Skip navigation

상단메뉴

글로벌메뉴

좌측메뉴

학술행사

검색

Seminars

Seminar
FIELD Mathematics
DATE January 26 (Tue), 2021
TIME 16:00-17:00
PLACE 1503
SPEAKER Hwang, Changha
HOST Keum, JongHae
INSTITUTE 동국대학교
TITLE Predicting Blood?Brain Barrier Penetration with Deep Learning Techniques
ABSTRACT The blood-brain-barrier (BBB) is a dynamic structure that separates brain tissue from the circulating blood to protect the central nervous system (CNS). In the paper we aim at qualitatively predicting whether compounds penetrate across the BBB or not. The ability to penetrate across the BBB, often expressed as BBB+ and BBB-, recently has become an important issue in the course of drug discovery and development. However, almost all classification models for penetration prediction have been built using imbalanced datasets. In addition, the existing prediction models have been usually developed using the data of the physical characteristics and chemical structure of compounds. To overcome the problem of imbalanced data we generated BBB- compounds using deep generative models such as conditional generative adversarial network (CGAN) and generative reinforcement learning (RL) system. Furthermore, to avoid using a lot of chemistry knowledge or molecular descriptors and fingerprints, we utilized deep convolutional neural networks (CNNs), which predict BBB penetration only with image data of 2-dimensional structures of compounds. In this respect, this paper is a novel attempt to predict drug penetration across the BBB. The validation results on 2358 compounds proved that a standard EfficientNet-B3 built based on balanced sets of compounds generated by CGAN and RL achieves better performance than the other existing methods. The accuracy of our deep learning (DL)-based method reaches 0.8685, the area under the curve (AUC) of the receiver operating characteristic reaches 0.9237, the area under the precision-recall curve (AUPRC) reaches 0.9606, the F1-score is 0.9104, the G-means is 0.8550, TNR is 0.9515 and TPR is 0.7683. The results proved that our DL-based method shows significantly good prediction performance of BBB penetration of compounds and it can be quite helpful in early drug discovery.
FILE  
  • list

date

~