(P03-01-26) Protein Language Model Embeddings and Proteomic Similarity Scores as Food Fingerprints for Nutritional Proteomics

Objectives: To evaluate whether protein language model (pLM) embeddings can serve as informative food fingerprints for computational nutrition applications. We compared ESM-3-derived embeddings with abundance-weighted sequence similarity for capturing proteome-level relationships across a large food database.

Methods: Proteomic data from ~500 foods generated by the Periodic Table of Food Initiative were analyzed. For each food, protein sequences and relative abundance values from mass-spectrometry–derived proteomes were used to compute food similarity scores based on abundance-weighted protein sequence identity. In parallel, multimodal embeddings were generated for each protein using the ESM-3 protein language model integrating amino acid sequence and AlphaFold-predicted structure. Protein embeddings were aggregated using abundance weighting to generate food-level proteomic fingerprint vectors. Food relationships were evaluated using hierarchical clustering, UMAP visualization, and nearest-neighbor analysis.

Results: Sequence-based similarity scores recapitulated biologically meaningful relationships across foods, with related species and tissues showing higher proteomic similarity than unrelated foods. Embedding-based fingerprints produced concordant clustering patterns and clearly separated plant and animal foods in embedding space. Plant foods occupied a broader region of the fingerprint space, whereas animal foods formed tighter clusters. Cross-kingdom comparisons identified plant foods with moderate similarity to common animal protein sources, highlighting opportunities for alternative protein discovery.

Conclusions: Proteome-level food fingerprints derived from protein language model embeddings and abundance-weighted sequence similarity provide a scalable framework for quantitative food comparison and classification in nutritional proteomics.

Funding Sources: Periodic Table of Food Initiative