Uncovering Physicochemical Determinants of Anticancer Peptides via Interpretable Ensemble Feature Importance Analysis

Fan Xiaoqin^*, Liu Tianci, Guan Suyang, Zheng Yuting

Suqian University

Abstract：

Anticancer peptides, or ACPs, are now more and more regarded as promising lead compounds that can be used for anticancer therapy, but the fact is that most computational predictors are not very interpretable, and this poor interpretability makes it difficult to carry out rational peptide design in a reasonable way. Deep‑learning models often give high prediction accuracy, however, they are like black boxes and are not transparent enough, so medicinal chemists cannot get clear guidance from them about which physical and chemical attributes actually matter. In the work reported here, we have used an interpretable framework that is based on ensemble learning, and we have tried to identify the sequence determinants as well as the physical‑chemical determinants of anticancer activity. We used the ACP-2.0 dataset for our analysis, and we integrated several types of features, including amino acid composition, k-spaced amino acid pair frequencies, and two physical‑chemical descriptors, namely average charge and average hydrophobicity; then we obtained feature importance scores by using an ensemble model that combined Random Forest and XGBoost. From the feature importance analysis, we found that lysine (K), average charge, leucine (L), and average hydrophobicity were the dominant factors, and these factors together pointed to the importance of cationic and amphipathic properties. These findings are in line with the well‑known mechanisms by which ACPs disrupt cell membranes, and they also offer some concrete design rules that can guide the development of new anticancer peptides.

Key Words：

anticancer peptides; interpretability; physicochemical property; feature importance; lysine; average charge; hydrophobicity