Programs
The Food Intelligence Lab builds shared datasets, benchmarks, algorithms, and practical tools that help researchers and companies develop better products faster.
AI is already transforming drug and materials discovery. Sustainable foods are next, but the necessary infrastructure that enabled AlphaFold, GNoME, and other impactful models does not yet exist for food.
The sustainable protein field lacks a dedicated, mission-driven, and open-source effort to apply and improve AI tools specifically for sustainable proteins. Individual companies cannot justify the investment to build these tools alone, and academic efforts remain slow and fragmented. What's missing is the equivalent of what PyTorch, ImageNet, and the Protein Data Bank did for their respective fields: shared tools and reference datasets that let everyone build and compare models.
By building public, shared infrastructure, we can accelerate discovery. Specifically, we work to establish common benchmarks against which the entire community can measure progress; predict sensory outcomes before physical prototyping; optimize formulations across cost, taste, and nutrition simultaneously; and reduce R&D cycle time.
We do this in three ways:
Data and Benchmarks
We build and curate large-scale datasets (e.g., sensory data from FSI’s program NECTAR, instrumental measurements of sustainable protein products, etc.) and establish public benchmarks to rigorously measure progress on key tasks such as sensory prediction and formulation design.
Algorithms
We develop algorithms to accelerate key prediction and design tasks for sustainable protein product design, drawing on and extending advances in other areas of AI for science (e.g., drug and materials discovery; olfaction). We open-source these algorithms and publish our findings.
Deployment
We work directly with companies, nonprofits, and academic labs to translate our models into real-world applications.
Flagship Work
The Food Intelligence Lab’s current flagship project is a Bezos Earth Fund-supported effort to develop open-source AI tools for sustainable protein design. The project builds on FSI’s sensory data from NECTAR to help predict sensory performance, improve formulations, and reduce the trial-and-error inefficiencies in product development.
The core goal is to help sustainable protein teams answer practical questions faster: Which products are likely to perform well with consumers? Which ingredients or formulations are promising starting points? What should a food scientist try next after receiving sensory feedback?
This work includes models for sensory prediction and formulation design, public benchmarks, scientific publications, and deployment pathways that allow tools to be tested with real users while preserving FSI’s public-interest commitment to open methods and field-wide learning.
TasteBench
Benchmarks and competitions, e.g. ImageNet, MMLU, GSM8K, etc. have driven significant progress in machine learning.
To formalize sensory prediction as a machine learning task, we have recently launched the TasteBench competition on Kaggle, with both food-level and molecular-level tasks.
Recent Publications and Preprints
The Food Intelligence Lab is committed to building the scientific foundations of AI for sustainable proteins through open, peer-reviewed research.
TasteBench: multimodal benchmark for sensory prediction, from molecules to sustainable foods. AI4Science@ICML, 2026.
Expert-guided Bayesian optimization for sustainable protein design. AI4Science@ICML, 2026.
Artificial intelligence for food innovation. Nature Food, 2026.
What can large language models do for sustainable food? ICML, 2025.
Leadership
The program is led by Anna Thomas, supported by Sohum Patnaik, and guided by an Advisory Board spanning research, engineering, and product development.
Get Involved
Contact us to learn more about the Food Intelligence Lab.
“The biggest AI breakthroughs were built on shared infrastructure: PyTorch, ImageNet, the Protein Data Bank. Sustainable protein still lacks that foundation. We’re building open tools and benchmarks for sustainable proteins, so the whole field can move faster than any one company could on its own.”
Past AI/ML Grants
FSI's previous AI/ML grantmaking work, led by Bill Meyer, supported open, field-building projects designed to create reusable datasets, benchmarks, and computational tools for the sustainable protein field.
Creating a protein thermostability dataset to accelerate cultured meat and seafood development
Awarded to Dr. Breanna Duffy's team at New Harvest, this project created an open-source dataset on growth factor thermostability. By combining public data, experimental melting-temperature measurements, and computational protein features, the dataset is designed to support growth factor optimization, culture media evaluation, and AI-enabled workflows for cultivated meat and seafood development.
Tasting stiffness: An open-source dataset for the mechanical features of alt protein products
Awarded to Dr. Ellen Kuhl's team at Stanford University, this project created an open-source dataset and computational framework for measuring the mechanical behavior of alternative protein products. By connecting instrumental texture measurements with human sensory texture surveys, the work gives researchers and companies better tools for understanding stiffness, mouthfeel, and texture in sustainable protein products.
Advisory Board
