Skip to content


Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time
Jul 16, 2019
Feb 8, 2022



This Python script is used to train, validate, test deep learning model for prediction of drug-target interaction (DTI) Deep learning model will be built by Keras with tensorflow. You can set almost hyper-parameters as you want, See below parameter description DTI, drug, target protein and their interaction data must be written as csv file format. And feature should be tab-delimited format for script to parse data. Basically, this script builds convolutional neural network on sequence. If you don't want convolutional neural network but traditional dense (fully connected) layers on provide protein feature, specify type of feature and feature length.


tensorflow > 1.0 and < 2.0
keras > 2.0 


  usage: [-h] [--test-name [TEST_NAME [TEST_NAME ...]]]
                     [--test-dti-dir [TEST_DTI_DIR [TEST_DTI_DIR ...]]]
                     [--test-drug-dir [TEST_DRUG_DIR [TEST_DRUG_DIR ...]]]
                     [--test-protein-dir [TEST_PROTEIN_DIR [TEST_PROTEIN_DIR ...]]]
                     [--with-label WITH_LABEL]
                     [--window-sizes [WINDOW_SIZES [WINDOW_SIZES ...]]]
                     [--protein-layers [PROTEIN_LAYERS [PROTEIN_LAYERS ...]]]
                     [--drug-layers [DRUG_LAYERS [DRUG_LAYERS ...]]]
                     [--fc-layers [FC_LAYERS [FC_LAYERS ...]]]
                     [--learning-rate LEARNING_RATE] [--n-epoch N_EPOCH]
                     [--prot-vec PROT_VEC] [--prot-len PROT_LEN]
                     [--drug-vec DRUG_VEC] [--drug-len DRUG_LEN]
                     [--activation ACTIVATION] [--dropout DROPOUT]
                     [--n-filters N_FILTERS] [--batch-size BATCH_SIZE]
                     [--decay DECAY] [--validation] [--predict]
                     [--save-model SAVE_MODEL] [--output OUTPUT]
                     dti_dir drug_dir protein_dir

Data Specification

All training, validation, test should follow specification to be parsed correctly by DeepConv-DTI

  • Model takes 3 types data as a set, Drug-target interaction data, target protein data, compound data.

  • They should be .csv format.

  • For feature column, each dimension of features in columns should be delimited with tab (\t)

After three data are correctly listed, target protein data and compound data will be joined with drug-target data, generating DTI feature.

Drug target interaction data

Drug target interaction data should be at least 2 columns Protein_ID and Compound_ID,

and should have Label column except --test case. Label colmun has to have label 0 as negative and 1 as positive.

Protein_ID Compound_ID Label
PID001 CID001 0
... ... ...
PID100 CID100 1

Target protein data

Because DeepConvDTI focuses on convolution on protein sequence, protein data specification is little different from other data.

If Sequence column is specified in data and --prot-vec is Convolution, it will execute convolution on protein.

Or if you specify other type of column with --prot-vec(i.e. Prot2Vec), it will construct dense (fully connected) network

Protein_ID column will be used as foreign key from Protein_ID from Drug-target interaction data.

Protein_ID Sequence Prot2Vec
PID001 MALAC....ACC 0.539\t-0.579\t...\t0.39

Compound data

Basically same with Target protein data, but no Convolution.

Compound_ID column will be used as forein key from Compound_ID from Drug-target interaction data.

Compound_ID morgan_r2
CID001 0\t1\t...\t0\t1

Parameter specification

Positional arguments

    dti_dir               Training DTI information [drug, target, label]
    drug_dir              Training drug information [drug, SMILES,[feature_name,
    protein_dir           Training protein information [protein, seq,

For training model, you should input 3 files, DTI information file, drug information file and target protein information file, as their format is specified above.

DTI information file for training should have Label column for training.

Optional arguments

    --validation          Excute validation with independent data, will give AUC
                          and AUPR (No prediction result)
    --predict             Predict interactions of independent test set

DeepConvDTI script has two mode, validation and predict mode.

In validation mode, performances (AUC, AUPR, threshold for AUC and AUPR) on each step and selected hyperparameters are recorded.

In test step, prediction results for given test dataset will be reported after training.

    --test-name [TEST_NAME [TEST_NAME ...]], -n [TEST_NAME [TEST_NAME ...]]
                          Name of test data sets
    --test-dti-dir [TEST_DTI_DIR [TEST_DTI_DIR ...]], -i [TEST_DTI_DIR [TEST_DTI_DIR ...]]
                          Test dti [drug, target, [label]]
    --test-drug-dir [TEST_DRUG_DIR [TEST_DRUG_DIR ...]], -d [TEST_DRUG_DIR [TEST_DRUG_DIR ...]]
                          Test drug information [drug, SMILES,[feature_name,
                          Test Protein information [protein, seq,
    --with-label WITH_LABEL, -W WITH_LABEL
                          Existence of label information in test DTI

You can input multiple datasets for validation or test with argument specifier.

In addition to DTI information file, drug information file and target protein information file, you need to name of validation or test datasets.

For test dataset, you can inform that test dataset has label or not with -W value

    --n-epoch N_EPOCH, -e N_EPOCH
                          The number of epochs for training or validation

The number of epoch for model training.

Validation will stop evaluating performance with specified epoch.

Test for given dataset will be executed after specified epoch.

