README.md

WAE-DTI: Ensemble-based architecture for drug-target interaction prediction using descriptors and embeddings

Getting Started

Clone the project from GitHub and install the necessary dependencies.

git clone https://github.com/tariqshaban/wae-dti
cd wae-dti
pip install -r requirements.txt

Pytorch needs to be installed (preferably utilizing GPU acceleration "CUDA-enabled").

The program will attempt to clone the ESM repository if it is not present automatically. Either install Git on your machine or manually download the repository and place it in ../esm.

Tip

You can also clone the ESM repository beforehand by running the following command:

git clone https://github.com/facebookresearch/esm ../esm

Project Structure

├── config
│   └── config.py                     <- Store terminal arguments from entry files.
├── data
│   ├── embeddings
│   │   ├── drug_embedding            <- Pre-trained drug embeddings (Parquet format).
│   │   └── target_embedding          <- Pre-trained target embeddings (Parquet format).
│   └── raw
│       ├── classification            <- Raw classification datasets for training and evaluation (Parquet format).
│       └── regression                <- Raw regression datasets for training and evaluation (Parquet format).
├── results                           <- Store trained model, predictions, and metrics.
├── saved_models
│   ├── classification                <- Trained models and their performance on the classification task.
│   └── regression                    <- Trained models and their performancee on the regression task.
├── src
│   ├── dti_dataset.py                <- Data preprocessing and mounting prior to training.
│   ├── dti_model.py                  <- Neural network definition for the WAE-DTI architecture.
│   ├── ensemble.py                   <- Implementation of the weighted average ensemble method.
│   ├── evaluate.py                   <- Mathematical methods to evaluate model performance.
│   ├── predict.py                    <- Provide predictions of a model given a dataloader.
│   ├── test.py                       <- Generate and evaluate predictions of the model given external examples.
│   └── train.py                      <- WAE-DTI model trainer.
├── utils
│   ├── embedding
│   │   ├── extractor
│   │   │   ├── drug_extractor.py     <- Extract and save embeddings from a list of drugs (Parquet format).
│   │   │   └── target_extractor.py   <- Extract and save embeddings from a list of targets (Parquet format).
│   │   └── loader
│   │       ├── drug_loader.py        <- Load drug embeddings from a Parquet file.
│   │       └── target_loader.py      <- Load target embeddings from a Parquet file.
│   ├── eda.py                        <- Perform exploratory data analysis (EDA) on all datasets.
│   └── terminal_command_runner.py    <- Run external terminal commands in real time.
├── entry_inference.py                <- Entry point for inference.
├── entry_train.py                    <- Entry point for training.
├── README.md                         <- README file and documentation.
└── requirements.txt                  <- List of dependencies required to run the project.

Usage

Training

entry_train.py [-h] [--use-wandb USE_WANDB] --task {classification,regression} [--dataset DATASET] [--seed SEED] [--epochs EPOCHS] [--patience PATIENCE]
               [--eda EDA] [--learning-rate LEARNING_RATE] [--batch-size BATCH_SIZE] [--torch-device TORCH_DEVICE]

Argument	Description	Default	Notes
-h, --help	Display a help message
--use-wandb USE_WANDB	Log training and validation metrics into W&B	False	If set to True, the code will initialize a project named by the selected dataset
--task	Supervised learning algorithm type		Must be set to either classification or regression
--dataset DATASET	Dataset used for training	'davis'	The dataset must be in the data/raw path with train and test folders containing Parquet files
--seed SEED	Ensures reproducibility (on the same device)
--epochs EPOCHS	Number of training epochs	3000
--patience PATIENCE	Number of epochs to elapse without improvement for the training to stop	200
--eda EDA	Conduct a quick EDA on startup	False	EDA is applied to all datasets regardless of the specified --dataset
--learning-rate LEARNING_RATE	Learning rate	5e-4
--batch-size BATCH_SIZE	Number of examples in each batch	1024
--torch-device TORCH_DEVICE	Device used for training (e.g. cuda:0, cpu)		If not specified, GPU will be utilized (if any)

Inference

entry_inference.py [-h] --models-path MODELS_PATH --input-file INPUT_FILE 
                   [--batch-size BATCH_SIZE] [--torch-device TORCH_DEVICE]

