Tahoe-100M-SCVI

Tahoe-100M-SCVI is a released scVI (single-cell variational inference) checkpoint trained by Tahoe Therapeutics on the Tahoe-100M drug-perturbation atlas. scVI is a probabilistic deep generative model — a variational autoencoder that learns a low-dimensional latent representation of gene expression while explicitly modeling count noise and batch effects — introduced by Romain Lopez, Jeffrey Regier, Michael Cole, Michael Jordan, and Nir Yosef in Nature Methods in 2018. This checkpoint applies that established method to a single, purpose-built corpus: the 100-million-cell Tahoe-100M compendium of cancer cell lines profiled under small-molecule treatment.

The model provides a compact, reusable embedding of Tahoe-100M cell states. It maps each single-cell transcriptome into a 10-dimensional latent space and can decode latent vectors back to gene expression, giving researchers a fixed representation of the atlas without retraining scVI from scratch. Because it is trained end-to-end on the drug-perturbation data itself, its latent space is tuned to the transcriptional variation present across the atlas's cell lines and treatments rather than to a general human reference.

Tahoe-100M-SCVI is best understood as a companion embedding model to the Tahoe-100M dataset (Zhang et al., 2025), which measures how roughly 1,100 small-molecule perturbations affect approximately 50 cancer cell lines. It is distinct from Tahoe-x1, the transformer-based single-cell foundation model that Tahoe Therapeutics later trained on the same atlas: Tahoe-100M-SCVI is a standard VAE baseline, not a compound-aware foundation model, and it does not take drug identity as an input.

#Key Features

• scVI embedding of Tahoe-100M: The model encodes single-cell transcriptomes into a 10-dimensional latent cell-state space learned directly from the drug-perturbation atlas, and can decode those latents back to gene expression.
• Count-appropriate generative model: As an scVI variational autoencoder, it models raw counts with a negative-binomial likelihood and captures technical variation, providing a posterior predictive distribution over expression rather than point estimates.
• Query adaptation: The checkpoint supports encoding new scRNA-seq data into the Tahoe-100M representation space and adapting to additional datasets, reusing the atlas-scale latent space instead of training from scratch.
• Compact and openly licensed: With roughly 40.4 million parameters and an MIT license, the checkpoint is lightweight to run and freely reusable through the scverse scvi-tools ecosystem.

#Technical Details

Tahoe-100M-SCVI is a single-hidden-layer scVI variational autoencoder with 128 hidden units and a 10-dimensional latent space, totaling 40,390,510 parameters across a vocabulary of 62,710 genes. It was trained on 95,624,334 cells from the Tahoe-100M atlas using the scvi-tools .train() method at fp32 precision, with one experimental plate (plate 14) held out for evaluation and reconstruction error monitored per minibatch rather than per epoch given the dataset scale. On its posterior predictive check, reconstructed expression contained the observed counts within the 95% confidence interval 97.7% of the time; the authors note that a trivial all-zero baseline reaches 97.4% on the same sparse-count metric, so calibration should be read alongside that reference point.

#Applications

The checkpoint is aimed at computational biologists analyzing drug-perturbation responses and at machine-learning researchers building drug-response prediction methods on top of the Tahoe-100M atlas. It offers a ready-made cell-state embedding for clustering, visualization, and integration of the atlas's cancer cell lines, and a starting representation for downstream perturbation-modeling work. Because it is trained specifically on cancer cell lines under treatment, it is not intended as a general reference for primary-tissue scRNA-seq.

#Impact

Tahoe-100M-SCVI provides a standardized, openly licensed baseline embedding for one of the largest single-cell perturbation datasets, lowering the barrier for groups that want a consistent scVI representation of Tahoe-100M without recomputing it. Its role is that of a well-documented reference checkpoint rather than a methodological advance: the architecture is the established scVI method, and its value comes from the scale and specificity of the corpus it is trained on. It complements the compound-aware Tahoe-x1 foundation model and the Tahoe-100M dataset release, and its stated cancer-cell-line focus and the modest margin of its calibration over a null baseline are honest limits on how far its representation should be extrapolated.