DeepCpG: Deep neural networks for predicting single-cell DNA methylation

DeepCpG [1] is a deep neural network for predicting the methylation state of CpG dinucleotides in multiple cells. It allows to accurately impute incomplete DNA methylation profiles, to discover predictive sequence motifs, and to quantify the effect of sequence mutations. (Angermueller et al, 2017).

Please help to improve DeepCpG, by reporting bugs, typos in notebooks and documentation, or any ideas on how to make things better. You can submit an issue or send me an email.

DeepCpG model architecture and applications.

(a) Sparse single-cell CpG profiles as obtained from scBS-seq or scRRBS-seq. Methylated CpG sites are denoted by ones, unmethylated CpG sites by zeros, and question marks denote CpG sites with unknown methylation state (missing data). (b) DeepCpG model architecture. The DNA model consists of two convolutional and pooling layers to identify predictive motifs from the local sequence context, and one fully connected layer to model motif interactions. The CpG model scans the CpG neighborhood of multiple cells (rows in b), using a bidirectional gated recurrent network (GRU), yielding compressed features in a vector of constant size. The Joint model learns interactions between higher-level features derived from the DNA- and CpG model to predict methylation states in all cells. (c, d) The trained DeepCpG model can be used for different downstream analyses, including genome-wide imputation of missing CpG sites (c) and the discovery of DNA sequence motifs that are associated with DNA methylation levels or cell-to-cell variability (d).

[1]	Angermueller, Christof, Heather Lee, Wolf Reik, and Oliver Stegle. Accurate Prediction of Single-Cell DNA Methylation States Using Deep Learning. http://biorxiv.org/content/early/2017/02/01/055715 bioRxiv, February 1, 2017, 55715. doi:10.1101/055715.

Table of contents

• News
• Installation
• Getting started
• Examples
• Model Zoo
• FAQ
• Content
• Changelog
• Contact

News

• 170406: A short description of all DeepCpG scripts!
• 170404: New guide on creating and analyzing DeepCpG data released!
• 170404: Training on continuous data, e.g. from bulk experiments, now supported!
• 170305: New documentation of DeepCpG model architectures released!
• 170302: New guide on DeepCpG model training released!
• 170228: New example shell scripts for building a DeepCpG pipeline released!

Installation

The easiest way to install DeepCpG is to use PyPI:

pip install deepcpg

Alternatively, you can checkout the repository,

git clone https://github.com/cangermueller/deepcpg.git

and then install DeepCpG using setup.py:

python setup.py install

Getting started

1. Store known CpG methylation states of each cell into a tab-delimted file with the following columns:

• Chromosome (without chr)
• Position of the CpG site on the chromosome starting with one
• Binary methylation state of the CpG sites (0=unmethylation, 1=methylated)

Example:

1   3000827   1.0
1   3001007   0.0
1   3001018   1.0
...
Y   90829839  1.0
Y   90829899  1.0
Y   90829918  0.0

2. Run dcpg_data.py to create the input data for DeepCpG:

dcpg_data.py
--cpg_profiles ./cpg/cell1.tsv ./cpg/cell2.tsv ./cpg/cell3.tsv
--dna_files ./dna/mm10
--cpg_wlen 50
--dna_wlen 1001
--out_dir ./data

./cpg/cell[123].tsv store the methylation data from step 1., ./dna contains the DNA database, e.g. mm10 for mouse or hg38 for human, and output data files will be stored in ./data.

3. Fine-tune a pre-trained model or train your own model from scratch with dcpg_train.py:

dcpg_train.py
  ./data/c{1,3,6,7,9}_*.h5
  --val_data ./data/c{13,14,15,16,17,18,19}_*.h5
  --dna_model CnnL2h128
  --cpg_model RnnL1
  --joint_model JointL2h512
  --nb_epoch 30
  --out_dir ./model

This command uses chromosomes 1-3 for training and 10-13 for validation. ---dna_model, --cpg_model, and --joint_model specify the architecture of the CpG, DNA, and Joint model, respectively (see manuscript for details). Training will stop after at most 30 epochs and model files will be stored in ./model.

4. Use dcpg_eval.py to impute methylation profiles and evaluate model performances.

dcpg_eval.py
  ./data/*.h5
  --model_files ./model/model.json ./model/model_weights_val.h5
  --out_data ./eval/data.h5
  --out_report ./eval/report.tsv

This command predicts missing methylation states on all chromosomes and evaluates prediction performances using known methylation states. Predicted states will be stored in ./eval/data.h5 and performance metrics in ./eval/report.tsv.

5. Export imputed methylation profiles to HDF5 or bedGraph files:

dcpg_eval_export.py
  ./eval/data.h5
  -o ./eval/hdf
  -f hdf

Examples

You can find example notebooks and scripts on how to use DeepCpG here.

Documentation

The DeepCpG documentation provides information on training, hyper-parameter selection, and model architectures.

Model Zoo

You can download pre-trained models from the DeepCpG model zoo.

FAQ

Why am I getting warnings 'No CpG site at position X!' when using `dcpg_data.py`? This means that some sites in --cpg_profile files are not CpG sites, i.e. there is no CG dinucleotide at the given position in the DNA sequence. Make sure that --dna_files point to the correct genome and CpG sites are correctly aligned. Since DeepCpG currently does not support allele-specific methylation, data from different alleles must be merged (recommended) or only one allele be used.

How can I train models on one or more GPUs? DeepCpG use the Keras deep learning library, which supports Theano or Tensorflow as backend. If you are using Tensorflow, DeepCpG will automatically run on all available GPUs. If you are using Theano, you have to set the flag device=GPU in the THEANO_FLAGS environment variable.

THEANO_FLAGS='device=gpu,floatX=float32'

You can find more information about Keras backends here, and about parallelization here.

Content

• /deepcpg/: Source code
• /docs: Documentation
• /examples/: Examples on how to use DeepCpG
• /script/: Executable DeepCpG scripts
• /tests: Test files

Changelog

1.0.3

Extends dcpg_data.py, updates documentation, and fixes minor bugs.

• Extends dcpg_data.py to support bedGraph and TSV input files.
• Enables training on continuous methylation states.
• Adds documentation about creating and analyzing Data.
• Updates documentation of scripts and library.

Contact

• Christof Angermueller
• [email protected]
• https://cangermueller.com
• @cangermueller