中文文档

This is a model for antimicrobial peptides recognition based on BERT which is proposed.
We pre-train a BERT model through amount of proteins sequences downloaded from UniPort.
We fine-tune the model on 4 datasets and evaluate its performance.
We merge all the datasets and train a comprehensive prediction model.

You should first clone the project by command

git clone [email protected]:BioSequenceAnalysis/Bert-Protein.git

Then you need to download models and datasets from the address:

https://pan.baidu.com/s/1RH_OeHkpzMFa-YBnTMHerA

The password is a1yi

Then you should unzip these zips and put them on the root of the project.

Before you create data, you should make sure the input file is properly formatted(just like this).

train 0   SEQUENCE

We also provide a sample code(fasta2tsv.py) for using the transformation format (which may require you to change something to suit your needs).

You should create data for pre-train by the command

sh create_data.sh

You should ensure the content of file pre_train.sh

input_file is your input data for pre-training whose format is tf_record.
output_dir is the dir of your output model.
bert_config_file defines the structure of the model.
train_batch_size should be change more little if your computer don't support so big batch size.
You can change the num_train_steps by yourself.

After ensuring the content, then you can pre-trian your model by the command:

sh pre_train.sh

When you ready to fine-tune the model or do other, you should run the following code

python ljy_run_classifier.py \
--do_eval False \
--do_save_model True \
--data_name AMPScan \
--batch_size 32 \
--num_train_epochs 50 \
--warmup_proportion 0.1 \
--learning_rate 2e-5 \
--using_tpu False \
--seq_length 128 \
--data_root ./dataset/1kmer_tfrecord/AMPScan/ \
--vocab_file ./vocab/vocab_1kmer.txt \
--init_checkpoint ./model/1kmer_model/model.ckpt \
--bert_config ./bert_config_1.json \
--save_path ./model/AMPScan_1kmer_model/model.ckpt

The meaning of each parameter is as follows, you should change these according to your needs. You can also open file ljy_run_classifier and change the codes in row 16-32 to modify the default values of these parameters.

do_eval: whether to evaluate the model after training
do_save_model: whether to save the model after training
data_name: the name of the training set
batch_size: batch size
num_train_epochs: training epochs
warmup_proportion: proportion of warmup
learning_rate: learning rate
using_tpu: Whether to use TPU
seq_length: sequence length
data_root: the location of the training set to be used
vocab_file: location of dictionary
init_checkpoint: initialization node of the model
bert_config: BERT configuration
save_path: where to save the trained model

You can predict your proteins data by command

python ljy_predict_AMP.py

You should change the codes in row 167-172 according to your needs.

data_name: location of the testing set
out_file: storage location of test results
model_path: the location of the trained model
step: word segmentation
config_file: BERT configuration
vocab_file: location of dictionary