Abstract / Summary
Photodynamic therapy (PDT) is a minimally invasive therapeutic approach based on the light-induced activation of a photosensitizer in the presence of molecular oxygen, resulting in the generation of reactive oxygen species, including singlet oxygen. Despite local selectivity and low systemic toxicity, reproducibility of PDT is restricted by non-uniform accumulation of the photosensitizer, aggregation and insufficient solubility of compounds, variability in the optical properties of tissues, hypoxia, oxygen consumption, photobleaching, and the lack of a universal measure of photodynamic dose. Artificial intelligence (AI), including machine learning and deep learning, graph neural networks, generative models, Bayesian optimization, and physics-informed neural networks, can integrate molecular, spectral, imaging, dosimetric, omics, and clinical data into decision-support systems. In this review, a closed AI–PDT cycle is considered, which includes the in silico design of photosensitizers, optimization of nanoformulations and delivery, personalized light exposure planning, real-time monitoring of photobleaching and oxygenation, and prediction of efficacy, toxicity, immune response, and long-term outcomes. Particular attention is given to data quality, external validation, interpretability, uncertainty assessment, and clinical readiness. This review emphasizes that the greatest potential of AI in PDT lies in its integration with mechanistic models of photochemistry, light transport, and oxygen dynamics, enabling the transition from empirical treatment protocols toward adaptive, reproducible, and personalized photomedicine. Unlike reviews focused on individual AI applications, this work considers the entire AI–PDT workflow as an interconnected system linking molecular design, drug delivery, dosimetry, treatment monitoring, and outcome prediction, while maintaining clinician oversight and the need for experimental and clinical validation.