Annotating inter-sample epigenetic heterogeneity in cancer using genome coordinate covariation analysis

Nathan Sheffield

www.databio.org/slides

Goal: understand variation among individuals

Supervised differential analysis

Supervised continuous analysis

Unsupervised analysis

Epigenomic data: high-dimensional
and low-interpretable

Dimensionality reduction

Even with known groups

How can we annotate the source of variation?

COCOA Overview

John Lawson

Coordinate Covariation Analysis

Quantify variation into a 'target variable'

Supervised (e.g. clincial variable).
Unsupervised (e.g. PCA)

Annotate target variable with region sets.

What is epigenetic signal covariation?

Covariation informs source of observed variation

1. Choose target variable

What is the variation we'd like to explain?

Supervised target

Unsupervised target

2. Quantify correlation with target variable

Permutation tests establish significance

Case studies

Breast cancer DNA methylation (Unsupervised)
Breast cancer ATAC-seq (Unsupervised)
Kidney cancer DNA methylation (Supervised)
Pan-cancer EZH2 analysis

Breast cancer DNA methylation PCA

COCOA results for PC1

ER-related regions have higher loadings on PC1

Raw DNA Methylation in ER binding regions

COCOA results for PC1-4

COCOA meta-region plots for PC1-4

Case studies

Breast cancer DNA methylation (Unsupervised)
Breast cancer ATAC-seq (Unsupervised)
Kidney cancer DNA methylation (Supervised)
Pan-cancer EZH2 analysis

Breast cancer ATAC-seq PCA

COCOA results for ATAC-seq

ER-related regions have higher loadings on PC1

COCOA results for ATAC-seq

Case studies

Breast cancer DNA methylation (Unsupervised)
Breast cancer ATAC-seq (Unsupervised)
Kidney cancer DNA methylation (Supervised)
Pan-cancer EZH2 analysis

Kidney cancer DNA methylation (Supervised)

Rank region sets for methylation
that correlates with cancer stage

COCOA results for cancer stage

Case studies

Breast cancer DNA methylation (Unsupervised)
Breast cancer ATAC-seq (Unsupervised)
Kidney cancer DNA methylation (Supervised)
Pan-cancer EZH2 analysis

DNA methylation in EZH2 regions and survival

DNA methylation in EZH2-binding regions most often positively correlated with risk of death.

#### Conclusions: COCOA... - can an interpret continuous regulatory variation - can use any signal that annotates genomic coordinates - can work on supervised and unsupervised data - is available from Bioconductor - depends critically on the database of region sets...

Thank You

Collaborators
Fran Garrett-Bakelman
Stefan Bekiranov

Sheffield lab
John Lawson
Jason Smith
Jianglin Feng
Michal Stolarczyk
Kristyna Kupkova
Aaron Gu
Jose Verdezoto
Tessa Danehy

Funding:

UVA Cancer Center

nsheff ·

databio.org ·

nsheffield@virginia.edu