---

CNN-peaks is a Convolution Neural Network(CNN) based ChIP-Seq peak calling software.

Install with Docker

We highly recommend that use Docker to install CNN-Peaks. CNN-Peaks easily can be installed by building docker-image using our Dockerfile. However, you might use Nvidia-docker to run GPU-accelerated CNN-Peaks.

### Docker run example ( In the cloned repository )
nvidia-docker build . -t cnnpeaks:test
nvidia-docker run -i -t -v <data directory>:/CNNpeaks/CNNPeaks/data cnnpeaks:test

After building Docker image for CNN-Peaks, you can use CNN-Peaks as described in Here.

Install

0. Overview

You can prepare to use CNN-Peaks via running a script "install.sh" in the CNN-Peaks directory.

chmod 775 install.sh
./install.sh

If you want to GPU-accelerated CNN-Peaks, please check "GPU-accelerated CNN-Peaks"

Details of the install are below ( you do not have to follow these descriptions since install.sh work well ):

1. Python packages

CNN-Peaks needs Python higher than 3.x. CNN-Peaks requires several Python packages including Tensorflow.

Required Python packages:

• numpy
• scipy
• sklearn
• tensorflow==1.X.X
• pandas
• progressbar2
• pysam

Install with pip

pip install numpy scipy sklearn tensorflow pandas progressbar2 pysam tensorflow-gpu

2. Samtools (htslib)

If you do not have samtools which is higher than 1.7 version you should install it.

Install samtools 1.8 :

cd dependencies
./install_samtools.sh

3. Extern modules of CNN-Peaks

After installing required Python packages and samtools ( htslib ) with above scripts, CNN-Peaks have to compile C, Cython extern modules using the below script:

./build.sh

4. GPU-accelerated CNN-Peaks

Note that if you want to GPU accelerated CNN-Peaks, your tensorflow should be configured to use GPU. Please check here for a description to configure GPU support for CNN-Peaks.

Install with Docker

You can build Docker image if you want to run CNN-peaks on your Docker container. However, you might use Nvidia-docker as long as you use our Dockerfile.

---

Quick start

Install for CNN-Peaks:

Examples of CNN-peaks command:

preprocessing:

python CNNpeaks -m preprocess -i testdir/

CNN model building:

python CNNpeaks -m buildModel -i testdir/

peak calling with trained model:

python CNNpeaks -m peakCall -i myOwnChIPSeq.bam

Note that you can use pre-trained models for CNN-Peaks calling. Please check "Using pre-trained models" for details.

peak calling with trained model:

python CNNpeaks -m peakCall -i myOwnChIPSeq.bam

Labeled data

CNN-peaks uses labeled data which has its format. All these approaches that use labeled data for marking are from [1]. Examples of labeled data are as the below. (It based on ASCII)

chr1:1,000,000-1,100,000 peaks K562

chr1:1,100,000-1,200,000 peakStart K562

chr1:1,250,000-1,300,000 peakEnd K562

chr2:10,000,000-10,002,000 peaks

In line 1, peaks, this means that cell K562 at least has one peak in a region (chr1:1,000,000-1,100,000). In line 2, 3,peakStart, peakEnd, it means that cell K562 just only one peak in the regions. In line 4, there is no peak in those regions about K562 or other cells because there is no matched cell line name at this raw. If you want to use this label data on other cells, all these line 1-4 mean noPeak because there is no cell name in the lines. If you want to know specific rules or methods of this labeling work, please look here.

Usage of CNN-peaks

1. Training data to build a CNN-peaks model.

Before you try to call peaks with your ChIP-Seq Data, CNN-peaks model should be trained by the internal module in CNN-peaks. CNN-peaks gives preprocessing module for generating actual training data samples with labeled data and bam alignment files of them. You can use our labeled data sample or use your own labeled data to generate training data. When you use its preprocessing module, the input of the module is a single directory which includes labeled data and bam files. For example, a directory which has name "./TestPreProcess" includes these files:

preprocessing:

python CNNpeaks -m preprocess -i TestPreProcess

./TestPreProcess:

• H3K36me3_None.txt
• H3K36me3_K562.bam
• H3K36me3_A549.bam

Each file must follow a format of the filename that is Target_CellName.bam. The case of labeled data, it must be like this: Target_AnyString.txt. The labeled data of this example H3K36me3_None.txt has data which includes both K562 and A549. As a result, directories that include labeled data, reference gene data and read depth data are generated.

