Similar to the case study on whole genome sequencing data, in this study we describe applying DeepVariant to a real exome sample using a single machine.
NOTE: This case study demonstrates an example of how to run DeepVariant end-to-end on one machine. This might not be the fastest or cheapest configuration for your needs. For more scalable execution of DeepVariant see the Docker-based exome pipeline created for Google Cloud Platform.
Starting from the 0.7 release, we use docker to run the binaries instead of copying binaries to local machines first. You can still read about the previous approach in the Exome Case Study in r0.6.
We recognize that there might be some overhead of using docker run. But using docker makes this case study easier to generalize to different versions of Linux systems. For example, we have verified that you can use docker to run DeepVariant on other Linux systems such as CentOS 7.
Script: https://github.com/google/deepvariant/blob/r0.7/scripts/run_wes_case_study_docker.sh
Get the script and run everything. This will install and download everything needed for this case study. And it will run DeepVariant to generate the output VCFs, and also run the evaluation as well.
Before you run the script, you can read through all sections to understand the details.
wget https://raw.githubusercontent.com/google/deepvariant/r0.7/scripts/run_wes_case_study_docker.sh -P ${HOME}
chmod +x ${HOME}/run_wes_case_study_docker.sh
${HOME}/run_wes_case_study_docker.shAny sufficiently capable machine will do. For this case study, we used a 64-core non-preemptible instance with 128GiB and no GPU.
If you need an example, see this section.
151002_7001448_0359_AC7F6GANXX_Sample_HG002-EEogPU_v02-KIT-Av5_AGATGTAC_L008.posiSrt.markDup.bam
Same as described in the case study for whole genome data
HG002_GRCh37_GIAB_highconf_CG-IllFB-IllGATKHC-Ion-10X-SOLID_CHROM1-22_v.3.3.2_highconf_*
are from NIST, as part of the Genomes in a Bottle
project. They are downloaded from
ftp://ftp-trace.ncbi.nlm.nih.gov/giab/ftp/release/AshkenazimTrio/HG002_NA24385_son/NISTv3.3.2/GRCh37/
According to the paper "Extensive sequencing of seven human genomes to
characterize benchmark reference
materials", the HG002 exome was
generated with Agilent SureSelect. In this case study we'll use the SureSelect
v5 BED (agilent_sureselect_human_all_exon_v5_b37_targets.bed) and intersect it
with the GIAB confident regions for evaluation.
More discussion can be found in the call_variants section in the case study.
More discussion can be found in the postprocess_variants section in the case study.
| Step | wall time |
|---|---|
make_examples |
13m 59s |
call_variants |
2m 1s |
postprocess_variants (no gVCF) |
0m 15s |
postprocess_variants (with gVCF) |
1m 23s |
| total time (single machine) | 16m 15s - 17m 23s |
We used the hap.py
(https://github.com/Illumina/hap.py)
program from Illumina to evaluate the resulting vcf file. This serves as a check
to ensure the three DeepVariant commands ran correctly and produced high-quality
results.
We evaluate against the capture region:
| Type | # FN | # FP | Recall | Precision | F1_Score |
|---|---|---|---|---|---|
| INDEL | 106 | 49 | 0.959231 | 0.980897 | 0.969943 |
| SNP | 48 | 16 | 0.998577 | 0.999525 | 0.999051 |
Starting from DeepVariant 0.5.* and later releases, we recommend a separate model for calling exome sequencing data. Here is how the exome model is trained: we used a WGS model as the starting checkpoint (instead of an ImageNet one), and trained only on examples created from exome data.