CELLama-Perturb：一种用于绘制肿瘤空间异质性中药物敏感性图谱的虚拟细胞建模方法

Background Predicting how anticancer drugs reshape cellular states in real tissue remains challenging. Perturbation datasets with transcriptomic profiling enable in vitro models of drug response, but how such models transfer to the human tumor microenvironment and capture spatial heterogeneity is unclear. There is a need for frameworks that integrate large-scale perturbation screens with spatial transcriptomics to map intratumoral variation in drug sensitivity. Method We developed CELLama-Perturb, a perturbation modeling framework that builds on CELLama-derived foundation embeddings generated by a sentence-transformer cell-embedding strategy. A memory-mapped data pipeline enabled large-scale training on the Tahoe-100M resource. Drug effects were learned from (i) DepMap PRISM drug-sensitivity profiles to model viability responses across cell lines and (ii) paired perturbed transcriptomes used to predict ranked lists of top differentially expressed genes. Drug identity, dose, and cellular state were embedded jointly, and a cross-attention architecture modeled drug-cell interactions. Trained models were then transferred to single-cell and spatial transcriptomics datasets from human tumors to infer spatially resolved drug-sensitivity maps from baseline gene expression. Results Across PRISM-derived models, CELLama embeddings captured biologically meaningful expression structure and enabled consistent prediction of drug sensitivity in held-out cell lines, with positive correlations between predicted and observed responses. For gene-expression perturbation tasks, the CELLama-based model accurately reconstructed ranked perturbed-gene profiles (nDCG@16=0.314, MRR=0.660, Recall@16=0.272), corresponding to ~4-5 true perturbed genes among the top 16 predictions out of 17,739 genes. Predicted top-gene sets aligned with experimentally observed differentially expressed genes, indicating that the model encodes transcriptional context relevant to downstream perturbation effects. Applied to lung cancer spatial transcriptomics, CELLama-Perturb produced high-resolution drug-sensitivity maps that revealed marked intratumoral heterogeneity; for example, regions predicted to be sensitive to tubulin inhibitors were spatially distinct from those sensitive to topoisomerase inhibitors, highlighting drug-class-specific vulnerability patterns within the same tumor. Conclusion CELLama-Perturb provides a foundation-model-based approach for projecting learned perturbation effects from in vitro systems into complex tissue environments. By generating spatially resolved drug-sensitivity maps, this framework enables virtual cell-to-tissue assessment of therapeutic response and supports refined decisions on drug modality selection, including antibody-drug conjugate payload prioritization.