PatHoST: weakly supervised reconstruction of pathogen spatial distributions from host transcriptomes

Dual host-pathogen spatial transcriptomics enables spatially resolved investigation of infection, yet the completeness of pathogen detection remains poorly characterized. Here, we generated pathogen-only spatial transcriptomic sections from a chronic Mycobacterium abscessus lung infection model to experimentally benchmark pathogen detection, revealing substantially increased pathogen transcript recovery compared with conventional dual profiling. Building on these data, we developed PatHoST, a weakly supervised computational framework that uses pathogen-derived information to generate silver-standard labels and learns host transcriptional programs predictive of pathogen presence. PatHoST reconstructed pathogen spatial distributions across independent tissue sections and enabled the identification of localized host regulatory programs beyond conventional cluster-based analyses. Application to human SARS-CoV-2 lung spatial transcriptomics supported generalization across host species and pathogen classes. Together, PatHoST combines purpose-built experimental benchmarking with weakly supervised inference to recover latent pathogen information and extend the analytical capabilities of dual host-pathogen spatial transcriptomics.