PIANO separates cell signals from batch effects in a 100M-cell sequencing atlas
PIANO, trained on 100 million cancer-cell profiles, separates cell-type signals from donor, lab, and instrument variation in single-cell RNA sequencing data and reproduces the known gene response to trametinib in preprint validation tests.
On August 12, the authors described PIANO in a bioRxiv preprint. The model trained on the Tahoe-100M atlas (more than 100 million cancer-cell profiles) in 21.8 hours on a single NVIDIA A100 GPU. PIANO separates the biological properties of each cell from the noise introduced by differences between donors, laboratories, and instruments.
PIANO works with single-cell RNA sequencing data: a method that shows which genes are active in each individual cell. Researchers use these measurements to build atlases of tissues, tumors, and the brain. The Eternal Search page linked below covers the method in more depth.
The problem: a single dataset may contain cells from different donors, laboratories, and instruments. In a direct comparison, differences caused by these conditions can easily be mistaken for differences between the cells themselves.
PIANO compresses the activity of thousands of genes into a compact numerical representation. When reconstructing a gene activity profile, the model receives experimental conditions as a separate input: donor, species, measurement platform, drug dose. The model attributes part of the variation to those conditions; the numerical representation encodes properties of the cell itself. A second network tries to predict the lab batch (a group of cells processed together in the same run) from the same representation. During training, PIANO weakens that link, so the representation becomes less tied to measurement conditions.
The authors tested the approach two ways. First, they removed the neuronal subtype eSPN from the nonhuman primate datasets while keeping it in human data. PIANO placed eSPN in the representation space separately from related neuronal types shared by all primates. Second, using the A549 cell line (a laboratory model of lung cancer), the model generated gene activity profiles at increasing doses of trametinib. Trametinib blocks the MEK protein in a signaling pathway controlling cell growth and division. In the generated profiles, activity of FOS, EGR1, and CCND1 decreased as the dose increased, while EGFR, which lies upstream of MEK in the same pathway, stayed unchanged. This matches what is known about trametinib's effect on gene expression: from sparse data, the model reproduced the expected pattern of changes.
The eSPN test shows that PIANO preserves cell-type distinctions when integrating data across species. The trametinib test shows the model reproduces known changes in gene expression at specified drug doses. The results are published as a preprint.
Open the related Eternal Search page
Sources
[1] biorxiv.org
The PIANO preprint addresses a core limitation of single-cell RNA sequencing: technical variation from donors, instruments, and lab batches can obscure biological signals at any scale. The Eternal Search glossary page on single-cell RNA sequencing gives readers the methodological grounding needed to evaluate both the problem PIANO targets and the significance of its validation results, making it a direct conceptual anchor for the post.