Skip to content

Latest commit

 

History

History
185 lines (138 loc) · 7.94 KB

deepvariant-hybrid-case-study.md

File metadata and controls

185 lines (138 loc) · 7.94 KB
Raw

Using the DeepVariant hybrid model for calling on PacBio and Illumina data together

In this case study we describe the hybrid DeepVariant model and apply it to the combination of two datasets:

  1. PacBio HiFi data on HG003 aligned with pbmm2.
  2. Illumina NovaSeq on HG003 aligned with BWA MEM.

The FASTQ files come from the PrecisionFDA Truth challenge v2.

They are merged together into a single bam file using samtools merge, and then a new index is created for this hybrid bam using samtools index. Note that the two original bam files must have the same sample name.

Finally, we assess the quality of the DeepVariant variant calls with hap.py.

To make it faster to run over this case study, we run only on chromosome 20.

Background on the hybrid model

This is what the pileup image looks like: The longer PacBio reads are shown at the top, followed by the shorter Illumina reads at the bottom.

Example of a hybrid pileup for one variant

A DeepVariant hybrid model was first trained for the PrecisionFDA Truth Challenge V2, and this release model is similar except it has been re-trained with additional datasets including the HG004 truth set that was held out during the challenge.

Interestingly, DeepVariant didn't strictly need any code changes to work on hybrid data -- it worked the first time we tried. But we knew from many previous experiments that Illumina reads benefit from being realigned to a haplotype graph, which is too time consuming and unnecessary for the PacBio long reads. We added a small code change to specifically realign all the short reads to the haplotype graph, while leaving longer reads with their original alignments. This created a small but measurable improvement, and was the only code change we made to enable the hybrid model, aside from training a dedicated hybrid model and exposing it for easy use through the --model_type parameter in run_deepvariant.py. Much of the work we put into DeepVariant is in experimenting with different approaches, training on more and better data, and carefully evaluating the models before releasing them. We did the same with this hybrid model.

Prepare environment

Tools

Docker will be used to run DeepVariant and hap.py,

Download Reference

We will be using GRCh38 for this case study.

mkdir -p reference

FTPDIR=ftp://ftp.ncbi.nlm.nih.gov/genomes/all/GCA/000/001/405/GCA_000001405.15_GRCh38/seqs_for_alignment_pipelines.ucsc_ids

curl ${FTPDIR}/GCA_000001405.15_GRCh38_no_alt_analysis_set.fna.gz | gunzip > reference/GRCh38_no_alt_analysis_set.fasta
curl ${FTPDIR}/GCA_000001405.15_GRCh38_no_alt_analysis_set.fna.fai > reference/GRCh38_no_alt_analysis_set.fasta.fai

Download Genome in a Bottle Benchmarks

We will benchmark our variant calls against v4.2 of the Genome in a Bottle small variant benchmarks for HG003.

mkdir -p benchmark

FTPDIR=ftp://ftp.ncbi.nlm.nih.gov//giab/ftp/data/AshkenazimTrio/analysis/NIST_v4.2_SmallVariantDraftBenchmark_07092020

curl ${FTPDIR}/HG003_GRCh38_1_22_v4.2_benchmark.bed > benchmark/HG003_GRCh38_1_22_v4.2_benchmark.bed
curl ${FTPDIR}/HG003_GRCh38_1_22_v4.2_benchmark.vcf.gz > benchmark/HG003_GRCh38_1_22_v4.2_benchmark.vcf.gz
curl ${FTPDIR}/HG003_GRCh38_1_22_v4.2_benchmark.vcf.gz.tbi > benchmark/HG003_GRCh38_1_22_v4.2_benchmark.vcf.gz.tbi

Download HG003 chr20 BAM

We'll use a HG003 BAM file that contains pacbio and illumina data merged together using samtools merge. See the top of this page for more information on those two datasets.

mkdir -p input
HTTPDIR=https://storage.googleapis.com/deepvariant/hybrid-case-study-testdata

curl ${HTTPDIR}/HG003_hybrid_35x_ilmn_35x_pacb.grch38.phased.chr20.bam > input/HG003_hybrid_35x_ilmn_35x_pacb.grch38.phased.chr20.bam
curl ${HTTPDIR}/HG003_hybrid_35x_ilmn_35x_pacb.grch38.phased.chr20.bam.bai > input/HG003_hybrid_35x_ilmn_35x_pacb.grch38.phased.chr20.bam.bai

Running DeepVariant

DeepVariant pipeline consists of 3 steps: make_examples, call_variants, and postprocess_variants. You can run DeepVariant with just one command using the run_deepvariant script.

