Machine Learning for Single-Cell 3D Epigenomics

on YouTube ↗35m

" summary

Presents a concrete pipeline: a hypergraph model where hyperedge = cell+bin+bin encodes each single-cell Hi‑C contact, a graph/neural embedding (Higashi) that clusters cell types and enables imputation, and a tensor-based fastHigashi for scale. Claims imputation outperforms HiCluster and is imputation validated against Zhuang imaging by downsampling imaging-derived distances. Imputed maps reveal cell-specific TAD boundary behaviors (on/off and gradual shifts) and separate prefrontal neuron subtypes previously indistinguishable. Also delivers Nucleome Browser integrates tracks, Hi‑C maps and microscopy with Jupyter; notes interpretability and scalability limits.

AI-written summary · about

" chapters (20)

1. 0:00Start
2. 2:04Multi-scale nuclear organization
3. 6:17SPIN infers spatial localization patterns
4. 7:19MOCHI connects structure and function
5. 9:08Challenges for single-cell 3D genome analysis using scHi-C data
6. 11:45Hypergraphs for higher-order interactions
7. 14:18Higashi-model scHi-C data as a hypergraph
8. 16:39Embeddings from Higashi reflect cell identity
9. 17:37Higashi robustly improves scHi-C imputation
10. 19:49Higashi imputation enables more reliable 3D genome feature analysis
11. 21:11Application to human prefrontal cortex
12. 22:17Cell type-specific 3D chromatin structures near marker genes
13. 23:23Limitations of Higashi
14. 24:47Fast-Higashi-More scalable and intertable
15. 27:20Visualization challenges for 4D Nucleome
16. 28:54Multimodal, multi-domain integration
17. 30:03Exploring data from different perspectives
18. 30:32Cross cell-type or single cell comparisons
19. 31:04Integration with Jupyter Notebook
20. 32:18Outlook | Algorithms for genome-scale discovery

" related talks

• 37mDynamics of 3D Enhancer-Gene Associations Across Diverse Human Cell Types and TissuesAnshul Kundaje · Network Biology 2016
• 38mInterpreting Deep Learning Models Of Functional Genomics Data To Decode Regulatory Sequence...Anshul Kundaje · Interpretable Machine Learning in Natural and Social Sciences (2022)
• 35mExploratory and Model-Based Analysis of ScHi-C DataSunduz Keles · Computational Challenges in Very Large-Scale 'Omics'
• 35mFrom cells to drug responses - machine learning in cancer researchJulian de Ruiter, Nanne Aben · PyData Amsterdam 2018
• 29mHarnessing Multimodal Single-Cell Sequencing Data for Integrative AnalysisElizabeth Purdom · Computational Challenges in Very Large-Scale 'Omics'
• 35mTranscriptional Enhancers - Looking out for the Genes and Each OtherSridhar Hannenhalli · Network Biology 2016
• 32mMulti-Omic Integration for Understanding DiseaseRon Shamir · Computational Challenges in Very Large-Scale 'Omics'
• 31mHacking the Human Genome Using Clojure and Similarity SearchArnoldo Jose Muller-Molina · Clojure/conj 2011