JaBbA (Junction Balance Analysis)

 _____         ___    _      _____ 
(___  )       (  _`\ ( )    (  _  )
    | |   _ _ | (_) )| |_   | (_) |
 _  | | /'_` )|  _ <'| '_`\ |  _  |
( )_| |( (_| || (_) )| |_) )| | | |
`\___/'`\__,_)(____/'(_,__/'(_) (_)

(Junction     Balance     Analysis)

JaBbA builds a genome graph based on junctions and read depth from whole genome sequencing, inferring optimal copy numbers for both vertices (DNA segments) and edges (bonds between segments). It can be used for discovering various patterns of structural variations.

If you use JaBbA in your work, please cite: Distinct Classes of Complex Structural Variation Uncovered across Thousands of Cancer Genome Graphs

Table of contents

• Table of contents
• Installation
• Usage
• Output
• Best practice
• FAQ
• The design of JaBbA
• Attributions
• Funding sources
• Fun fact

Installation

1. Install IBM ILOG CPLEX Studio. The software is proprietary, but can be obtained for free under IBM's academic initiative.

2. Set CPLEX_DIR variable to your CPLEX Studio installation directory

export CPLEX_DIR=/path/to/your/copy/of/CPLEX_Studio/

NOTE: if CPLEX_DIR is set correctly then $CPLEX_DIR/cplex/include and $CPLEX_DIR/cplex/lib should both exist.

3. Install JaBbA

devtools::install_github('mskilab/JaBbA')

4. For convenience, add jba executable to your PATH

$ JABBA_PATH=$(Rscript -e 'cat(paste0(installed.packages()["JaBbA", "LibPath"], "/JaBbA/extdata/"))')
$ export PATH=${PATH}:${JABBA_PATH}
$ jba ## to see usage

5. Test run jba executable on provided toy data

$ jba ${JABBA_PATH}/junctions.vcf ${JABBA_PATH}/coverage.txt 

Usage

 _____         ___    _      _____ 
(___  )       (  _`\ ( )    (  _  )
    | |   _ _ | (_) )| |_   | (_) |
 _  | | /'_` )|  _ <'| '_`\ |  _  |
( )_| |( (_| || (_) )| |_) )| | | |
`\___/'`\__,_)(____/'(_,__/'(_) (_)

(Junction     Balance     Analysis)

Usage: jba JUNCTIONS COVERAGE [options]
	JUNCTIONS can be BND style vcf, bedpe, rds of GrangesList
 	COVERAGE is a .wig, .bw, .bedgraph, .bed., .rds of a granges, or .tsv  .csv /.txt  file that is coercible to a GRanges
	use --field=FIELD argument so specify which column to use if specific meta field of a multi-column table


Options:
	--j.supp=J.SUPP
		supplement junctions in the same format as junctions

	--blacklist.junctions=BLACKLIST.JUNCTIONS
		rearrangement junctions to be excluded from consideration

	--whitelist.junctions=WHITELIST.JUNCTIONS
		rearrangement junctions to be forced to be incorporated

	--geno
		whether the junction has genotype information

	--indel=INDEL
		character of the decision to 'exclude' or 'include' small(< min.nbins * coverage bin width) isolated INDEL-like events into the model. Default NULL, do nothing.

	--cfield=CFIELD
		junction confidence meta data field in ra

	--tfield=TFIELD
		tier confidence meta data field in ra. tiers are 1 = must use, 2 = may use, 3 = use only in iteration>1 if near loose end. Default 'tier'.

	--iterate=ITERATE
		the number of extra re-iterations allowed, to rescue lower confidence junctions that are near loose end. Default 0. This requires junctions to be tiered via a metadata field tfield.

	--window=WINDOW
		window size in bp within which to look for lower confidence junctions. Default 1000.

	--nudgebalanced=NUDGEBALANCED
		whether to attempt to add a small incentive for chains of quasi-reciprocal junctions.

	--edgenudge=EDGENUDGE
		numeric hyper-parameter of how much to nudge or reward aberrant junction incorporation. Default 0.1 (should be several orders of magnitude lower than average 1/sd on individual segments), a nonzero value encourages incorporation of perfectly balanced rearrangements which would be equivalently optimal with 0 copies or more copies.

