Find research datasets worth reusing

VeloCycle-estimated cell cycle phases of single cells from a genome-scale perturb-seq performed in K562

<p>Continuous cell cycle phase position between 0 and 2&pi; estimated by <em>VeloCycle&nbsp;</em>(Lederer et al.,&nbsp;<em>Nature Methods</em> 2024) for single cells of the perturb-seq performed in the K562 CML cell line by Replogle et al., Cell 2022.&nbsp;</p> <p>Underlies the cell cycle imbalances inferred in Pulver &amp; Forey et al., 2024.</p> <p>Method manuscript exerpt:</p> <p>"For analysis on the genome-wide perturb-seq dataset of K562 cells (Replogle <em>et al.</em>, 2022), a transfer learning approach was applied. Condition-independent estimation of the periodic Fourier series components would be especially challenging on Perturb-seq knockdown conditions containing either (1) very few cells or (2) cells belonging to just one phase of the cell cycle. To infer accurate cell cycle phases for these cells, we first performed manifold-learning for 5,000 training steps to estimate the gene harmonic coefficients (&nu;0, &nu;1sin, &nu;1cos) on a larger set of non-targeting control (NT) K562 cells (75,328 cells), which are more evenly distributed throughout the various phases of the cell cycle. Next, we ran manifold-learning again for 5,000 training steps, but on the entire perturb-seq dataset of 1,971,608 cells and 4,127 gene knockdown conditions (with at least 75 cells per condition). This time, we conditioned <em>VeloCycle</em> on the gene harmonic coefficients learned in the first step. This allowed cells belonging to each stratified gene knockdown condition to be assigned to a position on the cell cycle manifold, while restricting those assignments such that they were based on gene expression patterns earned on a larger and more informative dataset (allowing for batch effect expression differences)."</p>