Spatial metabolomics-guided deep learning framework for dissecting immune evasion mechanisms in fungal infection microenvironments

Invasive fungal infections cause high mortality in immunocompromised patients, yet the spatially organized metabolic circuits that enable pathogens to evade host immunity remain poorly understood. Here we present an integrated computational–experimental framework coupling high-resolution spatial metabolomics with attention-enhanced deep learning to dissect immune evasion mechanisms within fungal infection microenvironments. We constructed a spatial metabolic atlas from MALDI-MSI and DESI-MSI data acquired on murine invasive pulmonary aspergillosis tissue sections, identifying 1,847 reproducible metabolite features across histologically distinct tissue compartments. A multi-scale convolutional network with fused channel and spatial attention modules was designed for automated segmentation of infection core, immune infiltration zone, and a transitional boundary zone, not previously resolved spatially in this infection model, that is enriched in immunosuppressive metabolites including kynurenine, prostaglandin E₂, and gliotoxin, achieving a mean Dice coefficient of 0.887 against expert pathologist annotations. To reconstruct the dynamic regulatory wiring underlying immune evasion, we developed a variational graph autoencoder coupled with neural ordinary differential equations, which inferred temporally evolving metabolite–immune interaction networks across 24, 48, and 72 h post-infection with an area under the precision-recall curve of 0.782 against KEGG and Reactome benchmarks. Network analysis ranked kynurenine as the highest-centrality hub metabolite at 48–72 h, and Neural ODE interpolation placed a putative metabolic tipping point between 36 and 48 h, marked by abrupt activation of tryptophan catabolism edges and a concurrent loss of Th1-associated signaling connectivity. Ablation experiments showed that the dual attention mechanism contributed the largest gain in segmentation and the second-largest in network inference, after the Neural ODE temporal coupling. This framework offers a hypothesis-generating, data-driven paradigm that recasts immune evasion as a spatially encoded and temporally dynamic process, and it yields prioritised candidate targets and therapeutic windows that remain to be tested by perturbation experiments before any antifungal intervention strategy can be built on them.