Skip to content

BD2KGenomics/hgvm-graph-bakeoff-evaluations

Repository files navigation

HGVM Graph Bakeoff Evaluations

Background

This repository contains scripts to validate and analyse reference variation graphs submitted to the graph bakeoff (Details). They should be sufficient to reproduce all analysis results.

Obtaining Submitted Graphs

Todo: I think we have multiple ways of getting the graphs. I've been piggybacking on graphs from Adam's scripts. Sean may be using something else. Worth moving into one place?

Dependencies

Submodules and/or Docker may be the way to go here?

Graph Properties Comparison

Obtaining GRC Alignments and Gene Definitions

For the GRC alignment concordance analysis, MAF files for the GRC alignments for the MHC, SMA, and LRC_KIR regions are required. Additionally, for the ortholog/paralog mapping analysis, BED files giving gene annotations for those regions are required.

These can both be obtained by running the fetchRegion.py script. For example, to get the BED and MAF files for the SMA region, deposited into an SMA directory under the current directory, run:

scripts/fetchRegion.py SMA

As of this writing, this ought to fetch the GENCODE v22 genes, and alignments and sequence from the GRCh38.p2 assembly.

This script requires the hgsql tool, available in the Kent tools, as well as a set-up and configured database containing the hg38 database and its knownGenes and kgXref tables. It may be possible to use UCSC's public SQL server; this has not been tested.

This script also requires a patched version of BioPython with MAF output support, which should have been installed via requirements.txt.

MAF files and gene annotations fetched with this script are also available in this archive.

Graph Mapping Comparison

One evaluation that we perform is to test how well each graph works as a read mapping target. This is accomplished by aligning the 1000 Genomes low coverage reads for each region against each graph, and evaluating the resulting alignments.

Obtaining Low Coverage Reads

To download the low coverage reads for the read mapping evaluation, you can use the getAltReads.py script. Note that this script will by default, in single machine mode, download in parallel as many samples as your system has cores, and that overall it wants to download 6 regions for 2,691 samples. To mitigate the load on your system, you can use --maxCores N to limit the number of cores used (and thus the number of parallel downloads), and --sampleLimit N to limit the number of samples retrieved to the first N found. (Note that this sample will not be representative; it will be biased towards the populations that are iterated over first.)

An example command invocation:

# Remove any previous Toil job store
rm -Rf tree
# Run the download
time scripts/getAltReads.py ./tree low_coverage_reads --batchSystem=singleMachine --sample_ftp_root "ftp://ftp.1000genomes.ebi.ac.uk/vol1/ftp/data_collections/1000_genomes_project/data/" --population_pattern="*" --sample_pattern "*" --file_pattern "*.low_coverage.cram"  2>&1 | tee log.txt

Running the Mapping Evaluation

To actually run the alignments for the downloaded reads, you should use the parallelMappingEvaluation.py script. An example invocation on a single machine is given below:

# Remove any previous Toil job store
rm -Rf tree
# Run the alignment
scripts/parallelMappingEvaluation.py ./tree graph_servers.tsv ./low_coverage_reads ./low_coverage_alignments --batchSystem=singleMachine --use_path_binaries --kmer_size=27 --edge_max=3 2>&1 | tee log.txt

Note that if you are running on all 2,691 samples retrieved in the previous step, the command given above will take an impractical amount of time to complete. One option to mitigate this is to only run some samples, with a --sampleLimit flag. Another is to use the script on a cluster, by changing the --batchSystem argument to any batch system supported by toil (and the job store of ./tree to a job store that will be accessible from your cluster nodes).

Note that, if your cluster does not have a shared file system, the only storage backend currently supported for the input and output files is Microsoft Azure storage, using a syntax similar to that used for Toil job stores (azure:<account>:<container>/<prefix>).

Graph Variant Calling Comparison

One evaluation that we perform is to test how well each graph works as a reference for calling variants against.

Performing Initial Alignments

This evaluation depends on having alignments of 1000 Genomes Project high-coverage reads to each graph.