./TestPreProcess:

• H3K36me3_None.txt
• H3K36me3_K562.bam
• H3K36me3_A549.bam
• H3K36me3_K562 (directory)
• H3K36me3_A549 (directory)

2. Build CNN-peaks model with preprocessed data

After you created your training data with our preprocess module, you can build CNN-peaks model by using our buildModel module. Results of running the module include visualization of peak predictions about test data, train and test sensitivity and specificity during the training process and trained models in a "models" directory. You can check those visualization results and saved tensorflow variables at the "models" directory in a path of CNNpeaks.

CNN model building:

python CNNpeaks -m buildModel -i testdir/

3. Peak calling with the trained model

If you finish building your CNNpeaks model, buildModel module generated 'k' numbers of saved tensorflow variables. All you have to do is just pick your saved model number and bam alignment file as an input of peak calling.

peak calling with trained model

python CNNpeaks -m peakCall -i myOwnChIPSeq.bam

Output format of CNN-Peaks

The output format of CNN-Peaks follows regular .bed format which is tab delimited. 1 to 5`th columns is same with a regular bed file format that can be displayed in an annotation track for visualization tools such as UCSC genome browser and IGV.

Field of CNN-Peaks output format is:

1. Chromosome – The name of chromosome
2. Chromosome Start – The starting position of features in the chromosome.
3. Chromosome End – The ending position of the feature in the chromosome.
4. Name – The randomly generated string of the peak.
5. Score – The score of peak signals determined by column 6 and 9.
6. Score2 – The score of peak signals determined by column 6 and 8.
7. Sigmoid activation value – The sigmoid activation value of the peak from CNN-Peaks model output.
8. P-value (Avg) – Average p-value in the peak region on Poisson distribution with a window.
9. P-value (Min) – Minimum p-value in the peak region on Poisson distribution with a window.

Filtering the peak calling result

After calling peaks by CNN-Peaks model, you can filter out some peaks with scoring metrics in CNN-Peaks result ( column 5 to 9 ) using a script in CNN-Peaks.

[filteringPeaks.sh -o<options> -t <threshold> -i <input> > <output>]

Options	Type	Description
[-o]	interval	Filtering peaks by those interval (3'rd col - 2'ed col)
	score	Filtering peaks by the score of peak regions ( 4'th col )
	scoreBroad	Filtering peaks by the score of peak regions ( 5'th col )
	sigVal	Filtering peaks by the sigmoid activation value of peak regions ( 6'th col )
	pValueBroad	Filtering peaks by the P-value of peak regions ( 7'th col )
	pValue	Filtering peaks by the P-value of peak regions ( 8'th col )
[-t]	Threshold Value	Threshold value for -o
[-i]	File Name	Input file name

Using pre-trained models

The pre-trained CNN-Peaks model is available. You do not have to train your model if you use this well-tuned model. To use the model, you just have to move "models" directory into your CNN-Peaks working directory which is outer most directory of CNN-Peaks.

Available data link

1. Alpha

The Alpha model had been trained with 2997 genomic segments with narrow histone modifications such as H3K4me3 and small amounts of transcription factors such as MAX. Note that it has 50 threshold outputs.

2. Beta

The Beta model had been trained with 3294 genomic segments with narrow histone modifications such as H3K4me3 and small amounts of transcription factors such as MAX. In addition, to give variaty for the model, it contains 297 ATAC-Seq genomic segments. Note that it has 10 threshold outputs.

---

CITATION

1. HOCKING, Toby Dylan, et al. Optimizing ChIP-seq peak detectors using visual labels and supervised machine learning. Bioinformatics, 2016, 33.4: 491-499.
2. Oh, D., et al. CNN-Peaks: Chip-Seq peak detection pipeline using convolutional neural networks that imitate human visual inspection. Scientific reports, 2020, 10(1), 1-12.