DeepVariant/README.md

# DeepVariant

![deepvariatn workflow](./picture/Screen Shot 2019-06-11 at 10.09.27 AM.png)

DeepVariant was trained on 8 whole genome replicates of NA12878 sequenced under a variety of conditions related to library preparation. These conditions include loading concentration, library size selection, and laboratory technician. Then the trained model can be generalized to a variety of new datasets and call variants.

Docker uploaded to Alibaba Cloud is downloaded from [dockerhub](<https://hub.docker.com/r/dajunluo/deepvariant>) (version r0.8.0).

WGS training datasets v0.8 include 12 HG001 PCR-free, 2 HG005 PCR-free, 4 HG001 PCR+. Sequencing platforms are all Illumina Hiseq. For Illumina Novaseq, BGISEQ-500 and BGISEQ-2000, we probably need to train customized small variants callers, more information can be found in [Github](<https://github.com/google/deepvariant/blob/r0.8/docs/deepvariant-tpu-training-case-study.md>). 

[CUSTOMIZED MODELS TO BE ADDED]

DeepVariant pipeline consist of 3 steps:

1. `make_examples` consumes reads and the refernece genome to create TensorFlow examples for evaluation with deep learning models. 
2. `call_variants`  (Multiple-threads) consums TFRecord files of tf.Examples protos created by `make_examples` and a deep learning model checkpoint and evaluates the model on each example in the TFRecord. The output here is a TFRecord of CallVariantOutput protos. Multiple-threads
3. `postprocess_variants` (Single-thread) reads all of the output TFRecord files from `call_variants`, it needs to see all of the outputs from `call_variants`for a single sample to merge into a final VCF. 



**Some tips for DeepVariant:**

1. Duplicate marking may be performed, there is almost no difference inaccuracy except lower (<20x) coverages.
2. Authors recommend that you do not perform BQSR. Running BQSR has a small decrease on accuracy.
3. It is not necessary to do any form of indel realignment, though there is not a difference in DeepVariant accuracy either way.


You can run with one common using the `run_deepvariant.py` script

```bash
python run_deepvariant.py --model_type=WGS \
--ref=../../data/"${REFERENCE_FILE}" \
--reads=../../data/"${BAM_FILE}" \
--regions "chr20:10,000,000-10,010,000" \
--output_vcf=../output/output.vcf.gz \
--output_gvcf=../output/output.g.vcf.gz
```

Four files are generated

```bash
output.vcf.gz
output.vcf.gz.tbi
output.g.vcf.gz
output.g.vcf.gz.tbi
```



Command used in choppy app

```bas
python run_deepvariant.py --model_type=WGS \
--ref=${ref_dir}/${fasta} \
--reads=${Dedup.bam} \
--output_vcf=${sample}_DP.vcf.gz
```

**Reference:**

1. DeepVariant Github <https://github.com/google/deepvariant>

2. DeepVariant paper  <https://www.nature.com/articles/nbt.4235>