BrainShuttle-ESM: A Multi-Stage Transformer Architecture for Predicting Blood-Brain Barrier-Penetrating Short...

The blood-brain barrier (BBB) is a critical protective semi-permeable membrane that separates the bloodstream from brain tissue, preventing harmful substances from entering the central nervous system and maintaining brain homeostasis. However, this protective function also restricts the permeation of most therapeutic molecules, making drug delivery to the brain a major challenge. Among the strategies explored to overcome this limitation, blood-brain barrier-penetrating peptides (B3PPs) have emerged as a particularly promising class of molecules. These peptides, especially shorter ones (5-20 amino acids), can function as both delivery shuttles and therapeutic agents. Their small size, inherent neuroprotective properties, and amenability to chemical modification make them attractive candidates. Despite growing interest, current computational approaches have not been explicitly designed to predict short BBB-penetrating peptides.

To address this gap, we developed BrainShuttle-ESM, a fine-tuned ESM-2-based model specifically trained to predict BBB penetration of short peptides. Using a multi-stage fine-tuning strategy incorporating balanced positive peptides and structurally diverse negative examples, we trained the model to classify peptides based on their BBB-penetrating potential. Attention-map analysis was subsequently employed to facilitate interpretation of the model's predictions and to highlight physicochemical features that may contribute to its performance.

The model demonstrated strong performance, achieving an AUC-ROC of 0.88_±0.007 and an F1-score of 0.83_±0.009. Attention-map analysis indicated that amphipathic residues consistently receive high attention weights, suggesting a potential role of amphipathicity in BBB permeability, consistent with prior experimental observations. The analysis also suggests that hydrophobicity and side-chain dihedral angles may be associated with membrane penetration. Additionally, we developed a user-friendly web server to allow researchers to screen peptide sequences for BBB-penetrating potential.

BrainShuttle-ESM is an LLM-based computational model for identifying short BBB-penetrating peptides. Residue-level attention analysis suggested potential associations with amphipathicity and hydrophobicity. The study emphasizes the importance of dual-functional peptides that combine BBB penetration with additional therapeutic properties to support peptide-based drug development for neurological disorders, with BrainShuttle-ESM offering potential to accelerate this process.