AI-Driven Prediction of Allergenicity and Toxicity in Novel Plant-Based Peptides

Abstract Peptides from plant sources have antihypertensive, antioxidative, and antimicrobial properties but also entail a risk of allergic and toxic reactions that cannot be detected in conventional wet-lab assays. This review examines deep learning techniques and protein language models for their prediction, from linear motifs screening to transformers (BERT, ESM). The best-performing tools achieve high sensitivity (pLM4Alg 96.2%, ToxinPred3 95.8%) but low specificity and positive predictive value (pLM4Alg 92.8% and 81.7%), meaning that most predictions are false positives. There are four major constraints on the generalization capabilities of these tools: scoring of native peptides while proteolytic processing yields post-translational modifications and neoantigens; training samples correspond to IgE binding, not clinical manifestations, and are biased toward well-characterized organisms; T-cell immunity, microbiome, and organ-specific toxicity are scarcely modeled; and regulatory agencies do not consider in silico predictions sufficient. AI currently serves best as a prioritization filter, not a replacement for biological assays.