Epigenomic enrichment analysis using Bioconductor EuroBioc 2019 - - PowerPoint PPT Presentation

EuroBioc 2019 – Brussels

Dario Righelli – PhD Istituto per le Applicazioni del Calcolo «M. Picone» – CNR - Napoli d.righelli@na.iac.cnr.it || dario.righelli@gmail.com

Epigenomic enrichment analysis using Bioconductor

drighelli

What’s the aim?

Compare methods and provide guidelines on epigenomic data analysis

ATAC-seq dataset

Yijing Su et al. 2017 - Nature Neuroscience - Neuronal activity modifies the chromatin accessibility landscape in the adult brain

Before Fear Induction Condition (E0) After Fear Induction Condition (E1)

} }

4 biological replicates 4 biological replicates } Catching differences in open chromatine regions

ChIP-seq dataset (NULL dataset)

Home Cage Controls - Histon 3, Lysine 9 Acetilation (H3K9ac)

}

9 biological replicates} How many random differences are we able to catch inside a control dataset?

BWA and Bowtie2 perform the same

• Most used aligners for

epigenomics data

• Correlation computed on

ChIP-seq data coverages

• used DeepTools

plotCorrelation tool

• Computed correlations on the

coverages of the same samples on BWA and Bowtie2 bams have value of 1.

• MACS2 (No Bioconductor)
• Most used peak caller
• Broad and Narrow peaks option
• DEScan2
• Has a peak detector in R
• Peak resolution -> bin size
• Can work with external peaks
• DiffBind
• No peak detection
• Fast on matrix construction
• Uses external peaks
• CSAW
• Starts from BAM files
• Computes matrix of bins x samples
• edgeR
• Widely used method
• Very flexible in usage

A Bioconductor Approach

Peak Callers DESCan2 MACS2 CSAW Narrow Broad Peak Consensus & Matrices DESCan2 DiffBind CSAW Differential Enrichment edgeR

Counts Normalization Affects Differentially Accessible Regions (DARs)

• Pay attention to the normalization

process

• One tryes to apply a classic RNA-Seq

normalization

• The process does not always give the

same results

• Maybe some more specific

normalization is required for this kind of data

ATAC-seq dataset

16652 11982 7956 7505 6597 6530 4491 976 976 654 599 523 514 344 282

5000 10000 15000

Intersection Size

DiffBindBroad CSAW DiffBindNarrow DEScan_Z10_K4_DARs 20000 40000

Set Size

ATAC−seq DARs

Comparing DARs across methods

• All the methods have the biggest
• verlap on the detected peaks
• CSAW and DiffBind show a big amount
• f not-overlapping regions
• DEScan2 shows the lowest number of

not-overlapping regions

• The big amount of not-overlapping

regions by CSAW and DiffBind suggests a possible high-level of false positive regions detected.

• Ad-hoc designed UpsetPlot on GRanges
• Based on findOverlaps method

Peaks contrasts on NULL dataset show no results

H3K9ac ChIP-seq dataset

• Compared performances on a null

dataset of ChIP-seq H3K9ac samples

• Performed 126 permutations of samples
• Samples are randomly divided in

two groups

• All the possible permutations on 9

samples (126)

• All the methods find mostly 0

Differential Enriched Peaks on the random conditions.

• Sometimes some differences have been

found

• With and without normalization

2 4 6 8 D E S c a n 2 _ N

• N
• r

m D E S c a n 2 _ N

• r

m D E S 2 _ M 2 _ B r

• a

_ N

• N
• r

m D E S 2 _ M 2 _ B r

• a

_ N

• r

m D E S 2 _ M 2 _ N a r r _ N

• N
• r

m D E S 2 _ M 2 _ N a r r _ N

• r

m D i f f B i n d _ N a r r _ N

• r

m D i f f B i n d _ B r

• a

d _ N

• r

m

method nElem normalized

NO YES

What’s Next?

On-going and future works

Some comparisons are still needed

• Compare CSAW on ChIP-seq
• Compare normalization methods with all epigenomics methods
• Explore in-silico biological functions of results
• Testing ATAC—seq Single Cell dataset

• Dr. Claudia Angelini – Istituto per le Applicazioni del Calcolo-CNR
• Dr. Davide Risso – Univeristy of Padua
• Dr. Lucia Peixoto – Elon S. Floyd College of Medicine, Washington State University
• Dr. Timothy Triche Jr. - Van Andel Research Institute
• Dr. Ben Johnson - Van Andel Research Institute
• Thank you for your Attention!

Acknowledgements

http://lists.moo.gs/mailman/listinfo/biocmeetup.naples napoli.r.bioc@gmail.com https://www.facebook.com/pg/NapoliRBiocMeetup

Napoli R/Bioconductor Meetup

• Since Nov 2018
• R Consortium Array Group
• At least 25 people any event with a good

turn-over of attendees

• Eight meetups until now
• R Package Creation
• scRNA-seq Analysis
• Differentially Methylated Regions Analysis
• Microscope Image Processing
• Chromosomal Copy Number Changes

Detection

• Bulk RNA-seq Differential Expression
• Hi-C data analysis using HiCeekR
• Metagenomics analysis workflow

Napoli R/Bioconductor Meetup

• Part of a wider idea
• Third city in the World
• Boston (USA)
• New York (USA)
• Napoli (IT)
• Useful to
• share ideas and workflows
• create new collaborations
• extend bioinfo community

Is there a best Aligner?

Bowtie2 vs BWA

Comparing DARs across methods (2)

ATAC-seq dataset

73600 60794 24236 16120 13994 5854 4376 2879 2424 2122 1913 1857 1457 1046722 703 466 410 351 350 327 241 137 100 93 89 9 5 5 3 2 20000 40000 60000 80000

Intersection Size

DEScanMACSNarrow DEScanMACSBroad DiffBindMACSBroad DEScanZ10K4 DiffBindMACSNarrow 25000 50000 75000 100000 125000

Set Size

ATAC−seq Regions Nar/Broa & DEScan2

• Ad-hoc designed UpsetPlot on Granges
• Based on findOverlaps method
• Results description

Duplicates Removal doesn’t impact peak detection

• Diagonal Correlations on

counts matrices show that there is no big differences between duplicates and no-duplicates samples

• rmDup with samtools
• DEScan2 counts matrices
• DiffBind counts matrices

noDup_E0_1 noDup_E0_2 noDup_E0_3 noDup_E0_4 noDup_E1_1 noDup_E1_2 noDup_E1_3 noDup_E1_4 withDup_E0_1 withDup_E0_2 withDup_E0_3 withDup_E0_4 withDup_E1_1 withDup_E1_2 withDup_E1_3 withDup_E1_4

DEScan2

noDup_E0_1 noDup_E0_2 noDup_E0_3 noDup_E0_4 noDup_E1_1 noDup_E1_2 noDup_E1_3 noDup_E1_4 withDup_E0_1 withDup_E0_2 withDup_E0_3 withDup_E0_4 withDup_E1_1 withDup_E1_2 withDup_E1_3 withDup_E1_4

DiffBind

DiffBind

DEScan2 – Differential Enriched Scan 2

• Filter out the peaks with a score lower than a

user-defined threshold

• Aligns the peaks over user-defined number of

samples

• Different thresholds produce different trends

in number of final peaks detected