DNAscope Machine Learning Model

A machine learning model for accurate and efficient germline small-variants detection

Sentieon DNAscope combines the robust and well-established preprocessing and assembly mathematics of the GATK’s HaplotypeCaller with a machine-learned genotyping model, achieving superb SNP and insertion/deletion accuracy as compared to state-of-the-art tools, while using much reduced computational cost.

Table of Contents

• Goal of a machine learning model in DNAscope
• Highlights
• DNAscope machine learning pipeline
• Performance on Whole Genome Sequencing (WGS) data

Goal of a machine learning model in DNAscope

From Sentieon software version 201808.01 onwards, DNAscope allows you to use a model to perform variant calling with higher accuracy by improving the candidate detection and filtering.

Sentieon can provide you with a model trained using a subset of the data from the GIAB truth-set found in https://github.com/genome-in-a-bottle. The model was created by processing samples HG001 and HG005 through a pipeline consisting of Sentieon BWA-mem alignment and Sentieon deduplication, and using the variant calling results to calibrate a model to fit the truth-set. Sentieon also provides DNAscope model training tool for you to create your model based on your own data.

Highlights

• High Accuracy
• Rapid Turnaround Time
• Cost Reduction
• Easy Deployment
• Customizable Model

DNAscope machine learning pipeline

Running DNAscope on-premises

FASTQ -> BAM -> VCF

1. Sentieon license

   Update Sentieon packages location and license file in dnascope.sh.

   export SENTIEON_INSTALL_DIR=/home/release/sentieon-genomics-201808.06 #your Sentieon package location
   export SENTIEON_LICENSE=/home/bundle/sentieon.lic #your license file location

   If you do not have a Sentieon license/package yet, please feel free to request free trial by filling out the form. Alternatively, you could run this pipeline on Google Cloud or AWS. A 14 days free trial license will be automatically applied to your account.

2. Location of input files

   Before running, you need to set the following variables in dnascope.sh.

   • fastq_folder: fastq file(s) folder

   • fastq_1: fastq file name

   • fastq_2: second fastq file, if using Illumina paired data

   • model: DNAscope model file

   • PCRFREE: boolean to indicate whether the sample is PCR Free or not. Set to true for PCR Free samples.

   • fasta: reference file

   • dbsnp: dbSNP file

3. Running the pipeline

   sh dnascope.sh

If you have any further question, please refer to Sentieon's Appnotes for DNAscope Machine Learning Model and DNAseq pipeline example script in the manual.

Running DNAscope in the cloud

Google Cloud Platform(GCP)

1. Set up

   Please follow steps in "Before you begin" section on Google Cloud page to set up your environment: Running a Sentieon DNAseq Pipeline.

   Right now, we are granting free-trial license to your account automatically. You will get 14 days free trial beginning when you first run a Sentieon pipeline.

2. Run the pipeline via gcloud alpha genomics API.

   Make necessary changes in gcp/run.sh and make sure the pipeline file gcp/dnascope_gcp.yaml is in your current working directory. With the current inputs, the command will run the DNAscope + ML pipeline on PrecisionFDA Truth Challenge HG002 sample, which is used to demonstrate model performance in the following section.

   In run.sh:

   BUCKET="<your bucket>"
   gcloud alpha genomics pipelines run \
     --pipeline-file dnascope_gcp.yaml \
     --inputs SENTIEON_VERSION=201808.07 \
     --inputs FQ1=gs://sentieon-dnascope-model/data/HG002-NA24385-50x_1.fastq.gz\
     --inputs FQ2=gs://sentieon-dnascope-model/data/HG002-NA24385-50x_2.fastq.gz\
     --inputs REF=gs://sentieon-test/pipeline_test/reference/hs37d5.* \
     --inputs DBSNP=gs://sentieon-test/pipeline_test/reference/dbsnp_138.b37.vcf.* \
     --inputs ML_MODEL=gs://sentieon-dnascope-model/models/SentieonModelBeta0.4a.model \
     --outputs outputPath=gs://$BUCKET/output/ \
     --logging gs://$BUCKET/output/logging \
     --disk-size datadisk:600 \
     --cpus 64 \
     --memory 56

3. Check job status

   You will get a run id after running the pipeline. You could run gcloud alpha genomics operations describe <YOUR-RUNID> to check the job status.

