TOPMed Variant Calling Pipeline (latest: Freeze 8)

Overview of this repository

This repository is intended to provide a collection of software tools used for producing TOPMed variant calls and genotypes with a comprehensive documentation that allows investigators to understand the methods and reproduce the variant calls from the same set of aligned sequence reads.

This repository reflects specific versions of software tools that are under active development in the Center for Statistical Genetics (CSG). Most of the latest version of these software tools can be accessed through multiple repositories linked as submodules, this repository is focused on a freeze of software tools that can reproduce a variant calls compatible to the latest TOPMed Freeze (Freeze 8 currently).

Outline of the variant calling procedure

Our GotCloud vt pipeline detects and genotype variants from a list of aligned sequence reads. Specifically, the pipeline consist of the following six key steps. Most of these procedure will be integrated into GotCloud software package later this year.

Our GotCloud vt pipeline detects and genotype variants from a list of aligned sequence reads. Specifically, the pipeline consist of the following six key steps. Most of these procedure will be integrated into GotCloud software package later this year.

1. Sample quality control : For each sequenced genome (in BAM/CRAMs), the genetic ancestry, sequence contamination, and biological sex are inferred using cramore cram-verify-bam and cramore vcf-normalize-depth.
2. Variant detection : For each sequenced genome (in BAM/CRAMs), candidate variants are detected by vt discover2 software tools, separated by each chromosome. The candidate variants are normalized by vt normalize algorithm.
3. Variant consolidation : For each chromosome, the called variant sites are merged across the genomes, accounting for overlap of variants between genomes, using cramore vcf-merge-candidate-variants, vt annotate_indels software tool.
4. Genotype and feature collection : For each 100kb chunk of genome, the genotyping module implemented in cramore dense-genotypes collects individual genotypes and variant features across the merged sites by iterating each sequence genome focusing on the selected region.
5. Variant filtering : We use the inferred pedigree of related and duplicated samples to calculate the Mendlian consistency statistics using king, vcf-infer-ped, vt milk-filter, and train variant classifier using Support Vector Machine (SVM) implemented in the libsvm software package and run-svm-filter software tool.

Installation

First, clone the repository by recursively cloning each submodule.

   $ git clone --recurse-submodules https://github.com/statgen/topmed_variant_calling
   (Use --recursive instead of --recurse-submodules for git version 2.12 or lower) 

Next, build each submodule using the following sets of commands

   $ cd libsvm/; make; cd ..
   $ cd apigenome; autoreconf -vfi; ./configure --prefix $PWD; make; make install; cd ..
   $ cd libStatGen; make; cd ..
   $ cd bamUtil; make; cd ..
   $ cd invNorm; make; cd ..
   $ cd htslib; autoheader; autoconf; ./configure; make; cd ..
   $ cd vt-topmed; make; cd ..
   $ cd cramore; autoreconf -vfi; ./configure; make; cd ..
   $ cd samtools; autoheader; autoconf -Wno-syntax; ./configure; make; cd ..
   $ cd bcftools; make; cd ..
   $ cd king; g++ -O3 -c *.cpp; g++ -O3 -o king *.o -lz; cd ..

Performing variant calling with example data

To produce variant calls using this pipeline, the following input files are neded:

1. Aligned sequenced reads in BAM or CRAM format. Each BAM and CRAM file should contain one sample per subject. It also must be indexed using samtools index or equivalent software tools.
2. A sequence index file. Each line should contain [Sample ID] [Full Path to the BAM/CRAM file]. See examples/index/list.107.local.crams.index for example.
3. Genomic resource files, such as FASTA, dbSNP, HapMap files. An example collection of such resources is hosted at ftp://share.sph.umich.edu/1000genomes/fullProject/hg38_resources

Here we use 107 public samples from the TOPMed project to document a reproducible variant calling pipeline that resembles the latest TOPMed variant calling pipeline. In order to do so, you need do download the following two sets of files.

