Blood–brain barrier–transport artificial intelligence: A hybrid sequence–structure machine-learning framework for predicting blood–brain barrier-penetrant peptides

Background Cell-penetrating peptides (CPPs) and blood–brain barrier (BBB) shuttles are promising vehicles for delivering therapeutics to the central nervous system; however, few BBB-permeable sequences have been experimentally validated. Existing prediction tools mainly rely on sequence motifs and rarely incorporate structural information or provide mechanistic insights. We present BBB–transport artificial intelligence (TransAI), a hybrid machine-learning pipeline that fuses classical peptide descriptors with structure-derived features from AlphaFold3 models to predict BBB permeability. Methods We collected a dataset of BBB + and BBB − peptides and extracted 10 best-performing families of sequence descriptors totaling 1133 features using iFeature. AlphaFold3-predicted three dimensional (3D) structures were used to compute the structural metrics for high confidence. Random Forest importance and correlation pruning reduced the sequence space to non-redundant descriptors, which were combined with the structural features for model training. Results The final Random Forest classifier achieved a receiver operating characteristic (ROC) area under the curve (AUC) of 0.73 on an independent test set; five-fold cross-validation yielded an increased mean ROC AUC of 0.81 ± 0.09. The feature importance highlighted hydrophobicity indices, polarizability scales, and secondary-structure propensities as key determinants. The structural analysis revealed that BBB-positive peptides exhibit a higher hydrophobic residue content, increased exposure of positively charged side chains, and intermediate compactness relative to negatives. Conclusions BBB-TransAI demonstrates that integrating AlphaFold3 structural descriptors with traditional sequence features enhances the prediction accuracy and yields mechanistic insights into BBB translocation. The framework is scalable to larger peptide libraries, can incorporate additional biophysical descriptors, and can be released as a Streamlit web server for the in silico screening of BBB shuttles. Our study also outlines the current limitations of annotation quality and structural uncertainty, and provides guidance for future community efforts in BBB peptide informatics.