CancerSTFormer实现spot分辨率空间转录组的多尺度分析并剖析靶向治疗的基因和免疫调控反应

Despite the success of immunotherapy, a significant number of cancer patients still do not respond to it. It is important to understand why treatment fails and to find better treatment for them. We believe that achieving the goal of increasing the efficacy of immunotherapy must start with understanding the spatial microenvironment and the mechanisms of immunosuppression, so that biology-driven treatment strategies can be developed. A growing number of studies have placed increasing importance to the concept of niche, a multicellular environment, and recognize that niches are the primary contributor of tumor phenotypes and clinical outcomes rather than individual cells or cell types. Spatial transcriptomics are adept at providing niche-level gene expression measurements. Given the importance of niches, we believe that a spatially aware foundation model for cancer ST data collections will realize the therapeutic and prognostic potential of these datasets. We propose CancerSTFormer, consisting of a pair of spatially aware transcriptomic foundation models at the 50µm-Local and 250µm-Extended scales. The models possess unique capabilities to recover ligand-target gene relationships, niche-specific differentially expressed genes, and revealing the gene and immune regulatory responses of immune-checkpoint blockade therapies, and other targeted cancer therapies, on patients' tumors given ST profiles through perturbation analysis. In the pre-trained setting, we sought to perturb each gene encoding PD-1, PDL1, and CTLA-4 proteins, respectively PDCD1, CD274, and CTLA4, on TNBC ST profiles to simulate and understand ICB response. In silico perturbations correctly recapitulated anti-tumor immune response in both the long distance (250um Extended model), and short-range settings (50um Local model) according to GSEA. In the fine-tuned setting, we show that the tool can refine anti-PD1, ganitumab, and trebanalib treatment-resistance and sensitivity signatures. From ISPY-2 trial, we derived PD1-responsive and PD1-resistant gene sets, which served as gold standards to train a gene-classifier within for predicting response-associated genes from ST data. We next applied in silico deletion of PDCD1 using this fine-tuned model, and evaluated the results using the PD1-sensitive/resistant gene sets of the Holdout group. When compared against a fine-tuned Geneformer and a norefinement control, both Local and Extended CancerSTFormer variants consistently achieved higher accuracy in predicting resistant and sensitive genes in holdout cohorts across all 3 treatments. Overall, this tool reuses ST data for understanding how gene perturbation impact spatial niches in cancer, while also providing ST-driven, gene-based refinement of treatment-resistance and sensitivity signatures derived from existing bulk transcriptomics.