Deep Learning Models for Predicting Human Cytochrome P450 Inhibition and Induction

Given the critical roles played by human cytochrome P450
enzymes
(CYPs) in drug metabolism, accurately predicting their potential inhibition
and induction by drugs and drug candidates is a key objective for
improving drug development and safety assessment. Traditional experimental
methods for identifying CYP modulators are labor-intensive and costly,
underscoring the need for efficient in silico prediction models. In
this study, we present an advanced deep learning model for predicting
CYP inhibition, with a primary focus on key enzymes involved in drug
metabolism: CYP3A4, CYP2D6, CYP1A2, CYP2C9, and CYP2C19. This model
integrates deep neural networks with principal component analysis
(PCA) and the synthetic minority oversampling technique (SMOTE), and
it demonstrates excellent predictive performance. Furthermore, we
developed a novel classification model capable of accurately distinguishing
compounds as strong inhibitors, moderate inhibitors, or noninhibitors
for these CYPs, achieving robust and reliable overall performance.
Through statistical analysis, we also identified structural alerts
(SAs) associated with CYP inhibition and strong CYP3A4 induction,
providing a more precise characterization than previous approaches.
Finally, we introduced a novel deep learning-based method specifically
designed to predict human pregnane X receptor (hPXR) activation, a
major mechanism responsible for CYP induction, which also achieved
good performance.

鉴于人类细胞色素P450酶（CYPs）在药物代谢中发挥的关键作用，精准预测药物及候选药物对其产生的潜在抑制与诱导作用，是提升药物开发效率与安全性评价水平的核心目标。传统的CYP调节剂识别实验方法不仅耗时耗力，且成本高昂，这凸显了开发高效虚拟预测模型的迫切需求。本研究提出了一种用于预测CYP抑制作用的先进深度学习模型，重点聚焦于药物代谢相关的关键酶类：CYP3A4、CYP2D6、CYP1A2、CYP2C9及CYP2C19。该模型将深度神经网络与主成分分析（PCA）、合成少数类过采样技术（SMOTE）相结合，展现出优异的预测性能。此外，我们开发了一种新型分类模型，可精准区分化合物对上述CYP酶的作用类型，将其划分为强效抑制剂、中效抑制剂或非抑制剂，整体性能稳健可靠。通过统计分析，我们还识别出与CYP抑制作用及强效CYP3A4诱导作用相关的结构警示（SAs），相较于此前的研究方法，实现了更为精准的特征刻画。最后，我们提出了一种专为预测人类孕烷X受体（hPXR）激活而设计的新型深度学习方法——该激活作用是介导CYP诱导作用的核心机制，该方法同样取得了良好的预测效果。