LoopAnchor

CCCTC-binding factor (CTCF) is a transcription regulator which is involved in many cellular processes. How CTCF recognizes DNA sequence to exert chromosome barrier or enhancer blocking effects remains to be fully interrogated. Despite many computational tools were developed to predict CTCF-mediated loops qualitatively or quantitatively, few could specially evaluate the regulatory potential of DNA sequence at CTCF binding sites (CBSs) and how it affects chromatin loop formation. Here, we developed a deep learning model, DeepAnchor, to precisely characterize the binding patterns for different types of CBSs. By incorporating base-wise genomic/epigenomic features, we revealed distinct chromatin and sequence features for CTCF-mediated insulation and looping at a high resolution, such as two sequence motifs flanking the core CTCF motif at loop-associated CBSs. Besides, we leveraged the predicted anchor score to optimize the loop extrusion model and achieved the best performance in predicting CTCF-anchored loops. We established a compendium of context-specific CTCF-anchored loops across 52 human tissue/cell types and found that genomic disruption of CTCF-anchored loops may represent a general causal mechanism of disease pathogenesis. These computational models, together with the established resource, could facilitate the mechanistic research on how the CTCF-mediated cis-regulatory elements (CREs) shapes context-specific gene regulation in cell development and disease progression.

Prepare data

To implement DeepAnchor, one should prepare four types of data:

1. Obtain CBSs by motif scan.
2. Download base-wise genomic/epigenomic features from CADD v1.3. Here we also provide preprocessed CADD features for CTCF binding sites.
3. CTCF ChIP-seq data for specific cell type (for example GM12878).
4. Cohesin ChIA-PET data for the same cell type.

Generate training data

1. Extract feature values

We totally use 44 genomic/epigenomic features from CADD database and 4 DNA sequence feature (A/T/G/C). For each CBS and each feature, ±500bp feature values are extracted. Therefore, the size of feature matrix for each CBS is 48 x 1000. Feature matrix is strandardized to facilitate downstream analyses. However, this step is very time consuming. So we provide the preprocessed feature matrix data which can be downloaded from here: cadd feature matrix and dna feature matrix.

2. Generate positive/negative CBSs

We need create a work_dir with structure as below and put five data files within it.

work_dir
    └── raw                   
        ├── CTCF_peak.bed.gz              
        ├── target.bed              
        ├── cbs.tsv
        ├── dna_feature.npz      
        └── cadd_feature.npz

More explainations of data requirement:

CTCF_peak.bed.gz: (1) the ChIP-seq narrowPeak file for CTCF. (2) Columns: chrom,chromStart,chromEnd,name,score,strand,signalValue,pValue,qValue,summit.

target.bed: (1) Targeted chromosome intervals for CTCF binding sites. (2) Columns: chrom,start,end.

cbs.tsv: (1) position of all CTCF binding sites. (2) non-cell-type specific. (3) can be found in data folder.

dna_feature.npz: (1) one-hot representation of DNA sequence for CTCF binding sites. (2) non-cell-type specific. (3) downloaded as addressed in last section.

cadd_feature.npz: (1) cadd features of DNA sequence for CTCF binding sites (2) non-cell-type specific (3) downloaded as addressed in last section.

To generate P/N dataset, you can simply run following command:

python DeepAnchor_input.py work_dir

Tips:

If you have loop bedpe file, you can use ./pp/loop_to_bed.py to convert the loops to target.bed file.

python ./pp/loop_to_bed.py file_loop target_bed

DeepAnchr_input.py will generate an output folder named DeepAnchor in work_dir:

work_dir
    └── DeepAnchor  
        ├── CTCF_peaks.bed           # intersect CTCF_peak.bed.gz and cbs.tsv
        ├── marked_cbs.tsv           # marked CBSs with CTCF ChIP-seq peaks          
        ├── train.npz                # train set (chr1-16)
        ├── valid.npz                # valid set (chr17-18)
        ├── test.npz                 # test set (chr19-X)                 
        └── total.npz                # feature data for all CBSs

3. Run DeepAnchor

DeepAnchor trains a classifier to distinguish insulator-related CBSs from others. After training, the model will be used to predict DeepAnchor score to show the possibility that each CBS belong to insulator-related CBSs.

To train the classifier model, run command:

python DeepAnchor.py  work_dir train

This will generate a DeepAnchor model which will be saved in DeepAnchor.model in work_dir.

To predict the DeepAnchor score for all CBSs, run command:

python DeepAnchor.py  work_dir predict

This will generate a file scored_cbs.tsv that contain all CBSs and their DeepAnchor score. We need to copy this file to ./data/ folder for downstream analyses.

The data columns of scored_cbs.tsv are shown below:

chrom	start	end	strand	score	anchor_score

score: the score for motif scan. anchor_score: the score predicted by DeepAnchor model.

4. Run LoopAnchor to make loop prediction

To use LoopAnchor for loop prediction, you should prepare input data arranged as follow:

work_dir
    └── raw                   
        └── CTCF_peak.bed.gz
        └── scored_cbs.tsv      

Run command:

python run_LoopAnchor_denovo.py  work_dir

In work_dir/LoopAnchor folder, you can find the result LoopAnchor_pred.bedpe which contains all the loops predicted by LoopAnchor.LoopAnchor files is arranged in bedpe format and the last column is the predicted loop intensity.

chrom1	start1	end1	chrom2	start2	end2	name	score	strand1	strand2	LoopAnchor

Here is a complete example. The data can be found in ./data/ folder, but you still need to download some files as shown before.

python DeepAnchor_input.py ./data/GM12878
python DeepAnchor.py ./data/GM12878 train
python DeepAnchor.py ./data/GM12878 predict
cp ./data/GM12878/scored_cbs.tsv ./data/K562/raw
python run_LoopAnchor_denovo.py ./data/K562

Other code resources

Considering that github may be not available in some regions, we also provide ReadTheDocs and Bitbuket resource for LoopAnchor method.

Landscape availability

We collected 764 available CTCF ChIP-seq data from ENCODE, CistromDB and ChIP-Atlas and use LoopAnchor to predict CTCF-anchored loops. The results are available at UCSC Track Data Hubs (https://genome.ucsc.edu/cgi-bin/hgHubConnect) by entering customized hub URLs https://raw.githubusercontent.com/mulinlab/LoopAnchor/master/loopanchor/data/hubs/hubs_landscape.txt or https://raw.githubusercontent.com/mulinlab/LoopAnchor/master/loopanchor/data/hubs/hubs_all.txt, respectively.

Citation

Xu, Hang, et al. "Inferring CTCF insulators and anchored loops across human tissues and cell types." bioRxiv (2022).