1. Download the resource files for the variant calling. The tarball package is available at ftp://share.sph.umich.edu/1000genomes/fullProject/hg38_resources. The resources/ directory is assumed to be present under examples/directory in our tutorial. To download the data via command line, you may use the following command. Note that the file size is 4.5GB, and it will take a significant amount ot time.

   $ cd examples/
   $ wget ftp://share.sph.umich.edu/1000genomes/fullProject/hg38_resources/topmed_variant_calling_example_resources.tar.gz
   $ tar xzvf topmed_variant_calling_example_resources.tar.gz

2. Download 107 CRAMs from the public GCS bucket. The CRAM files are publicly available at gs://topmed-irc-share/public. However, if you want to access the data outside the Google Cloud, it will not be free of charge. To download the files, you need to set the [PROJECT_ID] that is associated with a billing account, and use gsutil tool documented at https://cloud.google.com/storage/docs/gsutil. The total amount of CRAM files is 2.17TB, and the estimate egress charge is $256 assuming $0.12/GB rate available at https://cloud.google.com/compute/pricing#internet_egress

   gsutil -u [PROJECT_ID] -m cp -r gs://topmed-irc-share/public [DESTINATION_PATH]

Here in the tutorial, we will assume that the files are stored or symbolic linked in the examples/crams directory. The examples/index/list.107.local.crams.index file contains the sample ID and CRAM file path. The CRAM file and the corresponding indices (.cram.crai) must present before running the examples.

The TOPMed variant calling has been performed on the Google Cloud. The software tool cloudify supports running jobs on the Google Cloud. However, the tutorial examples here are configured to run in a local computer. Cloud-based example commands will also be available later.

Step 1. Sample QC and Variant Detection

First, make sure to change your current directory to examples , and run the following command.

 $ cd examples/

 $ ../apigenome/bin/cloudify --cmd ../scripts/run-discovery-local.cmd 

Then follow the instruction to run make with proper arguments to complete the step. This step performs the following things

• Run vt discover2 to detect potential candidate variants to generate per-sample BCF.
• Run cramore verify-bam (verifyBamID2) to jointly estimate genetic ancestry and DNA contamination.
• Run vcf-normalize-depth to calculate relative X/Y depth to determine the sex of sequenced genome.

Upon successful completion, we expect to see the following files for each sequenced genome represented by [NWD_ID].

• out/sm/[NWD_ID]/[NWD_ID].bcf
• out/sm/[NWD_ID]/[NWD_ID].bcf.csi
• out/sm/[NWD_ID]/[NWD_ID].vb2
• out/sm/[NWD_ID]/[NWD_ID].norm.xy

More technical details can be found by directly examining ../scripts/run-discovery-local.cmd. The cloudify script take this command file and iterate the command across all samples listed in the index file in an idempotent manner using GNU make.

After running the steps above, the verifyBamID2 results and X/Y depth results should be merged together into a single index file using the following command:

  $ mkdir --p out/index/
  $ ../apigenome/bin/cram-vb-xy-index --index index/list.107.local.crams.index --dir out/sm/ --out out/index/list.107.local.crams.vb_xy.index

Upon successful completion, we expect to see the file out/index/list.107.local.crams.vb_xy.index that has a header line and one line for each of 107 samples.

Step 2. Hierarchical merge of variant sites across all samples

Next step is to merge the candidate variant sites across all sequenced genomes. Even though the example file consist of only 107 samples, we used a batch of 20 to show examples how to process hundreds of thousands of genomes without opening all the files in a single process. We typically process 1,000 samples per batch, and merge hundreds of batches together later on.

  $ ../apigenome/bin/cloudify --cmd ../scripts/run-merge-sites-local.cmd 

This command will make a merged site list for each batch and each 10Mb interval. Upon successful completion, we expect to see the following files.

• out/union/[BATCH]/b[BATCH].chr[CHROM]_[BEGIN_10Mb_CHUNK]_[END_10Mb_CHUNK].merged.sites.bcf
• out/union/[BATCH]/b[BATCH].chr[CHROM]_[BEGIN_10Mb_CHUNK]_[END_10Mb_CHUNK].merged.sites.bcf.csi where [BATCH] represent the batch (e.g. 1, 21, .., 101 in our example data, as available at examples/index/seq.batches.by.20.txt ), and [BEGIN_10Mb_CHUNK] and [END_10Mb_CHUNK] are available at examples/index/intervals/b38.intervals.X.10Mb.10Mb.txt.