Obtaining High Coverage Reads

The alignments, in turn, are created from reads. Reads are downloaded from the 1000 Genomes FTP using the scripts/getAltReads.py script.

For example, to download all of the available high coverage samples to the high_coverage_reads directory, running only on the local machine, you can do this:

# Remove any previous Toil job store
rm -Rf tree
# Run the download
time scripts/getAltReads.py ./tree high_coverage_reads --batchSystem=singleMachine --sample_ftp_root "ftp://ftp.1000genomes.ebi.ac.uk/vol1/ftp/data_collections/hgsv_sv_discovery/data/" --population_pattern="*" --sample_pattern "*" --file_pattern "*.high_coverage.cram"  2>&1 | tee log.txt

Aligning High Coverage Reads

Currently the easiest way to produce the alignments needed for the variant calling evaluation is to run the read alignment evaluation script on the high coverage samples. If you used the command above to download the reads, this can be accomplished thusly:

# Remove any previous Toil job store
rm -Rf tree
# Run the alignment
scripts/parallelMappingEvaluation.py ./tree graph_servers.tsv ./high_coverage_reads ./high_coverage_alignments --batchSystem=singleMachine --use_path_binaries --kmer_size=27 --edge_max=3 2>&1 | tee log.txt

It might be wise to parallelize this step across a cluster; Microsoft Azure Storage is supported for the input and output directories, using a syntax similar to that used for Toil job stores (azure:<account>:<container>/<prefix>).

Calling Variants

Preparing Input Data

The mapping steps above will result make a directory, ./high_coverage_alignments. It will contain a mapping for every region for every sample for every tool: ./high_coverage_reads/alignments/<region>/<tool>/sample.gam It will also contain tarballs of graphs and indexes used: ./high_coverage_reads/indexes/. We begin by extracting these into a new directory

 scripts/extractGraphs.py ./high_coverage_alignments/indexes/*/*/*.tar.gz ./high_coverage_graphs

The variant calling scripts require the original FASTA inputs. Uncompress them as follows:

 tar xzf data/altRegions.tar.gz -C data

Fasta and vcf data is also required to make some baseline sample graphs for the 1000 Genomes data. This is prepared with the following commands:

  tar xzf data/g1kvcf.tar.gz -C data
  tar xzf data/platinum.tar.gz -C data
  # Index truth set for vcfeval
  PLATINUM_DIR="data/platinum"
  for REGION in BRCA1 BRCA2 LRC_KIR MHC SMA
  do
      for SAMPLE in NA12877 NA12878
      do
          bgzip ${PLATINUM_DIR}/${SAMPLE}/${REGION}.vcf -c > ${PLATINUM_DIR}/${SAMPLE}/${REGION}.vcf.gz
          tabix -f -p vcf ${PLATINUM_DIR}/${SAMPLE}/${REGION}.vcf.gz
      done
  done
  tar xzf data/gatk3.tar.gz -C data
  tar xzf data/platypus.tar.gz -C data
  tar xzf data/filters.tar.gz -C data
  scripts/downloadChromFa.py
  scripts/downloadChromFa.py --leaveChr --out_fa data/g1kvcf/chrom2.fa
  # For vcfeval
  rtg format -o data/g1kvcf/chrom.sdf data/g1kvcf/chrom.fa

Generate Precision-Recall Plot using Platinum Genomes NA12878 VCF

Run the following script to generate the precision-recall plots. It is important to edit the script to specify the toil parameters, scope and comparison types before hand.

 ./scripts/call_pr.sh high_coverage_graphs/ high_coverage_alignments/alignments/ call_comparison

The output will be in call_comparison/pr_plots

Generate Variant Call Statistics

To get general statistics on the calls, including breakdown on reference vs non-reference calls, their proporitional accuracy, and trio concordance, run

 ./scripts/callStats.py call_comparison/pr_plots.best > stats.tsv

Note: call_pr.sh must be edited so that --skipBaseline is removed from OPTS and all trio samples must be run (not just NA12878)

About

All code for reproducing reference graph bakeoff results

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 4

  •  
  •  
  •  
  •