Kavli Affiliate: Wei Min
| Authors: Beimin Liu, Chao Wang, Qingyang Zeng, Min Wei, Weilai Lu, Xueyan Gao, Jing Wan, Jie Feng and Yu Vincent Fu
| Summary:
Multidrug-resistant hypervirulent Klebsiella pneumoniae (MDR-hvKP) poses a severe global health threat. Phage therapy is a promising alternative, but requires precise matching of phage to the bacterial strain. Here, we present a proof-of-concept method that integrates single-cell Raman spectroscopy with deep learning to enable rapid and precise selection of lytic phages against MDR-hvKP. By profiling Raman signatures of strains across multiple KL-types (capsule locus types), we trained three deep learning architectures for phage-host matching. Among them, the CNN_MLP-Transformer achieved the best performance (99.7%), slightly outperforming CNN_MLP (99.2%) and CNN_MLP-Attention (99.5%). Validation using 10 hvKP clinical isolates yielded an average phage selection accuracy of 78.3%. These findings demonstrate the feasibility and clinical potential of AI- augmented Raman spectroscopy as a rapid, label-free, precise strategy for guiding phage therapy against MDR-hvKP infections. Teaser AI-guided single-cell Raman profiling enables rapid precision phage selection against multidrug-resistant K. pneumoniae.