Machine Learning-Guided Design of Rhenium Tricarbonyl Complexes for Next-Generation Antibiotics

The escalating prevalence of antibiotic-resistant bacteria and the increasing complexity of managing severe infections emphasize the critical need for novel and effective antibiotics. Herein, we present a novel computational strategy focused on metal-based antibiotics, specifically rhenium (Re) complexes, for the rational design of next-generation antibacterial agents. Our approach integrates machine learning (ML) classification models to predict antibacterial potency, particularly against multidrug-resistant pathogens. A recognized limitation of conventional ML-driven antibiotic discovery is its dependence on structural similarity to known antibiotics, which hinders the exploration of structurally diverse and innovative antibiotic classes. To address this, we developed predictive ML models based on multi-layer perceptron (MLP) and random forest (RF) algorithms to estimate the minimum inhibitory concentration (MIC) of Re complexes against methicillin-resistant (MRSA) and methicillin-sensitive (MSSA) Staphylococcus aureus strains. Utilizing structural descriptors, these models demonstrated strong predictive performance and were successfully applied to evaluate 26 novel Re complexes. Additionally, Shapley additive explanation (SHAP) analysis provided insights into the structural features influencing antibacterial activity predictions. The study’s outcomes affirm the effectiveness of our ML-guided approach as a promising pathway for the rational, de novo design of potent Re based antibiotics capable of combating antibiotic-resistant bacterial infections.随着抗生素耐药菌的流行率持续攀升，重症感染管理的复杂度日益提升，研发新型高效抗生素的迫切性愈发凸显。在此，我们提出一种针对金属基抗生素的全新计算策略，聚焦于铼（Re）配合物，用于合理设计下一代抗菌剂。我们的方案整合了机器学习（ML）分类模型，用于预测抗菌活性，尤其针对多重耐药病原体。传统基于机器学习的抗生素发现研究存在一项公认局限：其依赖于与已知抗生素的结构相似性，这极大阻碍了对结构多样的创新型抗生素类别的探索。为解决这一问题，我们基于多层感知机（MLP）与随机森林（RF）算法开发了预测性机器学习模型，用于估算铼配合物对耐甲氧西林金黄色葡萄球菌（MRSA）和甲氧西林敏感金黄色葡萄球菌（MSSA）菌株的最低抑菌浓度（MIC）。这些模型利用结构描述符，展现出优异的预测性能，并成功应用于26种新型铼配合物的活性评估。此外，夏普利可加解释（SHAP）分析为解析影响抗菌活性预测结果的结构特征提供了重要洞见。本研究结果证实，我们的机器学习引导方案是一条颇具前景的路径，可用于合理、从头设计能够对抗抗生素耐药菌感染的高效铼基抗生素。