Use of tree-based machine learning methods to screen affinitive peptides based on docking data.

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning in drug discovery

Background:

• Peptide-drug discovery relies on identifying high-affinity peptides.
• Machine learning offers a powerful tool for accurate peptide affinity screening.

Purpose of the Study:

• To compare the performance of four tree-based machine learning algorithms for predicting peptide affinity using virtual docking data.
• To identify the optimal algorithm for accurate and robust peptide affinity prediction.

Main Methods:

• Four tree-based algorithms (CART, C5.0, BAG, RF) were applied to virtual docking data.
• Datasets were pre-processed using scaling, centering, and Principal Component Analysis (PCA).
• Model performance was evaluated using metrics like Accuracy, Kappa, Sensitivity, Specificity, F1, MCC, and AUC.

Main Results:

• The optimized C5.0 model (C50O) demonstrated superior performance across multiple metrics on test and unknown datasets (80.4% Accuracy).
• Bagged CART (BAG) and optimized Random Forest (RFO) showed high prediction correlation with C50O, indicating stability.
• CART (CARTO) exhibited poor performance in test data validation, limiting its predictive utility.

Conclusions:

• The C5.0 decision tree model is highly effective for predicting peptide affinity from virtual docking data.
• This study provides a benchmark for tree-based models in peptide-protein interaction (PPI) research.
• The findings support the expanded application of machine learning in developing peptide therapeutics.