OrensteinLab / DeepZF

DeepZF

We present DeepZF, a two-step deep-learning-based pipeline for predicting binding ZFs and their DNA-binding preferences given only the amino acid sequence of a C2H2-ZF protein. To the best of our knowledge, we compiled the first in vivo dataset of binding and non-binding ZFs for training the first ZF-binding classifier; BindZFpredictor. We took advantage of both in vivo and in vitro datasets to learn the recognition code of ZF-DNA binding through transfer learning. Our newly developed model, PWMpredictor, is the first to utilize deep learning for the task.

BindZFpredictor

Prerequisites

BindZFpredictor architecture is based on ProteinBERT which was implemented using Tensorflow.

By installing ProteinBERT you will get all BindZFpredictor requirements:

pip install protein-bert

BindZFpredictor was run on python 3.6.8 with keras 2.6.0 and tensorflow 2.6.2

Usage

Predicting using BindZFpredictor

Create trained model file model.p by concatenating its splits:

cd BindZF_predictor/code/

cat x?? > model.p

Run under BindZFpredictor/code dir:

python3.6 main_bindzfpredictor_predict.py -in <input_file> -out <output_file> -m model.p -e encoder.p -r <gpu-0/1>

Example:

python3.6 main_bindzfpredictor_predict.py -in 40_zf_40_b.csv -out results.tsv -m model.p -e encoder.p -r 1

Note that you will need to identify the zinc fingers according to the regular expression by yourself. Then, append each finger by its adajcent 40aa residues on each side, and proive them as input to the command line above to obtain their DNA-binding probabilities.

Training BindZFpredictor

After installing ProteinBERT you can update finetunning.py as in this git for saving predictions.

cd path/to/BindZFpredictor/directory

2. Create saving folders

   data_name="${i}_zf_${i}_b"
   (where i = 10k k= [0,10] see Data/BindZFpredictor folder)
   f="path/to/BindZFpredictor/directory/${data_name}"
   mkdir -p $f
   mkdir -p ${f}/predictions

3. Run model

   python3.6 main_bindzfpredictor.py -b_n ${data_name} -b_d path/to/bemchmark_directory -m_d path/to/ProteinBERT_pretrained_model -r 1 -p_add ${f} >> out

Flags

   '-b_n', '--benchmark_name', help='zfs data and labels name ', type=str, required=True
   '-b_d', '--benchmark_dir', help='zfs data and labels directory ', type=str, required=True
    '-m_d', '--model_dir', help='ProteinBERT pretrained model directory', type=str, required=True
    '-r', '--run_gpu', help='equal 1 if should run on gpu', type=int, required=True
    '-p_add', '--pred_add', help='predictions saving folders add ', type=str, required=True

Creating predicted binding ZF dataframe and evaluate model

python3.6 create_zf_pred_df_and_cal_auc.py -p_add path/to/predicted ZF -m_p path/to/Data

Flags

'-p_add', '--pred_add', help='predictions saving folders add ', type=str, required=True
'-m_p', '--main_path', help='main path add ', type=str, required=True

PWMpredictor

Prerequisites

• python >= 3.6
• tensorflow >= 2.4.0
• natsort >= 8.4.0
• pandas >= 2.0.3

Usage

Predicting using PWMpredictor

Run under PWMpredictor/code dir:

python3.6 main_PWMpredictor.py -in <input_file> -out <output_file> -m <model_file>

Example:

python3.6 main_PWMpredictor.py -in ../../Data/PWMpredictor/c_rc_df.csv -out predictions.txt -m ../models/transfer_model.h5

Training PWMpredictor

1. Create saving folders:

	f="path/to/PWMpredictor_directory"

	mkdir -p $f
	
	mkdir -p ${f}/history
	
	mkdir -p ${f}/models
	
	mkdir -p ${f}/predictions

2. Run model:

	python3.6 main_loo_PWMpredictor.py -d_add /path_to_data/ -add ${f} -zf_p_df c_rc_df.csv -lr $lr -e $i -res_num 12 -r 0 -t_v retrain -ac_x False >> ${f}_out

Flags

'-d_add', '--data_folder_address', help='main data and labels folder', type=str, required=True)
'-add', '--folder_address', help='main folder address for savings', type=str, required=True
'-zf_p_df', '--pred_zf_df', help='predicted binding zinc fingers df', type=str, required=True
'-lr', '--learning_rate', help='learning rate of adam optimizer', type=float, required=True
'-e', '--epochs', help='number of epochs', type=int, required=True
'-res_num', '--residual_num', help='number of residuals to use: 4, 7, 12', type=int, required=True 
'-r', '--run_gpu', help='equal 1 if should run on gpu', type=int, required=True
'-t_v', '--transfer_version', help='last_layer or retrain', type=str, required=True
'-ac_x', '--amino_acid_x', help='use b1h data with amino acid x', type=str, required=True

Model evaluation:

For PWMpredictor evaluation, we computed the Pearson correlation of each quartet in the PWM matrix representing one position in the binding site. The fallowing script calcultes the Pearson correlation and saves the Pearson correlation scv file

python3.6 eval_PWMpredictor.py -p_add /predictions_folder/ --c_rc_add /path/to/c_rc_df.csv -s_add /path_for savings >> out

Flags

'-p_add', '--pred_folder_add', help='c_rc predictions folder add', type=str, required=True
'-c_rc_add', '--c_rc_add', help='c_rc data frame folder add', type=str, required=True
'-s_add', '--s_add', help='saving folder add', type=str, required=True

DeepZF

Prerequisites

1. Running the model
   • python >= 3.6
   • tensorflow >= 2.4.0
2. Model evaluation In addition to the above:
   • MoSBAT for DeepZF evaluation.

Model training

1. Same as stage 1 in PWMpredictor

2. Run model:

	python3.6 main_loo_PWMpredictor.py -d_add /path_to_data/ -add ${f} -zf_p_df zf_pred.csv -lr $lr -e $i -res_num 12 -r 0 -t_v retrain -ac_x False >> ${f}_out

Flags

Same as in PWMpredictor (-zf_p_df is different now)

Model evaluation:

For DeepZF evaluation we calculated the similarity of two motif pairs (on the predicted and experimentally based PWM), using MoSBAT. Run bash script: (update paths)

./eval_DeepZF.sh

The script above creates 2 folder:

1. mosbat_input: includes mosbat input txt files: ground ruth and predicted PWMs.
2. mosbat_output: includes:
   • results.energy.correl.txt
   • correlation data frame: Pearson correlation score for each protein.
   • out_eval_mosbat.txt: a txt file with mean and std score.

Cite

https://doi.org/10.1093/bioinformatics/btac469