Tahoe-x1 foundation model for perturbation prediction

Necessary Disclaimers and Legal

The user is responsible for reviewing and complying with the license requirements of the models, notebooks, and data referenced in this documentation.

Users are responsible for the costs associated with analyzing notebooks and its storage in their project spaces.

Instance type availability and pricing are subject to the contract between the user or the user’s organization and DNAnexus.

Citations for Tahoe-x1 model

Paper: https://www.biorxiv.org/content/10.1101/2025.10.23.683759v1

Website: https://www.tahoebio.ai/news/tahoe-x1-blog

Hugging face: https://huggingface.co/tahoebio/Tahoe-x1

Citations for demo dataset

The sci-Plex3 dataset used in these notebooks is sourced from:

• Srivatsan et al., Science 2019. Massively multiplex chemical transcriptomics at single-cell resolution.

• .h5ad files available via scPerturb Single-Cell Perturbation Data on Figshare.

Tahoe-x1 model overview

Tahoe-x1 (Tx1) is pretrained on approximately 266 million single-cell profiles drawn from three sources: Tahoe-100M, Arc scBaseCount, and CZ CELLxGENE. Among these, Tahoe-100M is the largest single-cell perturbation dataset to date, comprising over 100 million high-quality single-cell transcriptomes from 50 diverse cancer cell lines exposed to more than 1,100 small-molecule perturbations. Training on this perturbation-rich data enables Tx1 to learn causal relationships between genes, providing insight into how cells respond to perturbations.

Tx1 is a family of transformer-based foundation models with up to 3 billion parameters, trained using a masked gene-expression prediction task. Each cell is encoded as a fixed-length sequence of 2,050 tokens capturing the overall cell state, chemical perturbation via Morgan fingerprints, and individual gene expression levels. Token embeddings are processed by a 32-layer transformer backbone, which learns gene–gene interactions and context-dependent transcriptional patterns through attention mechanisms.

Therefore, Tx1's large-scale perturbation pretraining and expressive transformer architecture make it a powerful platform for predicting transcriptional responses and accelerating the discovery of therapeutic targets.

Using Tahoe-x1 models on DNAnexus

Copying demo notebooks and example dataset into a project

To explore the demo analyses, copy the notebooks and example dataset from the public project into your own project space:

1. Create a project for your Tahoe-x1 analyses, billed to your own organization. Tutorials on project setup can be found here.

2. Navigate to the Resources tab, locate the project titled "Public Datasets AWS US (East)," and select the "Tahoe-x1" folder.

3. Select the two demo notebooks (.ipynb) and two example datasets (.h5ad).

1. Click "Copy" in the top-right menu and select the project created in Step 1.

2. Return to your project space to begin exploring the notebooks

To run the notebooks, refer to the JupyterLab section of the Academy documentation, which includes instructions for the standard environment. We recommend using the prepared snapshot snapshot-jupyterlab-Tx1-inference-analysis.tar.gz, available in the Notebook_snapshot folder of the public project.

Demo analysis: Tx1-3B Inference and Downstream Analysis

Two demo notebooks are provided to walk through the full Tx1 workflow using a subset of the sci-Plex3 dataset (~159K cells: 141,857 HDAC inhibitor-treated cells across 32 drugs at 4 doses each, and 17,578 controls, from 3 cancer cell lines: A549, K562, and MCF7). The notebooks are designed to be run in order, as the second notebook depends on the output of the first. To facilitate exploration, we also include the precomputed inference output (sciplex3_hdac_with_Tx1-3b_embeddings.h5ad) so the analysis notebook can be run independently.

1. Inference notebook (tahoe_x1_sciplex3_hdac_3b_inference_only.ipynb):

2. This notebook covers the inference pipeline: loading the Tx1-3B model, running it on single-cell RNA-seq data, and saving the resulting per-cell embeddings.

3. Note: Inference takes approximately 11 hours to complete. We recommend running this notebook via non-interactive execution rather than interactively to avoid session timeouts.

4. Analysis notebook (tahoe_x1_sciplex3_hdac_3b_analyses_demo.ipynb):

5. This notebook uses the saved embeddings and demonstrates four downstream analyses: cell line clustering as a sanity check, drug potency estimation validated against published IC50 values, dose-response detection, and drug similarity with gene-level validation.

Running the notebooks

Both notebooks are developed for JupyterLab with Python, R, Stata, ML, Image Processing (version 2.11). When launching JupyterLab, select the ML feature and the snapshot snapshot-jupyterlab-Tx1-inference-analysis.tar.gz.

Notebook	Instance	Runtime
tahoe_x1_sciplex3_hdac_3b_inference_only.ipynb	mem2_ssd2_gpu4_v2_x48	11h30m
tahoe_x1_sciplex3_hdac_3b_analyses_demo.ipynb	mem1_ssd1_v2_x36	2h

The analysis notebook (tahoe_x1_sciplex3_hdac_3b_analyses_demo.ipynb) can be run interactively via JupyterLab.

• Click the Snapshot box and select: snapshot-jupyterlab-Tx1-inference-analysis.tar.gz.

• Choose ML in the Feature box then click Start Analysis.