Stack Cellxgene45M: Curated Single-Cell RNA-Seq Collection for Model Alignment

A curated subset of approximately 45 million cells from the CellxGene database, used to align the Stack model after pretraining. The collection was created by arcinstitute and last updated on January 9, 2026. It includes author-annotated coarse-grained cell type labels transferred to the dataset.

Use Cases

• Fine-tuning foundation models for single-cell biology based on the ~45 million cell collection.
• Analyzing cell type composition across datasets based on the transferred author_cell_type annotations.
• Benchmarking model performance on large-scale, multi-donor single-cell data based on the selection criteria of ≥50,000 cells and ≥5 donors per dataset.

Strengths

• Approximately 45 million cells, indicating a large-scale collection.
• Includes author-annotated cell type labels (author_cell_type) for supervised tasks.
• Applies selection criteria for dataset inclusion (≥50,000 cells and ≥5 donors per dataset), which may improve consistency.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Curated from the CellxGene database.

Collection Method

Subset selected based on cell count and donor number criteria; cell type labels were heuristically identified and transferred.

Freshness

Last updated 2026-01-09 13:48:35; freshness should be verified.