For accurate classification of silencers, we propose a CNN-based model named DeepSilencer. As illustrated in the following figure, DeepSilencer consists of four modules. First, a pre-processing module transforms DNA sequences into matrices of sequences and counts of kmers. Second, a CNN module uses a convolutional neural network (CNN) with multiple convolutional and pooling layers to extract features from matrices of DNA sequences. Third, an ANN module is adopted to sufficiently learn the characteristics of kmers. Finally, a joint module integrates outputs of the CNN and ANN modules to predict the probability.

Requiements:  
1. Python 3.5 or later version  
2. Packages:  
    numpy (>=1.15.1)  
    keras (2.3.1)  
    tensorflow(-gpu) (1.15.2)  
    hickle (>=3.4)
  
Package installation:
  
$ pip install -U numpy  
$ pip install keras == 2.3.1 
$ pip install tensorflow-gpu==1.15.2 #pip install tensorflow==1.15.2  
$ pip install -U hickle  
$ git clone https://github.com/xy-chen16/DeepSilencer.git   
$ cd DeepSilencer    

$ cd data 
$ mkdir -p genome/mm10 && cd genome/mm10
$ nohup wget https://hgdownload.cse.ucsc.edu/goldenPath/mm10/bigZips/chromFa.tar.gz
$ tar zvfx chromFa.tar.gz
$ cd ..
$ mkdir hg19 && cd hg19
$ nohup wget https://hgdownload.cse.ucsc.edu/goldenPath/hg19/bigZips/chromFa.tar.gz
$ tar zvfx chromFa.tar.gz 
$ cd ../../..

$ tar -xjvf data/open_region.tar.bz2 -C data
$ tar -xjvf result/result.tar.bz2 -C result

We collected the uncharacterized cis-regulatory elements (CREs) in K562 cells with MPRA provided by Jayavelu et al. Then we chose the top 2000 uncharacterized CREs sequences with the lowest MPRA activity as a positive set, and the bottom 2000 uncharacterized CREs with highest MPRA activity as a negative set. We also downloaded the uncharacterized CREs in homo sapiens and mus musculus without the value of MPRA, and we further use these sequences to find the candidate silencer.

For demonstrating the classification performance of DeepSilencer, we conducted the self-projection experiment and compared our method with the gapped k-mer SVM (gkmSVM). We randomly selected 80% of the data as trainning set and used the remaining 20% of data for testing the two models.

$ python code/run_self_projection.py

The performance of DeepSilencer was shown in the following Figure.

The performance of two methods was shown in the following table.

In order to find the candidate silencer elements in homo sapiens and mus musculus, we trained the DeepSilencer model based on the whole sequences using in the self-projection experiments. First, we trained the model:

$ python code/train_for_crossdata_projection.py

Predict the candidate silencer elements in homo sapiens (hg19):

$ python code/run_crossdata_projection_human.py

Predict the candidate silencer elements in mus musculus (mm10):

$ python code/run_crossdata_projection_mouse.py

Then you can check the results in the results folder.

We got the candidate silencers using the trained DeepSilencer model from the uncharacterized CREs in human and mouse. Then we did some comparative analyses between silencers predicted in the human K562 cell line by the gkmSVM-based model and our DeepSilencer model, and believed that there is almost no difference between them. The length distribution of silencers from the gkmSVM-based model and our DeepSilencer model is shown in Figure A. We found that there is no significant difference in the distance between silencers and the nearest coding genes (two-sided Wilcox test p-value=0.5209, Figure B), the GC content (two-sided Wilcox test p-value=0.9299, Figure C) and the chromatin accessibility (two-sided Wilcox test p-value=0.5699, Figure D).