Argument	Description	Default	Notes
-h, --help	Display a help message
--models-path MODELS_PATH	Folder path which contains the models trained on each drug fingerprint		You can use the pretrained models of any dataset within models/saved_models
--input-file INPUT_FILE	CSV file path containing "drug", "target", and "label" columns (label is optional)		If "label" column is specified, you must have enough examples to satisfy the concordance index calculation requirement
--task	Supervised learning algorithm type		Must be set to either classification or regression
--batch-size BATCH_SIZE	Number of examples in each batch	1024
--torch-device TORCH_DEVICE	Device used for training (e.g. cuda:0, cpu)		If not specified, GPU will be utilized (if any)

Note

When running entry_train.py and entry_inference.py, missing embeddings from drugs and targets are automatically extracted and saved into data/embeddings

Important

During training, you may notice that tqdm reports monotonically increasing total value, this is caused by taking into account the dynamic fluctuation of the early stopping counter. So, rather than always having the total value equal to the number of epochs, the progress bar adjusts to display the smallest number of epochs needed to finish the training, which translates to the following formula:

total = min(epochs, elapsed_epochs + patience - elapsed_patience)

Results

The following tables are the result of training the model using nine drug descriptors and one target embedding (ESM-2) on six datasets while repeating the experiment five times. The values represent the mean and standard deviation of each metric.

Model	Davis			Kiba
	MSE	CI	R2	MSE	CI	R2
Atom pair fingerprint	0.229 ± 0.006	0.892 ± 0.002	0.706 ± 0.007	0.155 ± 0.002	0.885 ± 0.001	0.751 ± 0.006
Avalon fingerprint	0.213 ± 0.005	0.897 ± 0.004	0.718 ± 0.011	0.150 ± 0.002	0.883 ± 0.001	0.772 ± 0.006
MACCS keys fingerprint	0.217 ± 0.002	0.895 ± 0.004	0.706 ± 0.007	0.171 ± 0.002	0.868 ± 0.003	0.732 ± 0.013
MH fingerprint	0.217 ± 0.007	0.894 ± 0.002	0.709 ± 0.014	0.158 ± 0.002	0.883 ± 0.001	0.757 ± 0.014
Morgan fingerprint	0.220 ± 0.005	0.895 ± 0.003	0.700 ± 0.011	0.157 ± 0.003	0.883 ± 0.001	0.753 ± 0.014
RDKit fingerprint	0.222 ± 0.008	0.895 ± 0.002	0.711 ± 0.011	0.154 ± 0.001	0.885 ± 0.002	0.751 ± 0.010
SEC fingerprint	0.219 ± 0.004	0.893 ± 0.004	0.711 ± 0.004	0.158 ± 0.002	0.883 ± 0.002	0.756 ± 0.009
Topological torsion fingerprint	0.218 ± 0.007	0.896 ± 0.002	0.704 ± 0.012	0.158 ± 0.004	0.883 ± 0.003	0.747 ± 0.015
LDP	0.310 ± 0.003	0.853 ± 0.003	0.597 ± 0.014	0.389 ± 0.003	0.756 ± 0.001	0.422 ± 0.007
Ensemble	0.190 ± 0.001	0.915 ± 0.001	0.745 ± 0.003	0.127 ± 0.001	0.899 ± 0.000	0.778 ± 0.003