	--strict
		if used will only include junctions that exactly overlap segs

	--allin
		if TRUE will put all tiers in the first round of iteration

	--field=FIELD
		name of the metadata column of coverage that contains the data. Default 'ratio' (coverage ratio between tumor and normal). If using dryclean, it is 'foreground'.

	-s SEG, --seg=SEG
		Path to .rds file of GRanges object of intervals corresponding to initial segmentation (required)

	--maxna=MAXNA
		Any node with more NA than this fraction will be ignored

	--blacklist.coverage=BLACKLIST.COVERAGE
		Path to .rds, BED, TXT, containing the blacklisted regions of the reference genome

	--nseg=NSEG
		Path to .rds file of GRanges object of intervals corresponding to normal tissue copy number, needs to have $cn field

	--hets=HETS
		Path to tab delimited hets file output of pileup with fields seqnames, start, end, alt.count.t, ref.count.t, alt.count.n, ref.count.n

	--ploidy=PLOIDY
		Ploidy guess, can be a length 2 range

	--purity=PURITY
		Purity guess, can be a length 2 range

	--ppmethod=PPMETHOD
		select from 'ppgrid', 'ppurple', and 'sequenza' to infer purity and ploidy if not both given. Default, 'sequenza'.

	--cnsignif=CNSIGNIF
		alpha value for CBS

	--slack=SLACK
		Slack penalty to apply per loose end

	--linear
		if TRUE will use L1 loose end penalty

	-t TILIM, --tilim=TILIM
		Time limit for JaBbA MIP

	--epgap=EPGAP
		threshold for calling convergence

	-o OUTDIR, --outdir=OUTDIR
		Directory to dump output into

	-n NAME, --name=NAME
		Sample / Individual name

	--cores=CORES
		Number of cores for JaBBa MIP
		
	--lp
		Run as LP
	
	--ism
		Include infinite sites constraints

	-v, --verbose
		verbose output

	-h, --help
		Show this help message and exitsage: jba JUNCTIONS COVERAGE [options]
        JUNCTIONS can be BND style vcf, bedpe, rds of GrangesList
        COVERAGE is a .wig, .bw, .bedgraph, .bed., .rds of a granges, or .tsv  .csv /.txt  file that is coercible to a GRanges
        use --field=FIELD argument so specify which column to use if specific meta field of a multi-column table

Output

1. jabba.simple.gg.rds

   The out put genome graph with integer copy numbers. For more information of how to analyze it please refer to gGnome tutorial

2. karyograph.rds.ppfit.png

   This plot illustrates the distribution of the raw segmental mean of the coverage signal, with red dashed vertical lines indicating the grid of integer copy number states. When the grid align well with the peaks in the underlying histogram, it indicates the purity/ploidy estimation is relatively successful.

3. jabba.seg.txt

   SEG format file of the final segmental copy numbers, compatible with IGV/ABSOLUTE/GISTIC and many more.

4. opt.report.rds

   This file contains an R data.table object of the convergence statistics of all the sub-problems (identified by "cl" column). The column "convergence" indicates the state of the final solution:

   • 1= converged quickly, within short time limit (input tilim/10) to the stringent epgap (input epgap/1000)
   • 2= converged roughly, within short time limit to the relaxed epgap (input epgap)
   • 3= converged after a second round, within long time limit to the relaxed epgap
   • 4= hardly converged after a second round, even after long time limit still above the relaxed epgap

   For detailed explanation of tilim and epgap please read our manuscript and CPLEX help doc.

Best practice

We are working on a best practice pipeline setting to take you from BAMs to good quality reconstructed genome graphs. For the time being please refer to FAQ for practical guidances.

FAQ

1. "CPLEX Error 1016: Community Edition. Problem size limits exceeded."

   You are using a free trial version of CPLEX, please contact IBM's academic initiative for a full license for academic use. We will work on supporting Gurobi in the future.

