The framework, called AlterProtX, is described in the open-access journal npj Science of Food by Eunyeong Lee and Hee Yang of the Department of Food and Nutrition at Kookmin University. Its central insight is that meaningful comparison of protein sources requires bridging two levels of biological organization that are usually studied separately. On one level, the behavior of individual protein molecules—their thermal stability, their surface properties, their tendency to interact with neighbors—determines how they respond to the heat, shear, and pH shifts of food processing. On another level, what matters for a burger patty or a whipped topping is the statistical distribution of those molecular behaviors across an entire proteome. AlterProtX was built to connect the two.

At the molecular level, the system uses AI-based prediction to estimate protein thermal stability, a property with direct processing relevance. Thermal stability governs how a protein unfolds when heated, and unfolding is the first step in the aggregation and network formation that produce gelation, the process by which liquid mixtures set into solid-like structures such as yogurts, tofu, and processed meats. Proteins that denature too readily or too reluctantly can both cause problems in formulation, so having a predicted stability value for every protein in a candidate organism gives formulators a molecular handle on how that ingredient will behave in a heated process line.

Thermal stability alone, however, cannot capture the intermolecular interactions that give protein mixtures their functional character. To address this, AlterProtX integrates six additional attributes describing how proteins interact with one another. These intermolecular interaction features, combined with the predicted stability values, are then aggregated not as simple averages but as distributions across the proteome. This distribution-based representation is the key technical move: instead of collapsing a proteome to a single number, it preserves the heterogeneity of the protein mixture, capturing the fact that two proteomes with similar mean properties might have very different spreads of behaviors—and therefore very different functional profiles in food.

With these multiscale representations in hand, the researchers could perform something that has been largely impractical until now: a mechanistic, proteome-wide comparison between animal and non-animal proteomes. By examining where the distributions of molecular features overlap and where they diverge, the framework reveals which molecular characteristics underlie functional similarity between, say, a muscle proteome and a legume proteome, and which differences explain why certain plant ingredients fail to replicate the texture and performance of animal proteins. This kind of interpretable comparison turns the search for alternatives from a trial-and-error exercise into a hypothesis-generating science, pointing formulators toward the specific molecular gaps they need to close.

The framework also confronts two practical filters that any candidate protein source must pass before it can reach the market. The first is allergenic potential. A proteome that functionally mimics egg or milk protein is of little commercial value if it carries a heavy load of known or predicted allergens, so AlterProtX incorporates allergenicity assessment into its evaluation pipeline. The second is nutritional adequacy, since a sustainable replacement must deliver the amino acids and protein quality that consumers expect from the foods it replaces. By integrating these two dimensions alongside the functional predictions, the system supports early-stage prioritization: candidate sources that fail on safety or nutrition can be screened out before expensive experimental characterization begins, while promising candidates rise to the top of the queue for laboratory testing.

The interpretability of the platform deserves particular emphasis, because it distinguishes AlterProtX from the black-box models that dominate much of modern protein AI. Most AI-based protein tools, including the structure-prediction systems that have transformed structural biology, focus on individual proteins in isolation. They can tell you what one molecule looks like or how one chain folds, but they say little about how a heterogeneous mixture of thousands of proteins will behave when blended, heated, and sheared in a food matrix. By building representations at the proteome level and grounding them in mechanistically meaningful features—stability and interaction attributes—AlterProtX produces outputs that food scientists can reason about, interrogate, and act upon, rather than opaque scores they must simply trust.

The implications for the food industry could be substantial. The global search for sustainable protein alternatives currently spans an enormous diversity of organisms: legumes, cereals, algae, fungi, insects, and microbial biomass produced through precision fermentation. Each candidate harbors a distinct proteome with its own distribution of functional properties, and characterizing even one candidate experimentally requires months of ingredient-level testing—extraction, purification, functional assays, and pilot-scale formulation. A framework that can rank and compare candidates computationally, before any bench work begins, could redirect scarce research resources toward the most promising sources and accelerate the timeline for bringing new ingredients to market. In a food system under pressure from climate change, population growth, and rising demand for protein, that acceleration matters.

The work also illustrates a broader trend in computational food science: the migration of machine-learning methods from medicine and structural biology into the study of food matrices, where the questions are different and the scale of analysis must be different too. Food functionality is an emergent property of mixtures, not molecules, and the field has long lacked tools that respect that fact. Lee and Yang’s distribution-based proteome representations represent one answer, and the authors position AlterProtX as a platform—something others can extend with additional molecular features, additional functional endpoints, or additional candidate organisms as predicted and experimental data accumulate.

For now, the framework stands as a proof of concept that proteome-scale, AI-enabled evaluation of food proteins is feasible and interpretable. The authors note that it provides an interpretable platform for proteome-level evaluation of food proteins, supporting the early-stage decisions that determine which sustainable alternatives get a chance to prove themselves in the lab and, eventually, on the plate. As the food system transformation accelerates, tools that can see whole proteomes rather than single molecules may prove essential to finding the proteins that future diets will depend on. The study was published open access, with funding from Korea’s Ministry of Education and National Research Foundation through the Leaders in INdustry-university Cooperation 3.0 Project and from the Rural Development Administration’s project on high-quality domestic plant-based protein materials.

Subject of Research: An AI-based proteome-scale framework for identifying sustainable alternative food proteins

Article Title: An AI-enabled proteome-scale framework for identifying sustainable protein alternatives for future food systems

Article References: Lee, E., & Yang, H. (2026). An AI-enabled proteome-scale framework for identifying sustainable protein alternatives for future food systems. npj Science of Food. https://doi.org/10.1038/s41538-026-01159-x

Image Credits: AI Generated

DOI: 10.1038/s41538-026-01159-x

Keywords: alternative proteins, artificial intelligence, proteomics, food science, sustainable food systems, thermal stability, gelation, allergenicity, plant-based protein, computational biology, food formulation, npj Science of Food