Skip to content
Switch branches/tags


Failed to load latest commit information.
Latest commit message
Commit time

Paragraph: a suite of graph-based genotyping tools


Accurate genotyping of known variants is a critical for the analysis of whole-genome sequencing data. Paragraph aims to facilitate this by providing an accurate genotyper for Structural Variations with short-read data.

Please reference Paragraph using:

Genotyping data in this paper can be found at paper-data/download-instructions.txt

For details of population genotyping, please also refer to:


Please check doc/ for system requirements and installation instructions.

Run Paragraph from VCF

Test example

After installation, run script from the build/bin directory on an example dataset as follows:

python3 bin/ -i share/test-data/round-trip-genotyping/candidates.vcf \
                          -m share/test-data/round-trip-genotyping/samples.txt \
                          -r share/test-data/round-trip-genotyping/dummy.fa \
                          -o test \

This runs a simple genotyping example for two test samples.

  • candidates.vcf: this specifies candidate SV events in a vcf format.
  • samples.txt: Manifest that specifies some test BAM files. Tab or comma delimited.
  • dummy.fa a short dummy reference which only contains chr1

The output folder test then contains gzipped json for final genotypes:

$ tree test
├── grmpy.log            #  main workflow log file
├── genotypes.vcf.gz     #  Output VCF with individual genotypes
├── genotypes.json.gz    #  More detailed output than genotypes.vcf.gz
├── variants.vcf.gz      #  The input VCF with unique ID from Paragraph
└── variants.json.gz     #  The converted graphs from input VCF (no genotypes)

If successful, the last 3 lines of genotypes.vcf.gz will the same as in expected file.

Input requirements

VCF format

paraGRAPH will independently genotype each entry of the input VCF. You can use either indel-style representation (full REF and ALT allele sequence in 4th and 5th columns) or symbolic alleles, as long as they meet the format requirement of VCF 4.0+.

Currently we support 4 symbolic alleles:

  • <DEL> for deletion
    • Must have END key in INFO field.
  • <INS> for insertion
    • Must have a key in INFO field for insertion sequence (without padding base). The default key is SEQ.
    • For blockwise swap, we strongly recommend using indel-style representation, other than symbolic alleles.
  • <DUP> for duplication
    • Must have END key in INFO field. paraGRAPH assumes the sequence between POS and END being duplicated for one more time in the alternative allele.
  • <INV> for inversion
    • Must have END key in INFO field. paraGRAPH assumes the sequence between POS and END being reverse-complemented in the alternative allele.

Sample Manifest

Must be tab-deliemited.

Required columns:

  • id: Each sample must have a unique ID. The output VCF will include genotypes for all samples in the manifest
  • path: Path to the BAM/CRAM file.
  • depth: Average depth across the genome. Can be calculated with bin/idxdepth (faster than samtools).
  • read length: Average read length (bp) across the genome.

Optional columns:

  • depth sd: Specify standard deviation for genome depth. Used for the normal test of breakpoint read depth. Default is sqrt(5*depth).
  • depth variance: Square of depth sd.
  • sex: Affects chrX and chrY genotyping. Allow "male" or "M", "female" or "F", and "unknown" (quotes shouldn't be included in the manifest). If not specified, the sample will be treated as unknown.

Run time

  • On a 30x HiSeqX sample, Paragraph typically takes 1-2 seconds to genotype a simple SV in confident regions.

  • If the SV is in a low-complexity region with abnormal read pileups, the running time could vary.

  • For efficiency, it is recommended to manually set the "-M" option (maximum allowed read count for a variant) to skip these high-depth regions. We recommend "-M" as 20 times of your mean sample depth.

Population-scale genotyping

To efficiently genotype SVs across a population, we recommend doing single-sample mode as follows:

  • Create a manifest for each single sample
  • Run for each manifest. Be sure to set "-M" option for each sample according to its depth.
  • Multithreading (option "-t") is highly recommended for population-scale genotyping
  • Merge all genotypes.vcf.gz to create a big VCF of all samples. You can use either bcftools merge or your custom script.

Run Paragraph on complex variants

For more complicated events (e.g. genotype a deletion together with its nearby SNP), you can provide a custimized JSON to paraGRAPH:

Please follow the pattern in example JSON and make sure all required keys are provided. Here is a visualization of this sample graph.

To obtain graph alignments for this graph (including all reads), run:

bin/paragraph -b <input BAM> \
              -r <reference fasta> \
              -g <input graph JSON> \
              -o <output JSON path> \
              -E 1

To obtain the algnment summary, genotypes of each breakpoint, and the whole graph, run:

bin/grmpy -m <input manifest> \
          -r <reference fasta> \
          -i <input graph JSON> \
          -o <output JSON path> \
          -E 1

If you have multiple events listed in the input JSON, can help you to run multiple grmpy jobs together.

Further Information

Please check github wiki for common usage questions and errors.


External links

  • The Illumina/Polaris repository gives the short-read sequencing data we used to test our method in population.


The LICENSE file contains information about libraries and other tools we use, and license information for these.