Single-Cell Biology · tahoebio
Tahoe-x1-embeddings
Pre-computed single-cell embeddings from the Tahoe-x1-3B foundation model across cancer cell lines and small molecules.
What Tahoe-x1-embeddings is
Tahoe-x1-embeddings is a dataset of pre-computed high-dimensional representations of single-cell transcriptomic profiles. Published by tahoebio, it applies the Tahoe-x1-3B single-cell foundation model to profiles from the Tahoe-100M dataset.
The dataset covers cellular states across 50 cancer cell lines representing multiple tissue types perturbed by approximately 1,100 small-molecule compounds with diverse mechanisms of action. It spans over 100 million single-cell profiles.
By providing pre-computed embeddings, the resource enables researchers to carry out downstream analyses without the computational overhead of reprocessing raw single-cell gene expression data.
How it was made
Cell representations are derived by passing single-cell transcriptomic profiles from the Tahoe-100M dataset through the Tahoe-x1-3B model, a 3 billion parameter foundation model trained on cellular perturbations.
The resulting data are packaged into Parquet format. Each entry contains identifiers for the drug compound, sample, Cellosaurus cell line, and sub-library barcode, alongside the cell embedding vector stored in the column named mosaicfm-3b-prod-cont-MFMv2.
Key facts
- Generating model
- Tahoe-x1-3B (3 billion parameters)
- Source data
- Tahoe-100M dataset
- Coverage
- 50 cancer cell lines, ~1,100 small-molecule compounds, 100+ million single-cell profiles
- Data format
- Parquet
- Embedding vector column
- mosaicfm-3b-prod-cont-MFMv2
- Primary task category
- Feature extraction
What you can use it for
- Drug response prediction
- Use high-dimensional cellular representations to build predictive models of how various cancer cell lines respond to small-molecule compounds.
- Cell state classification
- Train classifiers on pre-computed embedding vectors to identify cellular phenotypes across 50 distinct cancer cell lines.
- Perturbation effect analysis
- Characterise and compare transcriptional shifts driven by around 1,100 chemical compounds across multiple tissue types.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from tahoebio/Tahoe-x1-embeddings at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 1,494,131 rows of default/train
BARCODE_SUB_LIB_ID
string_text- min
- 19
- median
- 19
- mean
- 19
- max
- 19
cell_line
string_label- CVCL_05466%
- CVCL_04595%
- CVCL_04804%
drug
string_label- DMSO_TF2%
- Clonidine (hydrochloride)2%
- Quinestrol2%
sample
string_label- smp_18752%
- smp_18082%
- smp_18692%
Strengths and limits
Good at
- Eliminates the computational cost of running inference with a 3 billion parameter model over 100 million single-cell profiles.
- Spans a diverse range of biological conditions with 50 cancer cell lines and approximately 1,100 distinct small-molecule compounds.
- Stored in Parquet format, allowing memory-efficient streaming and loading of very large data partitions.
Watch out for
- Standard schema parsing may encounter errors when loading, requiring direct Parquet file loading.
- The dataset contains only cell-level embeddings and basic identifiers, requiring manual joins with Tahoe-100M to access full compound and cell line metadata.
Questions
Can I look inside Tahoe-x1-embeddings without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is Tahoe-x1-embeddings under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
What is the source of the cells and perturbations in this dataset?
The embeddings are calculated from the single-cell transcriptomic profiles in the Tahoe-100M dataset, which includes 50 cancer cell lines perturbed by about 1,100 small-molecule compounds.
Which model was used to compute the embeddings?
The embeddings were generated using Tahoe-x1-3B, a 3 billion parameter perturbation-trained foundation model.
What identifiers are included with each embedding vector?
Each row includes the drug compound name, sample identifier, Cellosaurus cell line identifier, and a 19-character sub-library barcode identifier.
How can additional metadata be linked to these embeddings?
The dataset can be joined with Tahoe-100M drug and cell line metadata tables using the drug name and cell line identifier fields.
Written from Tahoe-x1-embeddings's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.