Shoombuatong Watshara, Mookdarsanit Pakpoom, Schaduangrat Nalini, Mookdarsanit Lawankorn
Center for Research Innovation and Biomedical Informatics, Faculty of Medical Technology, Mahidol University, Bangkok, 10700, Thailand.
Computer Science and Artificial Intelligence, Faculty of Science, Chandrakasem Rajabhat University, Bangkok, 10900, Thailand.
Sci Rep. 2025 Jul 1;15(1):21402. doi: 10.1038/s41598-025-05839-8.
The glucocorticoid receptor (GR) is a critical nuclear receptor that regulates a broad spectrum of physiological functions, including stress adaptation, immune response, and metabolism. Given the association between aberrant GR signaling and various pathological conditions, this pathway represents a promising therapeutic target. Several GR antagonists have been developed to block glucocorticoid binding to the receptor, showing therapeutic potential in disorders characterized by heightened or dysregulated glucocorticoid signaling. Therefore, this study proposes an innovative deep learning-based hybrid framework (termed BGATT-GR) that leverages a data augmentation method, a bidirectional gated recurrent unit (BiGRU), and a self-attention mechanism (ATT) to attain more accurate identification of GR antagonists. In BGATT-GR, we first employed AP2D, CDKExt, KR, Morgan, and RDKIT to extract molecular descriptors of GR antagonists and combined these molecular descriptors to generate multi-view features. Second, we adopted a data augmentation method that combined both random under-sampling (RUS) and the synthetic minority over-sampling technique (SMOTE) to address the issue of class imbalance. Third, the BGATT architecture was constructed to enhance the utility of the multi-view features by generating informative feature embeddings. Finally, we applied principal component analysis (PCA) to reduce the dimensionality of these feature embeddings and fed the processed feature vectors into the final classifier. Extensive experimental results showed that BGATT-GR provided more stable performance in both cross-validation and independent tests. Furthermore, the independent test results revealed that BGATT-GR attained superior predictive performance compared with several conventional ML models, with a balanced accuracy of 0.957, an MCC of 0.853, and an AUPR of 0.962. In summary, our experimental results provide strong evidence to suggest that BGATT-GR is highly accurate and effective for identifying GR antagonists.
糖皮质激素受体（GR）是一种关键的核受体，可调节广泛的生理功能，包括应激适应、免疫反应和新陈代谢。鉴于异常的GR信号传导与各种病理状况之间的关联，该信号通路是一个有前景的治疗靶点。已经开发了几种GR拮抗剂来阻断糖皮质激素与受体的结合，在以糖皮质激素信号传导增强或失调为特征的疾病中显示出治疗潜力。因此，本研究提出了一种基于深度学习的创新混合框架（称为BGATT-GR），该框架利用数据增强方法、双向门控循环单元（BiGRU）和自注意力机制（ATT）来更准确地识别GR拮抗剂。在BGATT-GR中，我们首先使用AP2D、CDKExt、KR、Morgan和RDKIT来提取GR拮抗剂的分子描述符，并将这些分子描述符组合以生成多视图特征。其次，我们采用了一种结合随机欠采样（RUS）和合成少数过采样技术（SMOTE）的数据增强方法来解决类别不平衡问题。第三，构建BGATT架构以通过生成信息丰富的特征嵌入来增强多视图特征的效用。最后，我们应用主成分分析（PCA）来降低这些特征嵌入的维度，并将处理后的特征向量输入到最终分类器中。大量实验结果表明，BGATT-GR在交叉验证和独立测试中均提供了更稳定的性能。此外，独立测试结果表明，与几种传统机器学习模型相比，BGATT-GR具有卓越的预测性能，平衡准确率为0.957，马修斯相关系数为0.853，精确率均值为0.962。总之，我们的实验结果提供了有力证据，表明BGATT-GR在识别GR拮抗剂方面具有高度准确性和有效性。
