Food System Innovations launched the Food Intelligence Lab to build open-source AI tools for sustainable protein development, backed by a $2 million grant from the Bezos Earth Fund. The initiative targets the primary barrier to alternative protein adoption-taste and texture-by replacing slow trial-and-error lab work with predictive machine learning models.
The taste barrier in alternative proteins
Livestock supply chains account for 14.5% of global greenhouse gas emissions, making sustainable proteins critical for decarbonization. However, U.S. consumers have been slow to adopt alternatives. A 2026 industry report found that only 33% of participants liked dairy-free products, roughly half the approval rate of animal-based benchmarks.
Companies face long development cycles and fragmented data when trying to improve sensory performance. The Food Intelligence Lab aims to fix this by creating large-scale datasets that include human sensory data and instrumental measurements like texture profile analysis and pH levels.
Optimizing formulations with AI
The lab is already testing its infrastructure through a partnership with Proxy Foods AI. The teams co-developed an optimization system called Expert-Guided Bayesian Optimization (EGBO).
In a recent test, EGBO improved the sensory performance of a plant-based Greek-style yogurt by 29% across 10 formulation iterations over five days. The AI-optimized yogurt matched the animal-based benchmark for consistency, creaminess, and tanginess.
This approach to AI for Product Development allows teams to iterate faster and reduce reliance on physical lab testing.
"Food scientists shouldn't have to spend months on trial-and-error to get texture, mouthfeel, flavor, and aftertaste right," said Panos Kostopoulos, founder and CEO of Proxy Foods AI. He added that the goal is to eventually move 90% of formulation work in silico, reserving wet labs for final validation.
Benchmarking sensory predictions
Beyond formulation, the lab is addressing the high cost of running human sensory panels. Researchers developed TasteBench, a public benchmark and competition hosted on Kaggle, to evaluate how well AI models predict sensory similarity to target animal products.
The lab evaluated existing foundation models to advance AI for Science & Research, comparing their predictions against baseline methods. The best-performing AI model achieved 66.1% pairwise ranking accuracy, slightly outperforming the median human panelist at 65.0%.
Anna Thomas, director of machine learning at the lab, said food science currently lacks the shared data infrastructure found in drug and materials discovery.
"We believe AI can be a powerful accelerator for climate and nature solutions when it is paired with the right data, collaboration, and real-world applications, moving promising ideas into impact," said Dr. Amen Ra Mashariki, director of AI at the Bezos Earth Fund.
Why this matters for product development
Product development teams in the food sector can use these open-source models to predict texture and flavor outcomes before mixing physical ingredients. Access to standardized datasets like TasteBench removes the need to build proprietary sensory models from scratch, cutting months off the formulation cycle for alternative proteins.
Your membership also unlocks:
