Multi-Modal Foundation · tahoebio
Tahoe-100M-SCVI
A single-cell variational autoencoder mapping drug perturbations across cancer cell lines into a latent space.
What Tahoe-100M-SCVI is
Tahoe-100M-SCVI is a single-cell Variational Inference (scVI) probabilistic model developed by Vevo Tx (tahoebio). It is built to process and represent cellular transcriptomic states from the massive Tahoe-100M dataset, which spans over 100 million single-cell profiles across 50 cancer cell lines treated with roughly 1,100 pharmacological compounds.
The model addresses the challenge of compressing and modelling massive perturbation datasets at single-cell resolution. Computational biologists and machine learning researchers can use it to map scRNA-seq measurements into a compact 10-dimensional representation space or decode latent states back to normalised gene expression profiles for perturbation analysis and drug response modelling.
How it works
The model employs a standard single-cell Variational Inference (scVI) variational autoencoder architecture comprising one hidden layer with 128 units and a 10-dimensional normally distributed latent space. Gene expression count dispersion is modelled at the gene level with a negative binomial likelihood distribution and a 0.1 dropout rate. The full model contains 40,390,510 parameters.
Training was conducted in 32-bit floating point (fp32) precision on the Tahoe-100M single-cell atlas, encompassing 95,624,334 observed cells across 62,710 gene features. Because single training epochs were exceptionally large, reconstruction error (the negative log likelihood of observed counts) was evaluated on training data and an unobserved held-out plate (plate 14) every N minibatches, saving model state snapshots at each epoch.
Key facts
- Model type
- Single-cell Variational Inference (scVI) variational autoencoder
- Parameters
- 40,390,510
- Latent dimensions
- 10
- Hidden units
- 128
- Hidden layers
- 1
- Likelihood model
- Negative binomial (gene-level dispersion)
- Training data
- Tahoe-100M dataset (95,624,334 cells across 50 cancer lines and 1,100 drug perturbations)
- Precision
- fp32
What you can use it for
- Latent Space Cell Encoding
- Projects new or existing single-cell RNA sequencing data into the 10-dimensional Tahoe-100M latent representation space.
- Gene Expression Reconstruction
- Decodes latent cell representations back to normalised expression values across designated gene sets.
- Drug Response Modelling
- Serves as an operational foundation for downstream machine learning algorithms designed to predict cellular responses to therapeutic interventions.
- Model Adaptation
- Supports fine-tuning and further adaptation to incorporate supplementary scRNA-seq datasets.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Trained on an unprecedented dataset of more than 95 million single-cell transcriptomic profiles across 1,100 drug perturbations.
- Requires only around 1 GB of GPU memory to host and run inference with the model itself.
- Provides a calibrated posterior predictive distribution that encompasses observed counts within 95% confidence intervals 97.7% of the time.
Watch out for
- Reconstructed gene counts may be inaccurate, with calibration metrics (97.7%) only marginally beating a naive zero-count baseline (97.4%).
- Trained exclusively on cancer cell lines under chemical perturbation, meaning representations may not directly translate to primary tissues.
- Loading the accompanying minified AnnData reference structure demands 41 GB of storage and system RAM.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · mit
The mit licence carries terms worth reading before you depend on it.
Questions
Can I use Tahoe-100M-SCVI commercially?
Check the licence before commercial use. The mit licence carries terms worth reading before you depend on it.
What software frameworks are required to run Tahoe-100M-SCVI?
The model is built on scvi-tools (version 1.2.0) and uses AnnData (version 0.11.1).
What hardware resources are needed to use this model?
The model itself requires roughly 1 GB of GPU memory. However, loading the accompanying minified AnnData object requires 41 GB of disk storage in the cache directory and 41 GB of system RAM.
Can this model be applied to primary clinical samples?
Because it was trained specifically on immortalised cancer cell lines exposed to pharmacological compounds, it may not be directly applicable to primary cells.
How was the model evaluated during training?
A specific experimental plate (plate 14) was held out from training, and reconstruction error was evaluated periodically across minibatches using negative binomial negative log likelihood.
How to cite it
Zhang, Jesse, Airol A. Ubas, Richard de Borja, Valentine Svensson, Nicole Thomas, Neha Thakar, Ian Lai, et al. 2025. “Tahoe-100M: A Giga-Scale Single-Cell Perturbation Atlas for Context-Dependent Gene Function and Cellular Modeling.” bioRxiv. https://doi.org/10.1101/2025.02.20.639398.
Written from Tahoe-100M-SCVI's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.