henhaobio
Sign in

bioRxivGenomicsPreprint

Tahoe-100M: A Giga-Scale Single-Cell Perturbation Atlas for Context-Dependent Gene Function and Cellular Modeling

Jesse M. ZhangiD, Airol A. Ubas, Richard de BorjaiD, Valentine SvenssoniD, Nicole Thomas, Neha Thakar, Ian K. LaiiD, Aidan F. WintersiD, Umair KhaniD, Matthew G. JonesiD, John D. ThompsoniD, Vuong TraniD, Joseph Pangallo, Efthymia PapalexiiD, Ajay Sapre, Hoai Nguyen, Oliver Sanderson, Maria Nigos, Olivia Kaplan, Sarah Schroeder, Bryan HariadiiD, Simone Marrujo, Crina Curca Alec Salvino, Guillermo Gallareta Olivares, Ryan T. KoehleriD, , Johnny X. YuiD

PDFDOI
In-depth full-text digest available · Jump to digest ↓

Abstract

Building predictive models of the cell requires systematically mapping how perturbations reshape each cells state, function, and behavior. Here, we present Tahoe-100M, a giga-scale single-cell atlas of 100 million transcriptomic profiles measuring how each of 1,100 small-molecule perturbations impact cells across 50 cancer cell lines. Our high-throughput Mosaic platform, composed of a highly diverse and optimally balanced "cell village", reduces batch effects and enables parallel profiling of thousands of conditions at single-cell resolution at an unprecedented scale. As the largest single-cell dataset to date, Tahoe-100M enables artificial-intelligence (AI)-driven models to learn context-dependent functions, capturing fundamental principles of gene regulation and network dynamics. Although we leverage cancer models and pharmacological compounds to create this resource, Tahoe-100M is fundamentally designed as a broadly applicable perturbation atlas and supports deeper insights into cell biology across multiple tissues and contexts. By publicly releasing this atlas, we aim to accelerate the creation and development of robust AI frameworks for systems biology, ultimately improving our ability to predict and manipulate cellular behaviors across a wide range of applications.

In-depth summary

Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 3 (10 May 2025)

A 100-million-cell atlas: 50 cancer lines pooled as spheroid villages, treated with 379 drugs at several doses, profiled by combinatorial-barcoding scRNA-seq and embedded with a 10-dimensional scVI model.

Large single-cell perturbation data remain scarce, and pooling across published studies brings batch effects, mismatched protocols and too few cells per condition to resolve how subpopulations respond. Tahoe-100M measures 379 compounds across 47 usable cancer lines inside pooled spheroid villages, so each condition is read out in a shared batch with a median of 1,287 cells.

Key results

  1. ShownThe released resource holds 95.6 million quality-filtered single-cell transcriptomes covering 47 cancer cell lines, 379 compounds and 1,135 drug-dose combinations, with a median of 1,287 cells for each cell line-drug-dose condition.

    Fourteen 96-well plates of pooled spheroids yielded 1,786 sublibraries and roughly 1.4 trillion reads over 153 million cells (8.2-14.5 million per plate; mean 2,288 transcripts per cell, median 1,890); 100.6 million passed minimal and 95.6 million full filters. The atlas spans 52,886 cell line-drug-dose and 17,813 cell line-drug conditions, put at 31-fold more than a benchmark single-cell drug screen. · Generating the Tahoe-100M Atlas using Tahoe…; Cell line and treatment diversity in Tahoe-100M; The global transcriptomic landscape captured by…

  2. ShownCells separated by cell line and cell cycle phase rather than by plate, and a plate rerun as a biological replicate reproduced its pseudobulk profiles closely.

    Plate 14 repeated plate 6 (6,201,077 and 7,285,856 cells respectively; 4,796 shared cell line-treatment pairs): quartiles of Pearson correlation between log pseudobulk profiles were 0.97/0.975/0.98 for matched pairs against 0.892/0.915/0.929 for unmatched ones. Local inverse Simpson's index over 100 subsamples of 10,000 cells put cell line identity first and cell cycle phase second, with drug and dose weaker. · The global transcriptomic landscape captured by…; Drug-induced transcriptomic effects across cell…; Replication analysis; Single cell gene expression dimensionality…

  3. ShownPerturbation magnitude in the scVI latent space grew with dose and sat between the effects of knocking down non-essential and essential genes, with proteasome, HDAC and PI3K/AKT inhibitors among the strongest classes.

    E-distances from DMSO controls were computed per cell line in the 10-dimensional scVI space and set against Sciplex3 and the Replogle 2022 genome-scale CRISPRi screen, both reprocessed the same way from scPerturb; drug effects were intermediate relative to CRISPRi genes split by essentiality. Harringtonine, homoharringtonine and dinaciclib were outliers. Numeric values sit in figures absent from this text. · Drug-induced transcriptomic effects across cell…; E-distance

  4. ShownTranscriptional response tracked driver genotype: dabrafenib suppressed a RAS/RAF signature specifically in BRAF-V600E lines, while the pan-RAS inhibitor RMC-6236 moved it most in KRAS-mutant lines without reaching conventional significance.