These per-batch site list are further merged and consolidated using the following command.

  $ ../apigenome/bin/cloudify --cmd ../scripts/run-union-sites-local.cmd

As a result, there will be merged and consolidated site list across all samples for each 10Mb region at out/union/ directory. Upon completion, the following files are expected to be seen.

• out/union/union.chr[CHROM]_[BEGIN_10Mb_CHUNK]_[END_10Mb_CHUNK].sites.bcf
• out/union/union.chr[CHROM]_[BEGIN_10Mb_CHUNK]_[END_10Mb_CHUNK].sites.bcf.csi

Step 3. Hierarchical joint genotyping of merged variant sites

The merged site list can be jointly genotyped across the samples. However, joint genotyping of >100,000s samples is not straighforward. We again perform genotyping for 1,000 samples for eac 10Mb region. Next we hierarchically merge the genotypes across all the batches, but with much smaller region size (e.g. 100kb) to maintain the file size and running time manageable in a single machine. During this joint genotyping step, we use the DNA contamination rate, inferred genetic ancestries, and inferred sex for more accurate genotyping.

In this example, we used a batch size of 20, and use 1Mb region size when merging the genotypes across batches, to resemble the Freeze 8 calling pipeline.

First, generating per-batch genotypes for each 10Mb region can be acheived using the following command:

  $ ../apigenome/bin/cloudify --cmd ../scripts/run-batch-genotype-local.cmd 

Upon successful completion, we expect to see the following files.

• out/genotypes/batches/[BATCH]/b[BATCH].chr[CHROM]_[BEGIN_10Mb_CHUNK]_[END_10Mb_CHUNK].genotypes.bcf
• out/genotypes/batches/[BATCH]/b[BATCH].chr[CHROM]_[BEGIN_10Mb_CHUNK]_[END_10Mb_CHUNK].genotypes.bcf where [BATCH] represent the batch (e.g. 1, 21, .., 101 in our example data, as available at examples/index/seq.batches.by.20.txt ), and [BEGIN_10Mb_CHUNK] and [END_10Mb_CHUNK] are available at examples/index/intervals/b38.intervals.X.10Mb.10Mb.txt.

Next, pasting the genotypes across all batches while recalculating the variant features is achieved using the following command:

  $ ../apigenome/bin/cloudify --cmd ../scripts/run-paste-genotype-local.cmd

Upon successful completion, we expect to see the following files.

• out/genotypes/merged/[CHROM]/merged.chr[CHROM]_[BEGIN_1Mb_CHUNK]_[END_1Mb_CHUNK].genotypes.bcf
• out/genotypes/merged/[CHROM]/merged.chr[CHROM]_[BEGIN_1Mb_CHUNK]_[END_1Mb_CHUNK].genotypes.bcf.csi
• out/genotypes/minDP0/[CHROM]/merged.chr[CHROM]_[BEGIN_1Mb_CHUNK]_[END_1Mb_CHUNK].gtonly.minDPO.bcf
• out/genotypes/minDP0/[CHROM]/merged.chr[CHROM]_[BEGIN_1Mb_CHUNK]_[END_1Mb_CHUNK].gtonly.minDPO.bcf.csi
• out/genotypes/minDP10/[CHROM]/merged.chr[CHROM]_[BEGIN_1Mb_CHUNK]_[END_1Mb_CHUNK].gtonly.minDP10.bcf
• out/genotypes/minDP10/[CHROM]/merged.chr[CHROM]_[BEGIN_1Mb_CHUNK]_[END_1Mb_CHUNK].gtonly.minDP10.bcf.csi
• out/genotypes/hgdp/[CHROM]/merged.chr[CHROM]_[BEGIN_1Mb_CHUNK]_[END_1Mb_CHUNK].gtonly.minDP10.hgdp.bcf
• out/genotypes/hgdp/[CHROM]/merged.chr[CHROM]_[BEGIN_1Mb_CHUNK]_[END_1Mb_CHUNK].gtonly.minDP10.hgdp.bcf.csi where [BATCH] represent the batch (e.g. 1, 21, .., 101 in our example data, as available at examples/index/seq.batches.by.20.txt ), and [CHROM], [BEGIN_1Mb_CHUNK], and [END_1Mb_CHUNK] are available in 1st, 4th, and 5th column at examples/index/intervals/b38.intervals.X.10Mb.1Mb.txt.

