CPP2Vec Machine Learning Model Robustly Predicts Cell-Penetrating Peptides and PMO Delivery Efficiency

Effective intracellular delivery of therapeutic molecules, from small chemicals to nucleic acids like antisense oligonucleotides (ASOs), remains a significant challenge in drug development. Cell-penetrating peptides (CPPs) offer a promising solution by facilitating membrane translocation. Specifically, peptide nucleic acids (PNAs) and phosphorodiamidate morpholino oligomers (PMOs) are crucial ASOs being explored for treating severe genetic diseases like Duchenne Muscular Dystrophy (DMD), which causes progressive muscle degeneration. However, identifying and designing effective CPPs through traditional wet-lab experiments is costly and time-consuming, necessitating advanced in silico prediction methods to accelerate discovery and optimize delivery strategies.

Researchers developed CPP2Vec, a novel Word2Vec-based machine learning approach for representing amino acid sequences of peptides. To overcome limited sequence diversity and sparse biological data, they constructed CPP2Vec-GenSet, a hybrid dataset integrating computationally generated peptides with experimentally curated CPPs. This robust dataset served as the foundation for training three task-specific supervised machine learning models: one for CPP-Classification (determining if a peptide is a CPP), another for Uptake-Efficiency prediction, and a PMO-Delivery model to predict if a peptide enhances cellular delivery of a PMO-complex versus its naked form. Additionally, an alternative framework, CPP2LLM, was explored using pre-trained protein-based Large Language Models (ProtT5, ProtBERT, ESM-2) to generate peptide embeddings for similar task-specific models.

The study successfully developed CPP2Vec, a sophisticated Word2Vec-based machine learning framework designed to learn robust representations of peptide amino acid sequences. A key innovation was the creation of CPP2Vec-GenSet, a hybrid training dataset that significantly expands sequence diversity by combining both experimentally validated and computationally generated peptide data, thereby enhancing the model's learning capacity and cross-task performance. This framework enabled the development of three distinct predictive models: a CPP-Classification model to identify novel CPPs, an Uptake-Efficiency model to quantify their cellular penetration, and a specialized PMO-Delivery model to predict the efficacy of peptides in enhancing PMO-complex cellular uptake. > Benchmarking against existing state-of-the-art CPP prediction tools reportedly demonstrated that CPP2Vec achieves robust prediction capabilities, though specific quantitative performance metrics were not detailed in the provided abstract.