Data Con LA 2022 - Early cancer detection using higher-order genome architecture
•
0 likes•3 views
My (Angela) Chung, Data Enthusiast, San Jose State University Cancer is a complex disease which requires interactions between cell-intrinsic alterations and tumor microenvironment. The connection between epigenetics and genomic structure plays a key role in chromatin interactions and enhancer-promoter communications for transcriptional activities. Alterations of these components in oncogenic signaling pathway potentially cause cancer cell-intrinsic changes and inappropriate instructions to normal cell cycles, leading to abnormal cell growth. ' Topologically associating domains (TADs) and A/B compartments are the main structures of higher-order chromatin structure. These contact domains, chromatin states, super-enhancers, and histone modifications together regulate transcription and gene expression for normal/abnormal cell cycles. ' Several bioinformatics tools were utilized ' FANC for processing raw FASTQ data to Hi-C contact matrices, JuicerTools for obtaining the locations of contact domains on the entire genome, and CoolBox for visualizing chromatin contacts in different cell lines. ' High-resolution chromatin contacts showed dynamic interactions among chromosomal regions in different cell lines. ' Qualitative and quantitative features were comprehensively engineered from 3D chromatin folding and epigenetic regulators using available packages (scikit learn, pytorch, pandas, numpy, matplotlib, etc.). ' XGBoost multi-class classifier achieved the highest accuracy of 80.90% in classifying normal and cancer cell lines based on chromatin interactions, followed by Random Forest at 73.76% and TabNet classifier at 70.00%.
Recommended
Recommended
More Related Content
Similar to Data Con LA 2022 - Early cancer detection using higher-order genome architecture
Similar to Data Con LA 2022 - Early cancer detection using higher-order genome architecture (20)
More from Data Con LA
More from Data Con LA (20)
Recently uploaded
Recently uploaded (20)
Data Con LA 2022 - Early cancer detection using higher-order genome architecture
- 1. Effective Cancer Detection Using Higher-Order Genome Architecture and Chromatin Interactions My (Angela) Chung Master of Science in Bioinformatics
- 2. Conceptual question How do higher-order genomic structures and epigenetic regulations influence the chromatin interactions in normal and cancerous cell lines? 2
- 3. Roadmap 1 3 5 6 4 2 PIPELINE SUMMARY DECTECTION MODEL CONTACT MATRICES & REGULATORY TRACKS SCIENTIFIC BACKGROUND DATA ENGINEERING & VISUALIZATION MODEL EVALUATION 3
- 4. Background – clinical & scientific! 4
- 5. Chromatin structure Chang et al. (2018) 5 Higher-order Primary-order Genomic Structure
- 6. TAD structure Super-enhancer Rowley et al. (2018) Li et al. (2021) Accumulation of multiple enhancers at TAD domains 6 CTCF: the regulator of chromatin organization Cohesin: the structural maintenance of chromosome cohesin complex Genomic Structure Loop extrusion
- 7. Epigenetic marks & chromatin states 7 Epigenetics H3K27ac is ACTIVATOR ACTIVATORS for super-enhancer H3K27me3 is REPRESSOR Martire et al. (2017) & Chen et al. (2020)
- 8. Chromatin state in oncogenic signaling pathways 8 Hniz et al. (2015) & Li et al. (2021) Epigenetics – Histone modifications Pathway effects Transcriptional super-enhancers q Chromatin structure q Epigenetic regulation q Effects on cancer pathway
- 10. Cell morphology IMR90 Lung fibroblast Normal cell K562 Chronic myeloid leukemia Cancer cell GM12878 Lymphoblastoid Normal cell A549 Human lung adenocarcinoma Cancer cell 10