Step 4. Inferring Duplicated and Related Individuals

The step above not only pastes the genotypes across the samples, but also generate multiple versions of genotypes, such as minDP0 (no missing genotypes), minDP10 (genotypes marked missing if depth is 10 or less)., and hgdp (extract only HGDP-polymorphic sites).

The HGDP genotypes on autosomes can be merged together across all regions using the following commands:

   $ cut -f 1,4,5 index/intervals/b38.intervals.X.10Mb.1Mb.txt | grep -v ^chrX | awk '{print "out/genotypes/hgdp/"$1"/merged."$1"_"$2"_"$3".gtonly.minDP0.hgdp.bcf"}' > out/index/hgdp.auto.bcflist.txt

   $ ../bcftools/bcftools concat -n -f out/index/hgdp.auto.bcflist.txt -Ob -o out/genotypes/hgdp/merged.autosomes.gtonly.minDP0.hgdp.bcf

These HGDP-site BCF file can be convered into PLINK format, and pedigree can be inferred using king and vcf-infer-ped software tools as follows. Note that plink-1.9 is not the part of this repository and you need to obtain the software separately at https://www.cog-genomics.org/plink2 .

   $ plink-1.9 --bcf out/genotypes/hgdp/merged.autosomes.gtonly.minDP0.hgdp.bcf --make-bed --out out/genotypes/hgdp/merged.autosomes.gtonly.minDP0.hgdp.plink --allow-extra-chr

   $ ../king/king -b out/genotypes/hgdp/merged.autosomes.gtonly.minDP0.hgdp.plink.bed --degree 4 --kinship --prefix out/genotypes/hgdp/merged.autosomes.gtonly.minDP0.hgdp.king

   $ ../apigenome/bin/vcf-infer-ped --kin0 out/genotypes/hgdp/merged.autosomes.gtonly.minDP0.hgdp.king.kin0 --sex out/genotypes/merged/chr1/merged.chr1_1_1000000.sex_map.txt --out out/genotypes/hgdp/merged.autosomes.gtonly.minDP0.hgdp.king.inferred.ped

The inferred pedigree file using these procedure only contains nuclear families and duplicates in a specialized PED format. When a sample is duplicated, all sample IDs representing the duplicated sample (in the 2nd column) need to presented in a comma-separated way. In the 3rd and 4th column to represend their parents, only representative sample ID (first among comma-separated duplicate ID) is required.

Step 5. Run SVM variant filtering guided by the inferred pedigree.

Using the infered pedigree, we compute duplicate and Mendelian consistency, and use the information to aid variant filtering. First, duplicate and Mendelian consistency is compute using the following milk (Mendelian-inheritance under likelihood framework) command

   $ ../apigenome/bin/cloudify --cmd ../scripts/run-milk-local.cmd

• out/milk/[CHROM]/milk.chr[CHROM]_[BEGIN_1Mb_CHUNK]_[END_1Mb_CHUNK].full.vcf.gz
• out/milk/[CHROM]/milk.chr[CHROM]_[BEGIN_1Mb_CHUNK]_[END_1Mb_CHUNK].sites.vcf.gz
• out/milk/[CHROM]/milk.chr[CHROM]_[BEGIN_1Mb_CHUNK]_[END_1Mb_CHUNK].sites.vcf.gz.tbi

The results are merged across each chromosome in the following way.