    --prot-vec PROT_VEC, -v PROT_VEC
                          Type of protein feature, if Convolution, it will
                          execute conlvolution on sequeunce
    --prot-len PROT_LEN, -l PROT_LEN
                          Protein vector length
    --drug-vec DRUG_VEC, -V DRUG_VEC
                          Type of drug feature
    --drug-len DRUG_LEN, -L DRUG_LEN
                          Drug vector length

Parameters for parsing data.

Model will parse column with specified name of column for drug and target protein feature.

Also, you need to give length of feature, which will be parsed.

For special case, with -v Convolution model will build convolution layer on Sequence of dataset.

    --window-sizes [WINDOW_SIZES [WINDOW_SIZES ...]], -w [WINDOW_SIZES [WINDOW_SIZES ...]]
                          Window sizes for model (only works for Convolution)
                          Dense layers for protein
    --drug-layers [DRUG_LAYERS [DRUG_LAYERS ...]], -c [DRUG_LAYERS [DRUG_LAYERS ...]]
                          Dense layers for drugs
    --fc-layers [FC_LAYERS [FC_LAYERS ...]], -f [FC_LAYERS [FC_LAYERS ...]]
                          Dense layers for concatenated layers of drug and
                          target layer
    --n-filters N_FILTERS, -F N_FILTERS
                          Number of filters for convolution layer, only works
                          for Convolution

Hyperparameters which determine shape of neural network.

If you want to deeper layer, you can write as -c 128 32, which will construct two consecutive neural layer on input drug feature with 128 and 32 units

    --activation ACTIVATION, -a ACTIVATION
                          Activation function of model
    --dropout DROPOUT, -D DROPOUT
                          Dropout ratio

Other hyperparameters.

We don't recommend to use dropout

    --batch-size BATCH_SIZE, -b BATCH_SIZE
                          Batch size
    --learning-rate LEARNING_RATE, -r LEARNING_RATE
                          Learning late for training
    --decay DECAY, -y DECAY
                          Learning rate decay

Hyperparameters for training

    --save-model SAVE_MODEL, -m SAVE_MODEL
                          save model

Save model after training done.

    --output OUTPUT, -o OUTPUT
                          Prediction output

Directory of .csv format file.

In validation mode, it will contain results for every epoch

In test mode, it will contain prediction results for given DTI pairs.

Toy Examples

Training and validation of model

For line example, if you have training dataset, toy_examples/training_dataset/training_dti.csv, toy_examples/training_dataset/training_compound.csv and toy_examples/training_dataset/training_protein.csv with right specification. You can validate model with validation dataset, toy_examples/validation_dataset/validation_dti.csv, toy_examples/validation_dataset/validation_compound.csv and toy_examples/validation_dataset/validation_protein.csv by using this command line.

python ./toy_examples/training_dataset/training_dti.csv ./toy_examples/training_dataset/training_compound.csv ./toy_examples/training_dataset/training_protein.csv --validation -n validation_dataset -i ./toy_examples/validation_dataset/validation_dti.csv -d ./toy_examples/validation_dataset/validation_compound.csv -t ./toy_examples/validation_dataset/validation_protein.csv -W -c 512 128 -w 10 15 20 25 30 -p 128 -f 128 -r 0.0001 -n 30 -v Convolution -l 2500 -V morgan_fp_r2 -L 2048 -D 0 -a elu -F 128 -b 32 -y 0.0001 -o ./validation_output.csv -m ./model.model -e 1

This command will train model with given hyper-parameters for 1 epoch. (because of -e 1). And resulting in validation result ./validation_output.csv, and corresponding model ./model.model

Prediction by using model

There are two ways to predict DTIs with command line When you have toy_examples/test_dataset/test_dti.csv, toy_examples/test_dataset/test_compound.csv and toy_examples/test_dataset/test_protein.csv

python ./toy_examples/training_dataset/training_dti.csv ./toy_examples/training_dataset/training_compound.csv ./toy_examples/training_dataset/training_protein.csv --predict -n predict -i ./toy_examples/test_dataset/test_dti.csv -d ./toy_examples/test_dataset/test_compound.csv -t ./toy_examples/test_dataset/test_protein.csv -c 512 128 -w 10 15 20 25 30 -p 128 -f 128 -r 0.0001 -n 30 -v Convolution -l 2500 -V morgan_fp_r2 -L 2048 -D 0 -a elu -F 128 -b 32 -y 0.0001 -o ./test_output.csv -m ./model.model -e 15 -W

With this command, model will be trained with training dataset for 15 epochs and predict test dataset when it finished training, resulting in test_output.csv which have prediction score and its true label (because of -W) The second way of prediction is using If you have model which is saved from validation or something, you can use it.

python ./model.model -n predict -i ./toy_examples/test_dataset/test_dti.csv -d ./toy_examples/test_dataset/test_compound.csv -t ./toy_examples/test_dataset/test_protein.csv -v Convolution -l 2500 -V morgan_fp_r2 -L 2048 -W -o test_result.csv

this code will result in same result file with first command.

evaluation of performance of model

In addition, you can evaluate the performances of prediction results with label by using When you have an optimal threshold from validation and the names of the test dataset.

python test_result.csv -n predict -T 0.2

This command will report performances.


DeepConv-DTI follow GPL 3.0v license. Therefore, DeepConv-DTI is open source and free to use for everyone.

However, compounds which are found by using DeepConv-DTI follows CC-BY-NC-4.0. Thus, those compounds are freely available for academic purpose or individual research, but restricted for commecial use.



DeepConv-DTI: Prediction of drug-target interactions via deep learning with convolution on protein sequences







No releases published


No packages published