Predicting Transcription Factor Binding Sites using Deep Learning

Here we use ENCODE data viz. ChIP-seq peaks (narrowPeak files), genome segmentation data (BED files), Motif information (GFF files) and DNASE information (narrowPeak files) to predict the probability of a Transcription Factor (TF) binding in a 200 bp window (chrStart-chrEnd).

Files:
- data_pipeline.sh
- find_bed_range.py
- create_bp_windows.py
- mod_chrom.py
- mod_motif.py
- mod_dnase_score.py
- mod_peak.py
- dat2csv.py
- uniq.py
- rm_ambiguous.py
- prc.py
- auprc.r

1. data_pipeline.sh
    Creates the dataset to be given to the model as input:
    usage: ./data_pipeline.sh
    For command line arguments, refer to each .py file's documentation.

2. find_bed_range.py
    Outputs the ranges (low1, low2 and high2) for the 200 bp windows to be created.
    usage: python find_bed_range.py bed_file.bed
    Round low1 to the nearest multiple of 50.
    Put these (low1, low2 and high2) in the data_pipeline.sh file.
    Windows created:
        low1    low1+200
        ...
        ...
        low2    high2

3. create_bp_windows.py
    Outputs a tab-separated dataset in the format:
        chrStart  chrEnd  chrom_state  e2f4_motif_score  e2f6_motif_score  dnase_score  chipseq_peak
    chrStart, chrEnd -> integers
    chrom_state -> an integer between [1, 7], both inclusive
    e2f4_motif_score, e2f6_motif_score, dnase_score -> floats
    chipseq_peak -> binary [0 or 1]
    In the output file, all features except chrStart and chrEnd are 0.

4. mod_chrom.py
    Modifies the chromatin state. Input is a BED file.

5. mod_motif.py
    Modifies the E2F4 and E2F6 motif score values. Input is a GFF file.

6. mod_dnase_score.py
    Modifies the DNASE score. Input is a narrowPeak file.

7. mod_peak.py
    Modifies the peak values. Input is a tsv file.

The dataset created has an imbalanced class distribution (99% 0s and 1% 1s). It also has many duplicate instances. To change the class distribution or to clean up the data, proceed further. Else, stop.

8. dat2csv.py
    Creates a csv file from the previous output in the format:
        chrom_state,e2f4_motif_score,e2f6_motif_score,dnase_score,chipseq_peak
    [Optional] Now, we can can apply the SMOTE algorithm to change the class distribution. Use the Weka Software to apply SMOTE.

9. uniq.py
    Creates a csv file containing only the unique instances from the previous output file.

10. rm_ambiguous.py
    Removes all the ambiguous instances eg.
        7,8.872,0,0.0574,yes
        7,8.872,0,0.0574,no
    Instances of such type having class label as 'yes' are removed.

11. [Optional] Now, we can can apply the SMOTE algorithm to change the class distribution. Use the Weka Software to apply SMOTE.

12. ffnn_train.py
    Train the Feed Forward Neural Network model.
    Tune the hyperparameters to get the best accuracy, precision and recall.

    Setting the `cw` (class_weight):
        Count the number of 'no's and 'yes's in the final dataset.
        For example, if there are 1000000 no's and 10 yes's, `cw` will be:
        cw = {0:100000, 1:1} (The keys of this dictionary are the class labels and the corresponding values are the class weights)
    
    This file has options to change the input file format (.dat or .csv), save the learnt model and save the precision-recall values at different thresholds to a file.

13. ffnn_test.py
    Test the Feed Forward Neural Network model.
    It has options to change the input file format (.dat or .csv), show predictions and save them to a file and save the precision recall values to a file.

14. [optional] model.py
    This file has the LSTM Neural Network model.
    The above described options are also available in this file.

15. prc.py
    Plot the precision recall curve. Inputs are arrays of precision and recall values.

16. auprc.r
    Calculates the auPRC, given arrays of precision and recall values. Note: The recall values have to be in ascending order.