Abstract Cell‐penetrating peptides (CPPs) facilitate the intracellular delivery of therapeutic molecules. However, their accurate identification and design remain challenging because of the complexity of their structural and physicochemical characteristics. This study aimed to develop an interpretable predictive model that enables reliable CPP discovery and provides interpretable descriptors suggestive of the molecular properties underlying their activity. Peptide samples were represented using two‐dimensional descriptors generated by Mordred. A novel two‐stage feature selection method was introduced, combining a correlation‐based filter with Shapley Additive exPlanations (SHAP). The model was trained on the CPP1708 dataset and built using an ensemble learning strategy integrating multiple machine learning algorithms. The ensemble framework, combining Extreme Gradient Boosting and Light Gradient Boosting Machine, identified five Mordred descriptors—5‐ordered bonding information content (BIC5), Extended Topochemical Atom epsilon 5 (ETAₑpsilon₅), averaged and centered Moreau‐Broto autocorrelation of lag 0 weighted by ionization potential (AATSC0i), centered Moreau‐Broto autocorrelation of lag 2 weighted by mass (ATSC2m), and first highest eigenvalue of Burden matrix weighted by gasteiger charge (BCUTc‐1h) —as critical features for CPP prediction. The model achieved an accuracy of 82. 0% and an area under the curve of 87. 5% on the CPP1708 test set, outperforming existing predictors. This interpretable, high‐performing prediction model supports the rational design of CPPs and advances peptide‐based drug development. The SHAP‐guided feature selection framework improves both efficiency and interpretability, with potential applications across diverse peptide classes. Furthermore, the identification of five mechanistic descriptors offers deeper insight into the structural, electronic, and physicochemical properties underpinning CPP activity.
Kwon et al. (Fri,) studied this question.