Enhancing Statistical Rigor In Genomic Data Science

on YouTube ↗42m

" summary

Argues that ill‑posed p‑values arise from three concrete failures (treating two‑sample problems as one‑sample, poor parametric fits, and treating inferred covariates as observed) and offers practical fixes: a knockoff‑inspired FDR wrapper CLIPper that rescues MACS/HOMER/DESeq2/edgeR outputs and empirically brings actual FDR to target with modest power loss; a subsample+permute calibrator PseudotimeDE (80% subsamples, permutation nulls; ~1,000 rounds for 0.001 resolution) for single‑cell pseudotime; and a simulator SADesign3 using Gaussian/Vine copulas and GAM backbones that models correlations, outputs reads, handles continuous trajectories, spatial/ATAC/protein data, and generates realistic nulls.

AI-written summary · about

" chapters (14)

1. 0:00Start
2. 0:50My first encounter of genomics bioinformatics (2007) Volume 447 Issue 7146, 14 June 2007
3. 3:44Statistical rigor in multiple testing
4. 4:47Formulation of a two-sample test as a one-sample test
5. 9:09Specification of a parametric model that does not fit data well
6. 24:04Halftime summary
7. 24:28Clipper: p-value-free FDR control for genomics feature screening
8. 26:46Clipper rectifies FDR control
9. 27:35The last piece
10. 29:04scDesign3: an omnibus single-cell & spatial omics simulator
11. 30:44ClusterDE: scDesign3 for null data generation (preliminary)
12. 31:04ClusterDE: Clipper + scDesign3 (preliminary)
13. 33:37Summary: relevant publications
14. 35:27Acknowledgements