KAN-PROSPECT: a Kolmogorov-Arnold Networks-integrated framework for predicting the effects and adverse reactions of natural products via transfer learning.

The current reliance on wet-lab experiments for evaluating the efficacy and adverse effects of natural products remains a major obstacle in new drug discovery. We propose KAN-PROSPECT, a novel deep learning framework that integrates transfer learning with Kolmogorov-Arnold Networks (KAN). This model can simultaneously predict the efficacy and adverse effects of natural products based solely on molecular SMILES representations, thereby addressing the limited generalizability of existing models in predicting these aspects for natural products. Beyond methodological advances, KAN-PROSPECT also contributes to more sustainable and resource-efficient drug discovery. Leveraging a cross-modal transfer learning strategy pretrained on approximately 3,800 drugs and fine-tuned on 400 natural products, KAN-PROSPECT consistently outperforms baseline models in dual-label prediction tasks. Notably, it demonstrates exceptional robustness in addressing data scarcity, excelling particularly in few-shot and zero-shot scenarios. Through the transfer learning strategy, the model partially alleviates the data scarcity issue commonly encountered in natural product prediction tasks. In addition, the incorporation of KAN layers enhances the ability to model complex nonlinear relationships between molecular structures and associated pharmacological or adverse reaction profiles, contributing to improved predictive performance. Furthermore, the framework demonstrates strong generalization ability, enabling high-accuracy predictions for the efficacy and adverse effects of entirely new natural products. KAN-PROSPECT was further applied to comprehensively predict natural products from the MEC and NPASS databases, with Icaritin from Epimedium used as a representative case study. Overall, KAN-PROSPECT is the first framework to unify transfer learning with the KAN architecture for dual-label prediction of natural products, showing great potential for large-scale bioactivity and toxicity prediction, new drug development, and drug repositioning of natural products.