Derived source data and frozen embeddings for a foundation-model atlas of EBV-associated B-cell states

This dataset contains large derived objects associated with the study “A foundation-model atlas reveals context-constrained host-response geometry across EBV-associated B-cell states.” The deposit includes the frozen scGPT cell-embedding matrix and associated large atlas-derived objects that were intentionally excluded from the GitHub software repository because of file-size considerations. The corresponding software, canonical figure panels, analysis scripts, small derived source-data tables, metadata, and reproducibility documentation are archived separately under Zenodo DOI 10.5281/zenodo.21619106. Primary sequencing data are publicly available from the Gene Expression Omnibus under accession numbers GSE158275, GSE189141, and GSE272763. The locked atlas comprises 194,153 cells across 20 sample units:• GSE158275: 5,704 cells, 2 sample units• GSE189141: 105,830 cells, 8 sample units• GSE272763: 82,619 cells, 10 sample units This record contains derived data only and does not redistribute third-party pretrained model weights.