   $ cut -f 1,4,5 index/intervals/b38.intervals.X.10Mb.1Mb.txt | awk '{print "out/milk/"$1"/milk."$1"_"$2"_"$3".sites.vcf.gz"}' > out/index/milk.autoX.bcflist.txt

   $ (seq 1 22; echo X;) | xargs -I {} -P 10 bash -c "grep chr{}_ out/index/milk.autoX.bcflist.txt | ../bcftools/bcftools concat -f /dev/stdin -Oz -o out/milk/milk.chr{}.sites.vcf.gz"

   $ (seq 1 22; echo X;) | xargs -I {} -P 10 ../htslib/tabix -f -pvcf out/milk/milk.chr{}.sites.vcf.gz 

Finally, the SVM filtering step is performed using vcf-svm-milk-filter tool. The training is typically done with one chromosome, for example chr2.

   $ mkdir out/svm

   $ ../apigenome/bin/vcf-svm-milk-filter --in-vcf out/milk/milk.chr2.sites.vcf.gz --out out/svm/milk_svm.chr2 --ref resources/ref/hs38DH.fa --dbsnp resources/ref/dbsnp_142.b38.vcf.gz --posvcf resources/ref/hapmap_3.3.b38.sites.vcf.gz --posvcf resources/ref/1000G_omni2.5.b38.sites.PASS.vcf.gz --train --centromere resources/ref/hg38.centromere.bed.gz --bgzip ../htslib/bgzip --tabix ../htslib/tabix --invNorm $GC/bin/invNorm --svm-train $GC/bin/svm-train --svm-predict $GC/bin/svm-predict 

After finishing the training, the other chromosomes uses the same training model to perform SVM filtering.

   $ (seq 1 22; echo X;) | grep -v -w 2 | xargs -I {} -P 10 ../apigenome/bin/vcf-svm-milk-filter --in-vcf out/milk/milk.chr{}.sites.vcf.gz --out out/svm/milk_svm.chr{} --ref resources/ref/hs38DH.fa --dbsnp resources/ref/dbsnp_142.b38.vcf.gz --posvcf resources/ref/hapmap_3.3.b38.sites.vcf.gz --posvcf resources/ref/1000G_omni2.5.b38.sites.PASS.vcf.gz --model out/svm/milk_svm.chr2.svm.model --centromere resources/ref/hg38.centromere.bed.gz --bgzip ../htslib/bgzip --tabix ../htslib/tabix --invNorm $GC/bin/invNorm --svm-train ../libsvm/svm-train --svm-predict ../libsvm/svm-predict 

The details of each steps are elaborated below.

Variant Detection Details

Variant detection from each sequence (ang aligned) genome is performed by vt discover2 software tool. The script step-1-detect-variants.pl provide a mean to automate the variant detection across a large number of sequence genome.

The variant detection algorithm consider a variant as a potential candidate variant if there exists a mismatch between the aligned sequence reads and the reference genome. Because such a mismatch can easily occur by random errors, only potential candidate variants passing the following criteria are considered to be candidate variants in the next steps.

1. At least two identical evidence of variants must be observed from aligned sequence reads.
2. Each individual evidence will be normalized using the normalization algorithm implemented in vt normalize software tools.
3. Only evidence on the reads with mapping quality 20 or greater will be considered.
4. Duplicate reads, QC-passed reads, supplementary reads, secondary reads will be ignored.
5. Evidence of variant within overlapping fragments of read pairs will not be double counted. Either end of the overlapping read pair will be soft-clipped using bam clipOverlap software tool.
6. Assuming per-sample heterozygosity of 0.1%, the posterior probability of having variant at the position should be greater than 50%. The method is equivalent to the glfSingle model described in http://www.ncbi.nlm.nih.gov/pubmed/25884587

The variant detection step is required only once per sequenced genome, when multiple freezes of variant calls are produced over the course of time.

Variant Consolidation

Variants detected from the discovery step will be merged across all samples. This step is implemented in the step-2-detect-variants.pl scripts.

