The World's Largest Atlas of Cancer Cell Drug Responses Passes Peer Review and Appears in Cell
The World's Largest Atlas of Cancer Cell Drug Responses Passes Peer Review and Appears in Cell
On September 17, 2026, the journal Cell published a peer-reviewed paper on Tahoe-100M, an atlas of 100 million single-cell gene-activity profiles from cancer cells, as part of a special issue on modeling the human cell. The atlas had been available as an open but unreviewed preprint since February 2025 and had already become a benchmark dataset for AI models that predict how a cell will respond to a drug without running the experiment. Its collection method has now passed independent scientific review.
The company behind the atlas (then called Vevo Therapeutics, later renamed Tahoe) was solving a problem common at this scale: the more experiments you run, the harder it becomes to tell a drug's real effect apart from random differences between experimental batches, such as different days, reagents, or technicians. Conventionally, each cell line is tested separately, and these technical differences accumulate. The authors pooled cells from roughly fifty cancer lines into a single vessel (a "cell village") and treated the mixture with one drug. After RNA sequencing, they assigned each cell back to its line of origin using the cell's own genetic markers. Because all lines received identical treatment, the remaining differences between them reflect each line's genuine response to the drug.
Pooling cell lines is part of the Mosaic platform, which the same team described in 2024 as a solution to this same batch-effect problem. Lines and compounds were selected on a single principle: maximum diversity of mutations and mechanisms of action, so that AI models trained on these data would more accurately predict the response of a cell line they had never encountered.
Over 24 hours of treatment, the team collected data on how 379 drugs and drug candidates at multiple doses altered gene activity across 50 cell lines representing different cancer types. After quality filtering, 95.6 million cells and 17,813 line-compound combinations remained. This is 31 times more than the previous benchmark single-cell drug screen and 29 times more observations per condition than the leading genetic-knockdown screen. Re-running the same batch of cells yielded 97 to 98 percent concordance: the method is reproducible. The biological drug response itself is far less reproducible: an earlier test on this atlas found it to be near zero.
Dabrafenib, an inhibitor of the mutant BRAF protein, suppressed signaling most strongly in cells carrying BRAF mutations rather than mutations in the neighboring gene KRAS. The compound RMC-6236 showed the mirror pattern, suppressing the same pathway more strongly in KRAS-mutant cells. Both genes operate in the same signaling cascade, where KRAS sits upstream of BRAF, so each inhibitor acts only at its own node: a drug targeting the upstream step is ineffective where the downstream step is broken, and the reverse is equally true. For the first time, the atlas demonstrated this across 47 different genetic backgrounds simultaneously, in a single experiment.
"Our paper on Tahoe-100M is out today in Cell, in a special issue on modeling the human cell. We released Tahoe-100M last year, it has been downloaded more than 600,000 times, and it has become the foundation for most serious efforts to build cell-state models," wrote Tahoe's scientific director Johnny Yu, the paper's corresponding author, on X.
As early as February 2025, Tahoe-100M became the first dataset in the Arc Virtual Cell Atlas, an open project of the nonprofit Arc Institute. The atlas now comprises more than 600 million cell profiles. The same special issue of Cell also carried a paper on LongevityBench, an open AI benchmark for understanding aging biodata.