Protein Language Models, Human NAMs and Evidence Fusion for Allergenicity and Immunotoxicity Risk Assessment of Novel Food Proteins.

Novel food proteins from precision fermentation, edible insects, cell-cultured systems, microbial biomass and protein engineering are expanding faster than traditional toxicology workflows can evaluate them. Their principal safety question is not classical systemic toxicity alone, but whether sequence novelty, post-translational modification, processing, digestion, food-matrix exposure and immune recognition combine to create IgE-mediated allergenicity, coeliac-relevant peptides or broader immunotoxic concern. This review develops a next-generation risk assessment framework in which protein language models, structure prediction, digestomics, immunopeptidomics, human intestinal and immune new approach methodologies (NAMs), exposure modelling and physiologically based kinetic reasoning are integrated into a transparent weight-of-evidence process. The review is deliberately positioned beyond single-tool allergen prediction: it argues that sequence similarity, AI classification and in vitro digestibility become useful for regulatory decision-making only when anchored to an adverse outcome pathway, human-relevant assays, exposure context, uncertainty analysis and external validation. The central conclusion is that AI-enabled NAMs can transform novel protein assessment from a binary hazard screen into a mechanistically interpretable, exposure-aware and decision-oriented system, but only if datasets, benchmark design, applicability domains and reporting standards are improved. An author-proposed scoring rule and a retrospective EFSA yellow-mealworm case illustrate the workflow.