BayesTyper performs genotyping of all types of variation (including structural and complex variation) based on an input set of variants and read k-mer counts. Internally, BayesTyper uses exact alignment of k-mers to a graph representation of the input variation and reference sequence in combination with a probabilistic model of k-mer counts to do genotyping. The variant representation ensures that the resulting calls are not biased towards the reference sequence as is otherwise generally the case when basing calls only on mapped reads.

The BayesTyper was used to integrate mapping- and assembly-based calls in the GenomeDenmark project. A manuscript describing the method is currently in revision.

The BayesTyper is being developed by Jonas Andreas Sibbesen, Lasse Maretty and Anders Krogh at the Section for Computational and RNA Biology, Department of Biology, University of Copenhagen.

Sensitive calling of structural variation typically requires running multiple callers to ensure sensitivity yet this leads to the problem of integrating calls across call-sets. The BayesTyper can be used to produce a fully integrated call-set including SNVs, indels and complex variation from input variant candidates produced by a panel of methods; the panel must include standard SNV and indel calls e.g. from GATK, Freebayes or Platypus.

A signficant amount of both simple and complex variation is already known from large population-scale studies. As some of these variants may be missed in a study - even when running multiple methods - due to alignment bias, we provide a database containing common SNPs/indels together with complex variants that can be combined with in-sample calls (i.e. calls based only on the study data) to improve sensitivity. This approach can for instance be used to quickly augment a set of standard SNV and indel calls (e.g. from GATK) with structural variation by running BayesTyper on the SNV/indel calls combined with our variation database. For higher sensitivity, in-sample complex variation calls can be combined with the database to produce the final intergrated call-set.

BayesTyper can either be build from source or a static Linux x86_64 build can be downloaded under releases.

• gcc (c++11 support required. Tested with gcc 4.8 and 4.9)
• CMake (version 2.8.0 or higher)
• Boost (tested with version 1.55.0 and 1.56.0)

1. git clone https://github.com/bioinformatics-centre/BayesTyper.git
2. cd BayesTyper
3. mkdir build && cd build
4. cmake ..
5. make

The compiled bayesTyper and bayesTyperTools binaries are located in the bin directory.

The BayesTyper package contains bayesTyper, which does the genotyping, and bayesTyperTools, which is used to pre- and post-process VCF files for BayesTyper.

1. Count k-mers

   1. Run KMC3 on each sample: kmc -k55 sample_1.fq sample_1
      • This will output k-mer counts to sample_1.kmc_pre and sample_1.kmc_suf.
      • For low coverage data (<20X), include singleton k-mers by adding -ci1 to the kmc3 commandline.

2. Prepare variant input

   IMPORTANT: The variant input must contain simple variants (SNPs and short indels). These can be obtained by first running a standard tool like GATK, Platypus or Freebayes and then combine these variants with structural variants calls and/or prior as desired. At least 1 million simple variants are required.

   1. If required, convert allele IDs (e.g. <DEL>) to sequence: bayesTyperTools convertAlleleId -o sample_1_sv_calls_seq -v sample_1_sv_calls.vcf -g hg38.fa

      • Currently <DEL>, <DUP>, <CN[digit(s)]>, <CNV>, <INV>, <INS:ME:[sequence name]> are supported. The latter require a fasta file with the mobile element insertion sequences.
      • This step can be skipped if the variant sets does not include any allele IDs (e.g. GATK, Platypus and Freebayes output).

   2. Normalise variants using Bcftools: bcftools norm -o sample_1_gatk_norm.vcf -f hg38.fa sample_1_gatk.vcf

   3. Combine variant sets: bayesTyperTools combine -o bayesTyper_input -v gatk:sample_1_gatk_norm.vcf,gatk:sample_2_gatk_norm.vcf,gatk:sample_3_gatk_norm.vcf,varDB:SNP_dbSNP150common_SV_1000g_dbSNP150all_GDK_GoNL_GTEx_GRCh38.vcf

      • The contig fields in the headers need to be identical between variant sets and the variants sorted in the same order as the fields.
      • *IMPORTANT: The variant input must contain simple variants (SNPs and short indels). These can be obtained by first running a standard tool like GATK, Platypus or Freebayes

3. Genotype variants

   IMPORTANT: If you want to run BayesTyper on more than 30 samples, you should run BayesTyper in batches of 30 samples or less but using the full set of variants (i.e. across all individuals)

   1. Prepare sample information: Create tsv file with one sample per row with columns <sample_id>, <sex> and <path_to_kmc3_output> (example)

   2. Run BayesTyper: bayesTyper -o integrated_calls -s samples.tsv -v bayesTyper_input.vcf -g hg38.fa -p <threads> > bayesTyper_log.txt

      • Decoy sequences: BayesTyper can be provided with decoy sequences using '-d' to handle sequence similarities between genotyped regions and non-genotyped regions (e.g. the mitochondrial genome and unplaced contigs in the reference). Matching reference and decoy sequences are available for
        • GRCh37: Reference and decoy
        • GRCh38: Reference and decoy

4. Filter output

   1. Run filtering: bayesTyperTools filter -o integrated_calls_filtered -v integrated_calls.vcf -g hg38.fa --kmer-coverage-filename integrated_calls_kmer_coverage_estimates.txt
      • By default only genotypes with high confidence (posterior probability >= 0.99) are kept. If low confident genotypes are needed in a downstream analyses this can be changed using the option --min-genotype-posterior.

• BayesTyper_varDB_GRCh37
• BayesTyper_varDB_GRCh38

*Reference and alternative alleles containing ambiguous nucleotides were removed from all variant sources.

Third-party software used by BayesTyper (distributed together with the BayesTyper source code).

• Edlib
• Eigen
• KMC