Bloom unifies brain cell representations across modalities and contexts
Multimodal genomic datasets now profile mammalian brain cells across anatomical regions, developmental stages, disease states, and species. Learning shared representations that capture cellular diversity across these heterogeneous observations remains a central challenge. Here, we introduce Bloom, a transformer-based model that represents cells using gene expression, chromatin accessibility, or both. Bloom is trained on 300,000 adult mouse brain cells with paired measurements, randomly sampled with broad-subclass balancing--approximately 100-fold fewer cells than used by comparable pre-trained models. By combining genomic sequence embeddings with randomly sampled multimodal cell views, Bloom outperforms unimodal cell models in resolving brain-wide cell types and in annotation, regression, and label-transfer tasks. Its attention patterns reveal cell-type-dependent associations among genomic loci, including pairs separated by millions of base pairs. These results establish Bloom as a data-efficient framework for integrating multimodal brain cell measurements and generating hypotheses about cell-type-specific genomic relationships.