Unlocking component-level chemical structural information for AI-driven targeted nanoparticle design

The structural complexity and unpredictable in vivo fate of nanoparticles pose major barriers to rational design of targeted nanoparticles. The rapidly developing artificial intelligence (AI) technology brings new opportunities to establish structure-property relationships. However, the effectiveness of AI-based prediction models is still constrained by the absence of nanoparticle compositional representation. To address this, we innovatively developed component-level image-based structural descriptors (CISD) and integrated them into the nanoparticle biodistribution prediction model. Our framework addresses several challenges in generating representations of nanoparticle compositions, such as data scarcity, poor generalizability across different nanoparticle types, and limited interpretability. Therefore, our work provides a tool that bridges the chemical structural information of nanoparticle components with representations usable by machine learning. In independent cross-domain validation, our developed predictive model significantly outperformed the traditional frameworks that rely solely on nanoparticle characteristics and biodistribution experimental parameters, with an R² increase of 0.25 and an RMSE reduction of 3.22. Leveraging SHapley Additive exPlanations (SHAP) analysis and hook functions, we disentangled multi-level structure-property relationships, with CISD enhancing tunable factors in nanoparticle design. Building upon the extracted structure-property relationships and model projections, we achieved up to a 12.66-fold enhancement in predicted biodistribution values. In the future, synergistic integration with advanced algorithms like generative AI holds promise not merely for curtailing animal experimentation, but for pioneering closed-loop inverse design systems to rationally screen a wider range of potential structures of nanoparticle components and nanoparticle physicochemical attributes.