Scholarlink

<p>Blood-brain barrier (BBB) is an important element in drug discovery in central nervous system (CNS) since it limits the intake of most pharmacological agents into the brain. This paper is a deep learning BBB permeability prediction framework in the form of a graph model built on MoleculeNet BBBP data. Originally, there are 3120 molecules in the original dataset, which were canonicalized and followed by the elimination of salts and duplicates to produce 2013 unique compounds in the dataset. RDKit 2023.09.1 was used to construct molecular graphs and feature vectors of chemically informed atom-bond features, which were used to encode them. The stratified splitting, intensive hyperparameter optimization, and class-balanced training were used to train a gated Message Passing Neural Network (MPNN) that has a multi-head transformer readout. In order to put the proposed model into perspective, the baseline QSAR and classical machine learning (Random Forest, SVM, XGBoost, Logistic Regression) models were trained on the same dataset split. Marxia improved accuracy and discrimination (AUC-ROC = 0.9627; accuracy = 92.54%) compared to the baselines based on the descriptors, which is explained by the fact that the MPNN-Transformer architecture was more accurate and discriminative. These results indicate the benefit of graph-based learning as compared to rigid molecular fingerprints and descriptors. The model is suitable for integration into early-stage compound screening pipelines in CNS drug discovery.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Graph-based deep learning for blood–brain barrier permeability screening in central nervous system drug discovery

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月8日

查看全文
  • Ronith Lahoti,
  • Manisha Bhende

摘要

Blood-brain barrier (BBB) is an important element in drug discovery in central nervous system (CNS) since it limits the intake of most pharmacological agents into the brain. This paper is a deep learning BBB permeability prediction framework in the form of a graph model built on MoleculeNet BBBP data. Originally, there are 3120 molecules in the original dataset, which were canonicalized and followed by the elimination of salts and duplicates to produce 2013 unique compounds in the dataset. RDKit 2023.09.1 was used to construct molecular graphs and feature vectors of chemically informed atom-bond features, which were used to encode them. The stratified splitting, intensive hyperparameter optimization, and class-balanced training were used to train a gated Message Passing Neural Network (MPNN) that has a multi-head transformer readout. In order to put the proposed model into perspective, the baseline QSAR and classical machine learning (Random Forest, SVM, XGBoost, Logistic Regression) models were trained on the same dataset split. Marxia improved accuracy and discrimination (AUC-ROC = 0.9627; accuracy = 92.54%) compared to the baselines based on the descriptors, which is explained by the fact that the MPNN-Transformer architecture was more accurate and discriminative. These results indicate the benefit of graph-based learning as compared to rigid molecular fingerprints and descriptors. The model is suitable for integration into early-stage compound screening pipelines in CNS drug discovery.