    Differential Vision scores for the MSigDB KRAS.600.LUNG.BREAST.V1 signature, lines split into KRAS-mutant, BRAF-V600E and RAS/RAF wild-type: dabrafenib gave lower scores in BRAF-mutant lines at the lowest (P = 0.0013) and highest (P = 0.00023) dose by Welch's t-test, with little change in KRAS-mutant lines. For RMC-6236 the KRAS-mutant versus wild-type contrast reached p = 0.065 and p = 0.063. · Context-dependent inhibition of the RAS/RAF…; Gene-set signature analysis

  5. ShownCell cycle redistribution separated compounds nominally in the same class: palbociclib enriched G1 and dinaciclib G2/M, while several other CDK, HDAC and microtubule agents barely shifted the phase distribution or shifted it the other way.

    Log odds ratios of phase fractions versus DMSO, per cell line and plate, from Scanpy's score_genes_cell_cycle with the Regev lab signature: abemaciclib, AT5719 and ribociclib shifted little; among HDAC inhibitors belinostat and panobinostat drove G2/M, carbamazepine hardly moved and tucidinostat went negative; paclitaxel and Tubulin Inhibitor 6 were microtubule exceptions. Effect sizes are in missing figures. · Context-dependent inhibition of the RAS/RAF…; Cell cycle phase analysis

Methods, data and reproducibility

Fifty cancer cell lines were pooled into spheroid cell villages, dosed well by well for 24 h across 14 96-well plates, then fixed and profiled by Parse GigaLab combinatorial-barcoding scRNA-seq on Ultima sequencers. Cells were traced back to their line by SNP demultiplexing, embedded with a 10-dimensional scVI model, and summarised with E-distance, Vision gene-set scores and cell cycle scoring.

  • Tahoe-100M scRNA-seq profiles of drug-treated spheroid cell villages, after full QC filtering · human · n = 95,624,334 cells (100.6 million at minimal filters; 153 million sequenced)
  • Cancer cell lines from ATCC or NCI-43, drawn from 13 organs; 47 had enough representation for analysis and three were dropped for low counts · human · n = 50 cell lines
  • Compounds from MedChemExpress; 180 assigned to 25 mechanism-of-action classes and reportedly hitting 325 genes, 69% of them approved agents · n = 379 drugs, 1,135 drug-dose combinations
  • Sciplex3 single-cell drug screen, reprocessed from the scPerturb collection for the E-distance comparison
  • Replogle 2022 genome-scale CRISPRi screen, reprocessed the same way for the E-distance comparison
Methods & tools
Parse Biosciences GigaLab and Evercode fixation, split-pipe v1.4.0, demuxlet, scVI (scvi-tools), Vision, Scanpy, cuML t-SNE (NVIDIA RAPIDS), neural network multidimensional scaling, MSigDB v2023.1, GPT-4o, MedChemExpress, Ultima Genomics UG-100
Reagents & environment
A modified split-pipe v1.4.0 for GigaLab libraries, with gene models from Ensembl release 109 on genome build 38; scvi-tools, Scanpy, cuML from NVIDIA RAPIDS and MSigDB v2023.1. No container or requirements file is mentioned.
Assay conditions & compute
The scVI model spans 62,710 genes for 95,624,334 cells and was trained for 10 epochs on the 89,423,257 cells of plates 1-13, over 800 million single-cell examples; in minified form it takes 41 GB of storage. No hardware or runtime is stated.
Reproducibility gaps
Data are said to be on HuggingFace, but the link is missing from the extracted text and no code repository is named. Figures and result tables are also absent, so E-distance values, LISI estimates and cell cycle effect sizes cannot be quoted.
Limitations (5)
  • Coverage is uneven: three cell lines (NCI-H661, NCI-H596, NCI-H2122) were dropped for consistently low counts, and any cell line-drug condition with fewer than 50 cells was excluded. · stated by the authors
  • Genetic demultiplexing can call cells doublets when coverage or SNP support is low, so some discarded cells may be genuine singlets; the benchmark on known lines gave over 98% accuracy with more than 90% of cells called singlets. · stated by the authors
  • Cell line identity and cell cycle phase dominate the embedding over treatment and dose, so drug effects only emerge through differential expression analysis rather than directly in the latent space. · stated by the authors
  • One 24-hour timepoint in pooled spheroids of cancer cell lines: no primary cells, no time course and no genetic perturbations; reproducibility rests on a single duplicated plate (14 versus 6). · noted in this summary
  • Drug targets, mechanism classes and approval status were annotated by prompting GPT-4o over vendor descriptions, with only 97 of 380 compounds inspected by hand. · noted in this summary

Read this if you train or benchmark perturbation-response models and need to know how Tahoe-100M was built, what it covers, and where its annotations and controls are weakest.

Citations
108

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISGandhi '25Su '26Wang '26Yang '24Ho '24Li '26
63% semantic matchSystems Biology
Shreshth Gandhi, Farnoosh Javadi, Valentine Svensson+6 more·bioRxiv·Oct 23, 2025