Bloom unifies brain cell representations across modalities and contexts

Bloom, a transformer trained on 300,000 adult mouse brain cells, learns shared multimodal representations that improve brain-wide cell-type resolution and downstream tasks.

• Why it matters: Multimodal genomic profiling across regions, development, disease, and species creates heterogeneous data, and learning unified representations that preserve cellular diversity remains difficult.
• What they did: Bloom uses gene expression, chromatin accessibility, or both, trained with sequence embeddings and randomly sampled multimodal cell views with broad-subclass balancing.
• The result: Bloom outperforms unimodal cell models for cell-type resolution and for annotation, regression, and label-transfer, and its attention links cell types to genomic loci separated by millions of base pairs.

AI summary of the abstract · may contain errors · not medical advice · read the paper.