{
  "abstract": "Background The tumor microenvironment (TME) plays a crucial role in determining patient response to cancer immunotherapies. Cellular therapies, immune checkpoint inhibitors, and other modalities are limited for most solid tumors, especially those deemed ‘immune cold’. The TME suppresses the infiltration and activity of immune cells, and overcoming immunosuppressive TME signals is a major barrier to cancer immunotherapy. The landscape of immunotherapies is maturing and alternative strategies to discover and validate therapies in silico are needed.Spatial omics capture the spatial organization of cells, revealing their spatial localization with tens to thousands of molecular signals. T-cell infiltration is modulated by such signals as chemokines, adhesion molecules, and more. Recent computational work on multiplexed tumor images has focused on predicting patient-level phenotypes using spatial motifs from TMEs, but computational tools that can predict immune cell localization from spatial data are needed to systematically generate specific, feasible tumour perturbations to alter the TME and improve patient outcomes.Methods Counterfactual explanations (CFEs) can help clarify a machine learning image analysis model’s decisions by exploring the model’s response to alterations in the image. 1 In multiplexed tissue images where each pixel contains molecular information, image variations correspond to molecular interventions. Thus, spatial omics data enable the extension of CFEs from understanding to predicting actionable interventions. Here we introduce Morpheus,2 an integrated deep-learning framework that leverages spatial omics data to predict T-cell infiltration, and combines this prediction task with counterfactual optimization to propose tumor perturbations that are predicted to boost T-cell infiltration.Results We trained a convolutional neural network to predict T-cell infiltration using spatial protein data sets ( figure 1a). We applied Morpheus to melanoma, breast cancer, and colorectal cancer (CRC) with liver metastases data to discover perturbations that are predicted to support T-cell infiltration in tens to hundreds of patients (figure 1b). For patients with melanoma, Morpheus predicted that combinatorial perturbation to the CXCL9, CXCL10, CCL22 and CCL18 levels can convert immune-excluded tumours to immune-inflamed in a cohort of 69 patients. For CRC liver metastasis, Morpheus discovered two cohort-dependent therapeutic strategies consisting of blocking different subsets of CXCR4, PD-1, PD-L1 and CYR61 that were predicted to improve T-cell infiltration in a cohort of 30 patients. We experimentally validated Morpheus’ predictions (figure 1c) by showing that perturbing these targets substantially enhanced T-cell migration in vitro.Conclusions Our work provides a paradigm for counterfactual-based prediction and design of cancer therapeutics based on classification of immune system activity in spatial omics data.References Chang C-H, Creager E, Goldenberg A, Duvenaud D. Explaining image classifiers by counterfactual generation. In International Conference on Learning Representations (ICLR) (2019).Wang ZJ, Farooq AS, Chen YJ, Bhargava A, Xu AM, Thomson MW. Identifying perturbations that boost T-cell infiltration into tumours via counterfactual learning of their spatial proteomic profiles. Nat. Biomed. Eng. 2025;9:390–404.Abstract 1131 Figure 1Morpheus Strategy. A. Spatial data is used to train a classifier for predicting CD8+ T cell infiltration. B. The classifier helps predict molecular perturbations to increase immune infiltration. C. Morpheus predictions were validated in vitro",
  "authors": [
    {
      "affiliations": [
        "Westlake University, Zhejiang, Hangzhou, China"
      ],
      "name": "Zitong Wang"
    },
    {
      "affiliations": [
        "Caltech, Pasadena, CA, USA"
      ],
      "name": "Abdullah Farooq"
    },
    {
      "affiliations": [
        "Caltech, Pasadena, CA, USA"
      ],
      "name": "Yu-Jen Chen"
    },
    {
      "affiliations": [
        "Caltech, Pasadena, CA, USA"
      ],
      "name": "Aman Bhargava"
    },
    {
      "affiliations": [
        "University of Maryland, College Park, College Park, MD, USA"
      ],
      "name": "Alexander M Xu"
    },
    {
      "affiliations": [
        "Caltech, Pasadena, CA, USA"
      ],
      "name": "Matthew Thomson"
    }
  ],
  "title": "1131 Spatially-guided therapy predictions with morpheus",
  "uid": "f9b36082-55e5-5c26-be1e-10f24a0f7adb"
}