If you would like to run other sentieon pipelines on GCP, please refer to our sentieon-google-genomics repository for more examples scripts.

Performance on Whole Genome Sequencing (WGS) data

Here we demonstrate DNAscope's performance on PrecisionFDA Truth Challenge HG002 sample.

Software source

Sentieon packages/models are stored on AWS. You could get Sentieon tools by running:

 SENTIEON_VERSION="version-you-want" #For example 201808.07, Find released versions here https://support.sentieon.com/manual/appendix/releasenotes/
 INSTALL_DIR="your-install-dir"
 wget -nv -O - "https://s3.amazonaws.com/sentieon-release/software/sentieon-genomics-${SENTIEON_VERSION}.tar.gz" | tar -zxf - -C ${INSTALL_DIR} 

Sentieon Models:

https://s3.amazonaws.com/sentieon-release/other/SentieonDNAscopeModelBeta0.4a-201808.05.model

Data source

1. Reference Genome: hs37d5.fa.
2. Test Data: NIST HG002 from PrecisionFDA Truth Challenge.
3. Model File: Sentieon DNAscope Machine Learning Model version 0.5.
4. Truth VCF and BED file for evaluation: NIST GIAB project, version v3.3.2. (ftp://ftp-trace.ncbi.nlm.nih.gov/giab/ftp/release/AshkenazimTrio/HG002_NA24385_son/NISTv3.3.2/GRCh37/)

You could access these data from our google cloud bucket:

File	Location
Reference Genome	gs://sentieon-test/pipeline_test/reference/
PrecisionFDA Truth Challenge HG002 FASTQs	gs://sentieon-dnascope-model/data/
DNAscope Model	gs://sentieon-dnascope-model/models/SentieonModelBeta0.5.model
Truth VCF and Bed files	gs://sentieon-dnascope-model/truth/

Variant Calling Performance

We use RTG's vcfeval + hap.py (https://github.com/Illumina/hap.py) for variants evaluation, same comparison methodology as used in precisionFDA Truth Challenge.

HAPPY="/opt/hap.py/bin/hap.py"
OUTPREFIX="happy_eval"
$HAPPY truth.vcf query.vcf -f truth.bed -o $OUTPREFIX -r hs37d5.fa --engine=vcfeval --engine-vcfeval-template hs37d5.sdf

You could find DNAscope output as well as the hap.py evaluation results published on Google Storage Buckets gs://sentieon-dnascope-model/ under output/ and happy_eval/ directories.

Type	TP	FN	FP	Recall	Precision	F1_score	PrecisionFDA Truth Challenge Winning Fscore*
SNP	3046378	1459	700	0.999521	0.99977	99.9646%	99.9587%
INDEL	463754	1010	685	0.997827	0.998585	99.8206%	99.4009%

*PrecisionFDA Truth Challege Result taken from https://precision.fda.gov/challenges/truth/results

Runtime

PrecisionFDA HG002 50X sample on 64-vCPU Machine(n1-highcpu-64):

Pipeline	Step	Wall time
fastq -> bam	Sentieon-BWA-MEM + dedup	2h50m
bam -> VCF	DNAscope + ModelApply	1h26m

Tips:You could futher reduce the bam -> VCF stage runtime to ~1h by adding the nodecoy.bed interval file, to avoid computation on complex decoy regions that are usually not needed.

To summarize, based on whether you want to start from raw fastq files or already processed deduped bam files, the estimated runtime for a 50X WGS sample would be 4.5h(from fastq) or 1.5h(from bam).

If you are interested in other Sentieon Products, please visit www.sentieon.com for more information.