The purpose of the codes is to demonstrate the prediction of insulator-associated DNA-binding proteins performed in our paper entitled "Systematic discovery of regulatory motifs associated with human insulator sites reveals the role of directional bias".
In this demonstration, we calculate the DeepLIFT scores of DNA-binding sites (DBSs) of DNA-binding proteins (DBPs) in putative enhancers and promoters upstream and downstream of each transcript, taking into account the DBSs of known insulator-associated DBP CTCF (Fig. 2d and "Prediction of insulator-associated DBPs using DeepLIFT tool" in the Methods section of our paper). We used the same method to analyse other insulator-associated DBPs.
Due to file size limitations on GitHub, the dataset read from the codes is available from the Zenodo repository.
System environment
Ubuntu 20.04, Python 3.8.12, TensorFlow 2.6.3, NVIDIA CUDA toolkit 11.6, GeForce RTX 3080 GPU with 10 GB memory, Intel Core i7-10700 CPU with 128 GB memory.
(Originally in our paper, Python 3.8.10, TensorFlow 2.6.0, NVIDIA CUDA toolkit 11.4.2)
Requirements
Download the three data files (data.tgz, mydata1.tgz, mydata2.tgz) from the Zenodo repository, put them in the same directory as the codes in this GitHub repository, and extract the data with tar zxf *.tgz.
Python package and library (installed version)
- numpy (1.19.5)
- pandas (1.3.5)
- scipy (1.10.1)
- deeplift (0.6.12.0)
- yaml (6.0)
- h5py (3.1.0)
- tensorflow (2.6.3)
R package and library (installed version)
- tidyverse (2.0.0)
- ggpubr (0.6.0)
- broom (1.0.4)
DeepLIFT was developed based on TensorFlow 1. We have modified it to work with TensorFlow 2 for our analysis.
After installing the DeepLIFT package, the package was installed in our Python directory. We copied the original DeepLIFT directory, renamed it to deeplift_org, and modified the codes in it.
ls -l /home/user/.pyenv/versions/3.8.12/lib/python3.8/site-packages
drwxrwxr-x 6 user user 4096 3月 28 12:34 deeplift
drwxrwxr-x 2 user user 4096 3月 28 12:28 deeplift-0.6.13.0.dist-info
drwxrwxr-x 6 user user 4096 3月 28 12:34 deeplift_org
diff deeplift/conversion/kerasapi_conversion.py deeplift_org/conversion/kerasapi_conversion.py
386,387d385
+ < if 'input' in layer_name:
+ < continue
389,390c387,388
+ < ("Layer "+str(layer_name)+" is in the layer names but not in the "
+ < +" weights file which has layer names "+str(model_weights.keys()))
- > ("Layer "+layer_name+" is in the layer names but not in the "
- > +" weights file which has layer names "+model_weights.keys())diff deeplift/layers/convolutional.py deeplift_org/layers/convolutional.py
8,9d7
+ < tf.compat.v1.disable_v2_behavior()
+ < tf.compat.v1.disable_eager_execution()
93c91
+ < conv_without_bias = tf.compat.v1.nn.conv1d(
- > conv_without_bias = tf.nn.conv1d(
253c251
+ < conv_without_bias = tf.compat.v1.nn.conv2d(
- > conv_without_bias = tf.nn.conv2d(
281c279
+ < tf.compat.v1.nn.conv2d_transpose(
- > tf.nn.conv2d_transpose(
288c286
+ < +tf.compat.v1.nn.conv2d_transpose(
- > +tf.nn.conv2d_transpose(
296c294
+ < tf.compat.v1.nn.conv2d_transpose(
- > tf.nn.conv2d_transpose(
303c301
+ < +tf.compat.v1.nn.conv2d_transpose(
- > +tf.nn.conv2d_transpose(
310c308
+ < inp_mxts_increments += zero_inp_mask*tf.compat.v1.nn.conv2d_transpose(
- > inp_mxts_increments += zero_inp_mask*tf.nn.conv2d_transpose(diff deeplift/layers/core.py deeplift_org/layers/core.py
15,16c15
+ < tf.compat.v1.disable_v2_behavior()
+ < tf.compat.v1.disable_eager_execution()
66c65
+ < self._pos_mxts = tf.compat.v1.zeros_like(tensor=self.get_activation_vars(),
- > self._pos_mxts = tf.zeros_like(tensor=self.get_activation_vars(),
68c67
+ < self._neg_mxts = tf.compat.v1.zeros_like(tensor=self.get_activation_vars(),
- > self._neg_mxts = tf.zeros_like(tensor=self.get_activation_vars(),
208c207
+ < self._activation_vars = tf.compat.v1.placeholder(
- > self._activation_vars = tf.placeholder(
216c215
+ < return tf.compat.v1.placeholder(dtype=tf.float32,
- > return tf.placeholder(dtype=tf.float32,
456c455
+ < tf.compat.v1.variables_initializer([self.task_vector]))