Model	DTC			Metz
	MSE	CI	R2	MSE	CI	R2
Atom pair fingerprint	0.188 ± 0.004	0.883 ± 0.001	0.787 ± 0.009	0.357 ± 0.004	0.787 ± 0.002	0.570 ± 0.005
Avalon fingerprint	0.177 ± 0.003	0.883 ± 0.003	0.812 ± 0.013	0.325 ± 0.009	0.798 ± 0.005	0.616 ± 0.005
MACCS keys fingerprint	0.201 ± 0.003	0.867 ± 0.003	0.800 ± 0.004	0.337 ± 0.008	0.793 ± 0.003	0.610 ± 0.005
MH fingerprint	0.193 ± 0.006	0.877 ± 0.004	0.781 ± 0.010	0.358 ± 0.008	0.787 ± 0.002	0.575 ± 0.016
Morgan fingerprint	0.191 ± 0.004	0.880 ± 0.002	0.789 ± 0.015	0.365 ± 0.003	0.785 ± 0.001	0.561 ± 0.004
RDKit fingerprint	0.179 ± 0.005	0.885 ± 0.002	0.806 ± 0.017	0.330 ± 0.003	0.796 ± 0.001	0.610 ± 0.008
SEC fingerprint	0.191 ± 0.001	0.881 ± 0.001	0.789 ± 0.007	0.364 ± 0.005	0.785 ± 0.002	0.563 ± 0.009
Topological torsion fingerprint	0.189 ± 0.003	0.880 ± 0.001	0.786 ± 0.007	0.363 ± 0.002	0.787 ± 0.001	0.563 ± 0.006
LDP	0.434 ± 0.005	0.768 ± 0.001	0.565 ± 0.007	0.532 ± 0.009	0.723 ± 0.003	0.418 ± 0.011
Ensemble	0.143 ± 0.001	0.898 ± 0.001	0.839 ± 0.008	0.284 ± 0.004	0.813 ± 0.001	0.676 ± 0.006

Model	ToxCast			STITCH
	MSE	CI	R2	MSE	CI	R2
Atom pair fingerprint	0.326 ± 0.002	0.914 ± 0.002	0.553 ± 0.008	1.140 ± 0.003	0.751 ± 0.004	0.389 ± 0.004
Avalon fingerprint	0.321 ± 0.003	0.916 ± 0.001	0.559 ± 0.009	1.078 ± 0.006	0.740 ± 0.006	0.417 ± 0.003
MACCS keys fingerprint	0.324 ± 0.000	0.914 ± 0.001	0.561 ± 0.006	1.120 ± 0.006	0.708 ± 0.004	0.407 ± 0.002
MH fingerprint	very high MSE	very low CI	very low rm2	very high MSE	very low CI	very low rm2
Morgan fingerprint	0.333 ± 0.002	0.911 ± 0.002	0.540 ± 0.012	1.122 ± 0.006	0.763 ± 0.004	0.398 ± 0.005
RDKit fingerprint	0.323 ± 0.002	0.915 ± 0.001	0.553 ± 0.005	1.143 ± 0.008	0.743 ± 0.005	0.389 ± 0.006
SEC fingerprint	0.333 ± 0.003	0.911 ± 0.002	0.540 ± 0.009	1.167 ± 0.006	0.759 ± 0.005	0.374 ± 0.007
Topological torsion fingerprint	0.331 ± 0.002	0.913 ± 0.001	0.541 ± 0.006	1.163 ± 0.003	0.754 ± 0.004	0.378 ± 0.004
LDP	0.369 ± 0.001	0.898 ± 0.000	0.506 ± 0.002	1.542 ± 0.003	0.627 ± 0.002	0.201 ± 0.001
Ensemble	0.308 ± 0.001	0.922 ± 0.000	0.581 ± 0.003	0.934 ± 0.004	0.772 ± 0.001	0.488 ± 0.003

Classification Task Results

In addition to the mentioned results for the regression task, three datasets were used to evaluate the model for the classification task.

Model	BioSNAP
	Sensitivity	Specificity	AUC	AUPRC	Threshold
Atom pair fingerprint	0.824 ± 0.008	0.830 ± 0.019	0.899 ± 0.005	0.905 ± 0.002	0.560 ± 0.251
Avalon fingerprint	0.859 ± 0.008	0.872 ± 0.005	0.932 ± 0.002	0.933 ± 0.001	0.582 ± 0.121
MACCS keys fingerprint	0.873 ± 0.006	0.884 ± 0.006	0.942 ± 0.002	0.945 ± 0.002	0.557 ± 0.126
Morgan fingerprint	0.789 ± 0.008	0.830 ± 0.004	0.880 ± 0.003	0.892 ± 0.004	0.701 ± 0.082
RDKit fingerprint	0.841 ± 0.002	0.857 ± 0.009	0.915 ± 0.003	0.917 ± 0.003	0.670 ± 0.186
SEC fingerprint	0.795 ± 0.006	0.826 ± 0.011	0.879 ± 0.001	0.888 ± 0.002	0.682 ± 0.053
Topological torsion fingerprint	0.807 ± 0.014	0.834 ± 0.010	0.888 ± 0.004	0.896 ± 0.003	0.442 ± 0.098
LDP	0.767 ± 0.010	0.805 ± 0.012	0.858 ± 0.003	0.862 ± 0.001	0.537 ± 0.047
Ensemble	0.862 ± 0.003	0.871 ± 0.006	0.935 ± 0.001	0.943 ± 0.001	0.478 ± 0.018

