Interpretable Multimodal Models for Nanocarrier Biodistribution Using Physicochemical and Imaging Data Download PDF

Achieving predictable and favorable biodistribution is a central goal in nanomedicine. However, the relationships among nanoparticle design, imaging biomarkers, biological barriers, and organ accumulation remain complex and difficult to generalize. Current predictive approaches often operate as black boxes or depend on narrow feature sets. This makes it difficult for nanocarrier engineers to identify which modifiable particle properties drive undesirable uptake in clearance organs. This article proposes an interpretable multimodal machine learning framework for predicting organ-level nanocarrier biodistribution. The model is designed to connect physicochemical descriptors and imaging-derived features with transparent, organ-specific explanations. A multimodal architecture would combine an encoder for physicochemical features with an imaging-feature encoder. Explainability would be provided through post-hoc SHAP analysis or built-in attention mechanisms that decompose organ-uptake predictions into contributions from individual particle properties and imaging readouts.
Conceptually, the model could forecast liver, spleen, tumor, and kidney accumulation while highlighting the dominant drivers for each organ. For example, high liver uptake could be explained by low PEG density, positive zeta potential, or imaging signatures consistent with nonspecific reticuloendothelial accumulation. An interpretable multimodal approach could help close the feedback loop between prediction and nanocarrier redesign. By linking organ-level biodistribution patterns to actionable formulation variables, it could support more rational development of targeted nanomedicines.