Machine learning assisted classification of cell and brain penetrating peptides

Abstract

Crossing the blood-brain barrier (BBB) remains a major obstacle for central nervous system therapeutics. Short peptides have emerged as promising vectors, including cell-penetrating peptides (CPPs) and brain-penetrating peptides (BPPs). However, the structural and physicochemical features that distinguish CPPs from BPPs remain poorly understood, limiting rational design. Here, we compiled a curated dataset of 490 peptides, encompassing CPPs, BPPs, and non-BPP controls, and systematically analysed their amino acid composition, sequence distribution, and physicochemical descriptors. BPPs were found to exhibit a more balanced distribution of cationic, polar, and hydrophobic residues compared to CPPs, which were enriched in contiguous arginine and lysine blocks. Physicochemical analysis revealed that BPPs had lower charge density, greater stability, and reduced aromaticity relative to CPPs. Dimensionality reduction confirmed BPPs occupy an intermediate chemical space between CPPs and non-BPPs. Machine learning classification, particularly with Extra Trees models, achieved strong performance in discriminating peptide classes, with charge, instability index, and aromaticity identified as the most predictive features. These findings suggest that BBB penetration is not a simple extension of cell penetration but requires finely tuned physicochemical properties. This study provides mechanistic insights into BPP design and highlights machine learning as a valuable tool for engineering next-generation BBB-penetrating peptides and peptide-mimetic materials.