GPU-accelerated DeepSomatic

DeepSomatic builds on the deep learning-based variant caller DeepVariant. It processes aligned reads from tumor and normal samples (in BAM or CRAM format), generates pileup image tensors, classifies these tensors using a convolutional neural network, and outputs somatic variants in standard VCF or gVCF files.

DeepSomatic is designed for somatic variant calling using tumor-normal sequencing data.

Parabricks has enhanced Google DeepSomatic to leverage GPUs extensively. The Parabricks version of DeepSomatic operates similarly to other common command line tools: it accepts two BAM files and a reference file as inputs and generates variants in a VCF file as output.

See the deepsomatic Reference section for a detailed listing of all available options.

The commands below are the Google counterpart of the Parabricks command above. The output from these commands will be identical to the output from the above command. See the Output Comparison page for comparing the results.

            

            
docker run \
--interactve \
--tty \
--rm \
--volume ${INPUT_DIR}:${INPUT_DIR} \
--volume ${OUTPUT_DIR}:${OUTPUT_DIR} \
--workdir /workdir google/deepsomatic:1.9.0 \
run_deepsomatic \
--ref ${REFERENCE_FILE} \
--reads_tumor ${TUMOR_BAM} \
--reads_normal ${NORMAL_BAM}  \
--model_type=WGS \
--output_vcf ${OUTPUT_VCF} \
--make_examples_extra_args "ws_use_window_selector_model=true" \
--num_shards=$(nproc)
        

Parabricks DeepSomatic supports the following models:

While Parabricks DeepSomatic does not lose any accuracy in functionality when compared with Google DeepSomatic there is one reason that can result in different output files.

• CNN Inference

Google DeepSomatic uses a CNN (convolutional neural network) to predict the possibilities of each variant candidate. The model is trained, and does inference through, Keras. In Parabricks DeepSomatic we convert this Keras model to an engine file with TensorRT to perform accelerated deep learning inferencing on NVIDIA GPUs. Because of the optimizations from TensorRT there is a small difference in the final possibility scores after inferencing (10^-5), which could cause a few different variants in the final VCF output. Based on current observations the mismatches only happen to RefCalls with a quality score of zero.

Run DeepSomatic to convert BAM/CRAM to VCF.