Benchmarking, Inference, and in Silico Controls in Single-Cell and Spatial Omics Data Science

on YouTube ↗32m

" summary

scDesign3 is a unified probabilistic simulator that generates single-cell and spatial omics from discrete cell types and continuous trajectories using marginal count models plus a Gaussian copula (via distributional transform) to preserve gene–gene correlation; it fits an accessible likelihood so users can do model selection (BIC) and inspect interpretable parameters to make in silico negative and positive controls for FDR and power evaluation. scDesign3 pairs with scReadSim to emit sequence reads, and the CLIPper pipeline uses those synthetic nulls to control FDR where standard cluster-then-test pipelines inflate discoveries (works for ~2,000 selected genes).

AI-written summary · about

" chapters (22)

1. 0:00Intro
2. 0:28Single-cell and spatial omics data: statistical characteristics
3. 1:42Motivations
4. 3:45Use scDesign to benchmark doublet-detection methods
5. 4:32Exemplar scRNA-seq simulators and properties
6. 5:46sc Design2: notations
7. 6:43scDesign2: marginal distribution of each gene i
8. 7:35scDesign2: joint distribution of highly-expressed genes
9. 8:35scDesign2: distributional transform and the Gaussian copula
10. 9:28scDesign2: joint distribution fitting
11. 10:07scDesign2: data simulation
12. 10:42scDesign2: summary
13. 12:39scDesign3: an omnibus single-cell & spatial omics simulator
14. 14:57scDesign3 functionalities (simulation)
15. 15:12scDesign3 functionalities (interpretation)
16. 15:35Why need in silico controlled experiments?
17. 17:53scDesign3: in silico negative control
18. 18:32Clipper: a p-value-free FDR control framework
19. 20:46Clipper paper
20. 21:26Cluster DE: scDesign3 + Clipper (preliminary)
21. 22:40Take-home messages
22. 23:23Acknowledgements