2. How to prepare the genome-wide coverage input?

   We used fragCounter at 200bp resolution on hg19 reference genome in our paper (details in STAR Methods), which summarized the ratio between the numbers of reads mapped to a bin in tumor versus normal sample, and then corrected for GC content and mappability. Recently we've adopted dryclean (Deshpande et al., BioRxiv, 2019) to denoise coverage data using robust PCA with a panel of normal (PON). The default bin width is 1 kbp and the best practice protocol with dryclean is in preparation (expected Dec 2020).

   Please make sure that the field argument to JaBbA is set to the column name of the coverage data in your coverage input file.

3. How to prepare the junction input?

   We used SvABA (Wala et al., Genome Research, 2018), and we support all junctions callers whose output either conforms to BND style in VCF4.2 specifications or BEDPE format for junctions. Some popular junction callers that do not conform to either are supported too, namely Delly, Lumpy, Novobreak. Besides, the input format can also be a RDS file containing GRangesList of junctions.

   The tfield argument indicates the metadata column name in the junction input reflecting confidence in the call. Take advantage of this to customize your own consensus junction merging, e.g. set junctions called by all callers to tier 1 (highest confidence, JaBbA must use these junctions), called by at least 2 callers but not all to tier 2 (normal confidence, JaBbA will decide whether to use based on coverage change), and the junctions supported by only one caller to tier 3 (low confidence, JaBbA will only search for plausible candidates from them if close to a loose end, must set iteration>0).

4. How to prepare the segmentation input?

   Optional, as JaBbA will infer internally with CBS, but if needed you can also provide tumor and/or normal segmentation through seg and nseg arguments.

5. How to prepare the purity and ploidy input?

   Optional, as JaBbA will infer internally with one of ppgrid, sequenza, and Ppurple of your choice through ppmethod argument, depending on the availability of necessary input files. Purity and ploidy estimation is arguably the most influential hyper-parameter of a JaBbA run, as it dictates the relationship between coverage data to copy number space, so if you want to make high quality graphs, start with better estiamtions. It is a very hard problem disguised as an easy one. Other external tools that helps: ABSOLUTE, ASCAT-wgs, TITAN.

6. How to choose slack penalty?

   The default used in the paper is 100, which in theory correspond to a prior belief that 1 in 100 breakends have loose end, or unexplained copy number change without consistent junction. In practice though, there are many source of noise in the coverage data that could mix with the desired signal of copy number change. After running JaBbA, plot the output copy number alongside the input coverage, If your output segmentation looks too "rigid", i.e. missing obvious clean copy number change points, you might want to consider dropping the slack penalty, as it indicates that JaBbA is so reluctant to add a loose end that it ignores the true signal from the coverage. For dryclean coverage input, which is at 1kbp resolution and denoised, we currently recommend trying slack penalty around 20.

The design of JaBbA

The key to understanding what JaBbA is doing lies in its name, "balancing junctions". When analyzing SVs, junctions and segmental copy numbers have been treated separately in most if not all large scale WGS analyses, but what's not being addressed directly is that these are just two measurable features of the same DNA sequence. The structure of DNA tells us that it is a string, every segment in it are just joined in tandem, hence every copy of a segment should have exactly one upstream neighbor and one downstream neighbor. When adding all copies a segment has, the simple rule that cannot be broken is there must be same number of copies of up/downstream neighbors. If we treat segments as vertices and the 3'-5' phosphodiester bond between segments as edges, we get the junction balance constraints that couple them together.

Of course there would be copy number change points where we can't find a matching junction, and we fill them with loose ends. To make the copy number more correct without a junction we have to use these loose ends like placeholders so the junction balance constraint is still met. Then the construction of the objective function is clear: we want the segment copy number estiamte to be as close to the data's center as possible (minimize residual sum of square) while limit the places where we had to use loose ends for better fit (minimize the number of loose ends).

Attributions

Marcin Imielinski - Assistant Professor, Weill Cornell Medicine Core Member, New York Genome Center.

Xiaotong Yao - Graduate Research Assistant, Weill Cornell Medicine, New York Genome Center.

Funding sources

Fun fact

Why the name? Keith Bradnam started a parody award called JABBA and we thought we'd won. However, jokes aside, did you notice the palindrome there? That's what would happen to a genome if it were to undergo BFB cycles.