PepOSX-AI: CPP - an interpretable transformer-based deep learning model for prediction of cell-penetrating peptides.
TL;DR
PepOSX-AI: CPP provides a novel computational tool and theoretical basis for the screening of CPPs and effectively captured long-range dependencies in amino acid sequences through a self-attention mechanism, enabling the interpretability of attention patterns.
Abstract
Background
Cell-penetrating peptides (CPPs) are short-chain molecules capable of enhancing the transmembrane delivery of bioactive substances, displaying extensive application potential in the delivery of functional components and the improvement of their bioavailability. Traditional CPP discovery methods, however, rely on a tedious, step-by-step screening process involving cell and animal experiments, which is highly inefficient.
Methods
The deep-learning model, PepOSX-AI: CPP, was developed based on the Transformer architecture and integrative features of six physicochemical descriptors. This involved a series of explorations of model parameters and automated hyperparameter optimization. The model achieved a high area under the curve (AUC) of 0.914 on the dataset. Furthermore, the model effectively captured long-range dependencies in amino acid sequences through a self-attention mechanism, enabling the interpretability of attention patterns. This capability facilitates the identification of key sequence features influencing peptide penetration ability.
Significance
AND NOVELTY In comparison to some existing CPP prediction models, PepOSX-AI: CPP demonstrated an accuracy of 91.00% on the application test set, highlighting its acceptable predictive performance. It provides a novel computational tool and theoretical basis for the screening of CPPs. Based on these findings, an online platform was also developed to facilitate user application.