Train a model on FP vs. TP, check for informative attributes

Instead of a headerless tsv, create a json with the ability to use key-values.

Create a function that grabs status, timings, output paths etc. from cromwell_metadata.json

• Workflow Management with Cromwell and WDL (Ruchi Munshi):
  https://www.youtube.com/watch?v=mFzfeDTnDSk

For the VCF comparison read in the f1 and f2 arguments into cromwell_inputs.json

Examples:

DNA

bcbio

woof compare 
  --mode bcbio-dna
  --bc1 /path/to/run1/sampleA/final
  --bc2 /path/to/run2/sampleA/final 

umccrise

woof compare
  --mode umccrise
  --um1 /path/to/run1/sampleA/umccrised
  --um2 /path/to/run2/sampleA/umccrised

RNA:

woof compare 
  --mode bcbio-rna
  --bc1 /path/to/run1/sampleA/final
  --bc2 /path/to/run2/sampleA/final 

Questions

• How to detect genome assembly (hg37 vs. hg38)

Might be able to get that info from the bcbio config

See http://www.htslib.org/doc/htsfile.html for guessing file types

Fix recursive copy function. Getting the below when re-starting and the WDL files already exist:

Directory not copied. Error: [Errno 17] File exists : woof/work/wdl'

Get total counts, but only use filtered PASS calls for the evaluation stats

From Oliver (https://umccr.slack.com/archives/C025TLC7D/p1553825820038500):

Diff bcbio 1.1.3 / 1.1.4. Not surprised we're seeing those changes. A fair number of those 'false positives' (new MuTect2 calls) might actually now just be calls that VarDict/Strelka2 already had.
Which made me think of the easiest evaluation metric: how many of the 'new' (potential FP) calls for a caller have support from the other callers? And respectively, how many of the false negative calls (lost in the new version) are unique to that caller?

• AWS Batch detailed guide: https://stackify.com/aws-batch-guide/
• AWS Batch Fetch & Run: https://aws.amazon.com/blogs/compute/creating-a-simple-fetch-and-run-aws-batch-job/
• AWS 4-part blog series (2017):
  https://aws.amazon.com/blogs/compute/building-high-throughput-genomics-batch-workflows-on-aws-introduction-part-1-of-4/
• AWS newer blog series (2018):
  https://aws.amazon.com/blogs/compute/building-simpler-genomics-workflows-on-aws-step-functions/
• YouTube demo (2018): https://youtu.be/jcC3pz_K4gI
• AWS blog (2018): https://aws.amazon.com/blogs/compute/using-cromwell-with-aws-batch/
• AWS Genomics Workflows (2019):
  https://docs.opendata.aws/genomics-workflows/
• YouTube WDL/Cromwell webinar (Ruchi Munshi): https://www.youtube.com/watch?v=mFzfeDTnDSk

Move all code into different script

All you need is <f1> and <f2>.

Currently all happens withing Rmd report.
Workflow would go like:

• Manta1 + Manta2
• isec
• Eval Table, Circos with FP/FN

See https://github.com/ecerami/wdl_sandbox

https://docs.opendata.aws/genomics-workflows/orchestration/cromwell/cromwell-examples/

See https://bookdown.org/yihui/bookdown/github.html

• https://github.com/lvulliard/BioCircos.R
• https://github.com/nicgirault/circosJS

Check out attributes of FP/FN variants from ensemble-batch comparisons.
Things to look at:

• location (challenging regions, GC bed, Heng Li's problematic regions etc.)
• mapping quality (MQ)
• number of mismatches (NM)
• allele depth (DP)
• allele frequency (AF)

More at: http://bcb.io/2016/04/04/vardict-filtering/

From https://github.com/vladsaveliev/vcf_stuff/blob/master/README.md

1. Split multi-allelic variants into single sample records.

For instance, split one record

#CHROM  POS     ID      REF     ALT
1       10       .      A       T,C

Into 2 separate records

#CHROM  POS     ID      REF     ALT
1       10       .      A       T
1       10       .      A       C

For that, we are using vt tools:

vt decompose -s vcf_file

2. Decompose biallelic block substitutions.

For instance, split the following one records:

#CHROM  POS     ID      REF     ALT
1       20      .       AG      CT       

into 2 separate ones:

#CHROM  POS     ID      REF     ALT
1       20       .      A       C
1       20       .      G       T

We are using for that vcflib's vcfallelicprimitives:

vcfallelicprimitives -t DECOMPOSED --keep-geno vcf_file

3. Left-align and normalize indels, check if REF alleles match the reference.

For instance, given that the reference chromosome 1 starts with GCTCCG, split the following records

#CHROM  POS     ID      REF     ALT
1       2       .       CTCC    CCC,C,CCCC

into the following 3:

#CHROM  POS     ID      REF     ALT
1       1       .       GCTC    G
1       2       .       CT      C
1       3       .       T       C