DESSO (DEep Sequence and Shape mOtif)

DESSO is a deep learning-based framework that can be used to accurately identify both sequence and shape regulatory motifs from the human genome. The performance of DESSO was evaluated on the 690 ChIP-seq datasets.

Prerequisites and Dependencies

• Tensorflow 1.1.0 [Install]
• CUDA 8.0.44
• Python 2.7
• Biopython 1.7.0
• Scikit-learn
• Download GRCh37.p13.genome.fa and encode_101_background, then unzip them and put them into data/
• data/encode_101, data/encode_1001, and data/TfbsUniform_hg19_ENCODE only contain wgEncodeEH002288-related data as an example, owing to the file size limit. To access the source code and whole datasets (totally about 5.9GB) without additional manipulation, just click on code+whole data.

Model Training Based on Convolutional Neural Network (CNN)

Train CNN models on specified datasets:

cd code/
python train.py --start_index 0 --end_index 1 --peak_flank 50 --network CNN --feature_format Seq

Arguments	Description
--start_index	Start index of the 690 ENCODE ChIP-seq datasets
--end_index	END index of the 690 ENCODE ChIP-seq datasets
--peak_flank	Number of flanking base pairs at each side of peak summit (default is 50)
--network	Neural network used in model training (default is CNN)
--feature_format	Feature format of the input (default is Seq)

--start_index 0 --end_index 1 indicates the first dataset (i.e., wgEncodeEH002288). For example, to train models for the second and third datasets, use --start_index 1 --end_index 3
--peak_flank 50 indicates the peak length is (2 * 50 + 1) = 101 base pairs
--network indicates that CNN is used here
--feature_format can be Seq or DNAShape, where Seq indicates the input is DNA sequences, DNAShape indicates the input is the combination of four DNA shape features (i.e., HelT, MGW, ProT, and Roll).

Output

If --feature_format Seq was used, the trained model can be found at /output/encode_101/gc_match/wgEncodeEH002288/Seq/CNN, together with Test_result.txt indicating the area under the receiver operating characteristic curve (AUC) of the trained model in predicting TF-DNA binding specificity on the test data.
If --feature_format DNAShape was used, the trained model is located at /output/encode_101/gc_match/wgEncodeEH002288/DNAShape/CNN.

Motif Prediction

Obtain either sequence or shape motifs based on the trained models above:

cd code/
python predict.py --start_index 0 --end_index 1 --peak_flank 50 --network CNN --feature_format Seq --start_cutoff 0.01 --end_cutoff 1 --step_cutoff 0.03

Arguments	Description
--start_cutoff	Start of the motif cutoff interval (default is 0.01)
--end_cutoff	End of the motif cutoff interval (default is 1)
--step_cutoff	Increament of the cutoff (default is 0.03)

--feature_format Seq indicates that sequence motifs will be predicted. To identify shape motifs, use --feature_format DNAShape instead.

Output

For --feature_format Seq, the predicted sequence motifs are in output/encode_101/gc_match/wgEncodeEH002288/Seq/CNN/0.
For --feature_format DNAShape, four kinds of shape motifs would be predicted as shown in the following table:

Location	Type of predicted shape motif
output/encode_101/gc_match/wgEncodeEH002288/DNAShape/CNN/0	HelT motif
output/encode_101/gc_match/wgEncodeEH002288/DNAShape/CNN/1	MGW motif
output/encode_101/gc_match/wgEncodeEH002288/DNAShape/CNN/2	ProT motif
output/encode_101/gc_match/wgEncodeEH002288/DNAShape/CNN/3	Roll motif

Predict TF-DNA Binding Specicitity Using Gated-CNN (GCNN) and Long DNA Sequence

cd code/
python train.py --start_index 0 --end_index 1 --peak_flank 500 --network GCNN --feature_format Seq

--network GCNN indicates that GCNN is used for model training
--peak_flank 500 indicates that the peak length is (2 * 500 + 1) = 1001 base pairs

Output

The trained model and its AUC (Test_result.txt) on test data is located at output/encode_1001/gc_match/wgEncodeEH002288/Seq/GCNN.

Citation

If you use DESSO in your research, please cite the following paper:
Jinyu Yang, Adam D. Hoppe, Bingqiang Liu, Qin Ma, Systematic Prediction of Regulatory Motifs from Human ChIP-Sequencing Data Based on a Deep Learning Framework, bioRxiv 417378; doi: https://doi.org/10.1101/417378