1. Each non-reference allele normalized by vt normalize algorithm is merged across the samples, and unique alleles are printed as biallelic candidate variants. The algorithm is published at http://www.ncbi.nlm.nih.gov/pubmed/25701572
2. If there are alleles overlapping with other SNPs and Indels, overlap_snp and overlap_indel filters are added in the FILTER column of the corresponding variant.
3. If there are tandem repeats with 2 or more repeats with total repeat length of 6bp or longer, the variant is annotated as potential VNTR (Variant Number Tandem Repeat), and overlap_vntr filters are added to the variant overlapping with the repeat track of the putative VNTR.

Variant Genotyping and Feature Collection

The genotyping step iterate all the merged variant site across the sample. It iterates each BAM/CRAM files one at a time sequentially for each 1Mb chunk to perform contamination-adjusted genotyping and annotation of variant features for filtering. The following variant features are calculated during the genotyping procedure.

• AVGDP : Average read depth per sample
• AC : Non-reference allele count
• AN : Total number of alleles
• GC : Genotype count
• GN : Total genotype counts
• HWEAF : Allele frequency estimated from PL under HWE
• HWDAF : Genotype frequency estimated from PL under HWD
• IBC : [ Obs(Het) – Exp(Het) ] / Exp[Het]
• HWE_SLP : -log(HWE likelihood-ratio test p-value) ⨉ sign(IBC)
• ABE : Average fraction [#Ref Allele] across all heterozygotes
• ABZ : Z-score for tesing deviation of ABE from expected value (0.5)
• BQZ: Z-score testing association between allele and base qualities
• CYZ: Z-score testing association between allele and the sequencing cycle
• STZ : Z-score testing association between allele and strand
• NMZ : Z-score testing association between allele and per-read mismatches
• IOR : log [ Obs(non-ref, non-alt alleles) / Exp(non-ref, non-alt alleles) ]
• NM1 : Average per-read mismatches for non-reference alleles
• NM0 : Average per-read mismatches for reference alleles

The genotyping was done by adjusting for potential contamination. It uses adjusted genotype likelihood similar to the published method https://github.com/hyunminkang/cleancall, but does not use estimated population allele frequency for the sake of computational efficiency. It conservatively assumes that probability of observing non-reference read given homozygous reference genotype is equal to the half of the estimated contamination level, (or 1%, whichever is greater). The probability of observing reference reads given homozygous non-reference genotype is calculated in a similar way. This adjustment makes the heterozygous call more conservatively when the reference and non-reference allele is strongly imbalanced. For example, if 45 reference alleles and 5 non-reference alleles are observed at Q40, the new method calls it as homozygous reference genotype while the original method ignoring potential contamination calls it as heterozygous genotype. This adjustment improves the genotype quality of contaminated samples by reducing genotype errors by several folds.

Variant Filtering

The variant filtering in TOPMed Freeze 8 were performed by (1) first calculating Mendelian consistency scores using known familial relatedness and duplicates, and (2) training SVM classifier between the known variant site (positive labels) and the Mendelian inconsistent variants (negative labels).

The negative labels are defined if the Bayes Factor for Mendelian consistency quantified as Pr(Reads | HWE, Pedigree) / Pr(Reads | HWD, no Pedigree ) less than 0.001. Also variant is marked as negative labels if 3 or more samples show 20% of non-reference Mendelian discordance within families or genotype discordance between duplicated samples.

The positive labels are the SNPs found polymorphic either in the 1000 Genomes Omni2.5 array or in HapMap 3.3, with additional evidence of being polymorphic from the sequenced samples. Variants eligible to be marked both positive and negative labels are discarded from the labels. The SVM scores trained and predicgted by libSVM software tool will be annotated in the VCF file.

Two additional hard filtering was applied additionally. First is excessive heterozygosity filter (EXHET), if the Hardy-Weinberg disequilbrium p-value was less than 1e-6 in the direction of excessive heterozygosity. Another filter is Mendelian discordance filter (DISC), with 3 or more Mendelian discordance or duplicate discordance observed from the samples.

Questions

For further questions, pleast contact Hyun Min Kang (hmkang@umich.edu) and Jonathon LeFaive (lefaivej@umich.edu).