Title: Enhancing Antimicrobial Molecule Prediction via Dynamic Routing Capsule Networks and Multi-Source Molecular Embeddings Show Chinese title

Title: 通过动态路由胶囊网络和多源分子嵌入增强抗菌分子预测

Abstract: Antibiotics are a vital class of drugs closely associated with the prevention and treatment of bacterial infections. Accurate prediction of molecular antimicrobial activity remains a key challenge in the pursuit of novel antibiotic candidates. However, laboratory-based antimicrobial compounds identification is costly, time-consuming, and prone to rediscovering known antibiotics, highlighting the urgent need for efficient and accurate computational models. Recent advances in machine learning (ML) and deep learning (DL) have significantly enhanced the ability to explore chemical space and identify potential antimicrobial compounds. In this study, we particularly emphasize deep learning models and employ five chemistry language models tailored for chemical data to encode small molecules. Our model incorporates a unique capsule network architecture and introduces innovations in loss function selection and feature processing modules, demonstrating superior performance in predicting inhibitory activities against Escherichia coli and Acinetobacter baumannii. We conducted a series of ablation studies to elucidate the contributions of network design and input features. Case studies validated the usability and effectiveness of our model.To facilitate accessibility, we developed an intuitive web portal to disseminate this novel tool. Our results indicate that the proposed approach offers improved predictive accuracy and enhanced interpretability, underscoring the potential of interpretable artificial intelligence methods in accelerating antibiotic discovery and addressing the urgent challenge of antimicrobial resistance. Show Chinese abstract

Abstract: 抗生素是一类与预防和治疗细菌感染密切相关的关键药物。 准确预测分子的抗菌活性仍然是寻找新型抗生素候选物过程中的一个主要挑战。 然而，基于实验室的抗菌化合物鉴定成本高、耗时长，并且容易重新发现已知的抗生素，这突显了对高效且准确的计算模型的迫切需求。 机器学习（ML）和深度学习（DL）的最新进展显著增强了探索化学空间和识别潜在抗菌化合物的能力。 在本研究中，我们特别强调深度学习模型，并采用五种针对化学数据优化的化学语言模型来编码小分子。 我们的模型结合了独特的胶囊网络架构，并在损失函数选择和特征处理模块方面引入了创新，展示了在预测大肠杆菌和鲍曼不动杆菌抑制活性方面的优越性能。 我们进行了一系列消融实验，以阐明网络设计和输入特征的贡献。 案例研究验证了我们模型的可用性和有效性。为了便于使用，我们开发了一个直观的网络门户以传播这一新工具。 我们的结果表明，所提出的方法提供了改进的预测准确性并增强了可解释性，强调了可解释的人工智能方法在加速抗生素发现和应对抗菌药物耐受性这一紧迫挑战方面的潜力。