Alternative Neural Networks for DeepVariant

Homepage | Paper

This repository provides a workflow to

• train alternative models for DeepVariant
• run whole DeepVariant pipeline with implemented custom models

GPU Setup with Docker

Build Docker image and run GPU-enabled container based on the official google/deepvariant:1.6.1-gpu image. Then, open Jupyter environment in your browser using tokenized URL printed in the container logs.

docker build -t deepvariant:v1 .
docker run \
   -v /src/:/opt/deepvariant/src \
   -v /data/:/opt/deepvariant/data \
   --gpus all \
   -p 1234:8888 \
   deepvariant:v1

Alternative Model Training

From a Jupyter terminal or a shell inside the container, launch model training with train.py.

To configure the run, either modify dv_config.py, or override individual parameters inline using --config.* flags. You can choose different model architectures, specify datasets, and adjust batch size, learning rate, logging frequency, and more.

Support for EfficientNet models is now available, we encourage you to explore them or add your own custom model.

List of supported models:

Model	model_type
EfficientNet-B0	'efficientnetb0'
EfficientNet-B1	'efficientnetb1'
EfficientNet-B2	'efficientnetb2'
EfficientNet-B3	'efficientnetb3'
EfficientNet-B4	'efficientnetb4'
EfficientNet-B5	'efficientnetb5'
EfficientNet-B6	'efficientnetb6'
EfficientNet-B7	'efficientnetb7'

python train.py --config=dv_config.py:base \
   --config.model_type='efficientnetb3' \
   --config.batch_size=128 \
   --config.num_epochs=10 \
   --config.learning_rate=1e-3

Results are stored as:

current-run
└──training/
   ├── checkpoints/     # Trained model weights and states
   └── logs.log         # Detailed training metrics and logs

Custom Model Integration and Testing

You can now run and test full DeepVariant pipeline using recently trained model with test.sh. Provide current-run directory; model details will be automatically loaded from training/checkpoints/. You can also specify test dataset or restrict evaluation to specific chromosomes or regions. Variant call results will be saved to a VCF file, and hap.py testing will be launched automatically.

docker pull jmcdani20/hap.py:v0.3.12 # Pull hap.py image if not already available
sh test.sh current-run

Results are stored as:

current-run
└──testing/
   ├── test.vcf.gz              # Variant call results
   └── happy.output.summary     # Final evaluation metrics

Multi-Trial Training and Testing

You can orchestrate multiple training/testing runs using Optuna to:

• Automatically find optimal hyperparameters for training
• Perform cross-validation with multiple runs

Custom optuna_train.py script runs train() multiple times using parameters suggested by Optuna’s sampling strategy (e.g., different learning rates, optimizers, seeds, etc). Each trial is saved to a separate subdirectory under the current-run.

Example command to run hyperparameter tuning:

python optuna_train.py

Results are stored as:

current-run
├── optuna_trial_0/
│   ├── training/
│   └── testing/
├── optuna_trial_1/
└── optuna_trial_2/

---

Feel free to explore and experiment, your feedback and contributions are always welcome!