Gretel provides a command line tool for the recovery of haplotypes. We recommend the following protocol.

Read Alignment

Gretel requires your reads to be aligned to a common reference. This is to ensure that reads share a co-ordinate system, on which we can call for variants and recover haplotypes. The reference itself is of little consequence, though dropped reads will lead to evidence to be unavailable to Gretel.

Construction of a de novo consensus assembly for a metagenome is left as an exercise for the reader. Align the reads to your assembly (bowtie2, minimap2 etc.). Sort and index the alignment BAM.

Variant Calling

Gretel is robust to sequencing error and misalignment noise, thus the calling of variants need not be carefully conducted. Typically we have used samtools, but for our own Gretel pipeline, we have aggressively called all heterogenous sites in an alignment as a SNP using the snpper tool in our gretel-test repository.

For somewhat questionable reasoning, we currently require a compressed and indexed VCF:

bgzip <my.vcf>
tabix <my.vcf.gz>

Invocation of Gretel

As described in the README, Gretel is invoked as follows:

gretel <my.sort.bam> <my.vcf.gz> <contig> [-s 1startpos] [-e 1endpos] [--master master.fa] [-o output_dir]

You must provide your sorted BAM, compressed VCF, and the name of the contig on which to recover haplotypes. Use -s and -e to specify the positions on the aligned reads between which to recover haplotypes from your metagenome.

By default, Gretel will output a FASTA containing the recovered SNPs, in order, for each haplotype. Providing an optional “master” FASTA sequence will permit Gretel to “fill in” the non-SNP positions (i.e. the positions between -s and -e that do not appear in the VCF) with the nucleotide from the pseudo-reference.

Gretel Outputs


A FASTA containing each of the recovered sequences, in the order they were found. Each sequence is named <iteration>__-<log10 likelihood>. Sequences are not wrapped.


Additionally, Gretel outputs a whimsically named crumbs file, containing some potentially interesting metadata, as well as a record of each recovered haplotype. The first row is a comment containing the following (in order):

  • The number of SNPs across the region of interest
  • The number of ‘crumbs’: paired observations added to the Hansel matrix
  • The number of ‘slices’: reads with at least one observation added to the Hansel matrix
  • The chosen value of L for the L’th order Markov chain

The rest of the file contains tab-delimited metadata for each recovered haplotype:

  • The iteration number, starting from 0
  • The number of times this haplotype was returned
  • The weighted likelihood of the haplotype, given the Hansel matrix at the time the haplotype was recovered (comma-sep for each time the haplotype was returned)
  • The unweighted likelihood of the haplotype, given the Hansel matrix at the time the reads were parsed (comma-sep for each time the haplotype was returned)
  • The haplotype magnitude: total number of observations removed from the Hansel matrix by the reweighting mechanism

In practice, we rank with the weighted likelihoods to discern the haplotypes most likely to exist in the metagenome. One may attempt to use the unweighted likelihoods as a means to compare the abundance, or read support, between the returned haplotypes (i.e. not necessarily the metagenome as a whole).