DeepB3P: A transformer-based model for identifying blood-brain barrier penetrating peptides with data augmentation using feedback GAN.

Abstract

Introduction: The blood-brain barrier (BBB) serves as a critical structural barrier and impedes the entry of most neurotherapeutic drugs into the brain. This poses substantial challenges for central nervous system (CNS) drug development, as there is a lack of efficient drug delivery technologies to overcome this obstacle. BBB penetrating peptides (BBBPs) hold promise in overcoming the BBB and facilitating the delivery of drug molecules to the brain. Therefore, precise identification of BBBPs has become a crucial step in CNS drug development. However, most computational methods are designed based on conventional models that inadequately capture the intricate interaction between BBBPs and the BBB. Moreover, the performance of these methods was further hampered by unbalanced datasets.

Objectives: This study addresses the problem of unbalanced datasets in BBBP prediction and proposes a powerful predictor for efficiently and accurately identifying BBBPs, as well as generating analogous BBBPs.

Methods: A transformer-based deep learning model, DeepB³P, was proposed for predicting BBBP. The feedback generative adversarial network (FBGAN) model was employed to effectively generate analogous BBBPs, addressing data imbalance.

Results: The FBGAN model possesses the ability to generate novel BBBP-like peptides, effectively mitigating the data imbalance in BBBP prediction. Extensive experiments on benchmarking datasets demonstrated that DeepB³P outperforms other BBBP prediction models by approximately 9.09%, 4.55% and 9.41% in terms of specificity, accuracy, and Matthew's correlation coefficient, respectively. For accelerating the progress in BBBP identification and CNS drug design, the proposed DeepB³P was implemented as a webserver, which is accessible at http://cbcb.cdutcm.edu.cn/deepb3p/.

Conclusion: The interpretable analyses provided by DeepB³P offer valuable insights and enhance downstream analyses for BBBP identification. Moreover, the BBBP-like peptides generated by FBGAN hold potential as candidates for CNS drug development.