- > tf.variables_initializer([self.task_vector]))
472c471
+ < task_vector_update = tf.compat.v1.assign(self.task_vector,
- > task_vector_update = tf.assign(self.task_vector,
474c473
+ < task_vector_update = tf.compat.v1.scatter_update(
- > task_vector_update = tf.scatter_update(diff deeplift/layers/helper_functions.py deeplift_org/layers/helper_functions.py
2,4d1
+ < tf.compat.v1.disable_v2_behavior()
+ < tf.compat.v1.disable_eager_execution()
24c21
+ < return tf.squeeze(tf.compat.v1.nn.conv2d_transpose(
- > return tf.squeeze(tf.nn.conv2d_transpose(diff deeplift/layers/pooling.py deeplift_org/layers/pooling.py
10,11c10
+ < tf.compat.v1.disable_v2_behavior()
+ < tf.compat.v1.disable_eager_execution()
64c63
+ < tf.compat.v1.nn.max_pool(value=tf.expand_dims(input_act_vars,1),
- > tf.nn.max_pool(value=tf.expand_dims(input_act_vars,1),
74c73
+ < return tf.compat.v1.zeros_like(tensor=self.get_activation_vars(),
- > return tf.zeros_like(tensor=self.get_activation_vars(),
76c75
+ < tf.compat.v1.zeros_like(tensor=self.get_activation_vars(),
- > tf.zeros_like(tensor=self.get_activation_vars(),
117c116
+ < return tf.compat.v1.zeros_like(tensor=self.get_activation_vars(),
- > return tf.zeros_like(tensor=self.get_activation_vars(),
119c118
+ < tf.compat.v1.zeros_like(tensor=self.get_activation_vars(),
- > tf.zeros_like(tensor=self.get_activation_vars(),
196,197c195
+ < # width = self._get_input_activation_vars().get_shape().as_list()[1]
+ < width = self._get_input_activation_vars().get_shape().as_list()[2]
- > width = self._get_input_activation_vars().get_shape().as_list()[1]
266c264
+ < to_return = tf.compat.v1.nn.max_pool(value=input_act_vars,
- > to_return = tf.nn.max_pool(value=input_act_vars,
280c278
+ < return tf.compat.v1.zeros_like(tensor=self.get_activation_vars(),
- > return tf.zeros_like(tensor=self.get_activation_vars(),
282c280
+ < tf.compat.v1.zeros_like(tensor=self.get_activation_vars(),
- > tf.zeros_like(tensor=self.get_activation_vars(),bash calc_deepliftscore.sh
bash stat_deepliftscore.sh
perl stat2_deepliftscore.pl 1
perl stat2_deepliftscore.pl 2
We analysed two criteria of DNA-binding sites (DBSs) of CTCF: DBS Selection Criteria S1 and S2, which are shown in our paper (Fig. 4b). Absolute z scores for DBS Selection Criteria S2 are shown in Fig. 3d in the paper.
perl stat2_deepliftscore.pl 1
CTCF FR_vs._RF FR_vs._FF FR_vs.RR FR_vs.X
zscore 30.5086444636761 22.6982073925036 22.1370605812343 -4.15304647173935
pvalue 2.001291193103142e-204 4.666373842989266e-114 1.390119461104552e-108 3.280780828751686e-05
mean_difference 0.000129712500594734 0.000106661611432207 0.000122422833153355 1.53591758089962e-05
mean 0.0024391885341362956 0.0024119834276677098 0.0024181798004967595 0.0023663710668580524 0.002309476033541562 0.002305321816235503 0.002295756967343405 0.0023510118910490562
rvalue 0.02175830785424293 0.02137963064522097 0.02080993035315193 0.004149985121522093
#_scores 1352987 766181 769080 648057 1270983 737247 740290 688467
perl stat2_deepliftscore.pl 2
CTCF FR_vs._RF FR_vs._FF FR_vs.RR FR_vs.X
zscore 29.3514936805777 20.0532840167669 20.0667097837562 -10.908618128604
pvalue 2.286905905408163e-189 1.88951021507238e-89 1.442432336817605e-89 1.0483782537220221e-27
mean_difference 0.0001211693880644 9.09719290027194e-05 0.000110929932730617 4.12370877662651e-05
mean 0.0023831139855248745 0.0023536584654016545 0.0023575815683481555 0.0023206333968334067 0.002261944597460475 0.002262686536398935 0.0022466516356175383 0.0022793963090671416
rvalue 0.02214496692055956 0.02006341963250835 0.020128639347794408 0.011655044924477997
#_scores 1211945 684926 679036 513268 1133046 648080 646886 677371
- DEcode Tasaki, S., Gaiteri, C., Mostafavi, S. & Wang, Y. Deep learning decodes the principles of differential gene expression. Nature Machine Intelligence (2020) [link to paper] BSD3 license
- DeepLIFT Avanti Shrikumar, Peyton Greenside, Anshul Kundaje. Learning Important Features Through Propagating Activation Differences (2017) [link to paper] MIT license