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

Area of Science:

• Computational biology
• Drug discovery
• Neuroscience

Background:

• The blood-brain barrier (BBB) protects the brain but hinders drug delivery.
• Blood-brain barrier-penetrating peptides (B3PPs) show promise for brain-targeted therapeutics.
• Existing computational methods are insufficient for predicting short B3PPs.

Purpose of the Study:

• To develop a computational model for predicting short BBB-penetrating peptides.
• To identify key physicochemical features influencing BBB penetration.
• To create a user-friendly tool for researchers.

Main Methods:

• Fine-tuned an ESM-2 based model (BrainShuttle-ESM) using a multi-stage strategy.
• Trained the model on balanced positive and diverse negative peptide examples.
• Utilized attention-map analysis for model interpretability.

Main Results:

• BrainShuttle-ESM achieved high performance (AUC-ROC 0.88, F1-score 0.83).
• Attention analysis highlighted amphipathicity and hydrophobicity as key features for BBB penetration.
• A web server was created for easy screening of peptide sequences.

Conclusions:

• BrainShuttle-ESM is an effective LLM-based tool for identifying short B3PPs.
• The model's insights support the role of amphipathicity and hydrophobicity in BBB permeability.
• This tool can accelerate the development of peptide-based therapies for neurological conditions.