Prediction of plant food allergens using protein embeddings

The current trend to incorporate new protein sources to the diet, mostly from plants, requires a safety assessment of these polypeptides when used as food. Part of that safety evaluation involves assessing the potential allergenicity of the proteins. Such assessment typically follows a hierarchy of in silico, in vitro and in vivo assays. The computational prediction of protein allergenicity is the first step in the process and several predictors are available for this task. Due to the complexity and heterogeneity of the allergy phenomenon, the large differences between allergens in their mechanisms of exposure and action, and the role of their inherent biological and physicochemical properties, predictors focused on particular organisms or subsets of allergens are in principle better than general predictors.

In this work we present a predictor of protein allergenicity focused on plant food allergens. The predictor, trained on known plant food allergens and bona-fide plant non-allergens, represents proteins by their sequence embeddings, a deep-learning based representation known to capture important features of a protein's function and properties. The predictor outperforms the typical prediction based on sequence similarity to other allergens, especially in terms of sensitivity. In spite of being trained on plant food allergens, the predictor also performs well when applied to allergens from other sources.

The system is available as an open-source package at Github: https://github.com/Martinmendezlopez/ALLERGEN-PREDICTOR-FROM-ProtTrans/.