it means for food scientists and R&D teams, as well as the future of sustainable protein development,” Thomas concludes.