Can AI compare whole proteomes to find sustainable protein alternatives?

Similarity is measured across whole distributions

The paper's core claim is that food proteins work as mixtures, not as single molecules. Gelation, foaming and emulsification emerge from the collective behavior of heterogeneous protein mixtures, so comparing one representative protein per ingredient would miss the property that matters. AlterProtX therefore represents each source as a distribution of predicted protein properties rather than as a mean or median value.

The pipeline has four steps. First, each protein's sequence and structure are encoded and used to predict denaturation temperature (Td), the temperature at which a protein unfolds, which the paper treats as a processing-relevant indicator of thermal behavior. Second, predicted Td values and six intermolecular interaction attributes are aggregated across all proteins in a source using kernel density estimation, which produces a smooth distribution instead of a single summary number. Third, distributions for non-animal sources are compared with animal-derived reference proteomes using Wasserstein distance, a measure of how much one distribution must be shifted to match another. Fourth, high-ranking candidates are screened for allergenic potential and essential amino acid composition.

The key distinction for readers is between a protein-level prediction and a source-level ranking. The Td model output is a per-protein number; the ranking is a distance between distributions. A source can rank highly without every protein in it matching animal proteins, because the comparison is about the shape and range of the whole profile.

The Td model rests on 24,572 measured proteins

The thermal stability predictor was trained on experimentally validated Td values from the Meltome Atlas, a resource providing standardized thermal stability measurements across species. After excluding entries lacking full-length sequences or required annotation, the curated set contained 24,572 proteins from five biological categories: human, other animals, plants, fungi, and bacteria or archaea. Where multiple measurements existed for one protein, the median was used. The data were split 80% training, 10% validation and 10% test with stratified sampling to keep Td distributions balanced.

Each protein was represented as a graph: residues are nodes, and edges connect residues that are adjacent in sequence (within plus or minus three residues) or, in the 3D representation, whose Cα–Cα distance is at most 8 Å. Node features combine 1,024-dimensional ProtT5 sequence embeddings with physicochemical descriptors, giving a 1,030-dimensional vector per residue. Six input configurations combining different sequence and structure encodings were trained with either multilayer perceptrons or graph convolutional networks.

This matters because the framework's credibility depends on whether predicted Td is meaningful for food-relevant proteins. The paper reports that predicted Td trends were consistent with experimentally reported stability patterns in legume proteins, and argues that jointly using full primary sequences and atomic-level 3D context captures thermal behavior including the 40–50 °C range characteristic of many natural food proteins. That consistency check is a supporting argument, not a substitute for direct validation on food ingredients.

Fifty sources were ranked against five animal categories

The non-animal dataset covered 50 food-relevant sources: 15 grains, 14 seaweeds, 11 legumes and 10 mushrooms, selected using criteria from the Korea Ministry of Food and Drug Safety. Rather than compiling complete organismal proteomes, the authors curated entries from UniProt using food-relevant taxonomic identifiers and functional annotations, excluding non-food organisms and sources with fewer than 10 protein entries. The resulting set contained 1,345,631 proteins.

The animal references represented five food categories: egg white, egg yolk, milk, meat and fish. These were also curated rather than taken as whole organisms, using category-specific keywords for major food protein groups such as caseins, whey proteins, myofibrillar proteins, ovalbumin-related proteins and collagen.

For each animal category, the framework identified the 10 most similar non-animal sources. Similarity was visualized with Td-based Wasserstein distance on one axis and the mean rank across intermolecular interaction attributes on the other, so sources near the origin combined low Td distance with consistently favorable interaction-attribute rankings. The paper reports both category-specific matches and candidates that scored highly across multiple animal categories, which it interprets as broader potential for formulation and replacement. Sensitivity analyses that varied kernel density estimation bandwidths, evaluation grids and representative feature selection largely preserved the rankings, although some target categories were more sensitive to feature substitution.

Matching distribution shape separates close from weak candidates

The most informative result is not the ranking itself but the comparison of full attribute distributions. For egg white and meat, the paper plotted Td, molecular weight, isoelectric point and surface hydrophobicity for each animal proteome alongside its most similar and most dissimilar non-animal source. Close matches reproduced not only median values but the overall shapes and ranges of the animal distributions; mismatched sources showed distinctly different shapes.