- 11. High-throughput chromosome conformation capture (Hi-C) • Comprehensive mapping of chromatin contacts • Detect loops across the entire genome • DNA-DNA proximity ligation is combined with high-throughput sequencing • Chromosome spatially positions into 3D conformation Rao et al. (2014) 11 In situ Hi-C
- 12. 12
- 13. Chromatin contacts chr1: 20mb – 80mb chr1: 60mb – 80mb chr1: 60mb – 70mb chr1: 62mb – 67mb 13 q Cell line: A549 q Chromosome: 1 q Anatomy of contact matrices
- 14. Dynamic Chromatin contacts 14 q Cell lines and chromosomes q Resolution q Variations in genome architecture
- 15. Dynamic Chromatin contacts 15 q Cell lines and chromosomes q Resolution q Variations in genome architecture
- 16. Dynamic Chromatin contacts 16 q Cell lines and chromosomes q Resolution q Variations in genome architecture
- 17. Dynamic Chromatin contacts 17 q Cell lines and chromosomes q Resolution q Variations in genome architecture
- 18. Engineer and organize data for visualization and building a detection model. Data Engineering & Visualization
- 19. Flowchart for Data Engineering TADs CTCF Loops RAD21 Histone marks Insulation Directionality Index CTCF at both ends CTCF at TAD start CTCF at TAD end DI at TAD end DI at TAD start Sum of insulation scores No CTCF Number of histone marks Average location Histone marks are not present Start boundary score End boundary score CTCF loop Cohesin-independent loop Cohesin-associated loop Ordinary domain Loop extrusion Extrusion loss ChIP-Seq FANC Arrowhead HICCUPS ChIP-Seq FANC ChIP-Seq Within loop window Outside loop window CTCF loop loss Structure Effect Analysis Features Key Components 19
- 20. 20 Chromatin state distribution ACTIVATOR = H3K27ac
- 21. 21 Chromatin state distribution ACTIVATOR = H3K27ac
- 22. 22 Chromatin State vs Boundary Strength Distribution ACTIVATOR = H3K27ac
- 23. 23 Chromatin State vs Boundary Strength Distribution ACTIVATOR = H3K27ac
- 24. 24 Enhancers and promoters are potentially protected at strong TAD boundaries. Chromatin State vs Boundary Strength Distribution ACTIVATOR = H3K27ac
- 25. 25 Chromatin State vs Boundary Strength Distribution REPRESSOR = H3K27me3
- 26. 26 Normal cell Normal cell Chromatin State vs Boundary Strength Distribution REPRESSOR = H3K27me3
- 27. 27 Cancer cell Cancer cell Chromatin State vs Boundary Strength Distribution REPRESSOR = H3K27me3
- 28. 28 Lower density of weak boundaries and negative super-enhancers detected in cancerous cell lines. Chromatin State vs Boundary Strength Distribution REPRESSOR = H3K27me3
- 29. Detection Model Cell type classifer using machine learning algorithms
- 30. Machine learning models Random Forest 73.76% accuracy TabNet 70.00% XGBoost 80.90% accuracy
- 31. 31 Random Forest performance evaluation Model accuracy is 73.76% Max depth is unassigned for the highest performance Number of tree estimators is around 200 Numerical attributes represent feature importance K562 is best classified at 83%
- 32. 32 XGBoost performance evaluation Model accuracy is 80.90% Max depth is 8-12 for the highest performance Number of estimators is around 300 IMR90 and K562 are best classified at 83%
- 33. 33
- 34. Future direction 34 • Try different scalable and memory-efficient tools to remove biases and normalize Hi-C contact maps. • Consider gene expression profile (RNA-Seq) for chromatin state categorization. • Consider new visualization tools to further enhance image quality, especially for triangular heatmap. • Improve feature engineering to enhance model performance. • Consider Hi-C image classification to distinguish cell lines.
