Abstract

Oncolytic peptides (OPs) represent a promising class of cancer therapeutics capable of rapidly lysing tumor cells and activating antitumor immunity. However, accurate in silico identification of potent OPs remains challenging due to limited datasets and high false-positive rates. Here, we present MISPOP (Mechanism-Informed Screening Pipeline for Oncolytic Peptides), an integrated machine learning framework that combines eXtreme Gradient Boosting, deep neural networks, and transfer learning into a high-confidence ensemble model augmented with physicochemical priors. Applied to a natural peptide library of 1033 sequences, MISPOP prioritized 16 candidates, among which five were synthesized and evaluated across three tumor cell lines. Dermaseptin-S9 exhibited the most favorable therapeutic index. Molecular dynamics simulations revealed its deep insertion into lipid bilayers and stable peptide-membrane interactions, while in vitro assays confirmed pronounced membrane disruption and induction of immunogenic cell death. In a B16F10 melanoma model, Dermaseptin-S9 achieved over 92% tumor growth inhibition without evident systemic toxicity. Collectively, these findings demonstrate that embedding biochemical priors into ensemble learning can markedly improve predictive accuracy and enable the discovery of potent OPs, offering a generalizable paradigm for accelerating peptide-based oncotherapy development.