Deep learning prediction of intratumoral Fusobacterium from digital colorectal cancer H&E slides

Intratumoral Fusobacterium species are closely associated with colorectal cancer progression and molecular heterogeneity, yet current sequencing-based approaches lack spatial histopathological context. We aimed to determine whether routine H&E whole-slide images contain latent morphological signatures predictive of microbial colonization. A total of 1,301 CRC patients from four independent cohorts were included: TCGA-CRC ( n = 608), CPTAC-COAD ( n = 102), NCPDC ( n = 298), and CMU ( n = 293). Multiple pathology foundation models, including CONCH, Gigapath, UNI, UNI2, and Virchow_v2, were evaluated for predicting the Fusobacterium genus and five representative species from H&E whole-slide images. The primary microbial labels (> 0) were derived from metagenomic sequencing, and additional label-sensitivity analyses were performed using 0.001% and 0.01% relative abundance thresholds. Ensemble models were further constructed using an averaging fusion strategy. Model performance and ensemble robustness were evaluated using classification metrics, agreement analysis, and Shapley value attribution. UNI and Virchow_v2 demonstrated the most robust performance among single models, with the highest AUC reaching 83.0% for F. massiliense . Ensemble models showed consistently robust performance and high inter-model agreement across different Fusobacterium species, particularly for F. nucleatum (99.2%). Predicted Fusobacterium colonization was consistently enriched in mucinous adenocarcinoma and MSI tumors across multi-cohort association analysis cohorts. Grad-CAM analysis revealed that microbial-associated signals were mainly concentrated in tumor gland lumens and tumor–stroma interface regions, suggesting distinct histomorphological niches associated with microbial colonization. MIL attention heatmaps and representative pathological review indicated that high-attention regions frequently involved tumor-stroma interfaces, inflammatory cell-rich areas, mucinous components, and luminal debris or necrosis-like material, suggesting recurrent histomorphological contexts associated with Fusobacterium -related labels. Our findings demonstrate that routine histopathology contains latent spatial signals associated with intratumoral microbial ecology, highlighting the potential of foundation model–based computational pathology for microbiome characterization and biomarker discovery in colorectal cancer.