Model Overview

The ultimate use of this model is to denoise ATAC-seq data and create a higher-quality signal for a much cheaper cost compared to collecting high quality data.

Model Architecture

The model is a residual neural network consisting of 5 residual blocks that produce the denoised ATAC-seq signal, followed by 2 additional residual blocks and a sigmoid layer to produce classification (peak calling) output. All residual blocks are composed of 1-dimensional convolutional layers and ReLU activations.

Training

This model was trained on bulk atac-seq data from 4 blood cell types (B cells, NK cells, CD4+ and CD8+ T cells). Noisy data depth is 10 million. Clean data depth is 50 million. Data was obtained from Corces, M. R. et al. Lineage-specific and single-cell chromatin accessibility charts human hematopoiesis and leukemia evolution. Nature Genetics vol. 48 1193–1203 (2016).

Dataset

Noisy Data: Click on the File Browser tab above. Click on train_data -> noisy_data.

Clean Data: Click on the File Browser tab above. Click on train_data -> clean_data.

Performance

No performance information available at this time.

How to Use this Model

A step-by-step tutorial for denoising ATAC-seq data using a pre-trained model can be found at AtacWorks repository: https://github.com/clara-parabricks/AtacWorks/blob/master/tutorials/tutorial2.md Make sure to edit Steps 2,3 and 4 to change the download links of the files to this model.

Step-by-step tutorial for training a model can be found at AtacWorks repository: https://github.com/clara-parabricks/AtacWorks/blob/master/tutorials/tutorial1.md Make sure to edit Step 2 to change the download links of the files to this model.

Pretrained Models: The best performing model can be found under the File Broswer tab, models directory.

Configs: AtacWorks has the ability to take in config files that specify various command line options. This makes it easy to share the exact parameters the model was trained with. You can find the configs under the File Browser tab, configs folder.

Input

The input data for this model consists of noisy atac-seq coverage track. The labels to train the model are clean ATAC-seq coverage tracks and peak calls based on the clean coverage tracks.

Output

This model learned to denoise a 10 Million read depth atac-seq data to resemble a 50 Million read depth signal, and to call peaks from the noisy ATAC-seq signal. The output from atacworks is two bedGraph, optionally BigWig files. One containing denoised tracks and other containing denoised peaks.

Output Files: For the convenience of users, we have uploaded the output files generated by AtacWorks. They can be found under File Browser tab, output directory.

Limitations

This model was trained on data processed using the method in https://github.com/zchiang/atacworks_analysis. It may not deliver suitable results on data processed with different methods.

Reference

Lal, A., Chiang, Z.D., Yakovenko, N. et al. Deep learning-based enhancement of epigenomics data with AtacWorks. Nat Commun 12, 1507 (2021). https://doi.org/10.1038/s41467-021-21765-5

License

The model architecture, training and inference pipelines for this model were developed by the NVIDIA Genomics team. The relevant source code is open sourced under a custom NVIDIA license available here.