癌症的微生物组基础模型：MiFM衍生的连续轨迹和风险时钟用于泛癌种免疫治疗应答预测

Background: The human microbiome shapes cancer risk, progression and response to immunotherapy, but most existing microbiome models are task-specific and fail to generalize across cohorts and sequencing platforms. We developed MiFM (Microbiome Foundation Model), a self-supervised Transformer trained on ~2 million human, animal and environmental microbiome profiles. MiFM encodes communities as multi-level tokens (species, genus and functional pathways) and uses masked reconstruction plus contrastive learning to obtain transferable embeddings that capture taxonomic and functional structure. Here, we evaluated whether MiFM embeddings can support continuous, clinically interpretable readouts of cancer trajectories and immunotherapy response. Methods: From MiFM embeddings we derived two continuous metrics: (1) a disease pseudotime placing each patient along a microbiome-encoded cancer progression trajectory; and (2) a microbiome cancer risk clock estimating microbiome-derived cancer risk and immunotherapy non-response at the individual level. We applied this framework to colorectal cancer (CRC) cohorts spanning normal mucosa, adenoma and carcinoma, and to 12 independent pan-cancer immune checkpoint blockade cohorts (n =1237). We additionally defined a microbiome-based aging acceleration signature and evaluated its association with disease pseudotime, the risk clock and overall survival. Results: In CRC, disease pseudotime recapitulated the adenoma-carcinoma sequence, with median values of 0.15 (normal), 0.38 (adenoma) and 0.72 (carcinoma; p < 0.001), and separated disease states more clearly than conventional stage. The risk clock identified high-risk individuals within histologically normal mucosa whose microbiome embeddings resembled adenomas (AUC = 0.85) or carcinomas (AUC = 0.92), indicating a preclinical window for intervention. Across 15 CRC cohorts (n = 3739), baseline gut microbiome embeddings identified non-responders to immune checkpoint blockade with a median cross-cohort AUC of 0.84 and an AUC of 0.95 in a non-small cell lung cancer cohort. The aging acceleration signature correlated with both disease pseudotime and the risk clock and independently predicted overall survival (HR = 2.52, p < 0.001). Conclusions: To our knowledge, MiFM is the first large-scale microbiome foundation model applied to oncology, yielding continuous microbiome-derived readouts of cancer progression and immunotherapy response. The disease pseudotime and microbiome cancer risk clock capture cancer trajectories beyond conventional staging, enable ultra-early identification of high-risk lesions in normal-appearing mucosa and support robust prediction of immunotherapy resistance across cohorts. This framework delivers concrete microbiome biomarkers and a scalable platform for microbiome-guided precision oncology.