rabbitvar filter
A deep-learning-based variant filter for VarDict
There are three main steps in DeepFilter:
- DeepFilter uses a hard filter strategy to the intermediate results produced by VarDict. Variants that match these conditions will be filtered out.
- Then, the filtered data is the input to the network for inference.
- Finally, DeepFilter formats the filtered data into VCF file.
The IN_DATA below reference to the intermediate result of VarDict,
you can run VarDict like:
VarDict \
-G /path/to/hg19.fa \
-f $AF_THR -N sample_name \
-b "/path/to/tumor.bam|/path/to/normal.bam" \
-c 1 -S 2 -E 3 -g 4 \
/path/to/my.bed | VarDict/testsomatic.R > ${IN_DATA}- Filter INDEL variant and then format to VCF file
VARTYPE="INDEL"
python call_somatic.py \
--workspace /home/haoz/deepfilter/workspace \
--in_data ${IN_DATA} \
--nthread ${THREAD} \
--var_type ${VARTYPE} \
--trained_model ./models/checkpoint_indel_w1_24.adam.pth \
--out ${DEEPFILTER}/workspace/result/filtered_som_indel.vcf- Filter SNV variant and then format to VCF file
VARTYPE="INDEL"
python call_somatic.py \
--workspace /home/haoz/deepfilter/workspace \
--in_data ${IN_DATA} \
--nthread ${THREAD} \
--var_type ${VARTYPE} \
--trained_model ./models/checkpoint_snv_w1_24.adam.pth \
--out ${DEEPFILTER}/workspace/result/filtered_som_snv.vcf- step1: make .csv data
python make_data.py inter.txt groundtruth.vcf $TYPE train_data.tsv-
step2: modified the source code if you want to change the structure of the network or other strategies.
-
step3: re-train the model
python train_somatic.py \
--workspace . \
--train_data ${data_path}/data_indel_all.tsv \
--nthread 8 \
--var_type "INDEL" \
--weight ${weight} \
--out_model_path ./models/checkpoint_indel_w${weight}.adam.pthsage: call_somatic.py [-h] --workspace WORKSPACE
--in_data IN_DATA --truth_file TRUTH_FILE
[--model_out MODEL_OUT] --var_type VAR_TYPE
[--batch_size BATCH_SIZE] [--nthreads NTHREADS]
[--trained_model TRAINED_MODEL] --out OUT