Stack-DCL: A Density-Aware and Local-Geometry Stacking Ensemble for Cell-Penetrating and Quorum-Sensing Peptide Classification

Abstract

Cell-penetrating peptides (CPPs) and quorum-sensing peptides (QSPs) represent peptide functions related to intracellular delivery and microbial signaling. Sequence-based prediction remains difficult because benchmark embeddings often show uneven local support, overlapping class regions, and high dimensionality. A single classifier may therefore miss either local or global structure. We propose Stack-DCL, an out-of-fold stacking framework for CPP and QSP classification with fixed UniRep embeddings. Stack-DCL combines a density-aware weighted random forest, a cosine-neighborhood model, and an L2-regularized logistic regression classifier. The meta-learner is also logistic regression and fuses only positive-class probabilities generated from held-out folds, which reduces fusion-stage leakage. On benchmark train/test splits with five-fold cross-validation on training data and final held-out testing, Stack-DCL achieves independent-test AUC values of 0.980 for QSP and 0.974 for CPP. Ablation results indicate that the three branches capture local support, directional neighborhood similarity, and global linear separability as different representation-level signals. These results support Stack-DCL as a benchmark-level stacking design for the two evaluated peptide tasks.