Bi-MHA 双向交叉注意力模型预测细菌特异性药物代谢
Prediction of bacteria-specific drug metabolism using a cross-attention-based model
研究提出双向交叉注意力架构 Bi-MHA，整合预训练模型给出的原子级分子嵌入与蛋白级细菌嵌入，生成联合表示用于分类，以预测细菌特异性的药物代谢。模型在训练集上采用可靠性分层和 λ 缩放损失函数，在三个未见任务上表现一致：未见药物 ROC AUC 0.690、未见细菌 0.920、未见互作 0.949，在 rank-sum 指标上排名第一。注意力分析既指出此前研究实验验证过的酶促反应位点，也给出候选细菌蛋白，在原子和蛋白两个层面具备可解释性。
来源：Briefings in Bioinformatics · doi.org
Drug metabolism by gut microorganisms is bacteria-specific and can have clinically significant consequences, including prodrug activation, drug inactivation, and the generation of toxic metabolites. However, existing computational approaches have primarily focused on community-level predictions or drug-to-microbe effects, without resolving the metabolic contributions of individual bacterial strains or the bidirectional nature of drug–bacteria interactions, despite advances in computational frameworks. Here, we propose a bidirectional cross-attention architecture, multimodal Bi-MHA, that integrates atom-level molecular embeddings and protein-level bacterial embeddings from pretrained models to generate a combined representation for classification. The proposed model was developed using reliability stratification and a $\mathrm{\lambda}$-scaled loss function on the training dataset and achieved consistent performance across three unseen tasks: unseen drug (receiver operating characteristic area under the curve [ROC AUC]: 0.690), unseen bacteria (ROC AUC: 0.920), and unseen interaction (ROC AUC: 0.949), ranking first among all compared models in the rank-sum metric. Beyond prediction, attention-based analysis not only indicated enzymatic reaction sites that were experimentally validated in previous studies but also prioritized candidate bacterial proteins, demonstrating interpretability at both the atom and protein levels. The proposed model also achieved a 15 percentage point improvement in accuracy over the KNN baseline on an external validation dataset of 85 out-of-distribution drug-bacteria pairs. Furthermore, the model was extended to community-level prediction as a proof-of-concept application using a weighted-abundance pipeline on an independent microbiome dataset. These results demonstrate the model’s potential as an interpretable screening tool for drug safety evaluation and personalized medicine. The source code is available at https://github.com/Jaehyeon-O-Ob/Bacteria-Specific_Drug_Metabolism.