Running on a CPU-only machine

Here we specify --regions chr20 to run on just chromosome 20, saving time so you can run this case study within about half an hour (tested on 64 CPUs).

mkdir -p output
mkdir -p output/intermediate_results_dir

BIN_VERSION="1.1.0"

sudo docker run \
  -v "${PWD}/input":"/input" \
  -v "${PWD}/output":"/output" \
  -v "${PWD}/reference":"/reference" \
  google/deepvariant:"${BIN_VERSION}" \
  /opt/deepvariant/bin/run_deepvariant \
  --model_type "HYBRID_PACBIO_ILLUMINA" \
  --ref /reference/GRCh38_no_alt_analysis_set.fasta \
  --reads /input/HG003_hybrid_35x_ilmn_35x_pacb.grch38.phased.chr20.bam \
  --output_vcf /output/HG003.output.vcf.gz \
  --output_gvcf /output/HG003.output.g.vcf.gz \
  --num_shards $(nproc) \
  --regions chr20:10,000,000-10,100,000 \
  --intermediate_results_dir /output/intermediate_results_dir

By specifying --model_type HYBRID_PACBIO_ILLUMINA, you'll be using a model that is best suited for (and trained on) the combination of PacBio Hifi long reads and Illumina short reads.

--intermediate_results_dir flag is optional. By specifying it, the intermediate outputs of make_examples and call_variants stages can be found in the directory. After the command, you can find these files in the directory:

call_variants_output.tfrecord.gz
gvcf.tfrecord-?????-of-?????.gz
make_examples.tfrecord-?????-of-?????.gz

To see the pileup images visually, check out show_examples.

For running on GPU machines, or using Singularity instead of Docker, see Quick Start. Just make sure to use --model_type HYBRID_PACBIO_ILLUMINA when running on combined PacBio and Illumina data.

Benchmark with hap.py

See hap.py documentation for more details on the parameters and outputs.

mkdir -p happy

sudo docker run \
  -v "${PWD}/benchmark":"/benchmark" \
  -v "${PWD}/input":"/input" \
  -v "${PWD}/output":"/output" \
  -v "${PWD}/reference":"/reference" \
  -v "${PWD}/happy:/happy" \
  pkrusche/hap.py /opt/hap.py/bin/hap.py \
  /benchmark/HG003_GRCh38_1_22_v4.2_benchmark.vcf.gz \
  /output/HG003.output.vcf.gz \
  -f /benchmark/HG003_GRCh38_1_22_v4.2_benchmark.bed \
  -r /reference/GRCh38_no_alt_analysis_set.fasta \
  -o /happy/happy.output \
  --engine=vcfeval \
  -l chr20

Output:

Benchmarking Summary:
  Type Filter  TRUTH.TOTAL  TRUTH.TP  TRUTH.FN  QUERY.TOTAL  QUERY.FP  QUERY.UNK  FP.gt  METRIC.Recall  METRIC.Precision  METRIC.Frac_NA  METRIC.F1_Score  TRUTH.TOTAL.TiTv_ratio  QUERY.TOTAL.TiTv_ratio  TRUTH.TOTAL.het_hom_ratio  QUERY.TOTAL.het_hom_ratio
 INDEL    ALL        10634     10601        33        22758        52      11573     16       0.996897          0.995351        0.508524         0.996123                     NaN                     NaN                   1.749861                   2.587958
 INDEL   PASS        10634     10601        33        22758        52      11573     16       0.996897          0.995351        0.508524         0.996123                     NaN                     NaN                   1.749861                   2.587958
   SNP    ALL        70209     70186        23       101326        37      31088     13       0.999672          0.999473        0.306812         0.999573                2.297347                 1.84949                   1.884533                   2.081099
   SNP   PASS        70209     70186        23       101326        37      31088     13       0.999672          0.999473        0.306812         0.999573                2.297347                 1.84949                   1.884533                   2.081099

Notice that F1 scores are above 0.999 for SNPs and above 0.995 for indels!