Two examples carry the argument. Wheat (Triticum aestivum) closely matched egg white proteins across all four distributions, which the authors note is consistent with prior experimental reports of comparable foaming performance. Cauliflower mushroom (Sparassis crispa) showed strong alignment with meat-derived proteins, reflecting balanced surface hydrophobicity and moderate structural compactness.

The authors use these cases to argue that single-value summaries can mischaracterize functional equivalence, because two sources with the same average Td can still differ in how their proteins are spread across the temperature range. This is a mechanistic interpretation rather than a demonstrated functional measurement: the paper does not report new gelation or foaming experiments on these sources, and it explicitly states that no experimentally validated threshold exists for defining functional equivalence from proteome-level similarity.

Allergenicity screening reorders otherwise strong candidates

After similarity ranking, the framework predicted protein-level allergenicity with ALLERDET, a tool combining restricted Boltzmann machines with a decision tree classifier, labeling each sequence as probable allergen or non-allergen. Predictions were aggregated by organism: a source was classified as high allergenic potential when predicted allergens exceeded non-allergens, and low otherwise.

The results showed substantial variation across food categories, with many legume and grain species containing relatively high proportions of proteins predicted to be allergenic. Variation also appeared within a single source: although Oryza sativa was generally predicted to have higher allergenic potential, several subspecies showed markedly lower predicted allergenicity, which the authors present as an opportunity for varietal selection. For representative candidates, predicted allergenic proteins were compared with entries in the WHO/IUIS allergen database; wheat contained a substantially higher proportion of predicted allergenic proteins than quinoa or cauliflower mushroom.

The paper is explicit that this is a prioritization criterion, not a safety determination. It states that comprehensive allergenicity assessment would require protein digestibility, abundance, exposure routes and clinical validation before an ingredient could be considered suitable for practical use.

Nutrition can confirm a candidate or block it

The third filter compares individual essential amino acid contents, in mg per gram of protein, against the FAO/WHO reference amino acid pattern. Quinoa (Chenopodium quinoa), the top-ranked alternative for egg yolk proteins, met or exceeded the reference pattern across all nine scoring categories, and the paper combines this with its high proteome-level similarity to egg yolk and low predicted allergenic potential to support further evaluation.

Wheat, the top-ranked alternative for egg white proteins, showed lower lysine content than the reference pattern, consistent with lysine being the established limiting amino acid in wheat. Combined with its clinically documented allergenicity, the authors argue that nutritional quality and safety should be weighed alongside its favorable functional similarity.

Cauliflower mushroom, the top-ranked alternative for meat proteins, had high similarity and low predicted allergenic potential but no comprehensive amino acid composition data, so nutritional evaluation was not possible. The authors use this case to make a specific point: favorable functional similarity and safety evidence alone are insufficient when nutritional information is incomplete. The paper also notes that amino acid composition alone does not capture protein quality, because digestibility and bioavailability also shape nutritional value in vivo, and suggests future incorporation of digestibility-based metrics such as DIAAS.

The ranking is relative, not a functional guarantee

The framework's intended use is early-stage screening before experimental validation and application-specific testing. The authors describe it as supporting transparent, interpretable prioritization of candidate sources rather than definitive ingredient selection, and they frame the trade-offs among functionality, allergenicity and nutrition as evidence to be read together rather than independently.

Several limits bound that use. The organism-level proteomes come from public databases and do not fully represent edible ingredients, whose composition varies with tissue type, cultivar, environmental conditions, extraction and processing; the authors note that no standardized large-scale resource currently provides ingredient-level proteomes across diverse food sources. The allergenicity predictions are computational and require clinical and exposure evidence. The nutritional dimension relies on amino acid composition without digestibility. And because no experimentally validated threshold exists for functional equivalence, the rankings should be read as relative ordering within the analyzed search space.

The practical takeaway is a screening order rather than a winner. Quinoa illustrates a candidate that clears all three filters for egg yolk; wheat illustrates one whose functional match is offset by a lysine limitation and documented allergenicity; cauliflower mushroom illustrates one that cannot be fully assessed until amino acid data exist. The paper's contribution is making those trade-offs explicit and comparable across 50 sources, not declaring any single source a drop-in replacement.

Sources used in this answer