- 35. References Chang, P., Gohain, M., et al., Computational Methods for Assessing Chromatin Hierarchy. Comp Struc Biotech J., 16. 2018 Rowley, M.J., & Victor, G.C., Organization principles of 3D genome architecture. Nature Reviews: Genetics, 19, 2018 Li, G.H., Qu, Q. et al. Super-enhancers: a new frontier for epigenetic modifiers in cancer chemoresistance. J. Exp Clin Cancer Res, no. 40, 2021 Fudenberg G., Imakaev, M., et al., Formation of chromosomal domains by loop extrusion. Cell Reports, 15, 2038-2049, 2016 Kyrchanova, O., & Georgiev, P., Mechanisms of enhancer-promoter interactions in higher eukaryotes. Int J Mol Sci., vol. 22, no. 2, 2021 Rao, S., Huang, S.C., et al., Cohesin loss eliminates all loop domains. Cell, vol. 171, no. 2, 2017 Ong, C.T., & Corces V. G., Enhancer function: new insights into the regulation of tissue-specific gene expression. Nat Rev Genet., vol. 12, no. 4, 2011 Ernst J., Kheradpour, P., et al., Mapping and analysis of chromatin state dynamics in nine human cell types. Nature., 2011 Hnisz, D., Schuijers, J., et al., Convergence of developmental and oncogenic signaling pathways at transcriptional super- enhancers. Molecular Cell, vol. 58, no. 2, 2015 Rao, S. S. P., Huntley, M. H., Durand, N. C., et al., A 3D Map of the Human Genome at Kilobase Resolution Reveals Principles of Chromatin Looping. Cell, 159, 2014 35
- 36. Project Committee Members: Wendy Lee, Ph.D Department of Computer Science Carlos Rojas, Ph.D Department of Computer Engineering William Andreopoulos, Ph.D Department of Computer Science Thank you for your time and attention!
- 37. 37
- 38. 38
- 39. Correlation Diagrams TADs Cohesin- associated loops Chromatin states Histone modification RNA Pol II Gene expression Insulation score CTCF Cohesin- independent loops Inter- chromosomal links Gene expression Super- enhancer colocalization Up-regulated Down-regulated
- 40. CTCF and histone marks in contact domains q Cell lines: A549 and IMR90 q Chromosome: 8 q Resolution window: 80mb – 83mb q CTCF and epigenetic factors in genomic architecture
- 41. Basic mapping: • Paired-end reads were detected and split reads at ligation junctions by MboI or HindIII restriction enzyme. • Map reads (MAPQ > 30) to a reference genome hg19 using Bowtie2 indexing. Iterative mapping: • Eliminate unaligned reads Fragmentation: • assign mapped reads to restriction fragments (MboI or HindIII) Mapping to generate aligned reads & Fragmentation to assign to restriction fragments
- 42. Filtering read pairs to eliminate invalid read pairs from experimental procedures Distance filters: filter out self-ligation products <25kb Restriction distance: remove reads mapping >1000kb Strand filters (inward and outward ligations): • Inward same strand (un-ligated) <1kb • Outward same strand (self-ligated) <1kb Features of invalid reads: Unmapped Multi-mapping PCR duplicates Low map quality Ligation errors
- 43. Samples Total reads Valid pairs % Valid pairs Sum pairs filtered GM12878_577 87,569,310 75,182,513 85.9 16,519,508 GM12878_580 63,592,550 54,573,765 85.8 12,121,642 GM12878_581 31,678,266 27,537,842 86.9 5,570,781 GM12878_583 69,592,394 57,017,845 81.9 16,266,949 GM12878_586 59,628,563 49,030,434 82.2 13,838,160 IMR90_672 99,266,806 94,680,468 95.4 25,551,020 IMR90_673 107,407,597 90,914,417 84.6 21,623,779 IMR90_674 14,526,316 12,731,251 87.6 2,345,638 IMR90_675 66,649,091 56,971,226 85.5 26,162,938 IMR90_676 133,365,970 116,341,749 87.2 22,032,033 IMR90_677 118,965,610 94,694,521 79.6 34,048,741 Data processing summary Four cell lines: GM12878, IMR90, K562, and A549 Total 34 samples including both primary and replicates
- 44. Matrix balancing Goal: • Correct known and potentially unknown systemic biases • Generate true contact frequencies which preserve underlying architecture of the matrix. • Assure visibility for all loci Binning size: 10kb, 100kb, and more Normalization: (per chromosome level) Knight-Ruiz ICE Product of non-negative matrix and diagonal matrices D1 and D2 to obtain singular value P Expected contact frequency from biases and relative contact probabilities