Machine Learning-Assisted Design and Discovery of High-Performance Cyanine-Based Photosensitizers for Integrated Theranostic Applications.

Cyanine-based photosensitizers are attractive for photodynamic therapy (PDT) owing to their excellent near-infrared absorption and fluorescence. However, the low singlet oxygen quantum yield (ΦΔ) and structural optimization challenges hinder their further development. In this study, a machine learning (ML)-assisted molecular design framework based on the currently available dataset of cyanine molecules is provided. By integrating RDKit structural descriptors with quantum chemical descriptors, hybrid feature-based predictive models that accurately predict the ΦΔ and fluorescence quantum yield (ΦF) of cyanine derivatives (R2 >0.9) have been constructed for the first time. Based on this, a two-stage virtual screening strategy is developed to efficiently identify promising cyanine derivatives from a library of 2835 candidate structures. Three representative molecules are synthesized as proof-of-concept validation, confirming the predictive reliability and practical utility of the ML-guided workflow. The lead compound 1775 exhibits the highest performance (ΦΔ = 0.62) and performs well in cellular assays, supporting the robustness and applicability of the ML-assisted screening strategy in guiding experimental validation. This work establishes a data-driven paradigm bridging molecular modeling and experimental verification, offering a reliable and generalizable approach for the rational design and rapid evaluation of high-performance cyanine-based theranostic agents.