SSEL-CPP: A SHAP-based feature-selection ensemble learning framework identifies molecular properties of...

Area of Science:

• * Computational chemistry and bioinformatics
• * Drug discovery and development
• * Peptide science

Background:

• * Cell-penetrating peptides (CPPs) are crucial for delivering therapeutic molecules intracellularly.
• * Identifying and designing effective CPPs is hindered by their complex structural and physicochemical properties.
• * Predictive models are needed to streamline CPP discovery and understand activity determinants.

Purpose of the Study:

• * To develop an interpretable predictive model for reliable cell-penetrating peptide (CPP) discovery.
• * To identify key molecular descriptors that explain CPP activity.
• * To advance rational design strategies for peptide-based therapeutics.

Main Methods:

• * Peptide samples were represented using two-dimensional Mordred descriptors.
• * A two-stage feature selection method combined correlation-based filtering with Shapley Additive exPlanations (SHAP).
• * An ensemble learning framework integrated Extreme Gradient Boosting and Light Gradient Boosting Machine on the CPP1708 dataset.

Main Results:

• * The ensemble model achieved 82.0% accuracy and 87.5% AUC on the CPP1708 test set, surpassing existing predictors.
• * Five critical Mordred descriptors were identified: BIC5, ETA_epsilon_5, AATSC0i, ATSC2m, and BCUTc-1h.
• * The SHAP-guided framework enhanced model interpretability and efficiency.

Conclusions:

• * The developed interpretable model accurately predicts CPP activity and supports rational CPP design.
• * The identified mechanistic descriptors provide insights into structural and physicochemical properties governing CPP function.
• * This approach advances peptide-based drug development and has broad applications in peptide science.