fangyiming / atmtcr

ATMTCR

ATMTCR consists of two parts, the MHC encoder and the TCR encoder. The packages required for the python runtime environment are in the MHC-encoder and the requirements.txt in the root directory respectively . Please refer to our paper for more details: 'Attention-aware contrastive learning for predicting T cell receptor-antigen binding specificity.'

The CDR3 sequences used for the pre-training process are from the TCRdb database. (http://bioinfo.life.hust.edu.cn/TCRdb/#/). This is the first time that the model is pre-trained for 10 million CDR3 sequences, and the encoder obtained by the model, which can directly encode CDR3 sequences, is convenient for use in related tasks.

TCR-encoder:Guided Tutorial

Need to change hyperparameters such as dataset in the code. Pre-training with TCR encoder,The encoder file is 'TCR-encoder/results/model_transformer_state_dict.pkl' Comand :

python TCR-encoder/main_train.py

Encoding of CDR3 sequences for downstream tasks using pre-trained models.

Comand :

python TCR-encoder/main_con.py

Note: We use the mainstream model netMHCpan as the MHC encoder to ensure that the MHC as well as the antigenic peptide encoding can preserve richer features

Command:

python MHC-encoder/mhc_encoder.py -input input.csv -library library -output output_dir -output_log test/output/output.log

• input.csv: input csv file with 3 columns named as "CDR3,Antigen,HLA": TCR-beta CDR3 sequence, peptide sequence, and HLA allele.\
• library: diretory to the downloaded library
• output_dir : diretory you want to save the output
• output_log : local directory to log file with CDR, Antigen, HLA information and predicted binding rank.\

Main:Guided Tutorial

After encoding by two encoders, downstream prediction can be performed using the master function Comand :

python main.py

atmtcr's Issues

How the attention-tower get gard

I wonder how the attention-tower get gardint,I alearly try some methods but the gradint of attention-tower is 0 all time.