# BERT-Kcr: Prediction of protein lysine crotonylation (Kcr) sites by a transfer learning method with pre-trained BERT models


1. All data we used in this paper is in ./data/.

2. If you want to perform Kcr sites prediction, please follow these steps:
  (1) Predict Kcr sites using your fasta file. Please use tensorflow 1.X (tensorflow > 1.10) version to run: 
     wget http://zhulab.org.cn/BERT-Kcr_models/BERT-Kcr_models.tar.gz
     tar zxvf BERT-Kcr_models.tar.gz
     cd ./BERT-Kcr_models/codes
     python2 ./predict.py ./example.fasta
  (2) After you completed above steps, the final result are in file ./BERT-Kcr_models/codes/output/Results.txt.
  
3. If you want to use our data to train the BERT-Kcr model with all training dataset and test on the independent test dataset, please use tensorflow 1.X (tensorflow > 1.10) version to run:
     wget http://zhulab.org.cn/BERT-Kcr_models/BERT-Kcr_models.tar.gz
     tar zxvf BERT-Kcr_models.tar.gz
     cd ./BERT-Kcr_models/codes
     python2 ./extract-last-layer-features.py
     python2 ./train.py
     python2 ./test.py
     
     Note: Due to the randomness of the initial parameters in the process of feature extraction and training of the neural network, the final AUC result will have an error of about 0.001-0.01.