PIANO, a model for integrating single-cell measurements, separates differences between cells from differences between donors and instruments

PIANO, a model for integrating single-cell measurements, separates differences between cells from differences between donors and instruments

On August 12, the authors described PIANO in a bioRxiv preprint. The model integrates data on gene activity in individual cells. They trained it on the Tahoe-100M atlas, which contains more than 100 million cancer-cell profiles, in 21.8 hours using a single NVIDIA A100 graphics processor.

Single-cell RNA sequencing shows which genes are active in each cell. Researchers use these measurements to build atlases of tissues, tumors, and the brain. However, a single dataset may contain cells from different people, 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 short numerical representation. When it reconstructs a gene activity profile, it receives the experimental conditions separately, including the donor, species, measurement platform, or drug dose. This allows the model to attribute some of the variation to the experimental conditions and use the numerical representation for features of the cell itself. A second network tries to predict the laboratory batch from this representation. A batch is a group of cells processed together in the same run. During training, PIANO weakens this association.

The authors tested whether this approach keeps different cell types separate. They removed one neuronal subtype, eSPN, from the nonhuman primate datasets. eSPN remained in the human dataset, and PIANO kept it separate from related neuronal types shared by all primates.

Using Tahoe-100M, the authors tested whether the model could reproduce changes in gene activity at a specified drug dose. In the A549 cell line, a laboratory model of lung cancer, they simulated increasing doses of trametinib. This drug blocks the MEK protein in a signaling pathway that controls cell growth and division. In the generated profiles, FOS, EGR1, and CCND1 activity decreased as the dose increased, while the activity of EGFR, which lies upstream of MEK in the same pathway, remained unchanged. PIANO therefore reproduced the known gene response to trametinib from sparse data.