AI-Driven Prediction of Drug-Likeness for Peptides: An Evaluation of Newly Created MLP and k-NN Models

Peptide-based drug candidates frequently fall outside the boundaries of classical drug-likeness rules due to their high molecular weight, polarity, and structural flexibility. The aim of this study was to develop and evaluate machine learning (ML) models based on multilayer perceptron (MLP) and k-nearest neighbors (k-NN) algorithms for predicting violations of Lipinski’s Rule of Five (Ro5), beyond Rule of Five (bRo5), and the Muegge rule for structurally diverse peptide molecules. ML models were trained using physicochemical data from PubChem and applied to a data set of 26 tested peptide molecules. Predicted violations counts were compared with SwissADME, Molinspiration platforms, and manually calculated bRo5 values. MLP and k-NN models demonstrated excellent predictive performance, with accuracy, precision, recall, and F1-score values ranging from 0.98 to 0.99 for Ro5 and bRo5, and remaining above 0.94 for the Muegge rule. Prediction errors were limited to ±1 violation and occurred mainly for borderline structures near rule thresholds. The k-NN model exhibited higher consistency for bRo5 predictions, while MLP showed greater stability for Ro5. Descriptor-based ML models can reliably capture drug-likeness trends within peptide chemical space. The developed models complement ensemble approaches such as RF and represent efficient in silico tools for early assessment of peptide drug candidates.