Artificial intelligence can support the chemical design of multiepitope vaccine nanocarriers

Artificial intelligence (AI) has accelerated the prioritization of vaccine epitopes, but the translation of predicted epitopes into chemically viable delivery systems remains a major bottleneck. Multi-epitope vaccines are attractive because they can combine B-cell, CD4 + T-cell and CD8 + T-cell determinants into compact immunogens; however, peptide constructs often display poor solubility, aggregation, proteolytic instability, limited cellular uptake and release profiles that are not aligned with antigen presentation. Nanocarriers are therefore relevant not simply as passive vehicles, but as chemically tunable systems that can protect antigens, enable co-delivery of adjuvants, support multivalent display and modulate lymphatic trafficking. This critical review examines how AI is being used in epitope prediction and formulation optimization, while emphasizing the chemical descriptors that determine whether such predictions become experimentally actionable. Peptide-level descriptors include charge, hydrophobicity, molecular weight, solubility, aggregation propensity and protease-sensitive motifs, whereas carrier-level descriptors include size, polydispersity, zeta potential, composition, surface functionality, degradation rate, loading efficiency and stability. Extracellular vesicles are included as an emerging biologically derived platform, with attention to cargo loading, heterogeneity, purification and potency control. The review also discusses controlled-release modeling and physics-informed neural networks (PINNs) as emerging tools for connecting sparse release data with mechanistic transport assumptions, while noting that their maturity in vaccine nanocarriers remains limited. Pharmacokinetic (PK) and physiologically based pharmacokinetic (PBPK) modeling are considered in relation to biodistribution, lymph-node exposure and clearance. To make these links operational, the review introduces a descriptor-to-formulation mapping, a critical comparison of carrier and formulation choices, a study-level evidence map of descriptor use in published machine-learning models, and a minimum end-to-end reporting and validation framework. Future progress will require standardized datasets, external validation, uncertainty quantification, reproducible experiments and chemically interpretable models rather than isolated predictive scores.