MSF-CPMP: A novel multi-source feature fusion model for prediction of cyclic peptide membrane permeability.

Cyclic peptides are becoming attractive molecules for drug discovery because of their properties with inherent stability and structural diversity. However, the high potential of cyclic peptide drugs is challenged by the limited membrane permeability cross cell membrane. To predict cyclic peptide membrane permeability (CPMP), an increased number of computational models or tools are designed and used. But these existing algorithms or models do not appropriately capture feature diversity of cyclic peptides. In this study, we introduce a novel multi-source feature fusion model called MSF-CPMP, which aims to increase the accuracy of predicted CPMP. The MSF-CPMP model incorporates three features extracted from SMILES sequences, graph-based molecular structures, and physicochemical properties of cyclic peptides. By benchmarking with other non-deep-learning and deep learning-based methods, MSF-CPMP achieved the highest levels of the evaluation metrics such as accuracy of 0.9062 and AUROC of 0.9546 and further validated MSF-CPMP robustness in learning capabilities and efficacy of its multi-source fusion. Our result demonstrates that MSF-CPMP outperforms other methods in predicting CPMP, that provides also exemplifies the power of advanced deep learning methods in tackling complex biological challenges, offering contributions to computational biology and clinical treatment. Code is available at https://github.com/wanglabhku/MSF-CPMP.