Abstract Brain metastases (BrM) are the most common adult brain tumors, affecting hundreds of thousands of patients worldwide annually. International collaborations have greatly accelerated research into clinical trial design and emerging treatment modalities for BrM, e.g. the CIMARa, INTERNEO, and RENACER consortia. The Glioma Longitudinal AnalySiS consortium has demonstrated the power of longitudinal genomics to drive translational discovery in primary brain tumors. Inspired by these efforts, we endeavor to develop an international consortium of longitudinal genomics data derived from primary tumors and patient-matched BrMs. To that end, we profiled N=132 BrMs from 10 primary-cancer types from 90 patients at the Peking Union Medical College and 16 patients from the University of California, San Francisco via single-nucleus RNA and epigenetic sequencing. For 36 cases, we also profiled patient-matched primary tumors. Whole-exome and spatial profiling were performed for select cohorts. We fine-tuned a generative-AI foundation model on our single-cell data and found that it was capable of robust data integration, gene-perturbation analysis, subtype prediction, and other predictive tasks. Although the composition of BrMs varied across primary-tumor types, our model identified reproducible meta-programs of gene expression found in all BrMs. Subpopulations of BrM cells from all primary-tumor types adopted a neuronal gene-expression program (e.g., NAV2, NRCAM). Neuronal expression significantly correlated with immunosuppressive myeloid and fibroblast-like cells, and anticorrelated with cytotoxic T cells. The neuronal signature was prognostic in non-small cell lung cancer and HER2+ breast cancer, and negatively correlated with time to metastasis. The neuronal signature score was significantly associated with increased numbers of copy-number variants. We and others have recently demonstrated fibroblast-like cells in glioblastomas, enriched in a perivascular space and with an immunosuppressive phenotype. We demonstrate fibroblast-like cells in BrMs with similar phenotypes and spatial distribution. Moreover, these studies demonstrate the utility of foundation models for integrating data collected at multiple institutions.
Geng et al. (2025) studied this question.