EnDM-CPP: A Multi-view Explainable Framework Based on Deep Learning and Machine Learning for Identifying...

Area of Science:

• Biotechnology and Pharmaceutical Sciences
• Computational Biology and Cheminformatics

Background:

• Cell-penetrating peptides (CPPs) are vital for drug delivery.
• Laboratory synthesis of CPPs is time- and resource-intensive.
• Computational prediction of CPPs can accelerate therapeutic development.

Purpose of the Study:

• To develop an accurate computational method for predicting cell-penetrating peptides (CPPs).
• To enhance the efficiency of CPP discovery for therapeutic applications.

Main Methods:

• A hybrid model, EnDM-CPP, was developed, integrating Support Vector Machines (SVM), CatBoost, Convolutional Neural Networks (CNN), and TextCNN.
• Datasets from CPPsite 2.0, MLCPP 2.0, and CPP924 were merged for improved diversity and reduced homology.
• Transformer-based features (ProtT5, ESM-2) and sequence features (CPRS, Hybrid PseAAC, KSC) were utilized.
• Logistic Regression (LR) was employed for final decision prediction based on individual model outputs.

Main Results:

• Fusion features from ProtT5 and ESM-2 significantly improved CPP prediction accuracy.
• The combined model demonstrated superior performance compared to individual models.
• EnDM-CPP achieved 0.9495 accuracy and 0.9008 Matthews correlation coefficient on an independent test set.
• Performance improvements ranged from 2.23%-9.48% in accuracy and 4.32%-19.02% in MCC over state-of-the-art methods.

Conclusions:

• EnDM-CPP offers a highly accurate and efficient computational approach for identifying cell-penetrating peptides.
• The study highlights the effectiveness of combining diverse features and machine learning models for CPP prediction.
• The developed method can accelerate the discovery and development of CPPs for drug delivery applications.