目的 建立基于双向门控循环单元（bidirectional gated recurrent unit，BiGRU）与时序注意力机制的可解释深度学习模型，解析补阳还五汤干预脑缺血过程中多成分-多靶点的动态时空响应机制。方法 构建大鼠大脑中动脉闭塞（middle cerebral artery occlusion，MCAO）模型，于给药后多个时间点采集血浆和脑组织样本，分别采用UPLC-MS/MS测定14种补阳还五汤活性成分血药浓度，ELISA检测15个脑组织关键靶点表达水平，构建多成分-多靶点时序数据集。基于滑动窗口策略建立BiGRU-时序注意力网络，并结合留一批次交叉验证（leave-one-batch-out cross-validation，LOBO-CV）及Bootstrap重抽样评价模型性能与稳定性，同时解析模型提示的高贡献时间点、潜在相关成分及响应靶点特征。结果 模型在LOBO-CV中表现出良好的预测性能，各测试集R²均大于0.80，均方误差（mean squared error，MSE）均小于0.20。时间注意力分析识别给药后10 h为模型关注度较高的时间节点（注意力峰值为0.602 6）；梯度归因分析筛选出丹皮酚、毛蕊异黄酮和芒柄花黄素为主要驱动成分，脑源性神经营养因子（brain derived neurotrophic factor，BDNF）、蛋白激酶G II型（protein kinase G type II，PKG2）和钙调蛋白依赖性蛋白激酶II（calmodulin-dependent protein kinase II，CAMK2）为关键响应靶点。Bootstrap重抽样结果显示，10 h关键时间窗的识别频率达到90%，提示模型具有较好的稳定性。结论 建立了基于BiGRU与时序注意力机制的补阳还五汤多成分-多靶点动态时序解析方法，尝试为补阳还五汤抗脑缺血作用机制研究提供新的时序证据，并为中药复方动态药理机制研究提供参考。

Objective To establish an interpretable deep learning model based on a bidirectional gated recurrent unit (BiGRU) and a temporal attention mechanism to elucidate the dynamic spatiotemporal response mechanisms of the multi-component–multi-target interactions underlying the effects of Buyang Huanwu Decoction (BYHWD) against ischemic stroke (IS). Methods A rat model of middle cerebral artery occlusion (MCAO) was established. Plasma and brain tissue samples were collected at multiple time points after administration. Plasma concentrations of 14 active components of BYHWD were quantified by UPLC–MS/MS, and the expression levels of 15 key targets in brain tissue were determined by ELISA to construct a multi-component–multi-target time series dataset. A BiGRU-based temporal attention network was established using a sliding window strategy, and leave-one-batch-out cross-validation (LOBO-CV) combined with bootstrap resampling was employed to evaluate model performance and stability. The model was further interpreted to identify highly contributive time points, potential key components, and responsive target features. Results The model demonstrated good predictive performance in LOBO-CV, with an R² value greater than 0.80 and a mean squared error (MSE) below 0.20 across all test sets. Temporal attention analysis identified 10 h post-administration as a time point receiving high model attention, with an attention peak value of 0.602 6. Gradient attribution analysis identified paeonol, calycosin, and formononetin as the major driving components, while brain derived neurotrophic factor (BDNF), protein kinase G type II (PKG2), and calmodulin-dependent protein kinase II (CAMK2) were identified as the key responsive targets. Bootstrap resampling showed that the 10 h critical time window was identified with a frequency of 90%, indicating good model stability. Conclusion This study established a BiGRU and temporal attention-based approach for dynamic time series analysis of the multi-component–multi-target actions of BYHWD. The proposed method provides new temporal evidence for elucidating the mechanisms underlying the anti-IS effects of BYHWD and offers a reference for investigating the dynamic pharmacological mechanisms of traditional Chinese medicine formulas.

TP18;R285

国家自然科学基金项目（82274215）；湖南省自然科学基金项目（2022JJ30453）；湖南省研究生科研创新项目（CX20251190）；湖南中医药大学2022年度校级“揭榜挂帅”专项（2022-XJJB001）