This project is a big assignment on the UCAS Deep Learning course.We focus on the Identification of RBP(RNA-binding protein) binding sites based on convolutional neural networks.
RNA-binding proteins (RBPs) account for 5~10% of the eukaryotic proteome, which plays an important role in biological processes such as gene regulation. However, experimental prediction of RBP binding sites remains a time-consuming, costly task. Conversely, learning from existing annotation knowledge and predicting RBP binding sites is a quick method. In this project,we use a deep learning based model,which ensembles 2 convolutional neural networks (CNNs) to identify RBP binding sites. After training the two networks separately, the outputs of the two networks are combined to increase the accuracy of predicting RBP binding sites on the RNA sequence.The results show that the network we trained successfully captured experimentally validated binding motifs.
- python 3.6
- pytorch 1.1
- sklearn
The programs support GPUs and CPUs and are able to automatically detect the operating environment. If GPUs are present, it will run with the GPU first.
To ensure that the entire process works properly,verify that the directory structure all files are the same as follow: The directory structure of the training code and sample data is as follows:
│—— encoding.py
│—— motif_detection.py
│—— pre_processing.py
│—— train.py
│—— README.md
|—— workflow.png
|—— data
|—— ALKBH5_Baltz2012.ls.negatives.fa
|—— ALKBH5_Baltz2012.ls.positives.fa
|—— ALKBH5_Baltz2012.train.negatives.fa
|—— ALKBH5_Baltz2012.train.positives.fa
|—— ...
If you need to train other protein data, you need to put the training data in the data directory.
Other protein data links: https://pan.baidu.com/s/1DoNyS4hSPe5RtSpQTDTp9g
password：ughc
train.py
usage: train.py [-h] [-model_type MODEL_TYPE] [-protein PROTEIN]
[-epoch EPOCH]
optional arguments:
-h, --help show this help message and exit
-model_type MODEL_TYPE
it supports the following deep network
models:globalCNN,localCNN and ensemble, default=ensemble
-protein PROTEIN input the protein you want to train ,default=ALKBH5
-epoch EPOCH input the epoch you want to train ,default=1
motif_detection.py
usage: motif_detection.py [-h] [-protein PROTEIN]
optional arguments:
-h, --help show this help message and exit
-protein PROTEIN input the protein you have trained,i.e ALKBH5
- step1
run train.py
>>>python train.py -model_type ensemble -protein ALKBH5 -epoch 5
train.py can automatically recognizes the detection training environment. If the GPU is available, it will automatically switch to GPU execution.
The training effect will be printed in real time during the run of train.py.
After train.py is executed, the ALKBH5_param directory will be created, and the network parameters obtained by the training will be stored in this directory:
>>>ls ./ALKBH5_param
globalCNN_ALKBH5.pkl
localCNN_ALKBH5.pkl
- step2
runmotif_detection.py
>>>python motif_detection.py -protein ALKBH5
motif_detection.py recognizes potential Motifs and outputs the results in the 'ALKBH5_Motif` directory are as follow:
>>>ls ./ALKBH5_Motif
filter0_logo.fa
filter10_logo.fa
filter11_logo.fa
filter12_logo.fa
filter13_logo.fa
filter14_logo.fa
filter15_logo.fa
filter1_logo.fa
filter2_logo.fa
filter3_logo.fa
filter4_logo.fa
filter5_logo.fa
filter6_logo.fa
filter7_logo.fa
filter8_logo.fa
filter9_logo.fa
filters_meme.txt
The output of filters_meme.txt can be further compared with the experimentally validated Motif using the TOMTOM software (http://meme-suite.org/tools/tomtom).
filter i _logo.fa can be used as an input to AME (http://meme-suite.org/tools/ame) to detect the enrichment of Motifs.
Our work is strongly inspired by:
[1]Alipanahi B , Delong A , Weirauch M T , et al. Predicting the sequence specificities of DNA- and RNA-binding proteins by deep learning[J]. Nature Biotechnology, 2015.