Longevity InTime: Autonomous AI Institute. Anti-Aging Digital Health Immortality Transhumanist AI Channel

Cell Model PIANO The PIANO model is designed to distinguish differences between individual cells from differences between donors and instruments. On August 12, in a preprint on bioRxiv, authors described PIANO, a model that integrates data on gene activity in individual cells. The model was trained on the Tahoe-100M atlas with over 100 million cancer cell profiles in 21.8 hours on a single NVIDIA A100 graphics processor. Single-cell RNA sequencing shows which genes are active in each cell, and these measurements are used to build tissue, tumor, and brain atlases. However, a single dataset may contain cells from different people, laboratories, and instruments, which can be mistaken for cell differences when compared directly. PIANO compresses the activity of thousands of genes into a short numerical record. When restoring the gene activity profile, the model is separately informed of the experimental conditions: donor, species, measurement platform, or drug dose. The model can then attribute part of the differences to the experimental conditions and use the record for cell characteristics. Another network tries to guess the laboratory batch - a group of cells processed together in one run - from the record. During training, PIANO weakens this connection. The authors checked if this approach preserves different cell types separately, by removing a subset of neurons - eSPN - from non-human primate datasets. In the human dataset, eSPN remained, and PIANO preserved it separately from related neurons common to all primates, as described in Nature Aging, July 2026. On Tahoe-100M, the authors checked if the model can reproduce changes in gene activity at a given drug dose, using trametinib, which blocks the MEK protein in a signaling pathway that controls cell growth and division. 🔗 Read original →