Deep Learning For Variant Calling (DL4VC)

DL4VC is an advanced deep learning based variant caller for short-read based germline variant calling. It proposes a deep averaging network (DAN) designed specifically for variant calling. The model takes as input a tensor encoding aligned reads in a proposed variant region, a variant proposal, and outputs a softmax over three cateogies: {no variant, heterozygous variant, homozygous variant}. This model takes into account the independence of each short input read sequence by transforming individual reads through a series of 1D convolutional layers, limiting the communication between individual reads to averaging and concatenating operations, before passing them into a fully connected network.

Our purpose-built model achieves state of the art results on the precisionFDA germline variant calling dataset (compared post competition).

To facilitate future work, we release our code, trained models and pre-processed public domain datasets through this repo.

Accuracy Highlights

PrecisionFDA Truth Challenge results vs DL4VC

Variant Caller	Type	F1	Recall	Precision
rpoplin-dv42	Overall Indels SNPS	0.998597 0.989802 0.999587	0.998275 0.987883 0.999447	0.998919 0.991728 0.999728
dgrover-gatk	Overall Indels SNPS	0.998905 0.994008 0.999456	0.999005 0.993455 0.999631	0.998804 0.994561 0.999282
astatham-gatk	Overall Indels SNPS	0.995679 0.993422 0.995934	0.992122 0.992401 0.992091	0.999261 0.994446 0.999807
bgallagher-sentieon	Overall Indels SNPS	0.998626 0.992676 0.999296	0.998910 0.992140 0.999673	0.998342 0.993213 0.998919
DL4VC	Overall Indels SNPS	0.998924 0.992949 0.999596	0.999076 0.994708 0.999566	0.998772 0.991196 0.999625

Feature Highlights

PyTorch-based model training and inference
1D convolutional model with learned embeddings of bases
Variant proposal encodings
Down-sampling of easy examples to speed up training by 5x

Section	Description
Installation	System and code setup instructions
Data	Pre-processed datasets from precisionFDA to reproduce DL4VC results
Step by step guidelines	Instructions to train and run inference with DL4VC

Installation

The installation has been tested on bare metal as well as conda virtual environments. We recommend conda environments because they simplify the installation of non-python dependencies.

Core dependencies -

BCF Tools
Tabix
Python 3.5+ environment
vcfeval (optional, only needed for comparing with other VCFs)

Setup

git clone https://github.com/clara-genomics/DL4VC.git
cd DL4VC
pip install -r requirements.txt

Data

Please follow the dataset instructions in the Dataset Readme to download pre-processed data and model checkpoints from our experiment. The results mentioned in the Accuracy Highlights section can be reproduced using the same datasets.

Step By Step Guideline