Model	Davis
	Sensitivity	Specificity	AUC	AUPRC	Threshold
Atom pair fingerprint	0.892 ± 0.009	0.853 ± 0.021	0.929 ± 0.005	0.408 ± 0.014	0.559 ± 0.168
Avalon fingerprint	0.882 ± 0.013	0.847 ± 0.005	0.925 ± 0.005	0.431 ± 0.014	0.457 ± 0.082
MACCS keys fingerprint	0.890 ± 0.006	0.862 ± 0.009	0.929 ± 0.003	0.415 ± 0.021	0.564 ± 0.119
Morgan fingerprint	0.881 ± 0.019	0.855 ± 0.017	0.926 ± 0.003	0.390 ± 0.016	0.590 ± 0.289
RDKit fingerprint	0.876 ± 0.020	0.844 ± 0.011	0.924 ± 0.004	0.416 ± 0.002	0.485 ± 0.093
SEC fingerprint	0.879 ± 0.018	0.852 ± 0.009	0.926 ± 0.004	0.398 ± 0.022	0.561 ± 0.125
Topological torsion fingerprint	0.874 ± 0.022	0.851 ± 0.007	0.919 ± 0.005	0.383 ± 0.009	0.618 ± 0.048
LDP	0.797 ± 0.015	0.728 ± 0.014	0.827 ± 0.003	0.239 ± 0.003	0.537 ± 0.059
Ensemble	0.900 ± 0.013	0.867 ± 0.010	0.936 ± 0.002	0.474 ± 0.011	0.528 ± 0.055

Model	BindingDB
	Sensitivity	Specificity	AUC	AUPRC	Threshold
Atom pair fingerprint	0.869 ± 0.009	0.830 ± 0.015	0.909 ± 0.011	0.636 ± 0.014	0.481 ± 0.054
Avalon fingerprint	0.883 ± 0.007	0.834 ± 0.012	0.918 ± 0.003	0.651 ± 0.005	0.457 ± 0.045
MACCS keys fingerprint	0.886 ± 0.009	0.841 ± 0.006	0.923 ± 0.001	0.671 ± 0.004	0.372 ± 0.082
Morgan fingerprint	0.831 ± 0.005	0.794 ± 0.003	0.879 ± 0.002	0.597 ± 0.001	0.468 ± 0.052
RDKit fingerprint	0.885 ± 0.004	0.832 ± 0.007	0.914 ± 0.001	0.634 ± 0.007	0.490 ± 0.076
SEC fingerprint	0.844 ± 0.012	0.795 ± 0.006	0.881 ± 0.004	0.598 ± 0.002	0.500 ± 0.049
Topological torsion fingerprint	0.876 ± 0.015	0.822 ± 0.014	0.908 ± 0.013	0.615 ± 0.016	0.518 ± 0.056
LDP	0.818 ± 0.013	0.788 ± 0.019	0.873 ± 0.006	0.511 ± 0.008	0.489 ± 0.057
Ensemble	0.885 ± 0.006	0.853 ± 0.007	0.931 ± 0.001	0.707 ± 0.005	0.517 ± 0.022

Citation

Tariq Sha’ban, Ahmad M. Mustafa, Mostafa Z. Ali, WAE-DTI: Ensemble-based architecture for drug–target interaction prediction using descriptors and embeddings, Informatics in Medicine Unlocked, Volume 52, 2025, 101604, ISSN 2352-9148.

@article{SHABAN2025101604,
    title = {WAE-DTI: Ensemble-based architecture for drug–target interaction prediction using descriptors and embeddings},
    journal = {Informatics in Medicine Unlocked},
    volume = {52},
    pages = {101604},
    year = {2025},
    issn = {2352-9148},
    doi = {https://doi.org/10.1016/j.imu.2024.101604},
    url = {https://www.sciencedirect.com/science/article/pii/S2352914824001618},
    author = {Tariq Sha’ban and Ahmad M. Mustafa and Mostafa Z. Ali},
}

---