Exploratory and Model-Based Analysis of ScHi-C Data
summary
Simple per-band scaling that adds back an average band decay (BandNorm) runs in ~15 minutes and often matches more complex methods on cell-type separation, but it degrades at very high sparsity (100 kb resolution or rare cell types). A deep generative, band-aware model (scVI-3D) uses a zero-inflated gamma with locus-specific zeros, needs GPU and ~hour runs, and preserves performance with imputation. Gene-level scGAD scores correlate with expression, enable CCA integration with RNA and ChIP, and hundreds of cells produce believable pseudo-bulk contact maps.
AI-written summary · about
chapters (15)
- 0:00Start
- 1:38Hi-C for studying long-range interactions
- 2:55Hi-C meets single cell sequencing
- 3:23scHi-C data: sparsity
- 4:34scHi-C specific genomic distance effect
- 6:15BandNorm & SCVI-3D with band transformation of scHi-C data
- 9:48Normalize and de-noise scHi-C data
- 11:19Evaluation and impact on downstream
- 12:44Cell type separation: Ramani2017
- 13:17Cell type separation: Lee 2019
- 14:43Evaluation of unsupervised clustering
- 16:20Impact on downstream analysis: Pseudo- bulk ve bulk Hi-C
- 17:18Similarity of the pseudo-bulk to bulk
- 18:35Integrating scHi-C with other single cell data modalities
- 24:49Acknowledgements
related talks
35mMachine Learning for Single-Cell 3D Epigenomics
28mSome Current Computational Challenges in Bioinformatics
43mPanel on Interpretability in the Biological Sciences
57mBig Data and New Models Needed to Study DNA Variation in Evolution and Cancer
42mHow Novel Compute Technology Transforms Life Science Research
46mKeynote: Life Science Research at Scale
47mBANKSY: A Spatial Omics Algorithm that Unifies Cell Type Clustering and Tissue Domain Segmentation
26mShiva Amiri, CEO, BioSymetrics at MLconf Seattle 2017
