MT-DTI

An official Molecule Transformer Drug Target Interaction (MT-DTI) model

• Author: Bonggun Shin
• Paper: Shin, B., Park, S., Kang, K. & Ho, J.C.. (2019). Self-Attention Based Molecule Representation for Predicting Drug-Target Interaction. Proceedings of the 4th Machine Learning for Healthcare Conference, in PMLR 106:230-248

Required Files

• Download data.tar.gz

   cd mt-dti
   wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=16dTynXCKPPdvQq4BiXBdQwNuxilJbozR' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=16dTynXCKPPdvQq4BiXBdQwNuxilJbozR" -O data.tar.gz && rm -rf /tmp/cookies.txt
   tar -zxvf data.tar.gz
  

  • This includes;
    • Orginal KIBA dataset from DeepDTA
    • tfrecord for KIBA dataset
    • Pretrained weights of the molecule transformer
    • Finetuned weights of the MT-DTI model for KIBA fold0

• Unzip it (folder name is data) and place under the project root

cd mt-dti
# place the downloaded file (data.tar.gz) at "mt-dti"
tar xzfv data.tar.gz

• These files sholud be in the right places

mt-dti/data/chembl_to_cids.txt
mt-dti/data/CID_CHEMBL.tsv
mt-dti/data/kiba/*
mt-dti/data/kiba/folds/*
mt-dti/data/kiba/mbert_cnn_v1_lr0.0001_k12_k12_k12_fold0/*
mt-dti/data/kiba/tfrecord/*.tfrecord
mt-dti/data/pretrain/*
mt-dti/data/pretrain/mbert_6500k/*

VirtualEnv

• install mkvirtualenv
• create a dti env with the following commands

mkvirtualenv --python=`which python3` dti
pip install tensorflow-gpu==1.12.0

Preprocessing

• If downloaded data.tar.gz, then you can skip these preprocessings

• Transform kiba dataset into one pickle file

python kiba_to_pkl.py 

# Resulted files
mt-dti/data/kiba/kiba_b.cpkl

• Prepare Tensorflow Record files

cd src/preprocess
export PYTHONPATH='../../'
python tfrecord_writer.py 

# Resulted files
mt-dti/data/kiba/tfrecord/*.tfrecord

PreTraining

• Download Pubchem smiles

$ head CID-SMILES
1	CC(=O)OC(CC(=O)[O-])C[N+](C)(C)C
2	CC(=O)OC(CC(=O)O)C[N+](C)(C)C
3	C1=CC(C(C(=C1)C(=O)O)O)O
4	CC(CN)O
5	C(C(=O)COP(=O)(O)O)N
6	C1=CC(=C(C=C1[N+](=O)[O-])[N+](=O)[O-])Cl
7	CCN1C=NC2=C(N=CN=C21)N
8	CCC(C)(C(C(=O)O)O)O
9	C1(C(C(C(C(C1O)O)OP(=O)(O)O)O)O)O

• Split into several files

  • CID-SMILES -> smiles00.txt, smiles01.txt, ...
  • Place these files to

   mt-dti/data/pretrain/molecule/smiles*
  

• Make tfrecords for pretraining

cd src/pretrain
export PYTHONPATH='../../'
python tfrecord_smiles.py 

• This will create tfrecord files in the "output folder" of your google cloud storage

# for example
gs://your_gs/mbert/tfr/smiles.001
gs://your_gs/mbert/tfr/smiles.002
...

• Now pretrain (Need TPU in google cloud)

cd src/pretrain
export PYTHONPATH='../../'
python pretrain_smiles_tpu.py

• The resulting pretrained model will be stored at the checkpoint folder of your google cloud storage

# for example
gs://your_gs/mbert/pretrain-mini/model.ckpt-6500000.*

Result

INFO:tensorflow:Saving checkpoints for 6500000 into gs://bdti/mbert/pretrain/model.ckpt.
INFO:tensorflow:loss = 0.098096184, step = 6500000 (48.736 sec)
INFO:tensorflow:global_step/sec: 20.5185
INFO:tensorflow:examples/sec: 10505.5
INFO:tensorflow:Stop infeed thread controller
INFO:tensorflow:Shutting down InfeedController thread.
INFO:tensorflow:InfeedController received shutdown signal, stopping.
INFO:tensorflow:Infeed thread finished, shutting down.
INFO:tensorflow:infeed marked as finished
INFO:tensorflow:Stop output thread controller
INFO:tensorflow:Shutting down OutfeedController thread.
INFO:tensorflow:OutfeedController received shutdown signal, stopping.
INFO:tensorflow:Outfeed thread finished, shutting down.
INFO:tensorflow:outfeed marked as finished
INFO:tensorflow:Shutdown TPU system.
INFO:tensorflow:Loss for final step: 0.098096184.
INFO:tensorflow:training_loop marked as finished

mini model
INFO:tensorflow:***** Eval results *****
INFO:tensorflow:  global_step = 6500000
INFO:tensorflow:  loss = 0.15356757
INFO:tensorflow:  masked_lm_accuracy = 0.94406235
INFO:tensorflow:  masked_lm_loss = 0.1413514

FineTuning

• If downloaded data.tar.gz, then you can skip this finetuning

cd src/finetune
export PYTHONPATH='../../'
python finetune_demo.py 

Prediction

cd src/predict
export PYTHONPATH='../../'
python predict_demo.py