SSEL-CPP framework identifies five key molecular properties for cell-penetrating peptide activity with 82.0% accuracy

The intracellular delivery of therapeutic molecules is a major hurdle in drug development, often limited by cellular membrane impermeability. Cell-penetrating peptides (CPPs) offer a promising solution by facilitating cargo transport into cells. However, their rational design and accurate identification remain challenging due to the complex interplay of their structural and physicochemical characteristics. Current predictive models often lack interpretability, making it difficult to understand the underlying molecular properties driving CPP activity and hindering the development of more effective peptide-based drug delivery systems.

Researchers developed SSEL-CPP, a novel two-stage feature selection and ensemble learning framework, to predict and characterize cell-penetrating peptides. Peptide samples were initially represented using two-dimensional descriptors generated by Mordred. The framework employed a two-stage feature selection method, combining a correlation-based filter with Shapley Additive exPlanations (SHAP) to identify key molecular properties. The model was trained on the CPP1708 dataset and built using an ensemble learning strategy that integrated multiple machine learning algorithms, specifically Extreme Gradient Boosting and Light Gradient Boosting Machine, to enhance predictive performance and interpretability.

The SSEL-CPP ensemble framework achieved high predictive performance and identified five critical Mordred descriptors essential for CPP activity. The model demonstrated an accuracy of 82.0% and an area under the curve (AUC) of 87.5% on the CPP1708 test set, significantly outperforming existing predictors. These identified descriptors include: 5-ordered bonding information content (BIC5), Extended Topochemical Atom epsilon 5 (ETA_epsilon_5), 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).

The SSEL-CPP model achieved an accuracy of 82.0% and an AUC of 87.5% on the CPP1708 test set, identifying five key molecular descriptors for CPP prediction.